crime rates and local labor market opportunities in the united states

17
CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES: 1979–1997 Eric D. Gould, Bruce A. Weinberg, and David B. Mustard* Abstract —The labor market prospects of young, unskilled men fell dra- matically in the 1980s and improved in the 1990s. Crime rates show a reverse pattern: increasing during the 1980s and falling in the 1990s. Because young, unskilled men commit most crime, this paper seeks to establish a causal relationshi p between the two trends. Previous work on the relationship between labor markets and crime focused mainly on the relationshi p between the unemployment rate and crime, and found incon- clusive results. In contrast, this paper examines the impact of both wages and unemployment on crime, and uses instrumental variables to establish causality. We conclude that both wages and unemployment are signi - cantly related to crime, but that wages played a larger role in the crime trends over the last few decades. These results are robust to the inclusion of deterrenc e variables, controls for simultaneity, and controllin g for individual and family characteristics . I. Introduction T HIS paper examines the degree to which changes in crime rates for the United States from 1979 to 1997 can be explained by changes in the labor market opportunities for those most likely to commit crime. The labor market prospects of young, unskilled men fell dramatically in the 1980s and then improved in the 1990s. 1 Crime rates show a reverse pattern: increasing during the 1980s and falling in the 1990s. Since young, unskilled men commit most crime (Freeman, 1996), a connection between the two trends is suspected. 2 However, this paper is the rst to systematically examine whether various measures for the labor market conditions of unskilled men can be linked to the trends in crime. Economists typically explain crime rates by examining how the propensity to commit crime responds to the ex- pected costs and bene ts of illegal activity (Becker, 1968; Ehrlich, 1973, 1981, 1996; Levitt, 1997). This study focuses on the indirect costs to crime: the opportunity cost of working in the legal sector. The existing empirical literature has found moderate, but often inconclusive evidence that unemployment rates are positively associated with crime. 3 This paper differs from the existing literature in three ways. First, this paper is the rst to look at whether local crime rates are responsive to the labor market conditions of those most likely to commit crime—unskilled men—rather than looking at whether crime rates respond to the general economic conditions of the area. Second, instead of con- centrating only on the unemployment rate, we also measure the labor market prospects of potential criminals with the wages of low-skilled workers. Third, we establish a causal connection between crime and labor market conditions, which the existing literature fails to do. The fact that the effect of wages on crime has largely been ignored in the literature is surprising because wages may be a better measure for the labor market prospects of potential criminals. Unemployment is often short-lived and highly cyclical. Given the potentially long-lasting effects of incarceration and investing in human capital speci c to the criminal sector, crime should be more responsive to long- term changes in labor market conditions than to short-term uctuations. A secular decline in unskilled wages, as seen during the 1970s and 1980s, represents a decline in the “permanent” wages of uneducated workers, whereas cycli- cal unemployment uctuations have more temporary impli- cations. Although Freeman (1996), Wilson (1996), and Raphael and Winter-Ebmer (2001) speculated that the declining wages and employment opportunities of unskilled men contributed to their increasing involvement in crime, Grog- ger’s (1998) is the only paper to examine the relationship between wages and crime. 4 Grogger used a structural model with individual-level data from the NLSY, and estimates the relationship between the wage offer and the property crimes committed by the individual. In contrast, we focus on a variety of property and violent crimes, and use a nonstruc- tural approach that exploits the differences in the timing of wage changes across geographic areas to explain the timing of the changes in various types of crime. 5 Despite the Received for publication October 1, 1999. Revision accepted for pub- lication January 5, 2001. * Hebrew University, Ohio State University, and University of Georgia, respectively. We appreciate comments from Stephen Bronars, Richard Freeman, Rachel Friedberg, Saul Lach, Anne Piehl, two anonymous referees, and seminar participant s at the European Economic Association 1999, Amer- ican Society of Criminology (1999), Labor Studies Group at the NBER Summer Institute 1998, AEA Meetings in Chicago 1998, Hebrew Univer- sity, Tel Aviv University, University of Georgia, University of Akron, Ohio State University, and Vanderbilt University. We are responsible for any errors. 1 The wages and employment rates of unskilled men fell dramatically from the early 1970s until the early 1990s (Katz & Murphy, 1992; Juhn, 1992). 2 Wilson (1996) implicated the shifting wage and industrial structure of the economy as a possible explanatio n for the increasing trends in crime during the 1980s. 3 Freeman (1983, 1999) reviewed this literature. Freeman and Rodgers (1999) and Papps and Winkleman (2000) presented recent evidence. 4 Although the focus of their paper is not on wages, Cornwell and Trumbull (1994) included controls for wages in various sectors. However, their paper looks at only counties in North Carolina for seven years, and they aggregate all crimes into one category. We use counties throughou t the whole United States for nineteen years and analyze seven types of crime. Also, Lochner (1999) argued that labor market ability, even more than wages, affects crime. Fleisher (1966) and Hashimoto (1987) study the effect of income and the minimum wage on crime, respectively. 5 Topel (1994) showed that there are very signi cant difference s in local labor market conditions , whereas large variation in crime rates across areas has been shown by Glaeser and Sacerdote (1999), Glaeser, Sacer- dote, and Scheinkman (1996), and Levitt (1997). The Review of Economics and Statistics, February 2002, 84(1): 45–61 2002 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology

Upload: hoangkhanh

Post on 14-Feb-2017

217 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: crime rates and local labor market opportunities in the united states

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THEUNITED STATES 1979ndash1997

Eric D Gould Bruce A Weinberg and David B Mustard

AbstractmdashThe labor market prospects of young unskilled men fell dra-matically in the 1980s and improved in the 1990s Crime rates show areverse pattern increasing during the 1980s and falling in the 1990sBecause young unskilled men commit most crime this paper seeks toestablish a causal relationship between the two trends Previous work onthe relationship between labor markets and crime focused mainly on therelationship between the unemployment rate and crime and found incon-clusive results In contrast this paper examines the impact of both wagesand unemployment on crime and uses instrumenta l variables to establishcausality We conclude that both wages and unemployment are signi -cantly related to crime but that wages played a larger role in the crimetrends over the last few decades These results are robust to the inclusionof deterrence variables controls for simultaneity and controlling forindividua l and family characteristics

I Introduction

THIS paper examines the degree to which changes incrime rates for the United States from 1979 to 1997 can

be explained by changes in the labor market opportunitiesfor those most likely to commit crime The labor marketprospects of young unskilled men fell dramatically in the1980s and then improved in the 1990s1 Crime rates show areverse pattern increasing during the 1980s and falling inthe 1990s Since young unskilled men commit most crime(Freeman 1996) a connection between the two trends issuspected2 However this paper is the rst to systematicallyexamine whether various measures for the labor marketconditions of unskilled men can be linked to the trends incrime

Economists typically explain crime rates by examininghow the propensity to commit crime responds to the ex-pected costs and bene ts of illegal activity (Becker 1968Ehrlich 1973 1981 1996 Levitt 1997) This study focuseson the indirect costs to crime the opportunity cost ofworking in the legal sector The existing empirical literaturehas found moderate but often inconclusive evidence thatunemployment rates are positively associated with crime3

This paper differs from the existing literature in three waysFirst this paper is the rst to look at whether local crimerates are responsive to the labor market conditions of thosemost likely to commit crimemdashunskilled menmdashrather thanlooking at whether crime rates respond to the generaleconomic conditions of the area Second instead of con-centrating only on the unemployment rate we also measurethe labor market prospects of potential criminals with thewages of low-skilled workers Third we establish a causalconnection between crime and labor market conditionswhich the existing literature fails to do

The fact that the effect of wages on crime has largelybeen ignored in the literature is surprising because wagesmay be a better measure for the labor market prospects ofpotential criminals Unemployment is often short-lived andhighly cyclical Given the potentially long-lasting effects ofincarceration and investing in human capital speci c to thecriminal sector crime should be more responsive to long-term changes in labor market conditions than to short-term uctuations A secular decline in unskilled wages as seenduring the 1970s and 1980s represents a decline in theldquopermanentrdquo wages of uneducated workers whereas cycli-cal unemployment uctuations have more temporary impli-cations

Although Freeman (1996) Wilson (1996) and Raphaeland Winter-Ebmer (2001) speculated that the decliningwages and employment opportunities of unskilled mencontributed to their increasing involvement in crime Grog-gerrsquos (1998) is the only paper to examine the relationshipbetween wages and crime4 Grogger used a structural modelwith individual-level data from the NLSY and estimates therelationship between the wage offer and the property crimescommitted by the individual In contrast we focus on avariety of property and violent crimes and use a nonstruc-tural approach that exploits the differences in the timing ofwage changes across geographic areas to explain the timingof the changes in various types of crime5 Despite the

Received for publication October 1 1999 Revision accepted for pub-lication January 5 2001

Hebrew University Ohio State University and University of Georgiarespectively

We appreciate comments from Stephen Bronars Richard FreemanRachel Friedberg Saul Lach Anne Piehl two anonymous referees andseminar participant s at the European Economic Association 1999 Amer-ican Society of Criminology (1999) Labor Studies Group at the NBERSummer Institute 1998 AEA Meetings in Chicago 1998 Hebrew Univer-sity Tel Aviv University University of Georgia University of AkronOhio State University and Vanderbil t University We are responsibl e forany errors

1 The wages and employment rates of unskilled men fell dramaticallyfrom the early 1970s until the early 1990s (Katz amp Murphy 1992 Juhn1992)

2 Wilson (1996) implicated the shifting wage and industrial structure ofthe economy as a possible explanation for the increasing trends in crimeduring the 1980s

3 Freeman (1983 1999) reviewed this literature Freeman and Rodgers(1999) and Papps and Winkleman (2000) presented recent evidence

4 Although the focus of their paper is not on wages Cornwell andTrumbull (1994) included controls for wages in various sectors Howevertheir paper looks at only counties in North Carolina for seven years andthey aggregate all crimes into one category We use counties throughou tthe whole United States for nineteen years and analyze seven types ofcrime Also Lochner (1999) argued that labor market ability even morethan wages affects crime Fleisher (1966) and Hashimoto (1987) study theeffect of income and the minimum wage on crime respectivel y

5 Topel (1994) showed that there are very signi cant difference s in locallabor market conditions whereas large variation in crime rates acrossareas has been shown by Glaeser and Sacerdote (1999) Glaeser Sacer-dote and Scheinkman (1996) and Levitt (1997)

The Review of Economics and Statistics February 2002 84(1) 45ndash61rsquo 2002 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology

marked differences in methods Groggerrsquos results are gen-erally consistent with those reported here6

Our empirical work consists of three basic analyses the rst two use aggregated data at the county level and thethird incorporates individual-level data Our rst analysis isto run panel regressions using annual county-level datafrom 1979 to 1997 with county and time xed effectsBecause wages and unemployment rates for various demo-graphic groups are not available on an annual basis at thecounty level we explain the county-level crime rate byfocusing on the state-level average wage and unemploymentrate of non-college-educated men This approach exploitsyear-to-year variation in state-level labor market conditionsto explain year-to-year changes in the county-level crimerates

The second analysis explains the ten-year change (1979ndash1989) in the county crime rate by the ten-year change in theaverage wage and unemployment rate of non-college-educated men measured at the metropolitan area (MA)level This strategy exploits the low-frequency variation inthe data Given the long-term consequences of criminalactivity crime should be more responsive to low-frequencychanges in labor market conditions In addition the labormarket conditions of the MA are a closer match for the labormarket conditions of the county than are variables measuredat the state level This long-term regression approach alsoattenuates measurement error problems in panel regressionanalyses7

Our third analysis uses individual-level data from theNLSY79 to test whether local labor market conditions canexplain individual criminal activity The NLSY79 permitsus to control for a rich set of personal characteristics (suchas education ability and parental background) After con-trolling for these variables we exploit geographical varia-tion in the wages and unemployment rates of unskilledmen to explain the criminal behavior of individuals in oursample

All three strategies indicate that young unskilled men areresponsive to the opportunity costs of crime However ifspeci c types of workers or employers migrate in responseto increasing crime changes in the labor market conditionsof an area could be endogenous to the change in the localcrime rate To control for this potential endogeneity prob-lem we use instrumental variables based on the initialindustrial composition of the local area the aggregate in-dustrial trends and the demographic changes within indus-tries at the aggregate level8 Our IV results indicate thatendogeneity is not responsible for the signi cant relation-ship between the labor market conditions of unskilled work-

ers and the various crime rates Furthermore we nd thatthe long-term trends in crime are better explained by thelong-term trend in wages than by the trend in the unem-ployment rate simply because there is little long-term trendin the unemployment rate while wages for less educatedmen fell dramatically over the sample period

The paper is organized as follows Section II discussesthe general trends in crime rates wages and unemploymentSection III presents the panel regressions using annual dataSection IV analyzes the ten-year difference (1979ndash1989)regression speci cations Section V presents the individual-level data analysis and section VI concludes

II Trends in Crime Rates Wages and Unemployment

The aggregate crime data reported to the FBI by localpolice authorities come from the Uniform Crime ReportsCrime rates are offenses per 100000 people and the arrestrates are the ratios of arrests to offenses Offenses andarrests are reported for the individual violent crimes (mur-der rape robbery and aggravated assault) and propertycrimes (burglary larceny and auto theft) The violent andproperty crime indices aggregate their respective individualcrimes and the overall crime index aggregates all sevenindividual crimes The UCR data are described in moredetail in appendix A

There are many reasons to be wary of self-reported crimedata First not every crime is reported to the police and thisunder-reporting produces measurement error in the offenseand arrest rates which could vary by the type of crime orcounty of jurisdiction9 Also the methods of collecting andreporting data vary across local authorities Our inclusion ofcounty xed effects eliminates the effects of (time-invariant) cross-county variations in reporting methods10

Figure 1 shows the standardized log offense rates for theoverall property crime and violent crime indices for theentire United States The property crime index follows acyclical pattern that peaks in 1980 declines by 17 until1984 increases by 13 until 1991 and then declinesapproximately 24 until 1997 The global peak for propertycrime in 1980 was approximately 4 larger than the localpeak in 1991 Property crime increased through the latterhalf of the 1980s but the absolute levels were not extraor-dinary

Although violent crime is also cyclical the absolute levelis more than 24 larger in 1991 than at the local peak in

6 Grogger (1998) found that youth behavior is responsive to priceincentives and that falling real wages may have been an importantdeterminant of raising youth crime during the 1970s and 1980s

7 Griliches and Hausman (1986) and Levitt (1995) discussed advantagesof the ldquolong regressionrdquo in the presence of measurement error

8 This IV strategy is an extension of a strategy developed by Bartik(1991) and Blanchard and Katz (1992)

9 For example in 1994 the National Criminal Victimization Surveysindicates that 361 of rapes 407 of sexual assaults 554 of robber-ies 516 of aggravated assaults 268 of personal larcenies withoutcontact 505 of the household burglaries and 782 of motor vehiclethefts and theft attempts were reported Murder which has virtually nounder-reporting is not subject to this type of bias (Sourcebook of CriminalJustice Statistics 1995 table 338 p 250)

10 Ehrlich (1996) discussed reporting biases in the crime data Onemethod of addressing it is to work with the logarithms of the crime rateswhich are likely to be proportiona l to the true crime rates We use thisstrategy in this paper

THE REVIEW OF ECONOMICS AND STATISTICS46

1980 During the whole period violent crime rose by 32until 1991 and then steadily declined by 29 as of 1997Thus the pattern for violent crime is much more consistentwith the common perception of increasing crime throughthe 1980s and declining since the early 1990s11

In 1997 88 of all crime was property crime Thereforethe overall crime rate pattern in gure 1 is almost identicalto the property crime rate Consequently results for theoverall crime index are dominated by the results for theproperty crime index The property crime index is domi-nated by larceny (67) and burglary (213) and auto theftcomprises the remaining 117 Thus results for the prop-erty crime index will be heavily in uenced by larceny andburglary Violent crime is composed mainly of aggravatedassault (63) and robbery (305) whereas rape (6) andmurder (1) have only a minor in uence on the overallviolent crime rate However the seriousness of these lattertwo crimes gives them a disproportionate in uence oversocial welfare and public policy

The trends in our panel sample of 705 counties aredisplayed in gure 2 and are similar to the national trends in gure 1 The sample consists of all counties with a meanpopulation greater than 25000 between 1979 and 1997 andat least sixteen out of nineteen years of complete data Theaverage county population size is 248017 over the entireperiod which covers an average of almost 175 millionpeople per year The sample selection criteria were designedto capture a representative population while deleting thosecounties where the reporting accuracy is likely to be unre-liable The size of our sample and the trends displayed in

gure 2 (in comparison to gure 1) demonstrate that oursample is representative of the entire United States12

So far we have looked only at the raw crime data with noadjustments for changes in the demographic compositionswithin each county Figure 3 plots the property and violentcrime trends after adjusting for changes in the age sex andracial composition After controlling for these factors thetrends for both types of crime rose steadily throughout the1980s and declined after the early 1990s In 1994 theadjusted property crime rate hit a global peak at 23 higherthan the local peak in 1980 and 29 higher than it was atthe beginning of the period in 1979 The upward trend inunadjusted violent crime found before in gure 2 is now

11 Murder which has virtually no measurement error hit a global peakin 1980 at 102 murders per 100000 people and never got above 98which was the second peak in 1991

12 Levitt (1997) showed similar trends using the same data source for 59large cities

FIGURE 1mdashUNITED STATES NATIONAL TRENDS IN CRIME

INDICES 1979ndash1998

Plotted values are the log offense rate (offenses per 100000 people) relative to the year 1979 TheProperty Crime index is the sum of auto theft burglary and larceny The Violent Crime Index is the sumof aggravated assault robbery murder and rape The Overall Crime index is the sum of all property andviolent crimes Data come from the Uniform Crime Reports

FIGURE 2mdashCOUNTY SAMPLE TRENDS IN UNADJUSTED CRIME INDICES

Plotted values are the coef cients on the time dummies from regressions of the log offense rates ontime and county dummies for 705 counties with a mean population greater than 25000 and with at leastsixteen out of nineteen years of data (1979ndash1997) The county population means were used as weights

FIGURE 3mdashCOUNTY SAMPLE TRENDS IN ADJUSTED CRIME INDICES

Plotted values are the coef cients on the time dummies of regressions of the log offense rates on timedummies county xed effects and controls for age distribution (using the percentage of the populationin ve different age groups) the sex composition the percentage of the population that is black and thepercentage that is neither white nor black The sample consists of 705 counties with a mean populationgreater than 25000 and with at least sixteen out of nineteen years of data (1979ndash1997) The countypopulation means were used as weights

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 47

accentuated as the adjusted rate rose by over 47 until1993 and then declined 12 through 1997 The individualproperty and violent crimes (adjusted for changes in demo-graphics) are depicted in gure 4 and 5

Figure 3 demonstrates that changes in demographicsexplain much of the decline in both types of crime duringthe 1990s Without these controls the unadjusted propertyand violent crime indices peaked in 1991 and 1992 respec-tively With these controls they peaked later in 1994 and1993 respectively (at 29 and 47 higher than in 1979)From 1993 to 1997 adjusted property crime decreased by76 and adjusted violent crime by 123 Even as of 1997the adjusted property and violent crime rates were stilllarger (by 21 and 35 respectively) than they were in1979 Although the trends have reversed since the mid-1990s the secular trend in crime over the entire period isclearly upward While this was happening the labor marketprospects for young unskilled men deteriorated Figure 6plots the average wages of non-college-educated maleworkers (workers with only a high school degree or less)

over time The average wage of non-college-educated mendeclined by a total of 23 from 1979 to 1993 and thenrebounded somewhat until 1997 This overall pattern isalmost the mirror image of the crime patterns and there-fore this paper seeks to establish whether a causal connec-tion can be established

The theory behind such a connection is simple a declinein the wage offer increases the relative payoff of criminalactivity thus inducing workers to substitute away from thelegal sector towards the illegal sector In addition a lowerwage offer may produce an income effect by increasing theneed to seek additional sources of income in possibly lessdesirable and more dangerous ways A lower wage alsoreduces the opportunity cost of serving time in prison13 Thedegree to which legal alternatives affect criminal behaviormay however vary by the type of crime Some crimes (suchas robbery larceny burglary and auto theft) can be used forself-enrichment whereas other crimes (murder rape andassault) are much less likely to yield material gains to theoffender14 Offenders of the latter crimes are much morelikely to be motivated by nonpecuniary considerations15

13 Also Lott (1992) argued that reputationa l sanctions are positivelycorrelated with the wage

14 For example in 1992 the average monetary loss was $483 $840$1278 and $4713 for larceny robbery burglary and auto theft respec-tively compared with average monetary losses of $27 and $89 for rapeand murder as reported in Crime in the United States 1992

15 Offenders who commit the latter crimes are more likely to derivebene ts from interdependencie s in utility with the victim This notion ofinterdependenc e of utility between offender and victim for certain crimesis supported by the fact that murder rape and assault occur frequentlybetween people who know each other whereas the victim and offenderhave no relationship in the vast majority of property crimes For offensescommitted in 1993 the offenders were classi ed as non-stranger s to thevictims in 742 of rapes 519 of assaults and 199 of robberies(1994 Sourcebook of Criminal Justice Statistics table 311 p 235)Historically murder victims knew their offenders (Supplementary Homi-

FIGURE 4mdashCOUNTY SAMPLE TRENDS IN ADJUSTED PROPERTY CRIMES

See notes to Figure 3

FIGURE 5mdashCOUNTY SAMPLE TRENDS IN ADJUSTED VIOLENT CRIMES

See notes to Figure 3

FIGURE 6mdashSTANDARDIZED WAGES AND UNEMPLOYMENT RATES

OF NON-COLLEGE-EDUCATED MEN

The data were computed from the CPS Non-college-educated mean are de ned as full-time menbetween the age of 18 and 65 Residuals were computed after controlling for a quartic in potentialexperience years of school (within non-college) race (black and nonwhite nonblack) Hispanicbackground region of residence and marital status Wages de ated to 1982ndash1984 100 dollars Meanresiduals for each year were standardized to the base year 1979

