explaining demographic heterogeneity in cyclical …
TRANSCRIPT
Explaining Demographic Heterogeneity in CyclicalUnemployment
Eliza Forsythe∗
Jhih-Chian Wu†
November 4, 2020
Abstract
We investigate the sources of heterogeneity in the levels and cyclical sensitivityof unemployment rates across demographic groups. We develop a new methodologyto decompose cyclical and level differences in unemployment rates between groupsinto flows between three states (employment, unemployment, and out-of-the-labor-force). We find that increases in unemployment rates during recessions for young,non-white, and less-educated groups of workers are primarily driven by reductions inthe job-finding rates, which can explain more than 60% of cyclical fluctuations in theunemployment rate across demographic groups, compared with under 20% driven byseparations. However, separations are the most important factor in explaining thepersistent gap in unemployment rates between each disadvantaged group and theirrespective counterpart group, with important differences between groups. For less-educated workers, separation rates explain most of the unemployment gap, with 75%of the separation rate attributable to industry and occupation. Less-educated workersalso spend less time searching. For younger workers, we find separation rates explainall of the unemployment gap, while industry and occupation explain only 60% of theirelevated separation rates. For non-white workers, hiring explains almost half of theunemployment gap. Non-white workers search more intensely for work than othergroups, but spend less time interviewing per search time, suggesting that labor marketdiscrimination contributes to non-white workers’ persistently high unemployment rates.
JEL Classification: E24, J64, J63 Keywords: Unemployment rate, Unemploy-ment Gap, Gross worker flows, Job finding rate, Separation rate, Employment exitrate
∗School of Labor and Employment Relations and Department of Economics, Email: [email protected].†Department of Economics, National Chengchi University. Email: [email protected]. This project
received financial support from the University of Illinois Campus Research Board.
1 Introduction
Workers’ labor market outcomes vary substantially across demographic groups. During
the Great Recession the unemployment rate increased dramatically, from a pre-recession low
of 4.4% in May of 2007 to a high of 10% in October of 2009.1 During this period, young
workers, non-white workers, and those with no college education experienced increases in
unemployment rates of 7 to 8 percentage points, while prime age, white, and those with
some college education saw increases of 4 to 5 percentage points. During both expansions
and recessions, the unemployment rates for young workers, non-white workers, and those
with no college education are typically twice those of their respective counterpart groups.
A significant literature exists considering whether countercyclical increases in the unem-
ployment rate are due to inflows (e.g. job losses) or outflows (e.g. hiring).2 Recent method-
ological improvements have led to a consensus of sorts, with outflows appearing to dominate,
but with some debate remaining over the relative importance of inflows (e.g., M. Elsby et
al. (2009), Fujita and Ramey (2009), Shimer (2012)). While numerous researchers have in-
vestigated the sources of the high unemployment rates for non-whites (Barrett and Morgen-
stern (1974)), young workers (Clark and Summers (1982), Choi, Janiak, and Villena-Roldan
(2015)), and less-educated workers (Nickell (1979), Cairo and Cajner (2018)), it remains an
open question whether aggregate flows are similar across demographic groups, or if disad-
vantaged workers face special challenges.
In order to decompose unemployment rates into component flows for different demo-
graphic groups, we extend the methodology of gross flow decomposition. We follow Shimer
(2012) by decomposing the unemployment rate into gross flow transition rates. Further, we
build on Fujita and Ramey (2009) by decomposing cyclical variation in the unemployment
rate into cyclical variation of component flows, with two improvements. First, we extend the
two-state decomposition to include transition rates between three states: unemployment,
employed, and not-in-the-labor-force. This provides a more-flexible three-state decompo-
sition than M. W. L. Elsby, Michaels, and Ratner (2015), which is limited to measuring
unemployment fluctuations using the first order log-difference in the unemployment stock.3
Second, building on these techniques, we develop a new methodology to decompose
the difference in unemployment rates between two groups into the difference in transition
rates. This allows us to quantify how much of the differences in unemployment levels can
be attributed to differences in each flow between two groups. Although papers such as
M. W. Elsby, Hobijn, and Sahin (2010) and Cairo and Cajner (2018) consider the unemploy-
ment gaps between groups, to our knowledge we are the first to analytically decompose the
1Source: Bureau of Labor Statistics.2See M. Elsby, Michaels, and Solon (2009) for a survey.3Fujita and Ramey (2009) find this may be less reliable than other methods of measuring unemployment
fluctuations.
2
gap to explicitly measure the relative contributions of component flows.
We use matched monthly CPS data from 1978 through 2017, correcting for misclassifi-
cation errors following M. W. L. Elsby et al. (2015) and time aggregation errors following
Shimer (2012). We find that the relative importance of the component flows in explaining
cyclical unemployment remains consistent across demographic groups, with hiring (UE+IE)
always comprising the largest share, movements between unemployment and out of the labor
force (UI+IU) the next largest, and separations (EU+EI) the lowest share. In aggregate,
we find that 25% of fluctuations are due to the participation margin, slightly smaller than
M. W. L. Elsby et al. (2015).
The magnitudes of these flows, however, differ between groups. For young workers hiring
explains 70% of the variation, while for other groups it explains only 50–60%. Further, for
dominant groups (white, experienced, with college), separations explain about 20% of the
cyclical variation in unemployment, while for non-white and young workers separations only
explain about 10%. Thus, due to the greater relative importance of hiring, the reduction in
hiring during recessions is especially harmful for disadvantaged groups.
We then decompose the gap in unemployment rates between each disadvantaged group
and their respective counterpart, finding that separations are the biggest determinant of
the gap in each case, with subtle differences between groups. For non-white workers, hiring
is almost as important as separations. For young workers, the gap is entirely driven by
separations. For individuals with no college experience, separations explain 80% of the gap.
We next investigate mechanisms driving these gaps. We find that differences in industry
and occupation can explain around 60% of the gap in separation rates for young and non-
white workers and 75% of the gap in separation rates for workers with no college. We examine
the role of voluntary and involuntary separations and conclude that unemployment gaps are
primarily driven by involuntary separations.
Finally, sing time-use data from the American Time Use Survey and search data from
the CPS, we assess whether search effort can explain the cyclical and level differences in
hiring between demographic groups. We find that both young and non-white workers are
more likely to spend time searching than older and white workers, respectively, while workers
with no college are less likely to spend any time searching than college-educated workers.
Conditional on spending any time searching, younger workers and those with no college
spend less time searching than their counterpart groups, while non-white and white workers
show no difference. We find little evidence that disadvantaged workers spend comparatively
less time searching during recessions, although the specifications are under-powered. While
we cannot measure the returns to search in the ATUS, we construct a proxy based on the
time spent interviewing compared to time spent searching. Young workers spend 6% more
of their search time interviewing than older workers, while non-white workers spend 5% less
time interviewing than white workers.
3
Thus, although individuals in the three examined disadvantaged groups show similar
aggregate unemployment patterns, the mechanisms behind each differ significantly. For non-
white workers, the facts that the hiring margin explains almost half of the unemployment
gap and that they spend less time interviewing for time searching both suggest that hiring
discrimination likely plays a role. This is consistent with experimental evidence on racial
discrimination in interview success rates (e.g. Bertrand & Mullainathan, 2004).
Young workers’ high unemployment rates are entirely driven by high separation rates.
Although they have somewhat higher voluntary separation rates, which could be consistent
with lifecycle models of job hopping and experimentation, the vast majority of separations
are involuntary. These high separation rates are somewhat offset by young workers experi-
encing greater hiring rates than other demographic groups, but this leaves them especially
vulnerable to depressed hiring during recessions.
Finally, for individuals with no college, occupation and industry can explain 3/4 of the
gap in the separation rate compared with individuals with college experience. For these
workers elevated unemployment rates are largely due to differences in sorting across jobs,
although lower search effort could be related to the lower importance of hiring for these
workers.
In the Appendix we perform a variety of robustness checks. We show that our flow-
decomposition results are robust to alternative divisions of workers by race, age, and educa-
tion. In addition, we show our results are robust to common alternative specifications.
Our paper contributes to a large literature on flow-decomposition. Most closely related
is the work of M. W. Elsby et al. (2010), who perform a two-state decomposition on the
unemployment rate for different demographic groups, concluding that hiring is the most im-
portant cyclical feature for all demographic groups and separations can explain the difference
in unemployment rates between groups. We make several methodological improvements, in-
cluding three-state decomposition and estimating the relative contribution of each flow to
both cyclical and relative unemployment rates, which reveals substantive differences between
demographic groups.
Choi et al. (2015) focuses on explaining differences in unemployment and participation
rates over the lifecycle. These authors abstract from the cyclicality of flows and instead use a
Markov-chain simulation to construct lifecycle unemployment levels. Despite the differences
in methodology, they also find that separations are the most important factor in explaining
high youth unemployment rates.
Our results complement the literature on whether certain demographic groups are the
first to be fired during recessions or the last to be hired during recoveries. Couch and Fairlie
(2010) and Couch, Fairlie, and Xu (2016) test this hypothesis for non-white workers and
Xu and Couch (2017) for young workers, finding evidence that supports the “first-fired”
hypothesis but not the “last-hired” hypothesis. Our approach improves the methodology
4
by analyzing cyclical transition dynamics, allowing us to show that while separations are an
important part of the story, hiring dynamics play a larger role.
Finally, our paper also contributes to a literature on the racial employment gap. We find
that the hiring margin can explain a larger fraction of the unemployment gap for non-white
groups than for other disadvantaged groups. This is consistent with a large literature on
racial discrimination, including Lang and Lehmann (2012), Gobillon, Rupert, and Wasmer
(2014) , and Borowczyk-Martins, Bradley, and Tarasonis (2017).
The paper is organized as follows. The data and methodology are are explained in
Section 2. Section 3 explains the decomposition approach for unemployment fluctuations and
finds the estimated contributions for various transition flows. In Section 4 we use a similar
approach for unemployment gaps and discuss the results. Section 5 explores mechanisms
that may explain differences between groups. We conclude in Section 6 and discuss policy
implications.
2 Data and Methodology
We use monthly U.S. data from the Current Population Survey (CPS) spanning January
1978 through November 2017, retrieved from the IPUMS repository.4 This yields a total of
over 36 million individual monthly observations. In order to measure flows between labor
force statuses (employed (E), unemployed (U), and out-of-the-labor-force (I)), we match in-
dividuals between consecutive months.5 After matching the data, we correct misclassification
errors with a method based on that proposed by M. Elsby, Hobijn, and Sahin (2015). We
then construct transition rates and correct time aggregation errors, following Shimer (2012).
We divide our sample into three overlapping groups, according to potential labor market
experience, race, and educational attainment. We define young workers as those with less
than ten years of potential experience (age− education− six), who we compare with expe-
rienced workers, those with more than 10 years of potential experience.6 Dividing by race,
we compare non-white workers with white workers. We also divide workers with no college
education and those with at least some college.7
To discover the connection between the unemployment rates and the transition flows, we
4Sarah Flood, Miriam King, Steven Ruggles, and J. Robert Warren. Integrated Public Use MicrodataSeries, Current Population Survey: Version 5.0. [dataset]. Minneapolis: University of Minnesota, 2017.https://doi.org/10.18128/D030.V5.0.
5Specifically, we match individuals using gender and the IPUMS-CPS defined variable (CPSIDP) con-structed to uniquely identify individuals. By using CPSIDP, we simplify the matching process and avoidissues wherein the CPS identification number may not represent a unique individual. CPSIDP is consistentonly after 1978.
6In Appendix H we show results are similar using age instead of potential experience.7Table A.1 in the Appendix A shows the relative frequencies of each of these six demographic categories.
In Appendix I.4, we show that our results are not sensitive to the precise choice of cutoff.
5
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18Non-WhiteWhite
Non-White vs. White
1978 1983 1988 1993 1998 2003 2008 2013 20180.02
0.04
0.06
0.08
0.10
0.12
0.14YoungExperienced
Young vs. Experienced
1978 1983 1988 1993 1998 2003 2008 2013 20180.02
0.04
0.06
0.08
0.10
0.12
0.14 No College EducationCollege Education
No College vs. Any College
1978 1983 1988 1993 1998 2003 2008 2013 20180.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
0.11
FemaleMale
Female vs. Male
Figure 1: Unemployment Rate
Gray-shaded areas indicate NBER Recession periods. Data source: FRED & IPUMS-CPS.
first identify the dates when the unemployment rates are larger than the Hodrick-Prescott
(HP) filtered trend. We compare the difference between the unemployment rates and tran-
sition rates in these dates and those during periods where the unemployment rates are lower
than HP filter trend for different demographic groups. This exercise can help determine the
changes in the transition rates when the unemployment rates are significantly larger than
the trend.