THE REVIEW OF ECONOMICS AND STATISTICS48

However it is important to note that only the most severecrime is reported in the UCR data when multiple crimes arecommitted in the same incident Therefore pecuniary mo-tives may lie behind many of the reported assaults when aproperty crime was also involved Holding everything elseconstant a reduction in legal opportunities should make onemore likely to engage in any form of criminal activityregardless of motives due to the reduced legal earnings lostwhile engaging in a criminal career and potentially servingin jail

Including the modest increase in the 1990s gure 6indicates that the non-college-educated male wage was 20lower in 1997 than in 1979 This overall trend represents alarge long-term decline in the earning prospects of lesseducated men In contrast gure 6 shows that the unem-ployment rate of non-college-educated men did not suffer along-term deterioration throughout the period Althoughless educated workers suffer the most unemployment un-employment rates generally follow a cyclical pattern thatby de nition traces out the business cycle In gure 6 theunemployment rate is the same in 1997 as it was in 1979although there was variation in the intervening yearsClearly the unemployment rate affects the labor marketprospects of less educated men but it is hard to discern along-term deterioration in their legal opportunities by look-ing at the overall trend in the unemployment rate Theoverall decline in the labor market prospects of less edu-cated men however is clearly shown by their wages

The data clearly show that the propensity to commitcrime moved inversely to the trends in the labor marketconditions for unskilled men These trends seem to berelated particularly because young unskilled men are themost likely to commit crime16 The goal of the remainingsections is to establish empirically whether the relationshipis causal

III County-Level Panel Analysis 1979ndash1997

This section analyzes a panel sample of 705 counties overnineteen years The data and trends were described insection II In each regression county xed effects controlfor much of the cross-sectional variation and yearly timedummies control for the national trends The county xedeffects control for unobserved county-level heterogeneitythat might be correlated with the county crime rate Remov-ing the national trend allows us to abstract from any corre-lation between the aggregate trends in crime and some otherunobserved aggregate determinant of crime Including con-trols for national trends also controls for aggregate trends in

reporting practices Given the strong inverse relationshipbetween the wage trends of less skilled men and the aggre-gate crime trends eliminating the aggregate trend tends tobias the results against nding a relationship between thetwo phenomena We expect that the labor market variablescan help explain the cross-sectional variation and the na-tional trends but we identify the effects of these variablesfrom the within-county deviations from the national trendsto avoid any spurious correlations17 Each speci cation alsocontrols for changes in the age sex and race composition ofthe county

Because the wages and unemployment rates of less edu-cated men are not available at the county level we use thesevariables measured at the state level to explain the countycrime rates Our wage measure is the mean state residualafter regressing individual wages from the CPS on educa-tion experience experience squared and controls for raceand marital status The residual state unemployment ratewas calculated similarly The construction of these variablesis described in detail in appendix B Using the residualsallows us to abstract from changes in our measures due tochanges in observable characteristics of workers and thusmore accurately re ects changes in the structure of wagesand unemployment However very similar results are ob-tained by using the levels rather than the residuals of thesevariables

To control for the general level of prosperity in the areawe use log income per capita in the state As shown in gure7 income per capita increased steadily since the early1980s The impact of this trend on crime however istheoretically unclear If the level of prosperity increasesthere is more material wealth to steal so crime could

cide Reports) During the 1990s this relationship changed and nowslightly less than half of the murder victims know their offenders Forexample in 1993 477 of all murders were committed by people whowere known to the victim 140 were committed by strangers and in393 of the cases the relationship between victim and offender wasunknown (Crime in the United States 1993 table 212 p 20)

16 Freeman (1996) reports that two-thirds of prison inmates in 1991 hadnot graduated from high school

17 Very similar OLS and IV results are obtained when we do not controlfor county xed effects A notable exception is for the larceny category inwhich unobserved county heterogeneit y reverses the sign for the OLScoef cients on our two measures for the labor market prospects ofunskilled workers However the IV coef cients for larceny have theldquoexpectedrdquo sign and are statistically signi cant

FIGURE 7mdashRETAIL WAGES AND INCOME PER CAPITA OVER TIME

Both variables were computed by the procedure described in Figure 3

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 49

increase However higher income individuals invest morein self-protection from criminals so crime may decrease18

The overall effect therefore is an empirical question Byincluding this measure in our regressions we answer thisquestion and control for any correlation between the labormarket prospects of less educated men and the overalleconomic prosperity of the area

Table 1 displays the coef cient estimates for the eco-nomic variables in our ldquocorerdquo speci cation The standarderrors throughout the analysis are corrected for a commonunobserved factor underlying crime in each state in eachyear All three economic variables are very signi cant forthe property and violent crime indices and every individualcrime rate except for rape Furthermore the coef cientshave the expected signs increases in the wages of non-college-educated men reduce the crime rate and increasesin the unemployment rate of non-college-educated menincrease the crime rate19 The results for income per capitaare quite uniformly positive and signi cant indicating thatimprovements in the overall economic condition of the areaincrease the amount of material wealth available to stealthus increasing crime rates

Although the coef cients are statistically signi cant wewould like to know whether their magnitudes are econom-ically signi cant The numbers underneath each standarderror indicate the ldquopredictedrdquo effects of each independentvariable on the crime rate based on the coef cient estimateand the mean change in the independent variable over twodifferent time periods The rst number indicates the pre-dicted effects between 1979 and 1993 when the adjusted

crime indices increased (See section II) The second num-ber is the predicted effect during 1993ndash1997 when crimefell From table 1 the 233 fall in the wages of unskilledmen from 1979 to 1993 ldquopredictrdquo a 125 increase inproperty crime (the coef cient 054 multiplied by 233)and a 251 increase in violent crime (the coef cient 108multiplied by 233) The 305 increase in unemploymentduring this early period ldquopredictedrdquo a 71 (the coef cient233 times 305) increase in property crime and a 38 (thecoef cient 126 times 305) increase in violent crime There-fore the non-college-educated wage explains 43 of the29 increase in adjusted property crime during this timeperiod and 53 of the 472 increase in adjusted violentcrime The unemployment rate of non-college-educatedmen explains 24 of the total increase in property crimeand 8 of the increase in violent crime Clearly the long-term trend in wages was the dominant factor on crimeduring this time period

The declining crime trends in the 1993ndash1997 period arebetter explained by the unemployment rate The adjustedproperty and violent crime rates fell by 76 and 123respectively From table 1 the 31 increase in the wages ofnon-college-educated men predict a decrease of 17 inproperty crime and 33 in violent crime The comparablepredictions for the 31 decline in the unemployment rateare decreases of 75 for property crime and 40 forviolent crime20 Although the predicted effects are quitesimilar for violent crime the declining crime rates in the1993ndash1997 period were more in uenced by the unemploy-ment rate than the non-college-educated wage Whether this

18 Lott and Mustard (1997) and Ayres and Levitt (1998) showed thatself-protectio n lowers crime by carrying concealed weapons and purchas-ing Lojack (an auto-thef t prevention system) respectivel y

19 These results are robust to the inclusion of the mean state residualwage for all workers as an additiona l control variable

20 Focusing exclusively on the effect of declining unemployment ratesduring the 1990s on crimes per youth Freeman and Rodgers (1999) foundsimilar results They found that a decrease of one percentage point inunemployment lowers crimes per youth by 15 whereas we nd that itlowers crimes per capita by 233

TABLE 1mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

060(013)a 141b 19

054(014)a 125b 17

145(024)a 337b 44

060(017)a 140b 19

041(014)a 94b 12

108(015)a 251b 33

131(018)a 304b 40

095(025)a 221b 29

096(022)a 224b 30

010(017)a 22b 03

State unemploymen t rateresidual for non-college-educated men

222(028)a 68b 71

233(030)a 71b 75

085(041)a 23b 27

310(034)a 95b 100

233(031)a 71b 75

126(030)a 38b 40

108(032)a 33b 35

080(045)a 25b 26

212(044)a 65b 68

012(035)a 05b 04

Log state income per capita 054(018)a 11b 40

048(019)a 01b 36

072(030)a 14b 53

060(024)a 12b 44

051(018)a 10b 37

096(019)a 19b 71

118(021)a 24b 87

091(031)a 18b 67

120(031)a 24b 89

086(023)a 17b 64

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 004 002 005 004 001 001 004 000 001

indicates that the coef cient is signi cant at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Numbers after ldquoardquo represent the ldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independent variable

between 1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independen t variable during the latter period 1993ndash1997Observations are for 705 counties with at least 16 out of 19 years of data Regressions include county and time xed effects and county-level demographic controls (percentage of population age 10ndash19 age 20ndash29age 30ndash39 age 40ndash49 and age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions are weighted by the mean population size of each county Partial R2 are aftercontrolling for county and time xed effects and demographic controls

THE REVIEW OF ECONOMICS AND STATISTICS50

trend will continue is improbable because the recent cyclicaldrop is likely to be temporary and in the future unemploy-ment will continue to uctuate with the business cycle Thewage trends however can continue to improve and have alasting impact on the crime trends This is best exempli ed byexplaining the increases of 21 and 35 in the adjustedproperty and violent crime rates over the entire 1979ndash1997period The unemployment rate was virtually unchanged in1997 from 1979 and therefore explains none of the increase ineither crime index The 20 fall in non-college-educatedwages over the entire period predicts a 108 increase inproperty crime and a 216 increase in violent crime Thesepredictionsldquoexplainrdquo more than 50 of the long-term trend inboth indices illustrating just how much the long-term crimetrends are dominated by the wages of unskilled men as op-posed to their unemployment rate

To see if our wage and unemployment measures in table1 are picking up changes in the relative supplies of differenteducation groups we checked if the results are sensitive tothe inclusion of variables capturing the local educationdistribution The results are practically identical to those intable 1 and therefore are not presented21 Clearly theresults are not due to changes in the education distribution

The speci cation in table 2 includes our ldquocorerdquo economicvariables plus variables measuring the local level of crimedeterrence Three deterrence measures are used the county

arrest rate state expenditures per capita on police and statepolice employment per capita22 Missing values for arrestrates are more numerous than for offense rates so thesample is reduced to 371 counties that meet our sampleselection criteria In addition the police variables wereavailable for only the years 1979 to 1995 After includingthese deterrence variables the coef cients on the non-college-educated wage remain very signi cant for propertyand violent crime although the magnitudes drop a bit Theunemployment rate remains signi cant for property crimebut disappears for violent crime

The arrest rate has a large and signi cantly negative effectfor every classi cation of crime Because the numerator of thedependent variable appears in the denominator of the arrestrate (the arrest rate is de ned as the ratio of total arrests to totaloffenses) measurement error in the offense rate leads to adownward bias in the coef cient estimates of the arrest rates(ldquodivision biasrdquo)23 In addition the police size variables arelikely to be highly endogenous to the local crime rate exem-pli ed by the positive coef cient on police expenditures forevery crime Controlling for the endogeneity of these policevariables is quite complicated (Levitt 1997) and is not thefocus of this study Table 2 demonstrates that the resultsparticularly for the non-college-educated wage are generallyrobust to the inclusion of these deterrence measures as well asto the decrease in the sample To work with the broadestsample possible and avoid the endogeneity and ldquodivision-biasrdquoissues of these deterrence variables the remaining speci ca-tions exclude these variables

21 We included the percentage of male high school dropouts in the statepercentage of male high school graduates and percentage of men withsome college The property crime index coef cients (standard errors inparentheses ) were 053 (014) for the non-college-educate d male wageresidual and 215 (029) for the non-college-educate d unemployment rateresidual For the violent crime index the respective coef cients were

100 (015) and 129 (030) Compared to the results in table 1 whichexcluded the education distribution variables the coef cients are almostidentical in magnitude and signi cance as is also the case for theindividua l crime categories

22 Mustard (forthcoming ) showed that although conviction and sentenc-ing data are theoreticall y important they exist for only four or ve statesTherefore we cannot include such data in this analysis

23 Levitt (1995) analyzed this issue and why the relationship betweenarrest rates and offense rates is so strong

TABLE 2mdashOLS COUNTY-LEVEL PANEL REGRESSIONS USING THE ldquoCORE SPECIFICATIONrdquo PLUS ARREST RATES AND POLICE SIZE VARIABLES 1979ndash1995

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

052(015)

040(015)

140(027)

056(021)

022(015)

064(017)

085(022)

089(030)

069(023)

032(019)

State unemploymen t rateresidual for non-college-educated men

138(037)

132(037)

082(047)

219(043)

129(036)

018(033)

016(040)

100(057)

039(046)

029(037)

Log state income per capita 001(021)

007(021)

003(033)

005(027)

005(020)

002(020)

034(024)

114(030)

018(031)

042(024)

County arrest rate 0002(0002)

001(0002)

001(0001)

001(0003)

001(0001)

0004(00003)

0003(00003)

0002(00002)

0006(00005)

0004(0001)

Log state police expendituresper capita

028(010)

030(010)

051(0176)

034(011)

026(010)

036(011)

039(014)

023(016)

055(013)

004(010)

Log state police employmentper capita

004(014)

002(013)

026(020)

004(016)

007(012)

006(014)

001(016)

028(018)

014(018)

045(012)

Observations 5979 5979 5979 5979 5979 5979 5979 5979 5979 5979Partial R2 003 003 003 004 000 001 001 001 000 001

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 371 counties withat least sixteen out of seventeen years of data where the police force variables are available (1979ndash1995) County mean population is used as weights See notes to table 1 for other de nitions and controls usedin the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 51

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 2: crime rates and local labor market opportunities in the united states

marked differences in methods Groggerrsquos results are gen-erally consistent with those reported here6

Our empirical work consists of three basic analyses the rst two use aggregated data at the county level and thethird incorporates individual-level data Our rst analysis isto run panel regressions using annual county-level datafrom 1979 to 1997 with county and time xed effectsBecause wages and unemployment rates for various demo-graphic groups are not available on an annual basis at thecounty level we explain the county-level crime rate byfocusing on the state-level average wage and unemploymentrate of non-college-educated men This approach exploitsyear-to-year variation in state-level labor market conditionsto explain year-to-year changes in the county-level crimerates

The second analysis explains the ten-year change (1979ndash1989) in the county crime rate by the ten-year change in theaverage wage and unemployment rate of non-college-educated men measured at the metropolitan area (MA)level This strategy exploits the low-frequency variation inthe data Given the long-term consequences of criminalactivity crime should be more responsive to low-frequencychanges in labor market conditions In addition the labormarket conditions of the MA are a closer match for the labormarket conditions of the county than are variables measuredat the state level This long-term regression approach alsoattenuates measurement error problems in panel regressionanalyses7

Our third analysis uses individual-level data from theNLSY79 to test whether local labor market conditions canexplain individual criminal activity The NLSY79 permitsus to control for a rich set of personal characteristics (suchas education ability and parental background) After con-trolling for these variables we exploit geographical varia-tion in the wages and unemployment rates of unskilledmen to explain the criminal behavior of individuals in oursample

All three strategies indicate that young unskilled men areresponsive to the opportunity costs of crime However ifspeci c types of workers or employers migrate in responseto increasing crime changes in the labor market conditionsof an area could be endogenous to the change in the localcrime rate To control for this potential endogeneity prob-lem we use instrumental variables based on the initialindustrial composition of the local area the aggregate in-dustrial trends and the demographic changes within indus-tries at the aggregate level8 Our IV results indicate thatendogeneity is not responsible for the signi cant relation-ship between the labor market conditions of unskilled work-

ers and the various crime rates Furthermore we nd thatthe long-term trends in crime are better explained by thelong-term trend in wages than by the trend in the unem-ployment rate simply because there is little long-term trendin the unemployment rate while wages for less educatedmen fell dramatically over the sample period

The paper is organized as follows Section II discussesthe general trends in crime rates wages and unemploymentSection III presents the panel regressions using annual dataSection IV analyzes the ten-year difference (1979ndash1989)regression speci cations Section V presents the individual-level data analysis and section VI concludes

II Trends in Crime Rates Wages and Unemployment

The aggregate crime data reported to the FBI by localpolice authorities come from the Uniform Crime ReportsCrime rates are offenses per 100000 people and the arrestrates are the ratios of arrests to offenses Offenses andarrests are reported for the individual violent crimes (mur-der rape robbery and aggravated assault) and propertycrimes (burglary larceny and auto theft) The violent andproperty crime indices aggregate their respective individualcrimes and the overall crime index aggregates all sevenindividual crimes The UCR data are described in moredetail in appendix A

There are many reasons to be wary of self-reported crimedata First not every crime is reported to the police and thisunder-reporting produces measurement error in the offenseand arrest rates which could vary by the type of crime orcounty of jurisdiction9 Also the methods of collecting andreporting data vary across local authorities Our inclusion ofcounty xed effects eliminates the effects of (time-invariant) cross-county variations in reporting methods10

Figure 1 shows the standardized log offense rates for theoverall property crime and violent crime indices for theentire United States The property crime index follows acyclical pattern that peaks in 1980 declines by 17 until1984 increases by 13 until 1991 and then declinesapproximately 24 until 1997 The global peak for propertycrime in 1980 was approximately 4 larger than the localpeak in 1991 Property crime increased through the latterhalf of the 1980s but the absolute levels were not extraor-dinary

Although violent crime is also cyclical the absolute levelis more than 24 larger in 1991 than at the local peak in

6 Grogger (1998) found that youth behavior is responsive to priceincentives and that falling real wages may have been an importantdeterminant of raising youth crime during the 1970s and 1980s

7 Griliches and Hausman (1986) and Levitt (1995) discussed advantagesof the ldquolong regressionrdquo in the presence of measurement error

8 This IV strategy is an extension of a strategy developed by Bartik(1991) and Blanchard and Katz (1992)

9 For example in 1994 the National Criminal Victimization Surveysindicates that 361 of rapes 407 of sexual assaults 554 of robber-ies 516 of aggravated assaults 268 of personal larcenies withoutcontact 505 of the household burglaries and 782 of motor vehiclethefts and theft attempts were reported Murder which has virtually nounder-reporting is not subject to this type of bias (Sourcebook of CriminalJustice Statistics 1995 table 338 p 250)

10 Ehrlich (1996) discussed reporting biases in the crime data Onemethod of addressing it is to work with the logarithms of the crime rateswhich are likely to be proportiona l to the true crime rates We use thisstrategy in this paper

THE REVIEW OF ECONOMICS AND STATISTICS46

1980 During the whole period violent crime rose by 32until 1991 and then steadily declined by 29 as of 1997Thus the pattern for violent crime is much more consistentwith the common perception of increasing crime throughthe 1980s and declining since the early 1990s11

In 1997 88 of all crime was property crime Thereforethe overall crime rate pattern in gure 1 is almost identicalto the property crime rate Consequently results for theoverall crime index are dominated by the results for theproperty crime index The property crime index is domi-nated by larceny (67) and burglary (213) and auto theftcomprises the remaining 117 Thus results for the prop-erty crime index will be heavily in uenced by larceny andburglary Violent crime is composed mainly of aggravatedassault (63) and robbery (305) whereas rape (6) andmurder (1) have only a minor in uence on the overallviolent crime rate However the seriousness of these lattertwo crimes gives them a disproportionate in uence oversocial welfare and public policy

The trends in our panel sample of 705 counties aredisplayed in gure 2 and are similar to the national trends in gure 1 The sample consists of all counties with a meanpopulation greater than 25000 between 1979 and 1997 andat least sixteen out of nineteen years of complete data Theaverage county population size is 248017 over the entireperiod which covers an average of almost 175 millionpeople per year The sample selection criteria were designedto capture a representative population while deleting thosecounties where the reporting accuracy is likely to be unre-liable The size of our sample and the trends displayed in

gure 2 (in comparison to gure 1) demonstrate that oursample is representative of the entire United States12

So far we have looked only at the raw crime data with noadjustments for changes in the demographic compositionswithin each county Figure 3 plots the property and violentcrime trends after adjusting for changes in the age sex andracial composition After controlling for these factors thetrends for both types of crime rose steadily throughout the1980s and declined after the early 1990s In 1994 theadjusted property crime rate hit a global peak at 23 higherthan the local peak in 1980 and 29 higher than it was atthe beginning of the period in 1979 The upward trend inunadjusted violent crime found before in gure 2 is now

11 Murder which has virtually no measurement error hit a global peakin 1980 at 102 murders per 100000 people and never got above 98which was the second peak in 1991

12 Levitt (1997) showed similar trends using the same data source for 59large cities

FIGURE 1mdashUNITED STATES NATIONAL TRENDS IN CRIME

INDICES 1979ndash1998

Plotted values are the log offense rate (offenses per 100000 people) relative to the year 1979 TheProperty Crime index is the sum of auto theft burglary and larceny The Violent Crime Index is the sumof aggravated assault robbery murder and rape The Overall Crime index is the sum of all property andviolent crimes Data come from the Uniform Crime Reports

FIGURE 2mdashCOUNTY SAMPLE TRENDS IN UNADJUSTED CRIME INDICES

Plotted values are the coef cients on the time dummies from regressions of the log offense rates ontime and county dummies for 705 counties with a mean population greater than 25000 and with at leastsixteen out of nineteen years of data (1979ndash1997) The county population means were used as weights

FIGURE 3mdashCOUNTY SAMPLE TRENDS IN ADJUSTED CRIME INDICES

Plotted values are the coef cients on the time dummies of regressions of the log offense rates on timedummies county xed effects and controls for age distribution (using the percentage of the populationin ve different age groups) the sex composition the percentage of the population that is black and thepercentage that is neither white nor black The sample consists of 705 counties with a mean populationgreater than 25000 and with at least sixteen out of nineteen years of data (1979ndash1997) The countypopulation means were used as weights

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 47

accentuated as the adjusted rate rose by over 47 until1993 and then declined 12 through 1997 The individualproperty and violent crimes (adjusted for changes in demo-graphics) are depicted in gure 4 and 5

Figure 3 demonstrates that changes in demographicsexplain much of the decline in both types of crime duringthe 1990s Without these controls the unadjusted propertyand violent crime indices peaked in 1991 and 1992 respec-tively With these controls they peaked later in 1994 and1993 respectively (at 29 and 47 higher than in 1979)From 1993 to 1997 adjusted property crime decreased by76 and adjusted violent crime by 123 Even as of 1997the adjusted property and violent crime rates were stilllarger (by 21 and 35 respectively) than they were in1979 Although the trends have reversed since the mid-1990s the secular trend in crime over the entire period isclearly upward While this was happening the labor marketprospects for young unskilled men deteriorated Figure 6plots the average wages of non-college-educated maleworkers (workers with only a high school degree or less)

over time The average wage of non-college-educated mendeclined by a total of 23 from 1979 to 1993 and thenrebounded somewhat until 1997 This overall pattern isalmost the mirror image of the crime patterns and there-fore this paper seeks to establish whether a causal connec-tion can be established

The theory behind such a connection is simple a declinein the wage offer increases the relative payoff of criminalactivity thus inducing workers to substitute away from thelegal sector towards the illegal sector In addition a lowerwage offer may produce an income effect by increasing theneed to seek additional sources of income in possibly lessdesirable and more dangerous ways A lower wage alsoreduces the opportunity cost of serving time in prison13 Thedegree to which legal alternatives affect criminal behaviormay however vary by the type of crime Some crimes (suchas robbery larceny burglary and auto theft) can be used forself-enrichment whereas other crimes (murder rape andassault) are much less likely to yield material gains to theoffender14 Offenders of the latter crimes are much morelikely to be motivated by nonpecuniary considerations15

13 Also Lott (1992) argued that reputationa l sanctions are positivelycorrelated with the wage

14 For example in 1992 the average monetary loss was $483 $840$1278 and $4713 for larceny robbery burglary and auto theft respec-tively compared with average monetary losses of $27 and $89 for rapeand murder as reported in Crime in the United States 1992

15 Offenders who commit the latter crimes are more likely to derivebene ts from interdependencie s in utility with the victim This notion ofinterdependenc e of utility between offender and victim for certain crimesis supported by the fact that murder rape and assault occur frequentlybetween people who know each other whereas the victim and offenderhave no relationship in the vast majority of property crimes For offensescommitted in 1993 the offenders were classi ed as non-stranger s to thevictims in 742 of rapes 519 of assaults and 199 of robberies(1994 Sourcebook of Criminal Justice Statistics table 311 p 235)Historically murder victims knew their offenders (Supplementary Homi-

FIGURE 4mdashCOUNTY SAMPLE TRENDS IN ADJUSTED PROPERTY CRIMES

See notes to Figure 3

FIGURE 5mdashCOUNTY SAMPLE TRENDS IN ADJUSTED VIOLENT CRIMES

See notes to Figure 3

FIGURE 6mdashSTANDARDIZED WAGES AND UNEMPLOYMENT RATES

OF NON-COLLEGE-EDUCATED MEN

The data were computed from the CPS Non-college-educated mean are de ned as full-time menbetween the age of 18 and 65 Residuals were computed after controlling for a quartic in potentialexperience years of school (within non-college) race (black and nonwhite nonblack) Hispanicbackground region of residence and marital status Wages de ated to 1982ndash1984 100 dollars Meanresiduals for each year were standardized to the base year 1979

THE REVIEW OF ECONOMICS AND STATISTICS48

However it is important to note that only the most severecrime is reported in the UCR data when multiple crimes arecommitted in the same incident Therefore pecuniary mo-tives may lie behind many of the reported assaults when aproperty crime was also involved Holding everything elseconstant a reduction in legal opportunities should make onemore likely to engage in any form of criminal activityregardless of motives due to the reduced legal earnings lostwhile engaging in a criminal career and potentially servingin jail

Including the modest increase in the 1990s gure 6indicates that the non-college-educated male wage was 20lower in 1997 than in 1979 This overall trend represents alarge long-term decline in the earning prospects of lesseducated men In contrast gure 6 shows that the unem-ployment rate of non-college-educated men did not suffer along-term deterioration throughout the period Althoughless educated workers suffer the most unemployment un-employment rates generally follow a cyclical pattern thatby de nition traces out the business cycle In gure 6 theunemployment rate is the same in 1997 as it was in 1979although there was variation in the intervening yearsClearly the unemployment rate affects the labor marketprospects of less educated men but it is hard to discern along-term deterioration in their legal opportunities by look-ing at the overall trend in the unemployment rate Theoverall decline in the labor market prospects of less edu-cated men however is clearly shown by their wages

The data clearly show that the propensity to commitcrime moved inversely to the trends in the labor marketconditions for unskilled men These trends seem to berelated particularly because young unskilled men are themost likely to commit crime16 The goal of the remainingsections is to establish empirically whether the relationshipis causal

III County-Level Panel Analysis 1979ndash1997

This section analyzes a panel sample of 705 counties overnineteen years The data and trends were described insection II In each regression county xed effects controlfor much of the cross-sectional variation and yearly timedummies control for the national trends The county xedeffects control for unobserved county-level heterogeneitythat might be correlated with the county crime rate Remov-ing the national trend allows us to abstract from any corre-lation between the aggregate trends in crime and some otherunobserved aggregate determinant of crime Including con-trols for national trends also controls for aggregate trends in

reporting practices Given the strong inverse relationshipbetween the wage trends of less skilled men and the aggre-gate crime trends eliminating the aggregate trend tends tobias the results against nding a relationship between thetwo phenomena We expect that the labor market variablescan help explain the cross-sectional variation and the na-tional trends but we identify the effects of these variablesfrom the within-county deviations from the national trendsto avoid any spurious correlations17 Each speci cation alsocontrols for changes in the age sex and race composition ofthe county

Because the wages and unemployment rates of less edu-cated men are not available at the county level we use thesevariables measured at the state level to explain the countycrime rates Our wage measure is the mean state residualafter regressing individual wages from the CPS on educa-tion experience experience squared and controls for raceand marital status The residual state unemployment ratewas calculated similarly The construction of these variablesis described in detail in appendix B Using the residualsallows us to abstract from changes in our measures due tochanges in observable characteristics of workers and thusmore accurately re ects changes in the structure of wagesand unemployment However very similar results are ob-tained by using the levels rather than the residuals of thesevariables

To control for the general level of prosperity in the areawe use log income per capita in the state As shown in gure7 income per capita increased steadily since the early1980s The impact of this trend on crime however istheoretically unclear If the level of prosperity increasesthere is more material wealth to steal so crime could

cide Reports) During the 1990s this relationship changed and nowslightly less than half of the murder victims know their offenders Forexample in 1993 477 of all murders were committed by people whowere known to the victim 140 were committed by strangers and in393 of the cases the relationship between victim and offender wasunknown (Crime in the United States 1993 table 212 p 20)

16 Freeman (1996) reports that two-thirds of prison inmates in 1991 hadnot graduated from high school

17 Very similar OLS and IV results are obtained when we do not controlfor county xed effects A notable exception is for the larceny category inwhich unobserved county heterogeneit y reverses the sign for the OLScoef cients on our two measures for the labor market prospects ofunskilled workers However the IV coef cients for larceny have theldquoexpectedrdquo sign and are statistically signi cant

FIGURE 7mdashRETAIL WAGES AND INCOME PER CAPITA OVER TIME

Both variables were computed by the procedure described in Figure 3

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 49

increase However higher income individuals invest morein self-protection from criminals so crime may decrease18

The overall effect therefore is an empirical question Byincluding this measure in our regressions we answer thisquestion and control for any correlation between the labormarket prospects of less educated men and the overalleconomic prosperity of the area

Table 1 displays the coef cient estimates for the eco-nomic variables in our ldquocorerdquo speci cation The standarderrors throughout the analysis are corrected for a commonunobserved factor underlying crime in each state in eachyear All three economic variables are very signi cant forthe property and violent crime indices and every individualcrime rate except for rape Furthermore the coef cientshave the expected signs increases in the wages of non-college-educated men reduce the crime rate and increasesin the unemployment rate of non-college-educated menincrease the crime rate19 The results for income per capitaare quite uniformly positive and signi cant indicating thatimprovements in the overall economic condition of the areaincrease the amount of material wealth available to stealthus increasing crime rates

Although the coef cients are statistically signi cant wewould like to know whether their magnitudes are econom-ically signi cant The numbers underneath each standarderror indicate the ldquopredictedrdquo effects of each independentvariable on the crime rate based on the coef cient estimateand the mean change in the independent variable over twodifferent time periods The rst number indicates the pre-dicted effects between 1979 and 1993 when the adjusted

crime indices increased (See section II) The second num-ber is the predicted effect during 1993ndash1997 when crimefell From table 1 the 233 fall in the wages of unskilledmen from 1979 to 1993 ldquopredictrdquo a 125 increase inproperty crime (the coef cient 054 multiplied by 233)and a 251 increase in violent crime (the coef cient 108multiplied by 233) The 305 increase in unemploymentduring this early period ldquopredictedrdquo a 71 (the coef cient233 times 305) increase in property crime and a 38 (thecoef cient 126 times 305) increase in violent crime There-fore the non-college-educated wage explains 43 of the29 increase in adjusted property crime during this timeperiod and 53 of the 472 increase in adjusted violentcrime The unemployment rate of non-college-educatedmen explains 24 of the total increase in property crimeand 8 of the increase in violent crime Clearly the long-term trend in wages was the dominant factor on crimeduring this time period

The declining crime trends in the 1993ndash1997 period arebetter explained by the unemployment rate The adjustedproperty and violent crime rates fell by 76 and 123respectively From table 1 the 31 increase in the wages ofnon-college-educated men predict a decrease of 17 inproperty crime and 33 in violent crime The comparablepredictions for the 31 decline in the unemployment rateare decreases of 75 for property crime and 40 forviolent crime20 Although the predicted effects are quitesimilar for violent crime the declining crime rates in the1993ndash1997 period were more in uenced by the unemploy-ment rate than the non-college-educated wage Whether this

18 Lott and Mustard (1997) and Ayres and Levitt (1998) showed thatself-protectio n lowers crime by carrying concealed weapons and purchas-ing Lojack (an auto-thef t prevention system) respectivel y

19 These results are robust to the inclusion of the mean state residualwage for all workers as an additiona l control variable

20 Focusing exclusively on the effect of declining unemployment ratesduring the 1990s on crimes per youth Freeman and Rodgers (1999) foundsimilar results They found that a decrease of one percentage point inunemployment lowers crimes per youth by 15 whereas we nd that itlowers crimes per capita by 233

TABLE 1mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

060(013)a 141b 19

054(014)a 125b 17

145(024)a 337b 44

060(017)a 140b 19

041(014)a 94b 12

108(015)a 251b 33

131(018)a 304b 40

095(025)a 221b 29

096(022)a 224b 30

010(017)a 22b 03

State unemploymen t rateresidual for non-college-educated men

222(028)a 68b 71

233(030)a 71b 75

085(041)a 23b 27

310(034)a 95b 100

233(031)a 71b 75

126(030)a 38b 40

108(032)a 33b 35

080(045)a 25b 26

212(044)a 65b 68

012(035)a 05b 04

Log state income per capita 054(018)a 11b 40

048(019)a 01b 36

072(030)a 14b 53

060(024)a 12b 44

051(018)a 10b 37

096(019)a 19b 71

118(021)a 24b 87

091(031)a 18b 67

120(031)a 24b 89

086(023)a 17b 64

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 004 002 005 004 001 001 004 000 001

indicates that the coef cient is signi cant at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Numbers after ldquoardquo represent the ldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independent variable

between 1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independen t variable during the latter period 1993ndash1997Observations are for 705 counties with at least 16 out of 19 years of data Regressions include county and time xed effects and county-level demographic controls (percentage of population age 10ndash19 age 20ndash29age 30ndash39 age 40ndash49 and age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions are weighted by the mean population size of each county Partial R2 are aftercontrolling for county and time xed effects and demographic controls

THE REVIEW OF ECONOMICS AND STATISTICS50

trend will continue is improbable because the recent cyclicaldrop is likely to be temporary and in the future unemploy-ment will continue to uctuate with the business cycle Thewage trends however can continue to improve and have alasting impact on the crime trends This is best exempli ed byexplaining the increases of 21 and 35 in the adjustedproperty and violent crime rates over the entire 1979ndash1997period The unemployment rate was virtually unchanged in1997 from 1979 and therefore explains none of the increase ineither crime index The 20 fall in non-college-educatedwages over the entire period predicts a 108 increase inproperty crime and a 216 increase in violent crime Thesepredictionsldquoexplainrdquo more than 50 of the long-term trend inboth indices illustrating just how much the long-term crimetrends are dominated by the wages of unskilled men as op-posed to their unemployment rate

To see if our wage and unemployment measures in table1 are picking up changes in the relative supplies of differenteducation groups we checked if the results are sensitive tothe inclusion of variables capturing the local educationdistribution The results are practically identical to those intable 1 and therefore are not presented21 Clearly theresults are not due to changes in the education distribution

The speci cation in table 2 includes our ldquocorerdquo economicvariables plus variables measuring the local level of crimedeterrence Three deterrence measures are used the county

arrest rate state expenditures per capita on police and statepolice employment per capita22 Missing values for arrestrates are more numerous than for offense rates so thesample is reduced to 371 counties that meet our sampleselection criteria In addition the police variables wereavailable for only the years 1979 to 1995 After includingthese deterrence variables the coef cients on the non-college-educated wage remain very signi cant for propertyand violent crime although the magnitudes drop a bit Theunemployment rate remains signi cant for property crimebut disappears for violent crime

The arrest rate has a large and signi cantly negative effectfor every classi cation of crime Because the numerator of thedependent variable appears in the denominator of the arrestrate (the arrest rate is de ned as the ratio of total arrests to totaloffenses) measurement error in the offense rate leads to adownward bias in the coef cient estimates of the arrest rates(ldquodivision biasrdquo)23 In addition the police size variables arelikely to be highly endogenous to the local crime rate exem-pli ed by the positive coef cient on police expenditures forevery crime Controlling for the endogeneity of these policevariables is quite complicated (Levitt 1997) and is not thefocus of this study Table 2 demonstrates that the resultsparticularly for the non-college-educated wage are generallyrobust to the inclusion of these deterrence measures as well asto the decrease in the sample To work with the broadestsample possible and avoid the endogeneity and ldquodivision-biasrdquoissues of these deterrence variables the remaining speci ca-tions exclude these variables

21 We included the percentage of male high school dropouts in the statepercentage of male high school graduates and percentage of men withsome college The property crime index coef cients (standard errors inparentheses ) were 053 (014) for the non-college-educate d male wageresidual and 215 (029) for the non-college-educate d unemployment rateresidual For the violent crime index the respective coef cients were

100 (015) and 129 (030) Compared to the results in table 1 whichexcluded the education distribution variables the coef cients are almostidentical in magnitude and signi cance as is also the case for theindividua l crime categories

22 Mustard (forthcoming ) showed that although conviction and sentenc-ing data are theoreticall y important they exist for only four or ve statesTherefore we cannot include such data in this analysis

23 Levitt (1995) analyzed this issue and why the relationship betweenarrest rates and offense rates is so strong

TABLE 2mdashOLS COUNTY-LEVEL PANEL REGRESSIONS USING THE ldquoCORE SPECIFICATIONrdquo PLUS ARREST RATES AND POLICE SIZE VARIABLES 1979ndash1995

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

052(015)

040(015)

140(027)

056(021)

022(015)

064(017)

085(022)

089(030)

069(023)

032(019)

State unemploymen t rateresidual for non-college-educated men

138(037)

132(037)

082(047)

219(043)

129(036)

018(033)

016(040)

100(057)

039(046)

029(037)

Log state income per capita 001(021)

007(021)

003(033)

005(027)

005(020)

002(020)

034(024)

114(030)

018(031)

042(024)

County arrest rate 0002(0002)

001(0002)

001(0001)

001(0003)

001(0001)

0004(00003)

0003(00003)

0002(00002)

0006(00005)

0004(0001)

Log state police expendituresper capita

028(010)

030(010)

051(0176)

034(011)

026(010)

036(011)

039(014)

023(016)

055(013)

004(010)

Log state police employmentper capita

004(014)

002(013)

026(020)

004(016)

007(012)

006(014)

001(016)

028(018)

014(018)

045(012)

Observations 5979 5979 5979 5979 5979 5979 5979 5979 5979 5979Partial R2 003 003 003 004 000 001 001 001 000 001

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 371 counties withat least sixteen out of seventeen years of data where the police force variables are available (1979ndash1995) County mean population is used as weights See notes to table 1 for other de nitions and controls usedin the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 51

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 3: crime rates and local labor market opportunities in the united states

1980 During the whole period violent crime rose by 32until 1991 and then steadily declined by 29 as of 1997Thus the pattern for violent crime is much more consistentwith the common perception of increasing crime throughthe 1980s and declining since the early 1990s11

In 1997 88 of all crime was property crime Thereforethe overall crime rate pattern in gure 1 is almost identicalto the property crime rate Consequently results for theoverall crime index are dominated by the results for theproperty crime index The property crime index is domi-nated by larceny (67) and burglary (213) and auto theftcomprises the remaining 117 Thus results for the prop-erty crime index will be heavily in uenced by larceny andburglary Violent crime is composed mainly of aggravatedassault (63) and robbery (305) whereas rape (6) andmurder (1) have only a minor in uence on the overallviolent crime rate However the seriousness of these lattertwo crimes gives them a disproportionate in uence oversocial welfare and public policy

The trends in our panel sample of 705 counties aredisplayed in gure 2 and are similar to the national trends in gure 1 The sample consists of all counties with a meanpopulation greater than 25000 between 1979 and 1997 andat least sixteen out of nineteen years of complete data Theaverage county population size is 248017 over the entireperiod which covers an average of almost 175 millionpeople per year The sample selection criteria were designedto capture a representative population while deleting thosecounties where the reporting accuracy is likely to be unre-liable The size of our sample and the trends displayed in

gure 2 (in comparison to gure 1) demonstrate that oursample is representative of the entire United States12

So far we have looked only at the raw crime data with noadjustments for changes in the demographic compositionswithin each county Figure 3 plots the property and violentcrime trends after adjusting for changes in the age sex andracial composition After controlling for these factors thetrends for both types of crime rose steadily throughout the1980s and declined after the early 1990s In 1994 theadjusted property crime rate hit a global peak at 23 higherthan the local peak in 1980 and 29 higher than it was atthe beginning of the period in 1979 The upward trend inunadjusted violent crime found before in gure 2 is now

11 Murder which has virtually no measurement error hit a global peakin 1980 at 102 murders per 100000 people and never got above 98which was the second peak in 1991

12 Levitt (1997) showed similar trends using the same data source for 59large cities

FIGURE 1mdashUNITED STATES NATIONAL TRENDS IN CRIME

INDICES 1979ndash1998

Plotted values are the log offense rate (offenses per 100000 people) relative to the year 1979 TheProperty Crime index is the sum of auto theft burglary and larceny The Violent Crime Index is the sumof aggravated assault robbery murder and rape The Overall Crime index is the sum of all property andviolent crimes Data come from the Uniform Crime Reports

FIGURE 2mdashCOUNTY SAMPLE TRENDS IN UNADJUSTED CRIME INDICES

Plotted values are the coef cients on the time dummies from regressions of the log offense rates ontime and county dummies for 705 counties with a mean population greater than 25000 and with at leastsixteen out of nineteen years of data (1979ndash1997) The county population means were used as weights

FIGURE 3mdashCOUNTY SAMPLE TRENDS IN ADJUSTED CRIME INDICES

Plotted values are the coef cients on the time dummies of regressions of the log offense rates on timedummies county xed effects and controls for age distribution (using the percentage of the populationin ve different age groups) the sex composition the percentage of the population that is black and thepercentage that is neither white nor black The sample consists of 705 counties with a mean populationgreater than 25000 and with at least sixteen out of nineteen years of data (1979ndash1997) The countypopulation means were used as weights

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 47

accentuated as the adjusted rate rose by over 47 until1993 and then declined 12 through 1997 The individualproperty and violent crimes (adjusted for changes in demo-graphics) are depicted in gure 4 and 5

Figure 3 demonstrates that changes in demographicsexplain much of the decline in both types of crime duringthe 1990s Without these controls the unadjusted propertyand violent crime indices peaked in 1991 and 1992 respec-tively With these controls they peaked later in 1994 and1993 respectively (at 29 and 47 higher than in 1979)From 1993 to 1997 adjusted property crime decreased by76 and adjusted violent crime by 123 Even as of 1997the adjusted property and violent crime rates were stilllarger (by 21 and 35 respectively) than they were in1979 Although the trends have reversed since the mid-1990s the secular trend in crime over the entire period isclearly upward While this was happening the labor marketprospects for young unskilled men deteriorated Figure 6plots the average wages of non-college-educated maleworkers (workers with only a high school degree or less)

over time The average wage of non-college-educated mendeclined by a total of 23 from 1979 to 1993 and thenrebounded somewhat until 1997 This overall pattern isalmost the mirror image of the crime patterns and there-fore this paper seeks to establish whether a causal connec-tion can be established