2.1 Constructing Transition and Unemployment Rates
We begin by constructing the raw transition rates between the three labor force states and
the unemployment rates for each of our six demographic groups. Let 1ijh,t,g be an indicator
that captures whether individual h in demographic group g transitioned from labor force
state i ∈ {E,U, I} at time t − 1 to state j ∈ {E,U, I} at time t. Here t denotes a specific
6
month in a year. Therefore, the flow of individuals from state i to state j at time t can be
written
zijt,g =H∑h=1
1ijh,t,g × wh,t,g (1)
where wh,t,g is the CPS sampling weight for individual h in demographic group g at time t.8
Using this expression, we can construct the unemployment rate for workers in demographic
group g as
ugt =
∑Hh=1
∑i∈{E,U,I} z
iUt,g∑
i∈{E,U,I} ziUt,g +
∑i∈{E,U,I} z
iEt,g
. (2)
Figure 1 plots the group unemployment rates constructed using Equation (2) for eight
demographic groups: non-white, white, young, experienced, female, male, individuals with
no college education, and those with some college education. Note that non-white workers,
young workers, and individuals with no college education have substantially higher unemploy-
ment rates than their counterpart groups, and that the groups with elevated unemployment
rates also experience more dramatic increases in unemployment during recessions. This is
consistent with previous research, such as Hoynes, Miller, and Schaller (2012).
Table 1 measures the increase in the unemployment rates for each demographic group for
each of the last four recessionary periods.9 As we are interested in labor market outcomes,
we use a broader definition of recessionary periods than that given by the National Bureau
of Economic Research (NBER) Business Cycle Dating Committee. Specifically, we measure
recessions as trough to peak unemployment rate from the pre-recession series minimum un-
employment rate to the maximum unemployment following the recession. For the beginning
of a recession, we choose the month with the minimum unemployment rate at the beginning
of an NBER recessionary period. Similarly, we choose the month with the maximum unem-
ployment rate at the end of the NBER recessionary period as the end of a recession. We
see increases in the unemployment rates for non-white, young, and less-educated workers are
almost twice those of their respective comparison groups.
Due to persistently higher unemployment rates as well as cyclical changes in unemploy-
ment, we focus on three demographic categories as particularly hard-hit: young workers,
non-white workers, and those with no college education. In contrast, Figure 1 shows that
there are no persistent level differences between female workers’ unemployment rates and
those of male workers and Table 1 shows that male workers see higher increases in unem-
ployment rate than female workers during recessions. We therefore leave the analysis of
8Due to the rotation feature of CPS data, we have two CPS sample weights: the weight at the beginningof the month and another at the end. Following Shimer (2012), we use the average of these two weights aswh,t,g.
9Since the unemployment rate did not appreciably recover between the 1980 recession and the 1981recession, we combine these two recessions.
7
1980 1985 1990 1995 2000 2005 2010 2015
0.040
0.045
0.050
0.055
0.060
0.065
0.070 Non-White: λEU + λEI
White: λEU + λEI
Non-White vs. White
1980 1985 1990 1995 2000 2005 2010 2015
0.2
0.3
0.4
0.5
Non-White: λUEWhite: λUE
Non-White vs. White
1980 1985 1990 1995 2000 2005 2010 20150.035
0.040
0.045
0.050
0.055
0.060
0.065Non-White: λ IEWhite: λ IE
Non-White vs. White
1980 1985 1990 1995 2000 2005 2010 20150.03
0.04
0.05
0.06
0.07
0.08Young: λEU + λEI
Experienced: λEU + λEI
Young vs. Experienced
1980 1985 1990 1995 2000 2005 2010 2015
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55 Young: λUEExperienced: λUE
Young vs. Experienced
1980 1985 1990 1995 2000 2005 2010 2015
0.04
0.06
0.08
0.10
0.12
0.14Young: λ IEExperienced: λ IE
Young vs. Experienced
1980 1985 1990 1995 2000 2005 2010 2015
0.03
0.04
0.05
0.06
0.07No College Education: λEU + λEI
College Education: λEU + λEI
With vs. Without College
1980 1985 1990 1995 2000 2005 2010 2015
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55
No College Education: λUECollege Education: λUE
With vs. Without College
1980 1985 1990 1995 2000 2005 2010 2015
0.03
0.04
0.05
0.06
0.07No College Education: λ IECollege Education: λ IE
With vs. Without College
Figure 2: Between Employment and Unemployment: λEU and λUE
The left column shows the employment exit rate (λEU + λEI), the middle column the transition rate from
unemployment to employment (λUE), and the right column the transition rate from out-of-the-labor-force
to employment (λIE). Gray-shaded areas indicate NBER Recession periods. Data source: FRED & IPUMS-
CPS.
8
Table 1: Unemployment Rate: Minimum to Maximum
Date Non-White White Young Experienced
1980s Recession 8.87% 4.91% 6.69% 4.75%1990s Recession 3.24% 2.41% 3.57% 2.25%2000s Recession 3.39% 2.01% 3.28% 1.98%Great Recession 7.52% 5.21% 6.84% 5.27%
Date No College With College Female Male
1980s Recession 6.97% 3.24% 4.43% 6.20%1990s Recession 3.51% 2.15% 1.60% 3.42%2000s Recession 2.99% 1.84% 1.73% 2.76%Great Recession 7.94% 4.28% 4.28% 6.87%
1980s Recession 1990s Recession 2000s Recession Great Recession
Date 1979/07–1982/11 1989/03–1992/01 2000/01–2003/06 2007/03–2010/02
Changes in the unemployment rate for each type of worker during recessionary periods, as wellas dates for the recessions.
female workers for the appendix.10
To identify the sources of differences in unemployment rates between demographic groups,
we next decompose each group’s unemployment rate into component flows. From the def-
inition of the unemployment rate in Equation (2) the share of workers unemployed at any
moment in time is a function of the flows in and out of unemployment. Returning to the
notation from Equation (1), we can write the transition rate between labor market state i
and j for a worker in demographic group g as
λijt,g =zijt,g∑
i∈{E,U,I} zijt,g
. (3)
In Figure 2 we show the transition rates from employment to non-employment (the em-
ployment exit rate, λEU + λEI), from unemployment to employment (the job-finding rate,
λUE), and entering employment from out-of-the-labor-force λIE. During recessionary periods
employment exit rates increase and job-finding rates decrease for all six demographic groups.