The theory behind such a connection is simple a declinein the wage offer increases the relative payoff of criminalactivity thus inducing workers to substitute away from thelegal sector towards the illegal sector In addition a lowerwage offer may produce an income effect by increasing theneed to seek additional sources of income in possibly lessdesirable and more dangerous ways A lower wage alsoreduces the opportunity cost of serving time in prison13 Thedegree to which legal alternatives affect criminal behaviormay however vary by the type of crime Some crimes (suchas robbery larceny burglary and auto theft) can be used forself-enrichment whereas other crimes (murder rape andassault) are much less likely to yield material gains to theoffender14 Offenders of the latter crimes are much morelikely to be motivated by nonpecuniary considerations15

13 Also Lott (1992) argued that reputationa l sanctions are positivelycorrelated with the wage

14 For example in 1992 the average monetary loss was $483 $840$1278 and $4713 for larceny robbery burglary and auto theft respec-tively compared with average monetary losses of $27 and $89 for rapeand murder as reported in Crime in the United States 1992

15 Offenders who commit the latter crimes are more likely to derivebene ts from interdependencie s in utility with the victim This notion ofinterdependenc e of utility between offender and victim for certain crimesis supported by the fact that murder rape and assault occur frequentlybetween people who know each other whereas the victim and offenderhave no relationship in the vast majority of property crimes For offensescommitted in 1993 the offenders were classi ed as non-stranger s to thevictims in 742 of rapes 519 of assaults and 199 of robberies(1994 Sourcebook of Criminal Justice Statistics table 311 p 235)Historically murder victims knew their offenders (Supplementary Homi-

FIGURE 4mdashCOUNTY SAMPLE TRENDS IN ADJUSTED PROPERTY CRIMES

See notes to Figure 3

FIGURE 5mdashCOUNTY SAMPLE TRENDS IN ADJUSTED VIOLENT CRIMES

See notes to Figure 3

FIGURE 6mdashSTANDARDIZED WAGES AND UNEMPLOYMENT RATES

OF NON-COLLEGE-EDUCATED MEN

The data were computed from the CPS Non-college-educated mean are de ned as full-time menbetween the age of 18 and 65 Residuals were computed after controlling for a quartic in potentialexperience years of school (within non-college) race (black and nonwhite nonblack) Hispanicbackground region of residence and marital status Wages de ated to 1982ndash1984 100 dollars Meanresiduals for each year were standardized to the base year 1979

THE REVIEW OF ECONOMICS AND STATISTICS48

However it is important to note that only the most severecrime is reported in the UCR data when multiple crimes arecommitted in the same incident Therefore pecuniary mo-tives may lie behind many of the reported assaults when aproperty crime was also involved Holding everything elseconstant a reduction in legal opportunities should make onemore likely to engage in any form of criminal activityregardless of motives due to the reduced legal earnings lostwhile engaging in a criminal career and potentially servingin jail

Including the modest increase in the 1990s gure 6indicates that the non-college-educated male wage was 20lower in 1997 than in 1979 This overall trend represents alarge long-term decline in the earning prospects of lesseducated men In contrast gure 6 shows that the unem-ployment rate of non-college-educated men did not suffer along-term deterioration throughout the period Althoughless educated workers suffer the most unemployment un-employment rates generally follow a cyclical pattern thatby de nition traces out the business cycle In gure 6 theunemployment rate is the same in 1997 as it was in 1979although there was variation in the intervening yearsClearly the unemployment rate affects the labor marketprospects of less educated men but it is hard to discern along-term deterioration in their legal opportunities by look-ing at the overall trend in the unemployment rate Theoverall decline in the labor market prospects of less edu-cated men however is clearly shown by their wages

The data clearly show that the propensity to commitcrime moved inversely to the trends in the labor marketconditions for unskilled men These trends seem to berelated particularly because young unskilled men are themost likely to commit crime16 The goal of the remainingsections is to establish empirically whether the relationshipis causal

III County-Level Panel Analysis 1979ndash1997

This section analyzes a panel sample of 705 counties overnineteen years The data and trends were described insection II In each regression county xed effects controlfor much of the cross-sectional variation and yearly timedummies control for the national trends The county xedeffects control for unobserved county-level heterogeneitythat might be correlated with the county crime rate Remov-ing the national trend allows us to abstract from any corre-lation between the aggregate trends in crime and some otherunobserved aggregate determinant of crime Including con-trols for national trends also controls for aggregate trends in

reporting practices Given the strong inverse relationshipbetween the wage trends of less skilled men and the aggre-gate crime trends eliminating the aggregate trend tends tobias the results against nding a relationship between thetwo phenomena We expect that the labor market variablescan help explain the cross-sectional variation and the na-tional trends but we identify the effects of these variablesfrom the within-county deviations from the national trendsto avoid any spurious correlations17 Each speci cation alsocontrols for changes in the age sex and race composition ofthe county

Because the wages and unemployment rates of less edu-cated men are not available at the county level we use thesevariables measured at the state level to explain the countycrime rates Our wage measure is the mean state residualafter regressing individual wages from the CPS on educa-tion experience experience squared and controls for raceand marital status The residual state unemployment ratewas calculated similarly The construction of these variablesis described in detail in appendix B Using the residualsallows us to abstract from changes in our measures due tochanges in observable characteristics of workers and thusmore accurately re ects changes in the structure of wagesand unemployment However very similar results are ob-tained by using the levels rather than the residuals of thesevariables

To control for the general level of prosperity in the areawe use log income per capita in the state As shown in gure7 income per capita increased steadily since the early1980s The impact of this trend on crime however istheoretically unclear If the level of prosperity increasesthere is more material wealth to steal so crime could

cide Reports) During the 1990s this relationship changed and nowslightly less than half of the murder victims know their offenders Forexample in 1993 477 of all murders were committed by people whowere known to the victim 140 were committed by strangers and in393 of the cases the relationship between victim and offender wasunknown (Crime in the United States 1993 table 212 p 20)

16 Freeman (1996) reports that two-thirds of prison inmates in 1991 hadnot graduated from high school

17 Very similar OLS and IV results are obtained when we do not controlfor county xed effects A notable exception is for the larceny category inwhich unobserved county heterogeneit y reverses the sign for the OLScoef cients on our two measures for the labor market prospects ofunskilled workers However the IV coef cients for larceny have theldquoexpectedrdquo sign and are statistically signi cant

FIGURE 7mdashRETAIL WAGES AND INCOME PER CAPITA OVER TIME

Both variables were computed by the procedure described in Figure 3

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 49

increase However higher income individuals invest morein self-protection from criminals so crime may decrease18

The overall effect therefore is an empirical question Byincluding this measure in our regressions we answer thisquestion and control for any correlation between the labormarket prospects of less educated men and the overalleconomic prosperity of the area

Table 1 displays the coef cient estimates for the eco-nomic variables in our ldquocorerdquo speci cation The standarderrors throughout the analysis are corrected for a commonunobserved factor underlying crime in each state in eachyear All three economic variables are very signi cant forthe property and violent crime indices and every individualcrime rate except for rape Furthermore the coef cientshave the expected signs increases in the wages of non-college-educated men reduce the crime rate and increasesin the unemployment rate of non-college-educated menincrease the crime rate19 The results for income per capitaare quite uniformly positive and signi cant indicating thatimprovements in the overall economic condition of the areaincrease the amount of material wealth available to stealthus increasing crime rates

Although the coef cients are statistically signi cant wewould like to know whether their magnitudes are econom-ically signi cant The numbers underneath each standarderror indicate the ldquopredictedrdquo effects of each independentvariable on the crime rate based on the coef cient estimateand the mean change in the independent variable over twodifferent time periods The rst number indicates the pre-dicted effects between 1979 and 1993 when the adjusted

crime indices increased (See section II) The second num-ber is the predicted effect during 1993ndash1997 when crimefell From table 1 the 233 fall in the wages of unskilledmen from 1979 to 1993 ldquopredictrdquo a 125 increase inproperty crime (the coef cient 054 multiplied by 233)and a 251 increase in violent crime (the coef cient 108multiplied by 233) The 305 increase in unemploymentduring this early period ldquopredictedrdquo a 71 (the coef cient233 times 305) increase in property crime and a 38 (thecoef cient 126 times 305) increase in violent crime There-fore the non-college-educated wage explains 43 of the29 increase in adjusted property crime during this timeperiod and 53 of the 472 increase in adjusted violentcrime The unemployment rate of non-college-educatedmen explains 24 of the total increase in property crimeand 8 of the increase in violent crime Clearly the long-term trend in wages was the dominant factor on crimeduring this time period

The declining crime trends in the 1993ndash1997 period arebetter explained by the unemployment rate The adjustedproperty and violent crime rates fell by 76 and 123respectively From table 1 the 31 increase in the wages ofnon-college-educated men predict a decrease of 17 inproperty crime and 33 in violent crime The comparablepredictions for the 31 decline in the unemployment rateare decreases of 75 for property crime and 40 forviolent crime20 Although the predicted effects are quitesimilar for violent crime the declining crime rates in the1993ndash1997 period were more in uenced by the unemploy-ment rate than the non-college-educated wage Whether this

18 Lott and Mustard (1997) and Ayres and Levitt (1998) showed thatself-protectio n lowers crime by carrying concealed weapons and purchas-ing Lojack (an auto-thef t prevention system) respectivel y

19 These results are robust to the inclusion of the mean state residualwage for all workers as an additiona l control variable

20 Focusing exclusively on the effect of declining unemployment ratesduring the 1990s on crimes per youth Freeman and Rodgers (1999) foundsimilar results They found that a decrease of one percentage point inunemployment lowers crimes per youth by 15 whereas we nd that itlowers crimes per capita by 233

TABLE 1mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

060(013)a 141b 19

054(014)a 125b 17

145(024)a 337b 44

060(017)a 140b 19

041(014)a 94b 12

108(015)a 251b 33

131(018)a 304b 40

095(025)a 221b 29

096(022)a 224b 30

010(017)a 22b 03

State unemploymen t rateresidual for non-college-educated men

222(028)a 68b 71

233(030)a 71b 75

085(041)a 23b 27

310(034)a 95b 100

233(031)a 71b 75

126(030)a 38b 40

108(032)a 33b 35

080(045)a 25b 26

212(044)a 65b 68

012(035)a 05b 04

Log state income per capita 054(018)a 11b 40

048(019)a 01b 36

072(030)a 14b 53

060(024)a 12b 44

051(018)a 10b 37

096(019)a 19b 71

118(021)a 24b 87

091(031)a 18b 67

120(031)a 24b 89

086(023)a 17b 64

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 004 002 005 004 001 001 004 000 001

indicates that the coef cient is signi cant at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Numbers after ldquoardquo represent the ldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independent variable

between 1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independen t variable during the latter period 1993ndash1997Observations are for 705 counties with at least 16 out of 19 years of data Regressions include county and time xed effects and county-level demographic controls (percentage of population age 10ndash19 age 20ndash29age 30ndash39 age 40ndash49 and age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions are weighted by the mean population size of each county Partial R2 are aftercontrolling for county and time xed effects and demographic controls

THE REVIEW OF ECONOMICS AND STATISTICS50

trend will continue is improbable because the recent cyclicaldrop is likely to be temporary and in the future unemploy-ment will continue to uctuate with the business cycle Thewage trends however can continue to improve and have alasting impact on the crime trends This is best exempli ed byexplaining the increases of 21 and 35 in the adjustedproperty and violent crime rates over the entire 1979ndash1997period The unemployment rate was virtually unchanged in1997 from 1979 and therefore explains none of the increase ineither crime index The 20 fall in non-college-educatedwages over the entire period predicts a 108 increase inproperty crime and a 216 increase in violent crime Thesepredictionsldquoexplainrdquo more than 50 of the long-term trend inboth indices illustrating just how much the long-term crimetrends are dominated by the wages of unskilled men as op-posed to their unemployment rate

To see if our wage and unemployment measures in table1 are picking up changes in the relative supplies of differenteducation groups we checked if the results are sensitive tothe inclusion of variables capturing the local educationdistribution The results are practically identical to those intable 1 and therefore are not presented21 Clearly theresults are not due to changes in the education distribution

The speci cation in table 2 includes our ldquocorerdquo economicvariables plus variables measuring the local level of crimedeterrence Three deterrence measures are used the county

arrest rate state expenditures per capita on police and statepolice employment per capita22 Missing values for arrestrates are more numerous than for offense rates so thesample is reduced to 371 counties that meet our sampleselection criteria In addition the police variables wereavailable for only the years 1979 to 1995 After includingthese deterrence variables the coef cients on the non-college-educated wage remain very signi cant for propertyand violent crime although the magnitudes drop a bit Theunemployment rate remains signi cant for property crimebut disappears for violent crime

The arrest rate has a large and signi cantly negative effectfor every classi cation of crime Because the numerator of thedependent variable appears in the denominator of the arrestrate (the arrest rate is de ned as the ratio of total arrests to totaloffenses) measurement error in the offense rate leads to adownward bias in the coef cient estimates of the arrest rates(ldquodivision biasrdquo)23 In addition the police size variables arelikely to be highly endogenous to the local crime rate exem-pli ed by the positive coef cient on police expenditures forevery crime Controlling for the endogeneity of these policevariables is quite complicated (Levitt 1997) and is not thefocus of this study Table 2 demonstrates that the resultsparticularly for the non-college-educated wage are generallyrobust to the inclusion of these deterrence measures as well asto the decrease in the sample To work with the broadestsample possible and avoid the endogeneity and ldquodivision-biasrdquoissues of these deterrence variables the remaining speci ca-tions exclude these variables

21 We included the percentage of male high school dropouts in the statepercentage of male high school graduates and percentage of men withsome college The property crime index coef cients (standard errors inparentheses ) were 053 (014) for the non-college-educate d male wageresidual and 215 (029) for the non-college-educate d unemployment rateresidual For the violent crime index the respective coef cients were

100 (015) and 129 (030) Compared to the results in table 1 whichexcluded the education distribution variables the coef cients are almostidentical in magnitude and signi cance as is also the case for theindividua l crime categories

22 Mustard (forthcoming ) showed that although conviction and sentenc-ing data are theoreticall y important they exist for only four or ve statesTherefore we cannot include such data in this analysis

23 Levitt (1995) analyzed this issue and why the relationship betweenarrest rates and offense rates is so strong

TABLE 2mdashOLS COUNTY-LEVEL PANEL REGRESSIONS USING THE ldquoCORE SPECIFICATIONrdquo PLUS ARREST RATES AND POLICE SIZE VARIABLES 1979ndash1995

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

052(015)

040(015)

140(027)

056(021)

022(015)

064(017)

085(022)

089(030)

069(023)

032(019)

State unemploymen t rateresidual for non-college-educated men

138(037)

132(037)

082(047)

219(043)

129(036)

018(033)

016(040)

100(057)

039(046)

029(037)

Log state income per capita 001(021)

007(021)

003(033)

005(027)

005(020)

002(020)

034(024)

114(030)

018(031)

042(024)

County arrest rate 0002(0002)

001(0002)

001(0001)

001(0003)

001(0001)

0004(00003)

0003(00003)

0002(00002)

0006(00005)

0004(0001)

Log state police expendituresper capita

028(010)

030(010)

051(0176)

034(011)

026(010)

036(011)

039(014)

023(016)

055(013)

004(010)

Log state police employmentper capita

004(014)

002(013)

026(020)

004(016)

007(012)

006(014)

001(016)

028(018)

014(018)

045(012)

Observations 5979 5979 5979 5979 5979 5979 5979 5979 5979 5979Partial R2 003 003 003 004 000 001 001 001 000 001

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 371 counties withat least sixteen out of seventeen years of data where the police force variables are available (1979ndash1995) County mean population is used as weights See notes to table 1 for other de nitions and controls usedin the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 51

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 4: crime rates and local labor market opportunities in the united states

accentuated as the adjusted rate rose by over 47 until1993 and then declined 12 through 1997 The individualproperty and violent crimes (adjusted for changes in demo-graphics) are depicted in gure 4 and 5

Figure 3 demonstrates that changes in demographicsexplain much of the decline in both types of crime duringthe 1990s Without these controls the unadjusted propertyand violent crime indices peaked in 1991 and 1992 respec-tively With these controls they peaked later in 1994 and1993 respectively (at 29 and 47 higher than in 1979)From 1993 to 1997 adjusted property crime decreased by76 and adjusted violent crime by 123 Even as of 1997the adjusted property and violent crime rates were stilllarger (by 21 and 35 respectively) than they were in1979 Although the trends have reversed since the mid-1990s the secular trend in crime over the entire period isclearly upward While this was happening the labor marketprospects for young unskilled men deteriorated Figure 6plots the average wages of non-college-educated maleworkers (workers with only a high school degree or less)

over time The average wage of non-college-educated mendeclined by a total of 23 from 1979 to 1993 and thenrebounded somewhat until 1997 This overall pattern isalmost the mirror image of the crime patterns and there-fore this paper seeks to establish whether a causal connec-tion can be established

The theory behind such a connection is simple a declinein the wage offer increases the relative payoff of criminalactivity thus inducing workers to substitute away from thelegal sector towards the illegal sector In addition a lowerwage offer may produce an income effect by increasing theneed to seek additional sources of income in possibly lessdesirable and more dangerous ways A lower wage alsoreduces the opportunity cost of serving time in prison13 Thedegree to which legal alternatives affect criminal behaviormay however vary by the type of crime Some crimes (suchas robbery larceny burglary and auto theft) can be used forself-enrichment whereas other crimes (murder rape andassault) are much less likely to yield material gains to theoffender14 Offenders of the latter crimes are much morelikely to be motivated by nonpecuniary considerations15

13 Also Lott (1992) argued that reputationa l sanctions are positivelycorrelated with the wage

14 For example in 1992 the average monetary loss was $483 $840$1278 and $4713 for larceny robbery burglary and auto theft respec-tively compared with average monetary losses of $27 and $89 for rapeand murder as reported in Crime in the United States 1992

15 Offenders who commit the latter crimes are more likely to derivebene ts from interdependencie s in utility with the victim This notion ofinterdependenc e of utility between offender and victim for certain crimesis supported by the fact that murder rape and assault occur frequentlybetween people who know each other whereas the victim and offenderhave no relationship in the vast majority of property crimes For offensescommitted in 1993 the offenders were classi ed as non-stranger s to thevictims in 742 of rapes 519 of assaults and 199 of robberies(1994 Sourcebook of Criminal Justice Statistics table 311 p 235)Historically murder victims knew their offenders (Supplementary Homi-

FIGURE 4mdashCOUNTY SAMPLE TRENDS IN ADJUSTED PROPERTY CRIMES

See notes to Figure 3

FIGURE 5mdashCOUNTY SAMPLE TRENDS IN ADJUSTED VIOLENT CRIMES

See notes to Figure 3

FIGURE 6mdashSTANDARDIZED WAGES AND UNEMPLOYMENT RATES

OF NON-COLLEGE-EDUCATED MEN

The data were computed from the CPS Non-college-educated mean are de ned as full-time menbetween the age of 18 and 65 Residuals were computed after controlling for a quartic in potentialexperience years of school (within non-college) race (black and nonwhite nonblack) Hispanicbackground region of residence and marital status Wages de ated to 1982ndash1984 100 dollars Meanresiduals for each year were standardized to the base year 1979

THE REVIEW OF ECONOMICS AND STATISTICS48

However it is important to note that only the most severecrime is reported in the UCR data when multiple crimes arecommitted in the same incident Therefore pecuniary mo-tives may lie behind many of the reported assaults when aproperty crime was also involved Holding everything elseconstant a reduction in legal opportunities should make onemore likely to engage in any form of criminal activityregardless of motives due to the reduced legal earnings lostwhile engaging in a criminal career and potentially servingin jail

Including the modest increase in the 1990s gure 6indicates that the non-college-educated male wage was 20lower in 1997 than in 1979 This overall trend represents alarge long-term decline in the earning prospects of lesseducated men In contrast gure 6 shows that the unem-ployment rate of non-college-educated men did not suffer along-term deterioration throughout the period Althoughless educated workers suffer the most unemployment un-employment rates generally follow a cyclical pattern thatby de nition traces out the business cycle In gure 6 theunemployment rate is the same in 1997 as it was in 1979although there was variation in the intervening yearsClearly the unemployment rate affects the labor marketprospects of less educated men but it is hard to discern along-term deterioration in their legal opportunities by look-ing at the overall trend in the unemployment rate Theoverall decline in the labor market prospects of less edu-cated men however is clearly shown by their wages

The data clearly show that the propensity to commitcrime moved inversely to the trends in the labor marketconditions for unskilled men These trends seem to berelated particularly because young unskilled men are themost likely to commit crime16 The goal of the remainingsections is to establish empirically whether the relationshipis causal

III County-Level Panel Analysis 1979ndash1997

This section analyzes a panel sample of 705 counties overnineteen years The data and trends were described insection II In each regression county xed effects controlfor much of the cross-sectional variation and yearly timedummies control for the national trends The county xedeffects control for unobserved county-level heterogeneitythat might be correlated with the county crime rate Remov-ing the national trend allows us to abstract from any corre-lation between the aggregate trends in crime and some otherunobserved aggregate determinant of crime Including con-trols for national trends also controls for aggregate trends in

reporting practices Given the strong inverse relationshipbetween the wage trends of less skilled men and the aggre-gate crime trends eliminating the aggregate trend tends tobias the results against nding a relationship between thetwo phenomena We expect that the labor market variablescan help explain the cross-sectional variation and the na-tional trends but we identify the effects of these variablesfrom the within-county deviations from the national trendsto avoid any spurious correlations17 Each speci cation alsocontrols for changes in the age sex and race composition ofthe county

Because the wages and unemployment rates of less edu-cated men are not available at the county level we use thesevariables measured at the state level to explain the countycrime rates Our wage measure is the mean state residualafter regressing individual wages from the CPS on educa-tion experience experience squared and controls for raceand marital status The residual state unemployment ratewas calculated similarly The construction of these variablesis described in detail in appendix B Using the residualsallows us to abstract from changes in our measures due tochanges in observable characteristics of workers and thusmore accurately re ects changes in the structure of wagesand unemployment However very similar results are ob-tained by using the levels rather than the residuals of thesevariables