We also see that our three disadvantaged demographic groups (young workers, non-white
workers, and those with no college education) have higher employment exit rates compared
with their respective counterpart groups. These results suggest that, while both job finding
and employment exit rates play a role in elevated unemployment rates for all demographic
10The difference between female workers’ unemployment rate and male workers’ (i.e., the gender unem-ployment gap) was positive before the early 1980s, but disappeared thereafter. In the Appendix B, weconduct analysis on the causes of this disappearance and find that the transition rate from employment tounemployment accounts for the change. Our finding is similar to that of Albanesi and Sahin (2018). Wealso analyze the increase in male workers’ unemployment rates during recessions and find it is explained bythe decline in the job-finding rates.
9
groups during recessions, employment exit rates are likely a larger driver of the relative differ-
ences in unemployment rates between disadvantaged and counterpart demographic groups.11
For the job-finding rate λUE in Figure 2, we find that non-white workers and those
with no college have lower job-finding rates than their counterpart workers. In contrast,
young workers have similar or even higher job-finding rates than more-experienced workers.
Although workers with no college education have job-finding rates that are lower than those
of workers with some college, the difference is much smaller than the gap between non-white
and white workers. Figure 2 also shows employment entry rates from out-of-the-labor-force.
We find that non-white workers and young workers have relatively higher employment entry
rates from out-of-the-labor-force, while those with no college have relatively lower rates.
Young workers exhibit significantly higher employment entry rates from out-of-the-labor-
force than experienced workers.
As a first step in quantifying the differences in transition rates between demographic
groups, we estimate a series of linear regression models in which we regress the log of the
transition rate on the log of the unemployment rate interacted with demographic dummies.
Specifically, we run the following specification:
ln yt = β0 + βd1d + βr lnut + βdr (lnut × 1d) +∑m
am1m + εt (4)
where 1d is an indicator for the disadvantaged demographic group, ut is the unemployment
rate for time t, and 1m are month-by-year fixed effects. We use yt to represent either the
unemployment rate or the transition rates, which are constructed based on Equations (2)
and (3), respectively.
Table 2 gives the regression results for each of the three specifications: white versus
non-white (top), young versus experienced (middle), and those with no college versus those
with some college education (bottom).12 In the first column, we show that the disadvantaged
groups’ unemployment rates increase more slowly during recessions in log terms than those of
the counterpart groups, due to the higher baseline unemployment levels. This is consistent
with Figure 1. In addition, we consider the coefficient on the unemployment rate across
specifications, showing that exits to unemployment increase with the unemployment rate,
but exits to NILF are acyclical. Hires from both unemployment and NILF decrease with
the unemployment rate and transitions from unemployment to NILF decrease while flows
from NILF to unemployment increase. Thus, broadly, exits from employment increase, hires
decrease, and flows into unemployment increase.
11In Appendix A, Figures A.1 and A.2 show the remaining four transitions in and out of the labor force.We show that disadvantaged groups tend to have higher transition rates in and out of the labor marketcompared with their counterpart groups.
12In Appendix I we show these results are robust to controlling for state, gender, race, and education.However, we do find some differences by occupation, which we discuss in detail.
10
Table 2: Estimated Transition Rates
Non-White u λEU λEI λUE λUI λIE λIU
Non-White 0.670*** 0.448*** 0.365*** -0.139** 0.080 0.511*** 0.676***(0.043) (0.083) (0.061) (0.068) (0.054) (0.066) (0.070)
Unemployment 1.043*** 0.585*** -0.018 -0.577*** -0.371*** -0.179*** 0.553***(0.009) (0.024) (0.021) (0.026) (0.019) (0.022) (0.017)
Unemp. × Non-White -0.010 -0.024 -0.076** -0.113*** 0.093*** -0.229*** 0.016(0.024) (0.047) (0.031) (0.037) (0.029) (0.034) (0.038)
R2 0.96 0.71 0.52 0.75 0.66 0.43 0.89N 980 980 980 980 980 980 980
Young u λEU λEI λUE λUI λIE λIU
Young 0.979*** 0.763*** 0.713*** 0.078 0.072 1.245*** 1.625***(0.033) (0.066) (0.079) (0.068) (0.063) (0.095) (0.087)
Unemployment 1.044*** 0.546*** -0.030 -0.592*** -0.405*** -0.233*** 0.505***(0.014) (0.021) (0.030) (0.024) (0.022) (0.033) (0.035)
Unemp. × Young -0.177*** -0.065* -0.089** 0.010 0.108*** -0.089* -0.085*(0.018) (0.036) (0.042) (0.037) (0.034) (0.050) (0.048)
R2 0.97 0.87 0.72 0.65 0.60 0.85 0.95N 980 980 980 980 980 980 980
No College u λEU λEI λUE λUI λIE λIU
No College 0.842*** 0.682*** 0.698*** -0.034 0.388*** 0.037 0.418***(0.050) (0.057) (0.060) (0.078) (0.059) (0.075) (0.058)
Unemployment 0.948*** 0.468*** -0.008 -0.529*** -0.321*** -0.024 0.644***(0.024) (0.022) (0.022) (0.034) (0.025) (0.034) (0.025)
Unemp. × No College -0.049* 0.016 -0.140*** -0.058 -0.073** -0.255*** -0.186***(0.028) (0.031) (0.030) (0.043) (0.032) (0.040) (0.032)
R2 0.95 0.90 0.70 0.60 0.62 0.62 0.72N 980 980 980 980 980 980 980
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
11
How do these patterns differ for the disadvantaged groups? In Columns (2) and (3)
we see little difference in outflows to unemployment, while both young workers and those
with no college see a smaller percentage increase in exits to non-employment. In Column
(4) we see that only non-white workers have a substantially larger decrease in hiring from
unemployment compared to the counterpart group. In Column (6) we see that both non-
white workers and those with no college education see substantially larger decreases in hiring
from non-employment than their counterpart groups.