To control for the general level of prosperity in the areawe use log income per capita in the state As shown in gure7 income per capita increased steadily since the early1980s The impact of this trend on crime however istheoretically unclear If the level of prosperity increasesthere is more material wealth to steal so crime could

cide Reports) During the 1990s this relationship changed and nowslightly less than half of the murder victims know their offenders Forexample in 1993 477 of all murders were committed by people whowere known to the victim 140 were committed by strangers and in393 of the cases the relationship between victim and offender wasunknown (Crime in the United States 1993 table 212 p 20)

16 Freeman (1996) reports that two-thirds of prison inmates in 1991 hadnot graduated from high school

17 Very similar OLS and IV results are obtained when we do not controlfor county xed effects A notable exception is for the larceny category inwhich unobserved county heterogeneit y reverses the sign for the OLScoef cients on our two measures for the labor market prospects ofunskilled workers However the IV coef cients for larceny have theldquoexpectedrdquo sign and are statistically signi cant

FIGURE 7mdashRETAIL WAGES AND INCOME PER CAPITA OVER TIME

Both variables were computed by the procedure described in Figure 3

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 49

increase However higher income individuals invest morein self-protection from criminals so crime may decrease18

The overall effect therefore is an empirical question Byincluding this measure in our regressions we answer thisquestion and control for any correlation between the labormarket prospects of less educated men and the overalleconomic prosperity of the area

Table 1 displays the coef cient estimates for the eco-nomic variables in our ldquocorerdquo speci cation The standarderrors throughout the analysis are corrected for a commonunobserved factor underlying crime in each state in eachyear All three economic variables are very signi cant forthe property and violent crime indices and every individualcrime rate except for rape Furthermore the coef cientshave the expected signs increases in the wages of non-college-educated men reduce the crime rate and increasesin the unemployment rate of non-college-educated menincrease the crime rate19 The results for income per capitaare quite uniformly positive and signi cant indicating thatimprovements in the overall economic condition of the areaincrease the amount of material wealth available to stealthus increasing crime rates

Although the coef cients are statistically signi cant wewould like to know whether their magnitudes are econom-ically signi cant The numbers underneath each standarderror indicate the ldquopredictedrdquo effects of each independentvariable on the crime rate based on the coef cient estimateand the mean change in the independent variable over twodifferent time periods The rst number indicates the pre-dicted effects between 1979 and 1993 when the adjusted

crime indices increased (See section II) The second num-ber is the predicted effect during 1993ndash1997 when crimefell From table 1 the 233 fall in the wages of unskilledmen from 1979 to 1993 ldquopredictrdquo a 125 increase inproperty crime (the coef cient 054 multiplied by 233)and a 251 increase in violent crime (the coef cient 108multiplied by 233) The 305 increase in unemploymentduring this early period ldquopredictedrdquo a 71 (the coef cient233 times 305) increase in property crime and a 38 (thecoef cient 126 times 305) increase in violent crime There-fore the non-college-educated wage explains 43 of the29 increase in adjusted property crime during this timeperiod and 53 of the 472 increase in adjusted violentcrime The unemployment rate of non-college-educatedmen explains 24 of the total increase in property crimeand 8 of the increase in violent crime Clearly the long-term trend in wages was the dominant factor on crimeduring this time period

The declining crime trends in the 1993ndash1997 period arebetter explained by the unemployment rate The adjustedproperty and violent crime rates fell by 76 and 123respectively From table 1 the 31 increase in the wages ofnon-college-educated men predict a decrease of 17 inproperty crime and 33 in violent crime The comparablepredictions for the 31 decline in the unemployment rateare decreases of 75 for property crime and 40 forviolent crime20 Although the predicted effects are quitesimilar for violent crime the declining crime rates in the1993ndash1997 period were more in uenced by the unemploy-ment rate than the non-college-educated wage Whether this

18 Lott and Mustard (1997) and Ayres and Levitt (1998) showed thatself-protectio n lowers crime by carrying concealed weapons and purchas-ing Lojack (an auto-thef t prevention system) respectivel y

19 These results are robust to the inclusion of the mean state residualwage for all workers as an additiona l control variable

20 Focusing exclusively on the effect of declining unemployment ratesduring the 1990s on crimes per youth Freeman and Rodgers (1999) foundsimilar results They found that a decrease of one percentage point inunemployment lowers crimes per youth by 15 whereas we nd that itlowers crimes per capita by 233

TABLE 1mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

060(013)a 141b 19

054(014)a 125b 17

145(024)a 337b 44

060(017)a 140b 19

041(014)a 94b 12

108(015)a 251b 33

131(018)a 304b 40

095(025)a 221b 29

096(022)a 224b 30

010(017)a 22b 03

State unemploymen t rateresidual for non-college-educated men

222(028)a 68b 71

233(030)a 71b 75

085(041)a 23b 27

310(034)a 95b 100

233(031)a 71b 75

126(030)a 38b 40

108(032)a 33b 35

080(045)a 25b 26

212(044)a 65b 68

012(035)a 05b 04

Log state income per capita 054(018)a 11b 40

048(019)a 01b 36

072(030)a 14b 53

060(024)a 12b 44

051(018)a 10b 37

096(019)a 19b 71

118(021)a 24b 87

091(031)a 18b 67

120(031)a 24b 89

086(023)a 17b 64

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 004 002 005 004 001 001 004 000 001

indicates that the coef cient is signi cant at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Numbers after ldquoardquo represent the ldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independent variable

between 1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independen t variable during the latter period 1993ndash1997Observations are for 705 counties with at least 16 out of 19 years of data Regressions include county and time xed effects and county-level demographic controls (percentage of population age 10ndash19 age 20ndash29age 30ndash39 age 40ndash49 and age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions are weighted by the mean population size of each county Partial R2 are aftercontrolling for county and time xed effects and demographic controls

THE REVIEW OF ECONOMICS AND STATISTICS50

trend will continue is improbable because the recent cyclicaldrop is likely to be temporary and in the future unemploy-ment will continue to uctuate with the business cycle Thewage trends however can continue to improve and have alasting impact on the crime trends This is best exempli ed byexplaining the increases of 21 and 35 in the adjustedproperty and violent crime rates over the entire 1979ndash1997period The unemployment rate was virtually unchanged in1997 from 1979 and therefore explains none of the increase ineither crime index The 20 fall in non-college-educatedwages over the entire period predicts a 108 increase inproperty crime and a 216 increase in violent crime Thesepredictionsldquoexplainrdquo more than 50 of the long-term trend inboth indices illustrating just how much the long-term crimetrends are dominated by the wages of unskilled men as op-posed to their unemployment rate

To see if our wage and unemployment measures in table1 are picking up changes in the relative supplies of differenteducation groups we checked if the results are sensitive tothe inclusion of variables capturing the local educationdistribution The results are practically identical to those intable 1 and therefore are not presented21 Clearly theresults are not due to changes in the education distribution

The speci cation in table 2 includes our ldquocorerdquo economicvariables plus variables measuring the local level of crimedeterrence Three deterrence measures are used the county

arrest rate state expenditures per capita on police and statepolice employment per capita22 Missing values for arrestrates are more numerous than for offense rates so thesample is reduced to 371 counties that meet our sampleselection criteria In addition the police variables wereavailable for only the years 1979 to 1995 After includingthese deterrence variables the coef cients on the non-college-educated wage remain very signi cant for propertyand violent crime although the magnitudes drop a bit Theunemployment rate remains signi cant for property crimebut disappears for violent crime

The arrest rate has a large and signi cantly negative effectfor every classi cation of crime Because the numerator of thedependent variable appears in the denominator of the arrestrate (the arrest rate is de ned as the ratio of total arrests to totaloffenses) measurement error in the offense rate leads to adownward bias in the coef cient estimates of the arrest rates(ldquodivision biasrdquo)23 In addition the police size variables arelikely to be highly endogenous to the local crime rate exem-pli ed by the positive coef cient on police expenditures forevery crime Controlling for the endogeneity of these policevariables is quite complicated (Levitt 1997) and is not thefocus of this study Table 2 demonstrates that the resultsparticularly for the non-college-educated wage are generallyrobust to the inclusion of these deterrence measures as well asto the decrease in the sample To work with the broadestsample possible and avoid the endogeneity and ldquodivision-biasrdquoissues of these deterrence variables the remaining speci ca-tions exclude these variables

21 We included the percentage of male high school dropouts in the statepercentage of male high school graduates and percentage of men withsome college The property crime index coef cients (standard errors inparentheses ) were 053 (014) for the non-college-educate d male wageresidual and 215 (029) for the non-college-educate d unemployment rateresidual For the violent crime index the respective coef cients were

100 (015) and 129 (030) Compared to the results in table 1 whichexcluded the education distribution variables the coef cients are almostidentical in magnitude and signi cance as is also the case for theindividua l crime categories

22 Mustard (forthcoming ) showed that although conviction and sentenc-ing data are theoreticall y important they exist for only four or ve statesTherefore we cannot include such data in this analysis

23 Levitt (1995) analyzed this issue and why the relationship betweenarrest rates and offense rates is so strong

TABLE 2mdashOLS COUNTY-LEVEL PANEL REGRESSIONS USING THE ldquoCORE SPECIFICATIONrdquo PLUS ARREST RATES AND POLICE SIZE VARIABLES 1979ndash1995

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

052(015)

040(015)

140(027)

056(021)

022(015)

064(017)

085(022)

089(030)

069(023)

032(019)

State unemploymen t rateresidual for non-college-educated men

138(037)

132(037)

082(047)

219(043)

129(036)

018(033)

016(040)

100(057)

039(046)

029(037)

Log state income per capita 001(021)

007(021)

003(033)

005(027)

005(020)

002(020)

034(024)

114(030)

018(031)

042(024)

County arrest rate 0002(0002)

001(0002)

001(0001)

001(0003)

001(0001)

0004(00003)

0003(00003)

0002(00002)

0006(00005)

0004(0001)

Log state police expendituresper capita

028(010)

030(010)

051(0176)

034(011)

026(010)

036(011)

039(014)

023(016)

055(013)

004(010)

Log state police employmentper capita

004(014)

002(013)

026(020)

004(016)

007(012)

006(014)

001(016)

028(018)

014(018)

045(012)

Observations 5979 5979 5979 5979 5979 5979 5979 5979 5979 5979Partial R2 003 003 003 004 000 001 001 001 000 001

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 371 counties withat least sixteen out of seventeen years of data where the police force variables are available (1979ndash1995) County mean population is used as weights See notes to table 1 for other de nitions and controls usedin the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 51

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 5: crime rates and local labor market opportunities in the united states

However it is important to note that only the most severecrime is reported in the UCR data when multiple crimes arecommitted in the same incident Therefore pecuniary mo-tives may lie behind many of the reported assaults when aproperty crime was also involved Holding everything elseconstant a reduction in legal opportunities should make onemore likely to engage in any form of criminal activityregardless of motives due to the reduced legal earnings lostwhile engaging in a criminal career and potentially servingin jail

Including the modest increase in the 1990s gure 6indicates that the non-college-educated male wage was 20lower in 1997 than in 1979 This overall trend represents alarge long-term decline in the earning prospects of lesseducated men In contrast gure 6 shows that the unem-ployment rate of non-college-educated men did not suffer along-term deterioration throughout the period Althoughless educated workers suffer the most unemployment un-employment rates generally follow a cyclical pattern thatby de nition traces out the business cycle In gure 6 theunemployment rate is the same in 1997 as it was in 1979although there was variation in the intervening yearsClearly the unemployment rate affects the labor marketprospects of less educated men but it is hard to discern along-term deterioration in their legal opportunities by look-ing at the overall trend in the unemployment rate Theoverall decline in the labor market prospects of less edu-cated men however is clearly shown by their wages

The data clearly show that the propensity to commitcrime moved inversely to the trends in the labor marketconditions for unskilled men These trends seem to berelated particularly because young unskilled men are themost likely to commit crime16 The goal of the remainingsections is to establish empirically whether the relationshipis causal

III County-Level Panel Analysis 1979ndash1997

This section analyzes a panel sample of 705 counties overnineteen years The data and trends were described insection II In each regression county xed effects controlfor much of the cross-sectional variation and yearly timedummies control for the national trends The county xedeffects control for unobserved county-level heterogeneitythat might be correlated with the county crime rate Remov-ing the national trend allows us to abstract from any corre-lation between the aggregate trends in crime and some otherunobserved aggregate determinant of crime Including con-trols for national trends also controls for aggregate trends in

reporting practices Given the strong inverse relationshipbetween the wage trends of less skilled men and the aggre-gate crime trends eliminating the aggregate trend tends tobias the results against nding a relationship between thetwo phenomena We expect that the labor market variablescan help explain the cross-sectional variation and the na-tional trends but we identify the effects of these variablesfrom the within-county deviations from the national trendsto avoid any spurious correlations17 Each speci cation alsocontrols for changes in the age sex and race composition ofthe county

Because the wages and unemployment rates of less edu-cated men are not available at the county level we use thesevariables measured at the state level to explain the countycrime rates Our wage measure is the mean state residualafter regressing individual wages from the CPS on educa-tion experience experience squared and controls for raceand marital status The residual state unemployment ratewas calculated similarly The construction of these variablesis described in detail in appendix B Using the residualsallows us to abstract from changes in our measures due tochanges in observable characteristics of workers and thusmore accurately re ects changes in the structure of wagesand unemployment However very similar results are ob-tained by using the levels rather than the residuals of thesevariables

To control for the general level of prosperity in the areawe use log income per capita in the state As shown in gure7 income per capita increased steadily since the early1980s The impact of this trend on crime however istheoretically unclear If the level of prosperity increasesthere is more material wealth to steal so crime could

cide Reports) During the 1990s this relationship changed and nowslightly less than half of the murder victims know their offenders Forexample in 1993 477 of all murders were committed by people whowere known to the victim 140 were committed by strangers and in393 of the cases the relationship between victim and offender wasunknown (Crime in the United States 1993 table 212 p 20)

16 Freeman (1996) reports that two-thirds of prison inmates in 1991 hadnot graduated from high school

17 Very similar OLS and IV results are obtained when we do not controlfor county xed effects A notable exception is for the larceny category inwhich unobserved county heterogeneit y reverses the sign for the OLScoef cients on our two measures for the labor market prospects ofunskilled workers However the IV coef cients for larceny have theldquoexpectedrdquo sign and are statistically signi cant

FIGURE 7mdashRETAIL WAGES AND INCOME PER CAPITA OVER TIME

Both variables were computed by the procedure described in Figure 3

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 49

increase However higher income individuals invest morein self-protection from criminals so crime may decrease18

The overall effect therefore is an empirical question Byincluding this measure in our regressions we answer thisquestion and control for any correlation between the labormarket prospects of less educated men and the overalleconomic prosperity of the area

Table 1 displays the coef cient estimates for the eco-nomic variables in our ldquocorerdquo speci cation The standarderrors throughout the analysis are corrected for a commonunobserved factor underlying crime in each state in eachyear All three economic variables are very signi cant forthe property and violent crime indices and every individualcrime rate except for rape Furthermore the coef cientshave the expected signs increases in the wages of non-college-educated men reduce the crime rate and increasesin the unemployment rate of non-college-educated menincrease the crime rate19 The results for income per capitaare quite uniformly positive and signi cant indicating thatimprovements in the overall economic condition of the areaincrease the amount of material wealth available to stealthus increasing crime rates

Although the coef cients are statistically signi cant wewould like to know whether their magnitudes are econom-ically signi cant The numbers underneath each standarderror indicate the ldquopredictedrdquo effects of each independentvariable on the crime rate based on the coef cient estimateand the mean change in the independent variable over twodifferent time periods The rst number indicates the pre-dicted effects between 1979 and 1993 when the adjusted

crime indices increased (See section II) The second num-ber is the predicted effect during 1993ndash1997 when crimefell From table 1 the 233 fall in the wages of unskilledmen from 1979 to 1993 ldquopredictrdquo a 125 increase inproperty crime (the coef cient 054 multiplied by 233)and a 251 increase in violent crime (the coef cient 108multiplied by 233) The 305 increase in unemploymentduring this early period ldquopredictedrdquo a 71 (the coef cient233 times 305) increase in property crime and a 38 (thecoef cient 126 times 305) increase in violent crime There-fore the non-college-educated wage explains 43 of the29 increase in adjusted property crime during this timeperiod and 53 of the 472 increase in adjusted violentcrime The unemployment rate of non-college-educatedmen explains 24 of the total increase in property crimeand 8 of the increase in violent crime Clearly the long-term trend in wages was the dominant factor on crimeduring this time period

The declining crime trends in the 1993ndash1997 period arebetter explained by the unemployment rate The adjustedproperty and violent crime rates fell by 76 and 123respectively From table 1 the 31 increase in the wages ofnon-college-educated men predict a decrease of 17 inproperty crime and 33 in violent crime The comparablepredictions for the 31 decline in the unemployment rateare decreases of 75 for property crime and 40 forviolent crime20 Although the predicted effects are quitesimilar for violent crime the declining crime rates in the1993ndash1997 period were more in uenced by the unemploy-ment rate than the non-college-educated wage Whether this

18 Lott and Mustard (1997) and Ayres and Levitt (1998) showed thatself-protectio n lowers crime by carrying concealed weapons and purchas-ing Lojack (an auto-thef t prevention system) respectivel y

19 These results are robust to the inclusion of the mean state residualwage for all workers as an additiona l control variable

20 Focusing exclusively on the effect of declining unemployment ratesduring the 1990s on crimes per youth Freeman and Rodgers (1999) foundsimilar results They found that a decrease of one percentage point inunemployment lowers crimes per youth by 15 whereas we nd that itlowers crimes per capita by 233

TABLE 1mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

060(013)a 141b 19

054(014)a 125b 17

145(024)a 337b 44

060(017)a 140b 19

041(014)a 94b 12

108(015)a 251b 33

131(018)a 304b 40

095(025)a 221b 29

096(022)a 224b 30

010(017)a 22b 03

State unemploymen t rateresidual for non-college-educated men

222(028)a 68b 71

233(030)a 71b 75

085(041)a 23b 27

310(034)a 95b 100

233(031)a 71b 75

126(030)a 38b 40

108(032)a 33b 35

080(045)a 25b 26

212(044)a 65b 68

012(035)a 05b 04

Log state income per capita 054(018)a 11b 40

048(019)a 01b 36

072(030)a 14b 53

060(024)a 12b 44

051(018)a 10b 37

096(019)a 19b 71

118(021)a 24b 87

091(031)a 18b 67

120(031)a 24b 89

086(023)a 17b 64

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 004 002 005 004 001 001 004 000 001

indicates that the coef cient is signi cant at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Numbers after ldquoardquo represent the ldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independent variable

between 1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independen t variable during the latter period 1993ndash1997Observations are for 705 counties with at least 16 out of 19 years of data Regressions include county and time xed effects and county-level demographic controls (percentage of population age 10ndash19 age 20ndash29age 30ndash39 age 40ndash49 and age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions are weighted by the mean population size of each county Partial R2 are aftercontrolling for county and time xed effects and demographic controls

THE REVIEW OF ECONOMICS AND STATISTICS50

trend will continue is improbable because the recent cyclicaldrop is likely to be temporary and in the future unemploy-ment will continue to uctuate with the business cycle Thewage trends however can continue to improve and have alasting impact on the crime trends This is best exempli ed byexplaining the increases of 21 and 35 in the adjustedproperty and violent crime rates over the entire 1979ndash1997period The unemployment rate was virtually unchanged in1997 from 1979 and therefore explains none of the increase ineither crime index The 20 fall in non-college-educatedwages over the entire period predicts a 108 increase inproperty crime and a 216 increase in violent crime Thesepredictionsldquoexplainrdquo more than 50 of the long-term trend inboth indices illustrating just how much the long-term crimetrends are dominated by the wages of unskilled men as op-posed to their unemployment rate

To see if our wage and unemployment measures in table1 are picking up changes in the relative supplies of differenteducation groups we checked if the results are sensitive tothe inclusion of variables capturing the local educationdistribution The results are practically identical to those intable 1 and therefore are not presented21 Clearly theresults are not due to changes in the education distribution

The speci cation in table 2 includes our ldquocorerdquo economicvariables plus variables measuring the local level of crimedeterrence Three deterrence measures are used the county

arrest rate state expenditures per capita on police and statepolice employment per capita22 Missing values for arrestrates are more numerous than for offense rates so thesample is reduced to 371 counties that meet our sampleselection criteria In addition the police variables wereavailable for only the years 1979 to 1995 After includingthese deterrence variables the coef cients on the non-college-educated wage remain very signi cant for propertyand violent crime although the magnitudes drop a bit Theunemployment rate remains signi cant for property crimebut disappears for violent crime

The arrest rate has a large and signi cantly negative effectfor every classi cation of crime Because the numerator of thedependent variable appears in the denominator of the arrestrate (the arrest rate is de ned as the ratio of total arrests to totaloffenses) measurement error in the offense rate leads to adownward bias in the coef cient estimates of the arrest rates(ldquodivision biasrdquo)23 In addition the police size variables arelikely to be highly endogenous to the local crime rate exem-pli ed by the positive coef cient on police expenditures forevery crime Controlling for the endogeneity of these policevariables is quite complicated (Levitt 1997) and is not thefocus of this study Table 2 demonstrates that the resultsparticularly for the non-college-educated wage are generallyrobust to the inclusion of these deterrence measures as well asto the decrease in the sample To work with the broadestsample possible and avoid the endogeneity and ldquodivision-biasrdquoissues of these deterrence variables the remaining speci ca-tions exclude these variables

21 We included the percentage of male high school dropouts in the statepercentage of male high school graduates and percentage of men withsome college The property crime index coef cients (standard errors inparentheses ) were 053 (014) for the non-college-educate d male wageresidual and 215 (029) for the non-college-educate d unemployment rateresidual For the violent crime index the respective coef cients were

100 (015) and 129 (030) Compared to the results in table 1 whichexcluded the education distribution variables the coef cients are almostidentical in magnitude and signi cance as is also the case for theindividua l crime categories

22 Mustard (forthcoming ) showed that although conviction and sentenc-ing data are theoreticall y important they exist for only four or ve statesTherefore we cannot include such data in this analysis

23 Levitt (1995) analyzed this issue and why the relationship betweenarrest rates and offense rates is so strong

TABLE 2mdashOLS COUNTY-LEVEL PANEL REGRESSIONS USING THE ldquoCORE SPECIFICATIONrdquo PLUS ARREST RATES AND POLICE SIZE VARIABLES 1979ndash1995

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

052(015)

040(015)

140(027)

056(021)

022(015)

064(017)

085(022)

089(030)

069(023)

032(019)

State unemploymen t rateresidual for non-college-educated men

138(037)

132(037)

082(047)

219(043)

129(036)

018(033)

016(040)

100(057)

039(046)

029(037)

Log state income per capita 001(021)

007(021)

003(033)

005(027)

005(020)

002(020)

034(024)

114(030)

018(031)

042(024)

County arrest rate 0002(0002)

001(0002)

001(0001)

001(0003)

001(0001)

0004(00003)

0003(00003)

0002(00002)

0006(00005)

0004(0001)

Log state police expendituresper capita

028(010)

030(010)

051(0176)

034(011)

026(010)

036(011)

039(014)

023(016)

055(013)

004(010)

Log state police employmentper capita

004(014)

002(013)

026(020)

004(016)

007(012)

006(014)

001(016)

028(018)

014(018)

045(012)

Observations 5979 5979 5979 5979 5979 5979 5979 5979 5979 5979Partial R2 003 003 003 004 000 001 001 001 000 001

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 371 counties withat least sixteen out of seventeen years of data where the police force variables are available (1979ndash1995) County mean population is used as weights See notes to table 1 for other de nitions and controls usedin the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 51

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 6: crime rates and local labor market opportunities in the united states

increase However higher income individuals invest morein self-protection from criminals so crime may decrease18

The overall effect therefore is an empirical question Byincluding this measure in our regressions we answer thisquestion and control for any correlation between the labormarket prospects of less educated men and the overalleconomic prosperity of the area

Table 1 displays the coef cient estimates for the eco-nomic variables in our ldquocorerdquo speci cation The standarderrors throughout the analysis are corrected for a commonunobserved factor underlying crime in each state in eachyear All three economic variables are very signi cant forthe property and violent crime indices and every individualcrime rate except for rape Furthermore the coef cientshave the expected signs increases in the wages of non-college-educated men reduce the crime rate and increasesin the unemployment rate of non-college-educated menincrease the crime rate19 The results for income per capitaare quite uniformly positive and signi cant indicating thatimprovements in the overall economic condition of the areaincrease the amount of material wealth available to stealthus increasing crime rates

Although the coef cients are statistically signi cant wewould like to know whether their magnitudes are econom-ically signi cant The numbers underneath each standarderror indicate the ldquopredictedrdquo effects of each independentvariable on the crime rate based on the coef cient estimateand the mean change in the independent variable over twodifferent time periods The rst number indicates the pre-dicted effects between 1979 and 1993 when the adjusted

crime indices increased (See section II) The second num-ber is the predicted effect during 1993ndash1997 when crimefell From table 1 the 233 fall in the wages of unskilledmen from 1979 to 1993 ldquopredictrdquo a 125 increase inproperty crime (the coef cient 054 multiplied by 233)and a 251 increase in violent crime (the coef cient 108multiplied by 233) The 305 increase in unemploymentduring this early period ldquopredictedrdquo a 71 (the coef cient233 times 305) increase in property crime and a 38 (thecoef cient 126 times 305) increase in violent crime There-fore the non-college-educated wage explains 43 of the29 increase in adjusted property crime during this timeperiod and 53 of the 472 increase in adjusted violentcrime The unemployment rate of non-college-educatedmen explains 24 of the total increase in property crimeand 8 of the increase in violent crime Clearly the long-term trend in wages was the dominant factor on crimeduring this time period

The declining crime trends in the 1993ndash1997 period arebetter explained by the unemployment rate The adjustedproperty and violent crime rates fell by 76 and 123respectively From table 1 the 31 increase in the wages ofnon-college-educated men predict a decrease of 17 inproperty crime and 33 in violent crime The comparablepredictions for the 31 decline in the unemployment rateare decreases of 75 for property crime and 40 forviolent crime20 Although the predicted effects are quitesimilar for violent crime the declining crime rates in the1993ndash1997 period were more in uenced by the unemploy-ment rate than the non-college-educated wage Whether this

18 Lott and Mustard (1997) and Ayres and Levitt (1998) showed thatself-protectio n lowers crime by carrying concealed weapons and purchas-ing Lojack (an auto-thef t prevention system) respectivel y

19 These results are robust to the inclusion of the mean state residualwage for all workers as an additiona l control variable

20 Focusing exclusively on the effect of declining unemployment ratesduring the 1990s on crimes per youth Freeman and Rodgers (1999) foundsimilar results They found that a decrease of one percentage point inunemployment lowers crimes per youth by 15 whereas we nd that itlowers crimes per capita by 233

TABLE 1mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

060(013)a 141b 19

054(014)a 125b 17

145(024)a 337b 44

060(017)a 140b 19

041(014)a 94b 12

108(015)a 251b 33

131(018)a 304b 40

095(025)a 221b 29

096(022)a 224b 30

010(017)a 22b 03

State unemploymen t rateresidual for non-college-educated men

222(028)a 68b 71

233(030)a 71b 75

085(041)a 23b 27

310(034)a 95b 100

233(031)a 71b 75

126(030)a 38b 40

108(032)a 33b 35

080(045)a 25b 26

212(044)a 65b 68

012(035)a 05b 04

Log state income per capita 054(018)a 11b 40

048(019)a 01b 36

072(030)a 14b 53

060(024)a 12b 44

051(018)a 10b 37

096(019)a 19b 71

118(021)a 24b 87

091(031)a 18b 67

120(031)a 24b 89

086(023)a 17b 64

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 004 002 005 004 001 001 004 000 001

indicates that the coef cient is signi cant at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Numbers after ldquoardquo represent the ldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independent variable

between 1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independen t variable during the latter period 1993ndash1997Observations are for 705 counties with at least 16 out of 19 years of data Regressions include county and time xed effects and county-level demographic controls (percentage of population age 10ndash19 age 20ndash29age 30ndash39 age 40ndash49 and age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions are weighted by the mean population size of each county Partial R2 are aftercontrolling for county and time xed effects and demographic controls

THE REVIEW OF ECONOMICS AND STATISTICS50

trend will continue is improbable because the recent cyclicaldrop is likely to be temporary and in the future unemploy-ment will continue to uctuate with the business cycle Thewage trends however can continue to improve and have alasting impact on the crime trends This is best exempli ed byexplaining the increases of 21 and 35 in the adjustedproperty and violent crime rates over the entire 1979ndash1997period The unemployment rate was virtually unchanged in1997 from 1979 and therefore explains none of the increase ineither crime index The 20 fall in non-college-educatedwages over the entire period predicts a 108 increase inproperty crime and a 216 increase in violent crime Thesepredictionsldquoexplainrdquo more than 50 of the long-term trend inboth indices illustrating just how much the long-term crimetrends are dominated by the wages of unskilled men as op-posed to their unemployment rate

To see if our wage and unemployment measures in table1 are picking up changes in the relative supplies of differenteducation groups we checked if the results are sensitive tothe inclusion of variables capturing the local educationdistribution The results are practically identical to those intable 1 and therefore are not presented21 Clearly theresults are not due to changes in the education distribution

The speci cation in table 2 includes our ldquocorerdquo economicvariables plus variables measuring the local level of crimedeterrence Three deterrence measures are used the county

arrest rate state expenditures per capita on police and statepolice employment per capita22 Missing values for arrestrates are more numerous than for offense rates so thesample is reduced to 371 counties that meet our sampleselection criteria In addition the police variables wereavailable for only the years 1979 to 1995 After includingthese deterrence variables the coef cients on the non-college-educated wage remain very signi cant for propertyand violent crime although the magnitudes drop a bit Theunemployment rate remains signi cant for property crimebut disappears for violent crime

The arrest rate has a large and signi cantly negative effectfor every classi cation of crime Because the numerator of thedependent variable appears in the denominator of the arrestrate (the arrest rate is de ned as the ratio of total arrests to totaloffenses) measurement error in the offense rate leads to adownward bias in the coef cient estimates of the arrest rates(ldquodivision biasrdquo)23 In addition the police size variables arelikely to be highly endogenous to the local crime rate exem-pli ed by the positive coef cient on police expenditures forevery crime Controlling for the endogeneity of these policevariables is quite complicated (Levitt 1997) and is not thefocus of this study Table 2 demonstrates that the resultsparticularly for the non-college-educated wage are generallyrobust to the inclusion of these deterrence measures as well asto the decrease in the sample To work with the broadestsample possible and avoid the endogeneity and ldquodivision-biasrdquoissues of these deterrence variables the remaining speci ca-tions exclude these variables

21 We included the percentage of male high school dropouts in the statepercentage of male high school graduates and percentage of men withsome college The property crime index coef cients (standard errors inparentheses ) were 053 (014) for the non-college-educate d male wageresidual and 215 (029) for the non-college-educate d unemployment rateresidual For the violent crime index the respective coef cients were

100 (015) and 129 (030) Compared to the results in table 1 whichexcluded the education distribution variables the coef cients are almostidentical in magnitude and signi cance as is also the case for theindividua l crime categories

22 Mustard (forthcoming ) showed that although conviction and sentenc-ing data are theoreticall y important they exist for only four or ve statesTherefore we cannot include such data in this analysis

23 Levitt (1995) analyzed this issue and why the relationship betweenarrest rates and offense rates is so strong

TABLE 2mdashOLS COUNTY-LEVEL PANEL REGRESSIONS USING THE ldquoCORE SPECIFICATIONrdquo PLUS ARREST RATES AND POLICE SIZE VARIABLES 1979ndash1995

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

052(015)

040(015)

140(027)

056(021)

022(015)

064(017)

085(022)

089(030)

069(023)

032(019)

State unemploymen t rateresidual for non-college-educated men

138(037)

132(037)

082(047)

219(043)

129(036)

018(033)

016(040)

100(057)

039(046)

029(037)

Log state income per capita 001(021)

007(021)

003(033)

005(027)

005(020)

002(020)

034(024)

114(030)

018(031)

042(024)

County arrest rate 0002(0002)

001(0002)

001(0001)

001(0003)

001(0001)

0004(00003)

0003(00003)

0002(00002)

0006(00005)

0004(0001)

Log state police expendituresper capita

028(010)

030(010)

051(0176)

034(011)

026(010)

036(011)

039(014)

023(016)

055(013)

004(010)

Log state police employmentper capita

004(014)

002(013)

026(020)

004(016)

007(012)

006(014)

001(016)

028(018)

014(018)

045(012)

Observations 5979 5979 5979 5979 5979 5979 5979 5979 5979 5979Partial R2 003 003 003 004 000 001 001 001 000 001

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 371 counties withat least sixteen out of seventeen years of data where the police force variables are available (1979ndash1995) County mean population is used as weights See notes to table 1 for other de nitions and controls usedin the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 51

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 7: crime rates and local labor market opportunities in the united states

trend will continue is improbable because the recent cyclicaldrop is likely to be temporary and in the future unemploy-ment will continue to uctuate with the business cycle Thewage trends however can continue to improve and have alasting impact on the crime trends This is best exempli ed byexplaining the increases of 21 and 35 in the adjustedproperty and violent crime rates over the entire 1979ndash1997period The unemployment rate was virtually unchanged in1997 from 1979 and therefore explains none of the increase ineither crime index The 20 fall in non-college-educatedwages over the entire period predicts a 108 increase inproperty crime and a 216 increase in violent crime Thesepredictionsldquoexplainrdquo more than 50 of the long-term trend inboth indices illustrating just how much the long-term crimetrends are dominated by the wages of unskilled men as op-posed to their unemployment rate

To see if our wage and unemployment measures in table1 are picking up changes in the relative supplies of differenteducation groups we checked if the results are sensitive tothe inclusion of variables capturing the local educationdistribution The results are practically identical to those intable 1 and therefore are not presented21 Clearly theresults are not due to changes in the education distribution

The speci cation in table 2 includes our ldquocorerdquo economicvariables plus variables measuring the local level of crimedeterrence Three deterrence measures are used the county

arrest rate state expenditures per capita on police and statepolice employment per capita22 Missing values for arrestrates are more numerous than for offense rates so thesample is reduced to 371 counties that meet our sampleselection criteria In addition the police variables wereavailable for only the years 1979 to 1995 After includingthese deterrence variables the coef cients on the non-college-educated wage remain very signi cant for propertyand violent crime although the magnitudes drop a bit Theunemployment rate remains signi cant for property crimebut disappears for violent crime

The arrest rate has a large and signi cantly negative effectfor every classi cation of crime Because the numerator of thedependent variable appears in the denominator of the arrestrate (the arrest rate is de ned as the ratio of total arrests to totaloffenses) measurement error in the offense rate leads to adownward bias in the coef cient estimates of the arrest rates(ldquodivision biasrdquo)23 In addition the police size variables arelikely to be highly endogenous to the local crime rate exem-pli ed by the positive coef cient on police expenditures forevery crime Controlling for the endogeneity of these policevariables is quite complicated (Levitt 1997) and is not thefocus of this study Table 2 demonstrates that the resultsparticularly for the non-college-educated wage are generallyrobust to the inclusion of these deterrence measures as well asto the decrease in the sample To work with the broadestsample possible and avoid the endogeneity and ldquodivision-biasrdquoissues of these deterrence variables the remaining speci ca-tions exclude these variables

21 We included the percentage of male high school dropouts in the statepercentage of male high school graduates and percentage of men withsome college The property crime index coef cients (standard errors inparentheses ) were 053 (014) for the non-college-educate d male wageresidual and 215 (029) for the non-college-educate d unemployment rateresidual For the violent crime index the respective coef cients were

100 (015) and 129 (030) Compared to the results in table 1 whichexcluded the education distribution variables the coef cients are almostidentical in magnitude and signi cance as is also the case for theindividua l crime categories

22 Mustard (forthcoming ) showed that although conviction and sentenc-ing data are theoreticall y important they exist for only four or ve statesTherefore we cannot include such data in this analysis

23 Levitt (1995) analyzed this issue and why the relationship betweenarrest rates and offense rates is so strong

TABLE 2mdashOLS COUNTY-LEVEL PANEL REGRESSIONS USING THE ldquoCORE SPECIFICATIONrdquo PLUS ARREST RATES AND POLICE SIZE VARIABLES 1979ndash1995

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

052(015)

040(015)

140(027)

056(021)

022(015)

064(017)

085(022)

089(030)

069(023)

032(019)

State unemploymen t rateresidual for non-college-educated men

138(037)

132(037)

082(047)

219(043)

129(036)

018(033)

016(040)

100(057)

039(046)

029(037)

Log state income per capita 001(021)

007(021)

003(033)

005(027)

005(020)

002(020)

034(024)

114(030)

018(031)

042(024)

County arrest rate 0002(0002)

001(0002)

001(0001)

001(0003)

001(0001)

0004(00003)

0003(00003)

0002(00002)

0006(00005)

0004(0001)

Log state police expendituresper capita

028(010)

030(010)

051(0176)

034(011)

026(010)

036(011)

039(014)

023(016)

055(013)

004(010)

Log state police employmentper capita

004(014)

002(013)

026(020)

004(016)

007(012)

006(014)

001(016)

028(018)

014(018)

045(012)

Observations 5979 5979 5979 5979 5979 5979 5979 5979 5979 5979Partial R2 003 003 003 004 000 001 001 001 000 001

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 371 counties withat least sixteen out of seventeen years of data where the police force variables are available (1979ndash1995) County mean population is used as weights See notes to table 1 for other de nitions and controls usedin the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 51

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 8: crime rates and local labor market opportunities in the united states

Up to now our results may be contaminated by theendogeneity of crime and observed labor market conditionsat the county level Cullen and Levitt (1996) argued thathigh-income individuals or employers leave areas withhigher or increasing crime rates On the other hand Willis(1997) indicated that low-wage employers in the servicesector are more likely to relocate due to increasing crimerates In addition higher crime may force employers to payhigher wages as a compensating differential to workers(Roback 1982) Consequently the direction of the potentialbias is not clear However it is likely that crime-inducedmigration will occur mostly across county lines withinstates rather than across states High-wage earners mayleave the county because of increases in the crime rate buttheir decision to leave the state is likely to be exogenous toincreases in the local crime rate To the extent that this istrue our use of state-level wage and unemployment rates toexplain county-level crime rates should minimize endoge-neity problems On the other hand our measures of eco-nomic conditions may be estimated with error which wouldlead to downward-biased estimates To control for anyremaining sources of potential bias we employ an instru-mental variables strategy

Our instruments for the economic conditions in each statebuild on a strategy used by Bartik (1991) and Blanchard andKatz (1992) and interact three sources of variation that areexogenous to the change in crime within each state (i) theinitial industrial composition in the state (ii) the nationalindustrial composition trends in employment in each indus-try and (iii) biased technological change within each indus-try as measured by the changes in the demographic com-position within each industry at the national level24

An example with two industries provides the intuitionbehind the instruments Autos (computers) constitute a largeshare of employment in Michigan (California) The nationalemployment trends in these industries are markedly differ-ent Therefore the decline in the auto industryrsquos share ofnational employment will adversely affect Michiganrsquos de-mand for labor more than Californiarsquos Conversely thegrowth of the high-tech sector at the national level translatesinto a much larger positive effect on Californiarsquos demand forlabor than Michiganrsquos In addition if biased technologicalchange causes the auto industry to reduce its employment ofunskilled men this affects the demand for unskilled labor inMichigan more than in California A formal derivation ofthe instruments is in appendix D

We use eight instruments to identify exogenous variationin the three ldquocorerdquo labor market variables After controllingfor the demographic variables the partial R2 between ourset of instruments and the three labor market variables are016 for the non-college-educated wage 008 for the non-college-educated unemployment rate and 032 for stateincome per capita

Table 3 presents the IV results for our ldquocorerdquo speci -cation The coef cient estimates for all three variablesremain statistically signi cant for the property and vio-lent crime indices although many of the coef cients forthe individual crimes are not signi cant The coef cientsfor the non-college-educated wage tend to be larger thanwith OLS whereas the IV coef cients for the non-college-educated unemployment rate are generally a bitsmaller The standard errors are ampli ed compared tothe OLS results because we are using instruments that areonly partially correlated with our independent variablesThe results indicate that endogeneity issues are not re-24 Bartik (1991) and Blanchard and Katz (1992) interacted the rst two

sources of variation to develop instruments for aggregate labor demandBecause we instrument for labor market conditions of speci c demo-graphic groups we also exploit cross-industr y variations in the changes in

industria l shares of four demographic groups (gender interacted witheducational attainment)

TABLE 3mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE ldquoCORErdquo SPECIFICATION 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log state non-college-educated male weeklywage residual

135(072)a 314b 42

126(075)a 292b 39

235(164)a 548b 73

055(097)a 129b 17

083(070)a 193b 26

253(087)a 589b 78

404(107)a 940b 124

247(119)a 574b 76

115(132)a 268b 35

108(104)a 251b 33

State unemploymen t rateresidual for non-college-educated men

166(100)a 51b 53

168(104)a 51b 54

529(214)a 162b 170

219(131)a 67b 70

245(099)a 75b 79

160(108)a 49b 51

261(136)a 80b 84

047(170)a 14b 15

075(171)a 23b 24

219(138)a 67b 70

Log state income per capita 165(058)a 33b 122

154(059)a 31b 115

246(129)a 49b 182

127(074)a 26b 94

124(055)a 25b 92

234(067)a 47b 174

272(078)a 55b 201

223(094)a 45b 165

313(108)a 63b 232

136(078)a 27b 101

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS52

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 9: crime rates and local labor market opportunities in the united states

sponsible for our OLS results and that if there is a biasthe bias is towards zero for the non-college-educatedwage coef cients as would be expected if there is mea-surement error25

Because county-level wage measures for less educatedmen are not available on an annual basis we have beenusing the non-college-educated wage measured at the statelevel Table 4 uses the county-level retail wage instead ofthe state non-college-educated wage because this is the bestproxy available at the county level for the wages of lesseducated workers26 As shown in gure 7 the trend in theretail wage is very similar to the trend in non-college-educated wages in gure 6 Table 4 and 5 present the OLSand IV results using the retail wage The OLS and IV resultsusing this measure are very signi cant in both analyses

The overall panel results indicate that crime responds tolocal labor market conditions All three of our core eco-nomic variables are statistically signi cant and have mean-ingful effects on the levels of crime within a county The

long-term trends in various crimes however are mostlyin uenced by the declining wages of less educated menthroughout this period These results are robust to OLS andIV strategies and the inclusion of variables for the level oflocal deterrence and the education distribution

IV Analysis of Ten-Year Differences 1979ndash1989

This section studies the relationship between crime andeconomic conditions using changes in these variables over aten-year period (1979ndash1989) This strategy emphasizes thelow-frequency (long-term) variation in the crime and labormarket variables in order to achieve identi cation Giventhe long-term consequences of criminal activity includinghuman capital investments speci c to the illegal sector andthe potential for extended periods of incarceration crimeshould be more responsive to low-frequency changes inlabor market conditions Given measurement error in ourindependent variables long-term changes may suffer lessfrom attenuation bias than estimates based on annual data(Griliches amp Hausman 1986 Levitt 1995)