We see that all groups experienced sharp declines in both job-finding rates and transi-
tion rates from unemployment to out-of-the-labor-force during periods with relatively high
unemployment rates. We therefore expect that the declines in the job-finding rates and the
transition rates from unemployment to out-of-the-labor-force are key drivers of the increases
in unemployment rates during recessions across demographic groups.
In sum, disadvantaged workers tend to have larger outflows from employment (either
to unemployment or non-employment), but during recessions their outflows increase by a
smaller percent than those of counterpart groups. On the other hand, for both non-white
workers and those with no college, hiring rates fall at a substantially higher rate, while
young workers’ hiring rates are inconclusive. These results suggest that hiring is likely
to be a more important driver of the cyclical unemployment rate gap than separations.
However, determining which flows have a larger impact on the stock requires a formal flow
decomposition, which we pursue in the following sections.
So far we have shown that fluctuations in unemployment and transition rates differ be-
tween demographic groups. A related question is how the changing demographic composition
of the labor force has affected the aggregate unemployment rate and flows. To answer this, we
construct a counterfactual unemployment rate by fixing demographic shares at 1978 levels,
which we then compare to the true unemployment rate.13
We construct 16 demographic categories based on identity in each of our four binary
categories (white/non-white, male/female, young/experienced, no college/ college). For each
group we calculate their weight in the sample, transition flows, and their unemployment rate.
The weighted aggregate unemployment rate at time t can be expressed as
uwt =16∑j=1
αjtujt , (5)
where uwt is the weighted unemployment rate, αj the weight for subgroup j, and ujt the unem-
ployment rate in period t for subgroup j. To determine the importance of the composition of
workers, we construct the following counterfactual unemployment rates and transition rates
by setting the weight αj to the average level for the year 1978. Thus, the counterfactual
13We would like to thank a referee for suggesting this exercise.
12
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.04
0.06
0.08
0.10
0.12
Figure 3: Composition-Adjusted Unemployment Rate
The black solid line shows the true weighted unemployment rate while the green dashed line is the counter-
factual weighted unemployment rate holding the demographic composition fixed at 1978 shares.
Data source: FRED & IPUMS-CPS.
unemployment rate is
uwt =16∑j=1
αj1978ujt (6)
Figure 3 compares our observed weighted unemployment rates (black solid lines) to the
counterfactual (green dashed lines). If the demographic composition of the U.S. labor market
was fixed in 1978, the unemployment rate would have been 2 percentage plints higher after
1994. This is primarily driven by increases in college attainment and the aging population.
This exercise drives home the importance of understanding the demographic heterogeneity
underlying the aggregate unemployment rate. In Appendix I.3 we show that separations
from employment would have been 10 percentage points higher under this counterfactual.
Hence, we have shown how transition rates differ between demographic groups and
demonstrated that changes in composition have important implications for the aggregate
unemployment rate. Differences in transition rates between demographic groups imply that
the unemployment level and cyclicality may be attributable to different underlying flows.
In the next section we introduce decomposition methodology to formally investigate hetero-
geneity in the flow decomposition.
13
2.2 Linking Transition Rates and Unemployment Rates
In order to measure the contribution of each transition rate to the unemployment rate,
we begin by using methodology developed in Shimer (2012). Specifically, we construct a
steady-state approximation of the unemployment rate as a function of the transition rates.
We express a steady state accounting identity, in which outflows from each labor market
state must equal inflows, giving the following conditions:
(λEU + λEI)E = λUEU + λIEI,
(λUE + λUI)U = λEUE + λIUI,
(λIE + λIU)I = λUIU + λEIE.
(7)
By this steady-state identity, we can express the stock of employed and unemployed as:
U = C · (λEIλIU + λIEλEU + λIUλEU), (8)
E = C · (λIUλUE + λUIλIE + λIEλUE). (9)
Here C is a constant such that total population U + E + I can be normalized as a fixed
number.14 Given the unemployment and employment inflow in Equations (8) and (9), we
can express the steady state unemployment rate as
u =U
U + E=
λEIλIU + λIEλEU + λIUλEU
(λEIλIU + λIEλEU + λIUλEU) + (λUIλIE + λIEλUE + λIUλUE),
which describes how the unemployment rate can be rewritten as a function of six transition
rates. Shimer (2012) shows that the steady state unemployment rate can be written as
a function of the six different transition rates to approximate the time-varying observed
unemployment rate. The approximation formula can be written as
ug,ct =Ut
Et + Ut
=λEIt λ
IUt + λIE
t λEUt + λIU
t λEUt
(λEIt λ
IUt + λIE
t λEUt + λIU
t λEUt ) + (λUI
t λIEt + λIE
t λUEt + λIU
t λUEt )
.
(10)
Here ug,ct represents the constructed unemployment rate based on the steady-state identity
approximation for the observed unemployment rate ugt of workers in demographic group g,
where c denotes the constructed unemployment rate based on Equation (10). This expression
for the unemployment rate allows us to disentangle how transitions between employment,
unemployment, and out of the labor force contribute to the overall unemployment rate.
Equation (10) is based on a continuous-time model of employment dynamics. Since CPS
14Here, we can derive C = C/[λEU(λIE−λUE)+λIE(λUE+λUI)+λIU(λUE+λEI+λEU)+λUE(λEI+λEU)+λUI(λEI+λEU)], where C is the sum of the number of unemployment, employment and not-in-the-labor-force.
14
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
Non-White
White
Non-White vs. White
1978 1983 1988 1993 1998 2003 2008 2013 20180.02
0.04
0.06
0.08
0.10
0.12
0.14
Young
Experienced
Young vs. Experienced
1978 1983 1988 1993 1998 2003 2008 2013 20180.02
0.04
0.06
0.08
0.10
0.12
0.14
No College Education
College Education
No College vs. Any College
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.04
0.05
0.06
0.07
0.08
0.09
0.10
Corr= 0.978
All Workers
Figure 4: Observed vs. Constructed Unemployment Rate
Gray-shaded area indicates NBER Recession periods. Solid lines represent the observed unemployment rate
while dashed and dotted lines represent constructed unemployment rate based on Shimer’s steady state
identity for disadvantaged and counterpart workers, respectively. Data source: FRED & IPUMS-CPS.
data is monthly, we will miss any transitions occurring within a single month. To correct for
this time-aggregation bias, we follow Shimer (2012) and Gomes (2015) and explicitly map the
continuous frequency of flows into the monthly data. In addition, we follow M. W. L. Elsby
et al. (2015) in adjusting our measured transitions to account for misclassification error. In
Appendix D we discuss these error corrections in detail.
Figure 4 shows that our constructed unemployment rate ug,ct tracks closely with the
observed unemployment rate for each demographic subgroup (ugt ), while Table 3 shows that
the correlation coefficients and R2 between the constructed and observed unemployment
rates are all close to 1. This indicates that this method for expressing the unemployment
rate in terms of transition rates works well for all demographic subgroups. In Sections 3 and
4 we will estimate how each transition rate contributes to unemployment rate fluctuations
15
Table 3: Constructed and Observed Unemployment
Workers’ Type Non-White White Young Experienced
Correlation 0.964 0.977 0.969 0.972R2 0.929 0.955 0.939 0.945
Workers’ Type No College With College All
Correlation 0.973 0.958 0.978R2 0.946 0.917 0.956
Correlation and R2 between the observed (ut) and constructed (ug,ct ) unem-ployment rates, based on Shimer (2012).
and levels, respectively.
3 Decomposing Unemployment Fluctuations
For our first set of results, we focus on decomposing the cyclical fluctuations in the
unemployment rate into transition rates. In order to isolate the cyclical component of the
unemployment rate we log linearize the constructed unemployment rate, which is based on
the steady state approximation from Equation (10), around the time trend over the period
for each demographic group g. In particular, we use an HP-filter to derive the time trend
for the constructed unemployment rate.15
We use Λg,t to represent the vector of the six transition rates between employment,
unemployment, and not-in-the-labor-force for demographic group g. In other words, Λg,t
includes λEUg,t , λEI
g,t, λUEg,t , λUI
g,t, λIEg,t, and λIU
g,t. The log-linearization expression for the constructed
unemployment rate can be written as
lnug,ct ≈ ln ug,ct +∑x∈X
∂ lnug,ct∂λxg,t
∣∣∣∣Λg,t=Λg,t
× (λxg,t − λxg,t)
= ln ug,ct +∑x∈X
∂ lnug,ct∂λxg,t
∣∣∣∣Λg,t=Λg,t
× λxg,t · (lnλxg,t − ln λxg,t).
(11)
Here we use ug,ct and λxt to denote the time trends for the constructed unemployment rate
ug,ct and transition rates λxt , respectively, where x ∈ X = {EU,EI,UE,UI, IE, IU}. We
compute the time trend based on an HP-filter. Equation (11) shows that we can decompose
the unemployment fluctuations (i.e., the log-deviation from the time trend), lnug,ct − ln ug,ct ,
into the components that depend on transition events. Figure 5 compares the log-linearized
15After transforming the frequency of the data to quarterly, we follow Fujita and Ramey (2009) andShimer (2012) and use an HP-filter with parameter 105 to decompose the trend and cyclical components. InAppendix F we show our results are similar if we instead use the sample average as the mean.
16
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4 R 2 = 1.0
Non-White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4 R 2 = 0.999
Young
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4
0.6
R 2 = 1.0
Without any college education
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4
0.6
R 2 = 1.0
All
Figure 5: Observed Unemployment Fluctuation vs. Approximation
Gray-shaded areas indicate NBER Recession periods. Black solid line shows observed unemployment fluc-
tuations, green dashed line the approximated first-order log-linearized unemployment fluctuations. Data
source: FRED & IPUMS-CPS.
approximation of the unemployment fluctuations based on the right-hand-side of Equation
(11) and the observed fluctuations in the unemployment rate. We construct the observed
fluctuations in the unemployment rate by removing the time trend based on the HP-filter
from the observed unemployment rates, which are obtained according to Equation (2). Figure
5 shows that the log-linearized approximation based on the transition events can capture
more than 99% of the fluctuations in the observed unemployment rate, despite the fact that
we only include the first-order terms in the Taylor expansion in Equation (11). Thus, nearly
all of the cyclical variation in the unemployment rate can be attributed to fluctuations in
the individual transition rates.
In Appendix E we analytically express workers’ total unemployment fluctuations F tott =
17
lnugt − ln ugt in terms of six factors, each of which depend on a different transition event:
F tott = FEU
t + FEIt + FUE
t + FUIt + F IE
t + F IUt + εt. (12)
The first two factors are related to employment exit rates, the third and fourth to unem-
ployment outflows, and the last two workers’ labor force participation.16
We build on the decomposition approach developed by Fujita and Ramey (2009) to
analyze the source of the unemployment fluctuations for each demographic group. Based on
Equation (12), the variance of total unemployment fluctuation cov (F tott , F tot
t ) can be written
as
cov (F tott , F tot
t ) = var (F tott ) =
∑x∈X
cov (F tott , F x
t ) + cov (F tott , εt), (13)
which can be further be rewritten as
1 =∑x∈X
cov (F tott , F x
t )
var (F tott )
+cov (F tot
t , εt)
var (F tott )
=∑x∈X
βx + βε. (14)
Here X is the set for flows EU,UE,EI,UI, IU, and IE, as we specified in Equation (12).
Based on Equation (14) we normalize the total contributions to unity, so each β coefficient
represents the percentage of unemployment fluctuation that can be attributed to flows x or
the error term. In particular, we estimate the model F xt = a+ βxF tot
t + et to isolate βx and
its confidence interval. This allows us to compare the β coefficients between different groups
and across different time periods.
Now that we have decomposed the unemployment rate into cyclical components, we be-
gin by examining the cyclical fluctuations in separations (FEU) and hiring (FUE) for each of
our focal demographic groups. In Figure 6 we see that hiring closely tracks the unemploy-
ment rate over the business cycle for each group, while separations exhibit substantially less
cyclicality. Thus, it appears that fluctuations in the unemployment rate for disadvantaged
groups, including the large increases during the Great Recession, are primarily driven by
firm hiring behavior.