The use of two Census years (1979 and 1989) for ourendpoints has two further advantages First it is possible toestimate measures of labor market conditions for speci cdemographic groups more precisely from the Census than ispossible on an annual basis at the state level Second wecan better link each county to the appropriate local labormarket in which it resides In most cases the relevant labormarket does not line up precisely with the state of residenceeither because the state extends well beyond the local labormarket or because the local labor market crosses stateboundaries In the Census we estimate labor market condi-tions for each county using variables for the SMSASCSAin which it lies Consequently the sample in this analysis isrestricted to those that lie within metropolitan areas27 Over-all the sample which is otherwise similar to the sample

25 Our IV strategy would raise concerns if the initial industria l compo-sition is affected by the initial level of crime and if the change in crime iscorrelated with the initial crime level as would occur in the case of meanreversion in crime rates However we note that regional differences in theindustria l composition tend to be very stable over time and most likely donot respond highly to short-term ldquoshocksrdquo to the crime rate Weinberg(1999) reports a correlation of 069 for employment shares of two-digitindustries across MAs from 1940 to 1980 We explored this potentialproblem directly by including the initial crime level interacted with timeas an exogenous variable in our IV regressions The results are similar tothose in table 3 For the property crime index 191 and 208 are thewage and unemployment coef cients respectivel y (compared to 126and 168 in table 3) For the violent crime index the respective coef -cients are 315 and 162 (compared to 253 and 160 in table 3)Similar to table 3 the wage coef cients are signi cant for both indiceswhereas the unemployment rate is signi cant for property crime We alsoran IV regressions using instrument sets based on the 1960 and 1970industria l compositions again yielding results similar to table 3

26 To test whether the retail wage is a good proxy we performed aten-year difference regression (1979ndash1989) using Census data of theaverage wages of non-college-educate d men on the average retail wage atthe MA level The regression yielded a point estimate of 078 (standarderror 004) and an R2 of 071 Therefore changes in the retail wage area powerful proxy for changes in the wages of non-college-educate d menData on county-leve l income and employment in the retail sector comefrom the Regional Economic Information System (REIS) disk from theUS Department of Commerce

27 We also constructed labor market variables at the county group levelUnfortunately due to changes in the way the Census identi ed counties inboth years the 1980 and 1990 samples of county groups are not compa-rable introducing noise in our measures The sample in this section differsfrom that in the previous section in that it excludes counties that are not

TABLE 4mdashOLS COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES Using the Retail Wage 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

035(006)

036(006)

003(012)

071(007)

033(006)

023(007)

001(009)

005(011)

048(010)

068(009)

State unemploymen t rateresidual for non-college-educated men

212(028)

226(029)

031(041)

313(033)

229(030)

094(029)

058(032)

118(043)

193(042)

040(034)

Log state income per capita 028(016)

028(016)

035(027)

054(020)

038(016)

028(016)

020(017)

017(024)

074(026)

131(017)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769Partial R2 004 005 0002 007 000 0005 0001 0002 000 0019

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect The excluded education category is thepercentage of male college graduates in the state Observations are for 705 counties with at least sixteen out of nineteen years of data County mean population is used as weights See notes to table 1 for otherde nitions and controls used in the regressions

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 53

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 10: crime rates and local labor market opportunities in the united states

used in the annual analysis contains 564 counties in 198MAs28

As in the previous analysis we measure the labor marketprospects of potential criminals with the wage and unem-ployment rate residuals of non-college-educated men Tocontrol for changes in the standard of living on criminalopportunities the mean log household income in the MA isincluded The construction of these variables is discussed inappendix C The regressions also control for the same set ofdemographic variables included in the previous section Theestimates presented here are for the ten-year differences ofthe dependent variables on similar differences in the inde-pendent variables Thus analogous to the previous sectionour estimates are based on cross-county variations in thechanges in economic conditions after eliminating time andcounty xed effects

Table 6 presents the OLS results for the indices andindividual crimes We focus on property crimes beforeconsidering violent crimes The wages of non-college-educated men have a large negative effect on propertycrimes The estimated elasticities range from 0940 forlarceny to 2396 for auto theft The 23 drop in wages fornon-college-educated men between 1979 and 1993 predictsa 27 increase in overall property crimes which is virtuallyall of the 29 increase in these years The unemploymentrate among non-college men has a large positive effect onproperty crimes The estimated responses to an increase ofone percentage point in unemployment range between 2310and 2648 percentage points However unemployment in-creased by only 3 over this period so changes in unem-ployment rates are responsible for much smaller changes incrime rates approximately a 10 increase The large cycli-cal drop in unemployment from the end of the recession in1993 to 1997 accounts for a 10 drop in property crimes

compared to the actual decline of 76 The estimatesindicate a strong positive effect of household income oncrime rates which is consistent with household income as ameasure of criminal opportunities

With violent crime the estimates for aggravated assaultand robbery are quite similar to those for the propertycrimes Given the pecuniary motives for robbery this sim-ilarity is expected and resembles the results in the previoussection Some assaults may occur during property crimesleading them to share some of the characteristics of propertycrimes Because assault and robbery constitute 94 ofviolent crimes the violent crime index follows the samepattern As expected the crimes with the weakest pecuniarymotive (murder and rape) show the weakest relationshipbetween crime and economic conditions In general theweak relationship between our economic variables andmurder in both analyses (and rape in this analysis) suggeststhat our conclusions are not due to a spurious correlationbetween economic conditions and crime rates generally Thedecline in wages for less educated men predicts a 19increase in violent crime between 1979 and 1993 or 40 ofthe observed 472 increase and increases in their unem-ployment rate predict a 26 increase Each variable ex-plains between two and three percentage points of the123 decline in violent crime from 1993 to 1997

Changes in the demographic makeup of the metropolitanareas that are correlated with changes in labor marketconditions will bias our estimates To explore this possibil-ity the speci cations in table 7 include the change in thepercentage of households that are headed by women thechange in the poverty rate and measures for the maleeducation distribution in the metropolitan area Increases infemale-headed households are associated with higher crimerates but the effect is not consistently statistically signi -cant Including these variables does little to change thecoef cients on the economic variables

Using the same IV strategy employed in the previoussection we control for the potential endogeneity of localcriminal activity and labor market conditions in table 8After controlling for the demographic variables the partialR2 between our set of instruments and the three labor market

in MAs (or are in MAs that are not identi ed on both the 1980 and 1990PUMS 5 samples)

28 The annual analysis included counties with data for at least sixteen ofnineteen years any county missing data in either 1979 or 1989 wasdeleted from this sample Results using this sample for an annual panel-level analysis (1979ndash1997) are similar to those in the previous sectionalthough several counties were deleted due to having missing data formore than three years

TABLE 5mdashIV COUNTY-LEVEL PANEL REGRESSIONS FOR VARIOUS OFFENSE RATES USING THE COUNTY RETAIL WAGE 1979ndash1997

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Log county retail income perworker

102(032)

098(033)

108(064)

121(043)

069(031)

160(041)

142(051)

069(054)

177(054)

185(048)

State unemploymen t rateresidual for non-college-educated men

200(093)

201(097)

537(202)

307(121)

272(095)

193(100)

208(128)

002(165)

187(153)

342(134)

Log state income per capita 123(031)

117(032)

139(062)

147(039)

101(029)

140(035)

068(039)

090(054)

319(053)

150(039)

Observations 12769 12769 12769 12769 12769 12769 12769 12769 12769 12769

indicates signi cance at the 5 level indicates signi cance at the 10 level The standard errors in parentheses have been corrected for a common state-year effect Observations are for 705 counties withat least sixteen out of nineteen years of data As described in the text and in the appendix the instruments are based on the county-leve l industrial composition at the beginning of the period All three ldquoeconomicrdquoindependent variables in the table were instrumented County mean population is used as weights See notes to table 1 for other de nitions and controls used in the regressions

THE REVIEW OF ECONOMICS AND STATISTICS54

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 11: crime rates and local labor market opportunities in the united states

variables ranges from 0246 to 033029 The IV estimates intable 8 for the wage and unemployment rates of non-college-educated men are quite similar to the OLS esti-mates indicating that endogeneity is not responsible forthese effects Overall these IV estimates like those fromthe annual sample in the previous section show strongeffects of economic conditions on crime30 Among theviolent crimes the IV estimates for robbery and aggravatedassault show sign patterns and magnitudes that are generallyconsistent with the OLS results although the larger standarderrors prevent the estimates from being statistically signif-icant The IV estimates show no effect of household incomeon crime whereas the OLS estimates show a strong rela-tionship as do the OLS and IV estimates from the annualanalysis

To summarize this section the use of ten-year changesenables us to exploit the low-frequency relation betweenwages and crime Increases in the unemployment rate ofnon-college-educated men increase property crime whereasincreases in their wages reduce property crime Violentcrimes are less sensitive to economic conditions than areproperty crimes Including extensive controls for changes incounty characteristics and using IV methods to control forendogeneity has little effect on the relationship between thewages and unemployment rates for less skilled men andcrime rates OLS estimates for ten-year differences arelarger than those from annual data but IV estimates are ofsimilar magnitudes suggesting that measurement error in

the economic variables in the annual analysis may bias theseestimates down Our estimates imply that declines in labormarket opportunities of less skilled men were responsiblefor substantial increases in property crime from 1979 to1993 and for declines in crime in the following years

V Analysis Using Individual-Level Data

The results at the county and MA levels have shown thataggregate crime rates are highly responsive to labor marketconditions Aggregate crime data are attractive because theyshow how the criminal behavior of the entire local popula-tion responds to changes in labor market conditions In thissection we link individual data on criminal behavior ofmale youths from the NLSY79 to labor market conditionsmeasured at the state level The use of individual-level datapermits us to include a rich set of individual control vari-ables such as education cognitive ability and parentalbackground which were not included in the aggregateanalysis The goal is to see whether local labor marketconditions still have an effect on each individual even aftercontrolling for these individual characteristics

The analysis explains criminal activities by each malesuch as shoplifting theft of goods worth less than and morethan $50 robbery (ldquousing force to obtain thingsrdquo) and thefraction of individual income from crime31 We focus onthese offenses because the NLSY does not ask about murderand rape and our previous results indicate that crimes witha monetary incentive are more sensitive to changes in wagesand employment than other types of crime The data come29 To address the possibility that crime may both affect the initial

industria l composition and be correlated with the change in crime wetried including the initial crime rate as a control variable in each regres-sion yielding similar results reported here

30 The speci cation in table 8 instruments for all three labor marketvariables The coef cient estimates are very similar if we instrument onlyfor either one of the three variables

31 The means for the crime variables are 124 times for shoplifting 1time for stealing goods worth less than $50 041 times for stealing goodsworth $50 or more and 032 times for robbery On average 5 of therespondent rsquos income was from crime

TABLE 6mdashOLS TEN-YEAR DIFFERENCE REGRESSION USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1125(0380)a 262b 35

1141(0403)a 265b 35

2396(0582)a 557b 74

1076(0415)a 250b 33

0940(0412)a 219b 29

0823(0368)a 191b 25

0874(0419)a 203b 27

0103(0749)a 24b 30

1040(0632)a 242b 32

0797(0497)a 185b 25

Change in unemploymentrate of non-college-educated men in MA(residuals)

2346(0615)a 72b 75

2507(0644)a 76b 80

2310(1155)a 70b 74

2638(0738)a 80b 85

2648(0637)a 81b 85

0856(0649)a 26b 27

0827(0921)a 25b 27

0844(1230)a 26b 27

2306(0982)a 70b 74

1624(0921)a 50b 52

Change in mean loghousehold income in MA

0711(0352)a 14b 53

0694(0365)a 14b 51

2092(0418)a 42b 155

0212(0416)a 04b 16

0603(0381)a 12b 45

0775(0364)a 16b 57

1066(0382)a 21b 79

0648(0661)a 13b 48

0785(0577)a 16b 58

0885(0445)a 18b 66

Observations 564 564 564 564 564 564 564 564 564 564R2 0094 0097 0115 0085 0083 0021 0019 0005 0024 0014

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common unobserved MA effect Numbers after ldquoa rdquo represent theldquopredictedrdquo percentage increase of the crime rate due to the mean change in the independent variable computed by multiplying the coef cient estimate by the mean change in the independen t variable between1979ndash1993 (multiplied by 100) Numbers after ldquob rdquo represent the ldquopredictedrdquo percentage increase in the crime rate based on the mean change in the independent variable during the latter period 1993ndash1997Dependent variable is log change in the county crime rate from 1979 to 1989 Sample consists of 564 counties in 198 MAs Regressions include the change in demographic characteristics (percentage of populationage 10ndash19 age 20ndash29 age 30ndash39 age 40ndash49 age 50 and over percentage male percentage black and percentage nonblack and nonwhite) Regressions weighted by mean of population size of each county Wageand unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 55

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 12: crime rates and local labor market opportunities in the united states

from self-reporting of the number of times individualsengaged in various forms of crime during the twelve monthsprior to the 1980 interview Although we expect economicconditions to have the greatest effect on less skilled indi-viduals we calculate average wages and unemploymentrates for non-college- and college-educated men in eachstate (from the 1980 Census) and see if they can explain thecriminal activity of each individual To allow the effects tovary by the education of the individual we interact labormarket conditions (and household income which is in-cluded to capture criminal opportunities) with the respon-dentrsquos educational attainment The level of criminal activityof person i is modeled as follows

Crimei HS1HS iW siHS

HS2HS iUsiHS

HS3HSiHouse Incsi COL1COLiWsiCOL

COL2COLiUsiCOL

COL3COLiHouse Incsi Zi Xsi

i

HS i and COL i are indicator variables for whether therespondent had no more than a high school diploma or somecollege or more as of May 1 1979 House Incsi denotes themean log household income in the respondentrsquos state ofbirth (we use state of birth to avoid endogenous migration)and W si

HS UsiHS (W si

COL UsiCOL) denote the regression-adjusted

mean log wage and unemployment rate of high school(college) men in the respondentrsquos state of birth Our mea-sures of individual characteristics Zi include years ofschool (within college and non-college) AFQT motherrsquoseducation family income family size age race and His-panic background To control for differences across statesthat may affect both wages and crime we also includecontrols for the demographic characteristics of the respon-dentrsquos state Xsi These controls consist of the same statedemographic controls used throughout the past two sectionsplus variables capturing the state male education distribu-tion (used in table 7)

Table 9 shows that for shoplifting and both measures oftheft the economic variables have the expected signs andare generally statistically signi cant among less educatedworkers Thus lower wages and higher unemployment ratesfor less educated men raise property crime as do higherhousehold incomes Again the implied effects of thechanges in economic conditions on the dependent variablesare reported beneath the estimates and standard errors32 Thepatterns are quite similar to those reported in the previousanalyses the predicted wage effects are much larger than

32 The magnitudes of the implied effects on the dependen t variablesreported here are not directly comparable to the ones previously reporteddue to the change in the dependent variable In the previous analyses thedependen t variable was the crime rate in a county whereas here it is eitherthe number of times a responden t would commit each crime or therespondent rsquos share of income from crime

TABLE 7mdashOLS TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION PLUS THE CHANGE IN FEMALE-HEADED HOUSEHOLDTHE POVERTY RATE AND THE MALE EDUCATION DISTRIBUTION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1858(0425)

1931(0445)

2606(0730)

1918(0478)

1858(0446)

1126(0464)

0977(0543)

0338(0948)

1964(0724)

0193(0604)

Change in unemploymen trate of non-college-educated men in MA(residuals)

3430(0510)

3682(0660)

3470(1281)

3777(0786)

3865(0627)

0935(0737)

0169(0932)

1685(1329)

3782(1081)

1135(0956)

Change in mean loghousehold income in MA

0809(0374)

0800(0247)

1637(0552)

0449(0439)

0811(0354)

0962(0498)

1107(0636)

0737(0927)

0987(0720)

0061(0576)

Change in percentage ofhouseholds female headed

2161(1360)

2314(1450)

2209(2752)

3747(1348)

3076(1488)

1612(1475)

1000(1906)

2067(3607)

2338(2393)

5231(2345)

Change in percentage inpoverty

5202(1567)

5522(1621)

4621(2690)

5946(1852)

5772(1649)

1852(1551)

0522(1797)

2388(3392)

6670(2538)

1436(2051)

Change in percentage malehigh school dropout inMA

0531(0682)

0496(0721)

1697(1123)

0476(0774)

0298(0748)

0662(0809)

0901(0971)

2438(1538)

1572(1149)

0945(1268)

Change in percentage malehigh school graduate inMA

1232(0753)

1266(0763)

0543(1388)

1209(1035)

1484(0779)

1402(1001)

0008(1278)

4002(1922)

2616(1377)

2939(1678)

Change in percentage malewith some college in MA

2698(1054)

2865(1078)

4039(1860)

2787(1371)

2768(1067)

0628(0430)

3179(1809)

4786(2979)

4868(2050)

3279(2176)

rw15rt Observations 564 564 564 564 564 564 564 564 564 564R2 0138 0147 0091 0116 0143 0023 0014 0004 0039 0002

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weighted by mean of population sizeof each county The excluded education category is the percentage of male college graduates in the MA See notes to table 6 for other de nitions and controls

THE REVIEW OF ECONOMICS AND STATISTICS56

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 13: crime rates and local labor market opportunities in the united states

the predicted unemployment effects for the earlier 1979ndash1993 period but the unemployment effects are larger in the1993ndash1997 period As expected economic conditions haveno effect on criminal activity for more highly educatedworkers33 The estimates are typically insigni cant andoften have unexpected signs The estimates explaining rob-bery are insigni cant and have the wrong signs howeverrobbery is the least common crime in the sample Theresults explaining the fraction of total income from crimeshow the expected pattern for less educated workers and theestimates are statistically signi cant A weaker labor marketlowers income from legal sources (the denominator) as wellas increasing income from crime (the numerator) bothfactors may contribute to the observed results Among theother covariates (not reported) age and education are typi-cally signi cant crime increases with age in this youngsample and decreases with education34 The other covariatesare generally not signi cant

Using the individual characteristics that are available inthe NLSY79 it is also possible to assess whether theestimates in the county-level analysis are biased by the lackof individual controls To explore this issue we drop manyof the individual controls from the NLSY79 analysis (wedrop AFQT motherrsquos education family income and familysize) Although one would expect larger effects for theeconomic variables if a failure to control for individualcharacteristics is responsible for the estimated relationshipbetween labor markets and crime the estimates for eco-nomic conditions remain quite similar35 Thus it appears

that our inability to include individual controls in thecounty-level analyses is not responsible for the relationshipbetween labor market conditions and crime

The analysis in this section strongly supports our previ-ous ndings that labor market conditions are importantdeterminants of criminal behavior Low-skilled workers areclearly the most affected by the changes in labor marketopportunities and these results are robust to controlling fora wealth of personal and family characteristics

VI Conclusion

This paper studies the relationship between crime and thelabor market conditions for those most likely to commitcrime less educated men From 1979 to 1997 the wages ofunskilled men fell by 20 and despite declines after 1993the property and violent crime rates (adjusted for changes indemographic characteristics) increased by 21 and 35respectively We employ a variety of strategies to investi-gate whether these trends can be linked to one another Firstwe use a panel data set of counties to examine both theannual changes in crime from 1979 to 1997 and the ten-yearchanges between the 1980 and 1990 Censuses Both of theseanalyses control for county and time xed effects as well aspotential endogeneity using instrumental variables We alsoexplain the criminal activity of individuals using microdatafrom the NLSY79 allowing us to control for individualcharacteristics that are likely to affect criminal behavior

Our OLS analysis using annual data from 1979 to 1997shows that the wage trends explain more than 50 of theincrease in both the property and violent crime indices overthe sample period Although the decrease of 31 percentage

33 When labor market conditions for low-education workers are used forboth groups the estimates for college workers remain weak indicatingthat the difference between the two groups is not due to the use of separatelabor market variables Estimates based on educationa l attainment at age25 are similar to those reported

34 Estimates are similar when the sample is restricted to respondent s ageeighteen and over

35 Among less educated workers for the percentage of income fromcrime the estimate for the non-college-educate d male wage drops in

magnitude from 0210 to 0204 the non-college-educate d male un-employment rate coef cient increases from 0460 to 0466 and the statehousehold income coef cient declines from 0188 to 0179 It is worthnoting that the dropped variables account for more than a quarter of theexplanatory power of the model the R2 declines from 0043 to 0030 whenthese variables are excluded

TABLE 8mdashIV TEN-YEAR DIFFERENCE REGRESSIONS USING THE ldquoCORErdquo SPECIFICATION 1979ndash1989

OverallCrimeIndex

Property Crime Violent Crime

PropertyCrimeIndex

AutoTheft Burglary Larceny

ViolentCrimeIndex

AggravatedAssault Murder Robbery Rape

Change in mean log weeklywage of non-college-educated men in MA(residuals)

1056(0592)a 246b 33

1073(0602)a 250b 33

2883(0842)a 671b 89

1147(0746)a 267b 35

0724(0588)a 168b 22

0471(0730)a 110b 15

0671(0806)a 156b 21

0550(1297)a 128b 17

1555(1092)a 362b 48

2408(1048)a 560b 74

Change in unemploymen trate of non-college-educated men in MA(residuals)