Next we analytically compare the magnitude of these fluctuations between different
groups, to accomplish which we estimate the β coefficients derived in Equation (14). The
results are shown in Table 4. Each β represents the share of fluctuations in the unemploy-
ment rate for the particular demographic group captured by the transition of interest. Here
we see several common trends across the six demographic groups. First, the contribution of
hiring margins FUE is significantly larger than that of separation FEU for all groups. Sec-
ond, the total contribution of the flows between unemployment and out-of-the-labor-force,
16Alternatively, the unemployment rate can be decomposed into compound transitions, such as employ-ment to out-of-the-labor-force to unemployment. We show in Appendix E that our results are similar usingeither methodology.
18
1980 1985 1990 1995 2000 2005 2010 2015
0.4
0.2
0.0
0.2
0.4
λUE + λ IE λEU + λEI
Non-White Workers
1980 1985 1990 1995 2000 2005 2010 2015
0.4
0.2
0.0
0.2
0.4
λUE + λ IE λEU + λEI
Young Workers
1980 1985 1990 1995 2000 2005 2010 2015
0.4
0.2
0.0
0.2
0.4
0.6
λUE + λ IE λEU + λEI
No College Education
Figure 6: Sources of Unemployment Fluctuation (Figure 5)
Gray-shaded areas indicate NBER Recession periods. The black solid line represents observed unemployment
fluctuations. The green dashed line represents the factor depending on the job-finding rate, while the red
dotted line represents the factor depending on the separation rate. Data source: FRED & IPUMS-CPS.
F IU +FUI, is larger than that of the separation margin for all groups. The magnitude of the
participation margin is consistent with M. Elsby et al. (2015).
Although we see that the magnitudes of the components are broadly similar across demo-
graphic groups, there is one exception. The contribution of the hiring margin is substantially
larger for young workers than for experienced workers, indicating that the decline in job-
finding rates contributes relatively more to the increase in the unemployment rate during
recessions for young workers than for experienced workers.
In Appendix C we estimate the β coefficients separately for recessionary periods (in Table
C.6). Here we see similar results to Table 4, as the hiring margin explains a substantially
larger component of the cyclical fluctuations than the separation margin. Thus, despite the
fact that the magnitude of unemployment fluctuations are larger for disadvantaged groups,
19
Table 4: β Coefficient: Unemployment Fluctuations,Overall
Non-White Young No College All
λEU 0.167 0.158 0.213 0.206(0.014) (0.013) (0.012) (0.01)
λEI -0.054 -0.056 -0.05 -0.037(0.009) (0.009) (0.006) (0.006)
λUE 0.49 0.57 0.503 0.512(0.013) (0.014) (0.012) (0.01)
λUI 0.155 0.117 0.139 0.138(0.01) (0.009) (0.008) (0.007)
λIE 0.139 0.145 0.1 0.075(0.009) (0.007) (0.006) (0.005)
λIU 0.11 0.073 0.102 0.111(0.011) (0.011) (0.008) (0.008)
ε -0.006 -0.006 -0.007 -0.005(0.001) (0.001) (0.001) (0.001)
White Experienced With College All
λEU 0.218 0.246 0.216 0.206(0.011) (0.012) (0.012) (0.01)
λEI -0.03 -0.024 -0.017 -0.037(0.005) (0.005) (0.007) (0.006)
λUE 0.522 0.464 0.523 0.512(0.011) (0.01) (0.013) (0.01)
λUI 0.128 0.149 0.12 0.138(0.007) (0.007) (0.008) (0.007)
λIE 0.058 0.052 0.042 0.075(0.005) (0.005) (0.006) (0.005)
λIU 0.108 0.118 0.12 0.111(0.008) (0.008) (0.009) (0.008)
ε -0.005 -0.005 -0.004 -0.005(0.001) (0.001) (0.001) (0.001)
Standard errors in parentheses.
we see that hiring can explain the largest share of the cyclical variation in the unemployment
rate across worker demographic groups.
4 Decomposing Level Differences between Groups
Now that we have determined that cyclical fluctuations in the unemployment rate are
primarily driven by movements into employment, we turn to examining unemployment rate
differences between groups. Recall from Figure 1 that non-white, young, and workers with
no college have substantially higher unemployment rates at all phases of the business cycle
than their counterpart demographic groups. In order to determine which transition rates
can explain these gaps, we return to the unemployment rate decomposition that we derived
20
in Section 3. As in Equation (11), the (log) unemployment gap between the disadvantaged
demographic group g and its counterpart demographic group g can be written as follows:
lnuc,gt − lnuc,gt ≈∑x∈X
∂ lnuct∂λxt
∣∣∣∣Λct=Λg
t
× (λx,gt − λx,gt )
≈∑x∈X
∂ lnuct∂λxt
∣∣∣∣Λgt =Λg
t
× λx,gt · (lnλx,gt − lnλx,gt ).
(15)
Here, as before, we use c to denote that the unemployment rates are constructed according
to Equation (10) and X to represent the set of flows, EU,UE,EI,UI, IE, and IU. The decom-
position approach in Equation (15) is similar to Equation (11) and shows that the difference
between the unemployment rates of demographic groups g and g can be decomposed into
the difference in the transition rates of the two groups.
Thus, the level difference in the unemployment rate between two demographic groups
depends on six transition events as follows:
F gapt = FEU
t + FEIt + FUE
t + FUIt + F IE
t + F IUt + εt,
where F gapt is equal to the observed log unemployment gap, lnuc,gt − lnuc,gt . On the right-
hand-side, for example, FEU represents the proportion of the unemployment gap driven by
the difference between the separation rates of the two demographic groups, λEU.
In order to evaluate how well the first-order log-linearization approximation based on
Equation (15) performs in capturing the observed gap in the unemployment rate for pairs
of demographic groups, we obtain the approximated unemployment gap based on the right-
hand-side of Equation (15). Figure 7 compares the approximated unemployment gaps with
the observed gaps. We see that our approximation method captures 99% of the difference in
unemployment rates for each pair of demographic groups. This indicates the decomposition
approach we propose in Equation (15) is reliable.