2710(0974)a 83b 87

2981(0116)a 91b 96

2810(1806)a 86b 90

3003(1454)a 92b 96

3071(1104)a 94b 99

1311(1277)a 40b 42

1920(1876)a 59b 62

0147(2676)a 04b 05

2854(1930)a 87b 92

1599(2072)a 49b 51

Change in mean loghousehold income in MA

0093(0553)a 02b 07

0073(0562)a 01b 05

1852(0933)a 37b 137

0695(0684)a 14b 51

0041(0537)a 01b 03

0069(0688)a 01b 05

0589(0776)a 12b 44

0818(1408)a 16b 61

0059(1004)a 770b 04

3219(1090)a 65b 239

Observations 564 564 564 564 564 564 564 564 564 564

indicates the coef cient is signi cant at the 5 signi cance level indicates signi cance at the 10 level Standard errors in parentheses have been corrected for a common MA effect Regressions weightedby mean of population size of each county The coef cients on all three presented independen t variables are IV estimates using augmented Bartik-Blanchard-Katz instruments for the change in total labor demandand in labor demand for four gender-education groups See notes to table 6 for other de nitions and controls

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 57

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 14: crime rates and local labor market opportunities in the united states

points in the unemployment rate of non-college-educatedmen after 1993 lowered crime rates during this period morethan the increase in the wages of non-college-educated menthe long-term crime trend has been unaffected by the un-employment rate because there has been no long-term trendin the unemployment rate By contrast the wages of un-skilled men show a long-term secular decline over thesample period Therefore although crime rates are found tobe signi cantly determined by both the wages and unem-ployment rates of less educated males our results indicatethat a sustained long-term decrease in crime rates willdepend on whether the wages of less skilled men continue toimprove These results are robust to the inclusion of deter-rence variables (arrest rates and police expenditures) con-trols for simultaneity using instrumental variables both ouraggregate and microdata analyses and controlling for indi-vidual and family characteristics

REFERENCES

Ayres Ian and Steven D Levitt ldquoMeasuring Positive Externalities fromUnobservabl e Victim Precaution An Empirical Analysis of Lo-jackrdquo Quarterly Journal of Economics 113 (February 1998) 43ndash77

Bartik Timothy J Who Bene ts from State and Local Economic Devel-opment Policies (Kalamazoo MI W E Upjohn Institute forEmployment Research 1991)

Becker Gary ldquoCrime and Punishment An Economic Approachrdquo Journalof Political Economy 762 (1968) 169ndash217

Blanchard Oliver Jean and Lawrence F Katz ldquoRegional EvolutionsrdquoBrookings Papers on Economic Activity 01 (1992) 1ndash69

Cornwell Christopher and William N Trumbull ldquoEstimating the Eco-nomic Model of Crime with Panel Datardquo this REVIEW 762 (1994)360ndash366

Crime in the United States (Washington DC US Department of JusticeFederal Bureau of Investigation 1992ndash1993)

Cullen Julie Berry and Steven D Levitt ldquoCrime Urban Flight and theConsequence s for Citiesrdquo NBER working paper no 5737 (Sep-tember 1996)

DiIulio John Jr ldquoHelp Wanted Economists Crime and Public PolicyrdquoJournal of Economic Perspectives 10 (Winter 1996) 3ndash24

Ehrlich Isaac ldquoParticipation in Illegitimate Activities A Theoretical andEmpirical Investigation rdquo Journal of Political Economy 813(1973) 521ndash565ldquoOn the Usefulness of Controlling Individuals An EconomicAnalysis of Rehabilitation Incapacitation and Deterrencerdquo Amer-ican Economic Review (June 1981) 307ndash322ldquoCrime Punishment and the Market for Offensesrdquo Journal ofEconomic Perspectives 10 (Winter 1996) 43ndash68

Fleisher Belton M ldquoThe Effect of Income on Delinquencyrdquo AmericanEconomic Review (March 1966) 118ndash137

Freeman Richard B ldquoCrime and Unemploymentrdquo (pp 89ndash106) in Crimeand Public Policy James Q Wilson (San Francisco Institute forContemporary Studies Press 1983)ldquoWhy Do So Many Young American Men Commit Crimes andWhat Might We Do About Itrdquo Journal of Economic Perspectives10 (Winter 1996) 25ndash42

TABLE 9mdashINDIVIDUAL-LEVEL ANALYSIS USING THE NLSY79

Dependent variable Number of times committedeach crime Shoplifted

Stole Propertyless than $50

Stole Propertymore than $50 Robbery

Share of IncomeFrom Crime

Mean log weekly wage of non-college-educate d men in state(residuals) respondent HS graduate or less

9853(2736)a 2292b 0303

8387(2600)a 1951b 0258

2688(1578)a 0625b 0083

0726(1098)a 0169b 0022

0210(0049)a 0049b 0006

Unemploymen t rate of non-college-educate d men in state(residuals) respondent HS graduate or less

17900(5927)a 0546b 0575

12956(4738)a 0395b 0416

5929(3827)a 0181b 0190

3589(2639)a 0109b 0115

0460(0080)a 0014b 0015

Mean log household income in state respondent HSgraduate or less

6913(2054)a 0139b 0512

5601(2203)a 0113b 0415

1171(1078)a 0024b 0087

1594(0975)a 0032b 0118

0188(0043)a 0004b 0014

Mean log weekly wage of men with some college in state(residuals) respondent some college or more

2045(3702)a 0209b 0088

7520(3962)a 0769b 0325

2017(1028)a 0206b 0087

0071(1435)a 0007b 0003

0112(0100)a 0011b 0005

Unemploymen t rate of men with some college in state(residuals) respondent some college or more

9522(17784)a 0078b 0155

11662(21867)a 0096b 0190

7390(5794)a 0061b 0120

0430(5606)a 0004b 0007

0477(0400)a 0004b 0008

Mean log household income in state respondent somecollege or more

2519(2108)a 0051b 0187

5154(1988)a 0104b 0382

0516(1120)a 0010b 0038

0772(1171)a 0016b 0057

0020(0064)a 00004b 0002

Observations 4585 4585 4585 4585 4585R2 0011 0012 0024 0008 0043

indicates signi cance at the 5 level indicates signi cance at the 10 level Standard errors correcting for within-state correlation in errors in parentheses Numbers after ldquoa rdquo represent the ldquopredictedrdquoincrease in the number of times each crime is committed (or the share of income from crime) due to the mean change in the independent variable computed by multiplying the coef cient estimate by the meanchange in the independen t variable between 1979ndash1993 Numbers after ldquob rdquo represent the ldquopredictedrdquo increase in the number of times each crime is committed (the share of income from crime) based on the meanchange in the independen t variable during the latter period 1993ndash1997 Dependent variables are self-reports of the number of times each crime was committedfraction of income from crime in the past year Additionalindividual-leve l controls include years of completed school (and a dummy variable for some college or more) AFQT motherrsquos education log of a three-year average of family income (and a dummy variable forfamily income missing) log of family size a quadratic in age and dummy variables for black and Hispanic background Additional state-level controls include percentage of population age 10ndash19 age 20ndash29 age30ndash39 age 40ndash49 age 50 and over percentage male percentage black percentage nonblack and nonwhite and the percentage of adult men that are high school dropouts high school graduates and have somecollege Wage and unemployment residuals control for educational attainment a quartic in potentia l experience Hispanic background black and marital status

THE REVIEW OF ECONOMICS AND STATISTICS58

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 15: crime rates and local labor market opportunities in the united states

ldquoThe Economics of Crimerdquo Handbook of Labor Economics vol 3(chapter 52 in O Ashenfelter and D Card (Eds) Elsevier Science1999)

Freeman Richard B and William M Rodgers III ldquoArea EconomicConditions and the Labor Market Outcomes of Young Men in the1990s Expansionrdquo NBER working paper no 7073 (April 1999)

Glaeser Edward and Bruce Sacerdote ldquoWhy Is There More Crime inCitiesrdquo Journal of Political Economy 1076 part 2 (1999) S225ndashS258

Glaeser Edward Bruce Sacerdote and Jose Scheinkman ldquoCrime andSocial Interactions rdquo Quarterly Journal of Economics 111445(1996) 507ndash548

Griliches Zvi and Jerry Hausman ldquoErrors in Variables in Panel DatardquoJournal of Econometrics 31 (1986) 93ndash118

Grogger Jeff ldquoMarket Wages and Youth Crimerdquo Journal of Labor andEconomics 164 (1998) 756ndash791

Hashimoto Masanori ldquoThe Minimum Wage Law and Youth CrimesTime Series Evidencerdquo Journal of Law and Economics 30 (1987)443ndash464

Juhn Chinhui ldquoDecline of Male Labor Market Participation The Role ofDeclining Market Opportunities rdquo Quarterly Journal of Economics107 (February 1992) 79ndash121

Katz Lawrence F and Kevin M Murphy ldquoChanges in Relative Wages1965ndash1987 Supply and Demand Factorsrdquo Quarterly Journal ofEconomics 107 (February 1992) 35ndash78

Levitt Steven ldquoWhy Do Increased Arrest Rates Appear to Reduce CrimeDeterrence Incapacitation or Measurement Errorrdquo NBER work-ing paper no 5268 (1995)ldquoUsing Electoral Cycles in Police Hiring to Estimate the Effect ofPolice on Crimerdquo American Economic Review 87 (June 1997)270ndash290

Lochner Lance ldquoEducation Work and Crime Theory and EvidencerdquoUniversity of Rochester working paper (September 1999)

Lott John R Jr ldquoAn Attempt at Measuring the Total Monetary Penaltyfrom Drug Convictions The Importance of an Individua lrsquos Repu-tationrdquo Journal of Legal Studies 21 (January 1992) 159ndash187

Lott John R Jr and David B Mustard ldquoCrime Deterrence andRight-to-Carry Concealed Handgunsrdquo Journal of Legal Studies 26(January 1997) 1ndash68

Mustard David ldquoRe-examining Criminal Behavior The Importance ofOmitted Variable Biasrdquo This REVIEW forthcoming

Papps Kerry and Rainer Winkelmann ldquoUnemployment and Crime NewEvidence for an Old Questionrdquo University of Canterbury workingpaper (January 2000)

Raphael Steven and Rudolf Winter-Ebmer ldquoIdentifying the Effect ofUnemployment on Crimerdquo Journal of Law and Economics 44(1)(April 2001) 259ndash283

Roback Jennifer ldquoWages Rents and the Quality of Liferdquo Journal ofPolitical Economy 906 (1982) 1257ndash1278

Sourcebook of Criminal Justice Statistics (Washington DC US Depart-ment of Justice Bureau of Justice Statistics 1994ndash1997)

Topel Robert H ldquoWage Inequality and Regional Labour Market Perfor-mance in the USrdquo (pp 93ndash127) in Toshiaki Tachibanak i (Ed)Labour Market and Economic Performance Europe Japan andthe USA (New York St Martinrsquos Press 1994)

Uniform Crime Reports (Washington DC Federal Bureau of Investiga-tion 1979ndash1997)

Weinberg Bruce A ldquoTesting the Spatial Mismatch Hypothesis usingInter-City Variations in Industria l Compositionrdquo Ohio State Uni-versity working paper (August 1999)

Willis Michael ldquoThe Relationship Between Crime and Jobsrdquo Universityof CaliforniandashSanta Barbara working paper (May 1997)

Wilson William Julius When Work Disappears The World of the NewUrban Poor (New York Alfred A Knopf 1996)

APPENDIX A

The UCR Crime Data

The number of arrests and offenses from 1979 to 1997 was obtainedfrom the Federal Bureau of Investigatio nrsquos Uniform Crime ReportingProgram a cooperative statistical effort of more than 16000 city county

and state law enforcement agencies These agencies voluntarily report theoffenses and arrests in their respective jurisdictions For each crime theagencies record only the most serious offense during the crime Forinstance if a murder is committed during a bank robbery only the murderis recorded

Robbery burglary and larceny are often mistaken for each otherRobbery which includes attempted robbery is the stealing taking orattempting to take anything of value from the care custody or control ofa person or persons by force threat of force or violence andor by puttingthe victim in fear There are seven types of robbery street and highwaycommercial house residence convenienc e store gas or service stationbank and miscellaneous Burglary is the unlawful entry of a structure tocommit a felony or theft There are three types of burglary forcible entryunlawful entry where no force is used and attempted forcible entryLarceny is the unlawful taking carrying leading or riding away ofproperty or articles of value from the possession or constructiv e posses-sion of another Larceny is not committed by force violence or fraudAttempted larcenies are included Embezzlement ldquoconrdquo games forgeryand worthless checks are excluded There are nine types of larceny itemstaken from motor vehicles shoplifting taking motor vehicle accessories taking from buildings bicycle theft pocket picking purse snatching theftfrom coin-operated vending machines and all others36

When zero crimes were reported for a given crime type the crime ratewas counted as missing and was deleted from the sample for that yeareven though sometimes the ICPSR has been unable to distinguish theFBIrsquos legitimate values of 0 from values of 0 that should be missing Theresults were similar if we changed these missing values to 01 beforetaking the natural log of the crime rate and including them in theregression The results were also similar if we deleted any county from oursample that had a missing (or zero reported) crime in any year

APPENDIX B

Description of the CPS Data

Data from the CPS were used to estimate the wages and unemploymentrates for less educated men for the annual county-leve l analysis Toconstruct the CPS data set we used the merged outgoing rotation group les for 1979ndash1997 The data on each survey correspond to the week priorto the survey We employed these data rather than the March CPS becausethe outgoing rotation groups contain approximatel y three times as manyobservation s as the March CPS Unlike the March CPS nonlabor incomeis not available on the outgoing rotation groups surveys which precludesgenerating a measure that is directly analogous to the household incomevariable available in the Census

To estimate the wages of non-college-educate d men we use the logweekly wages after controlling for observable characteristics We estimatewages for non-college-educate d men who worked or held a job in theweek prior to the survey The sample was restricted to those who werebetween 18 and 65 years old usually work 35 or more hours a week andwere working in the private sector (not self-employed ) or for government(the universe for the earnings questions) To estimate weekly wages weused the edited earnings per week for workers paid weekly and used theproduct of usual weekly hours and the hourly wage for those paid hourlyThose with top-coded weekly earnings were assumed to have earnings 15times the top-code value All earnings gures were de ated using theCPI-U to 1982ndash1984 100 Workers whose earnings were beneath $35per week in 1982ndash1984 terms were deleted from the sample as were thosewith imputed values for the earnings questions Unemployment wasestimated using employment status in the week prior to the surveyIndividuals who worked or held a job were classi ed as employedIndividuals who were out of the labor force were deleted from the sampleso only those who were unemployed without a job were classi ed asunemployed

To control for changes in the human capital stock of the workforce wecontrol for observable worker characteristic s when estimating the wages

36 Appendix II of Crime in the United States contains these de nitions andthe de nitions of the other offenses

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 59

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 16: crime rates and local labor market opportunities in the united states

and unemployment rates of less skilled men To do this we ran linearregressions of log weekly wages or in the case of unemployment statusa dummy variable equal to 1 if the person was unemployed uponindividua l worker characteristics years of completed schooling a quarticin potential experience and dummy variables for Hispanic black andmarital status We estimate a separate model for each year which permitsthe effects of each explanatory variable to change over time Adjustedwages and unemployment rates in each state were estimated using thestate mean residual from these regressions An advantage of this procedureis that it ensures that our estimates are not affected by national changes inthe returns to skill

The Outgoing Rotation Group data were also used to construct instru-ments for the wage and unemployment variables as well as the per capitaincome variable for the 1979ndash1997 panel analysis This required estimatesof industry employment shares at the national and state levels and theemployment shares for each gender-educatio n group within each industryThe sample included all employed persons between 18 and 65 Ourclassi cations include 69 industries at roughly the two-digit level of theSIC Individual s were weighted using the earnings weight To minimizesampling error the initial industry employment shares for each state wereestimated using the average of data for 1979 1980 and 1981

APPENDIX C

Description of the Census Data

For the ten-year difference analysis the 5 sample of the 1980 and1990 Census were used to estimate the mean log weekly wages ofnon-college-educate d men the unemployment rate of non-college -educated men and the mean log household income in each MA for 1979and 1989 The Census was also used to estimate the industria l compositioninstruments for the ten-year difference analysis Wage information is fromthe wage and salary income in the year prior to the survey For 1980 werestrict the sample to persons between 18 and 65 who worked at least oneweek were in the labor force for 40 or more weeks and usually worked 35or more hours per week The 1990 census does not provide data on weeksunemployed To generate an equivalen t sample of high labor forceattachment individuals we restrict the sample to people who worked 20 ormore weeks in 1989 and who usually worked 35 or more hours per weekPeople currently enrolled in school were eliminated from the sample inboth years Individual s with positive farm or nonfarm self-employmen tincome were excluded from the sample Workers who earned less than $40per week in 1979 dollars and those whose weekly earnings exceeded$2500 per week were excluded In the 1980 census people with top-coded earnings were assumed to have earnings 145 times the top-codedvalue The 1990 Census imputes individual s with top-coded earnings tothe median value for those with top-coded earnings in the state Thesevalues were used Individual s with imputed earnings (non top-coded) labor force status or individua l characteristic s were excluded from thesample

The mean log household income was estimated using the incomereported for the year prior to the survey for persons not living in groupquarters We estimate the employment status of non-college-educate d menusing the current employment status because the 1990 Census provides noinformation about weeks unemployed in 1989 The sample is restricted topeople between age 18 and 65 not currently enrolled in school As with theCPS individual s who worked or who held a job were classi ed asemployed and people who were out of the labor force were deleted fromthe sample so only individual s who were unemployed without a job wereclassi ed as unemployed The procedures used to control for individua lcharacteristic s when estimating wages and unemployment rates in the CPSwere also used for the Census data37

The industry composition instruments require industry employmentshares at the national and MA levels and the employment shares of eachgender-educatio n group within each industry These were estimated fromthe Census The sample included all persons between age 18 and 65 notcurrently enrolled in school who resided in MAs and worked or held a job

in the week prior to the survey Individual s with imputed industryaf liations were dropped from the sample Our classi cation has 69industries at roughly the two-digit level of the SIC Individual s wereweighted using the person weight in the 1990 Census The 1980 Censusis a at sample

APPENDIX D

Construction of the State and MA-Level Instruments

This section outlines the construction of the instruments for labordemand Two separate sets of instruments were generated one for eco-nomic conditions at the state level for the annual analysis and a second setfor economic conditions at the MA level for the ten-year differenceanalysis We exploit interstate and intercity variations in industria l com-position interacted with industria l difference s in growth and technologica lchange favoring particular groups to construct instruments for the changein demand for labor of all workers and workers in particular groups at thestate and MA levels We describe the construction of the MA-levelinstruments although the construction of the state-leve l instruments isanalogous

Let f i ct denote industry irsquos share of the employment at time t in city cThis expression can be read as the employment share of industry iconditional on the city and time period Let fi t denote industry irsquos share ofthe employment at time t for the nation The growth in industry irsquosemployment nationally between times 0 and 1 is given by

GROWi

f i 1

f i 01

Our instrument for the change in total labor demand in city c is

GROW TOTALc i fi c0 GROW i

We estimate the growth in total labor demand in city c by taking theweighted average of the national industry growth rates The weights foreach city correspond to the initial industry employment shares in the cityThese instruments are analogous to those in Bartik (1991) and Blanchardand Katz (1992)

We also construct instruments for the change in demand for labor infour demographic groups These instruments extend those developed byBartik (1991) and Blanchard and Katz (1992) by using changes in thedemographic composition within industries to estimate biased technolog-ical change toward particular groups Our groups are de ned on the basisof gender and education (non-college-educate d and college-educated) Letfg cti denote demographic group grsquos share of the employment in industry iat time t in city c ( fg t for the whole nation) Group grsquos share of theemployment in city c at time t is given by

fg ct i fg cti f t ci

The change in group grsquos share of employment between times 0 and 1 canbe decomposed as

fg c1 fg c0 i fg c0i f i c1 fi c0 i fg c1i fg c0i f i c1

The rst term re ects the effects of industry growth rates and the secondterm re ects changes in each grouprsquos share of employment within indus-tries The latter can be thought of as arising from industria l difference s inbiased technologica l change

In estimating each term we replace the MA-speci c variables withanalogs constructed from national data All cross-MA variation in theinstruments is due to cross-MA variations in initial industry employmentshares In estimating the effects of industry growth on the demand forlabor of each group we replace the MA-speci c employment shares( fg c0 i) with national employment shares ( fg 0 i) We also replace the actual

37 The 1990 Census categorize s schooling according to the degree earnedDummy variables were included for each educationa l category

THE REVIEW OF ECONOMICS AND STATISTICS60

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61

Page 17: crime rates and local labor market opportunities in the united states

end of period shares ( f i c1) with estimates ( fi ci) Our estimate of thegrowth term is

GROWgc i fg 01 f i c1 fi c0

The date 1 industry employment shares for each MA are estimated usingthe industryrsquos initial employment share in the MA and the industryrsquosemployment growth nationally

fi c1

fi c0 GROW i

j f j c0 GROWj

To estimate the effects of biased technology change we take the weightedaverage of the changes in each grouprsquos national employment share

TECHgc i fg 1i fg 0i f i c0

The weights correspond to the industryrsquos initial share of employment inthe MA

APPENDIX E

NLSY79 Sample

This section describes the NLSY79 sample The sample included allmale respondent s with valid responses for the variables used in theanalysis (the crime questions education in 1979 AFQT motherrsquoseducation family size age and black and Hispanic background adummy variable for family income missing was included to includerespondent s without valid data for family income) The number oftimes each crime was committed was reported in bracketed intervalsOur codes were as follows 0 for no times 1 for one time 2 for twotimes 4 for three to ve times 8 for six to ten times 20 for eleven to fty times and 50 for fty or more times Following Grogger (1998)we code the fraction of income from crime as 0 for none 01 for verylittle 025 for about one-quarter 05 for about one-half 075 for aboutthree-quarters and 09 for almost all

CRIME RATES AND LOCAL LABOR MARKET OPPORTUNITIES IN THE UNITED STATES 1979ndash1997 61