To assess how much of the difference in the unemployment rate can be accounted for by
differences in transition rates, we use the following identity:
1 =FEUt
F gapt
+FEIt
F gapt
+FUEt
F gapt
+FUIt
F gapt
+F IEt
F gapt
+F IUt
F gapt
+εtF gapt
=∑x∈X
rkt + rεt (16)
where X is again the set of flows, EU,UE,EI,UI, IE, and IU. By estimating each ratio rktand constructing confidence intervals for rkt we can compare the relative contribution of each
transition rate in explaining the total difference in unemployment rates between groups.17
17The β coefficients can only reveal the importance of each transition event in explaining variation inthe unemployment gap, so we compute the ratio r here. In Appendix G we apply Equation (14) to theunemployment gap and give these estimated β coefficients.
21
1978 1983 1988 1993 1998 2003 2008 2013 20180.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
R 2 = 0.99
Non-White Workers
1978 1983 1988 1993 1998 2003 2008 2013 20180.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
R 2 = 0.919
Young Workers
1978 1983 1988 1993 1998 2003 2008 2013 20180.5
0.6
0.7
0.8
0.9
1.0R 2 = 0.977
No College Education
Figure 7: Observed Unemployment Gap vs. Approximation
Gray-shaded areas indicate NBER Recession periods. Black solid line represents observed unemployment gap
and green dashed line represents approximated first-order log-linearized unemployment gap. Data source:
FRED & IPUMS-CPS.
We begin by examining the contributions of separation rates λEU (component FEU) and
job-finding rates λUE (component FUE) in the differences in unemployment rates between
each pair of demographic groups. Figure 8 shows that for young workers and those with
no college the separation rate explains the largest share of the differences in unemployment
rates, while the job-finding rate explains none of the difference for young workers and little
of the unemployment gap for workers with no college. However, for non-white workers we
see that separation and job-finding rates have similar magnitudes.
To compare these demographic differences analytically, Table 5 shows rk, the estimated
fraction of the difference in the unemployment rate that can be attributed to transition
flow event λk. Positive values indicate that the flow contributes to a larger gap in the
unemployment rate between the two groups, while a negative value indicates the flow serves
22
1980 1985 1990 1995 2000 2005 2010 2015
0.0
0.2
0.4
0.6
0.8
1.0
lnug − lnu g λUE + λ IE λEU + λEI
Non-White Workers
1980 1985 1990 1995 2000 2005 2010 20150.6
0.4
0.2
0.0
0.2
0.4
0.6
0.8
1.0 lnug − lnu g λUE + λ IE λEU + λEI
Young Workers
1980 1985 1990 1995 2000 2005 2010 2015
0.0
0.2
0.4
0.6
0.8
1.0
lnug − lnu g λUE + λ IE λEU + λEI
No College Education
Figure 8: Sources of Unemployment Gap
Percentage of the total unemployment gap that can be accounted for by job-finding and separation rates for
each type of disadvantaged worker. Data source: FRED & IPUMS-CPS.
to lessen differences in the unemployment rate between the two groups. We first note that the
signs on the direct flows in and out of unemployment are consistent with the regression results
in Table 2. In particular, faster inflow rates from employment and not-in-the-labor-force for
disadvantaged groups serve to increase the unemployment gap, while slower outflows from
unemployment to not-in-the-labor-force serve to decrease the gap. Flows from unemployment
to employment differ across demographic groups: for non-white workers and those with no
college this flow increases the unemployment rate gap, since these groups have relatively
lower hiring rates from unemployment. On the other hand, for young workers the flow from
unemployment to employment decreases the unemployment rate gap, since young workers
have higher hiring rates from unemployment than more-experienced individuals.
In the regression results in Table 2, it is hard to evaluate the impact of flows between
employment and not-in-the-labor-force on the unemployment rate. From Table 2 we see
23
that the disadvantaged groups have relatively higher flow rates between employment and
not-in-the-labor-force, except for individuals with no college, who are relatively less likely to
be hired directly from out-of-the-labor-force. In Table 5 we see that flows from employment
to not-in-the-labor-force increase the unemployment rate gap for all three groups, while flows
from not-in-the-labor-force to employment reduce the gaps for non-white and young workers,
but increase the gap for workers with no college.
In addition to evaluating whether each flow increases or decreases the unemployment
rate gap, we can also use the magnitudes of the rk estimates to rank the relative importance
of each flow. We find somewhat different patterns for each of the three disadvantaged
demographic groups, so we evaluate each in turn.
For non-white workers, flows from unemployment to employment are the largest factor
(0.44), followed by employment to unemployment (0.35), then not-in-the-labor-force to un-
employment (0.29). All other flows are relatively small in magnitude. These results indicate
that the lower hiring rates from unemployment for non-white workers compared with white
workers represents the most important factor in explaining the elevated unemployment rates
for non-white workers. This is in contrast to young workers and those with no college, for
whom the hiring margin plays a small role (for those with no college) or reduces the un-
employment rate gap (young workers). The importance of hiring for non-white workers is
consistent with evidence (see, e.g., Freeman, 1973 and Bertrand & Mullainathan, 2004) that
non-white job applicants face dramatically lower callback rates compared with otherwise
identical white job applicants.
In contrast, for young workers the largest component of the unemployment rate gap is
flows from not-in-the-labor-force to unemployment (0.86), followed by employment to unem-
ployment (0.78), and then employment to not-in-the-labor-force (0.35). This is consistent
with young individuals entering the labor market from schooling or child-rearing.18 However,
the higher rate of exits from employment to unemployment also explains a large fraction of
the unemployment rate gap for young workers. Finally, several transitions serve to reduce
the unemployment rate gap: direct hires from not-in-the-labor-force (-0.6), movement from
unemployment to not-in-the-labor-force (-0.19), and hires from unemployment (-0.19).
Finally, for workers with no college movements from employment to unemployment are
by far the largest contributor to the unemployment rate gap (0.61). The next two largest
flows are substantially smaller in magnitude: employment to not-in-the-labor-force (0.19)
and not-in-the-labor-force to employment (0.18), both of which only indirectly impact the
unemployment rate.
We can also evaluate whether inflows to unemployment or outflows from unemployment
can explain a larger share of the unemployment rate gap between disadvantaged groups and
their counterpart demographic groups. For all groups, flows into unemployment explain a
18See Guo (2018) for an analysis of young workers’ educational decisions over the business cycle.
24
Table 5: r Ratio: Compositions of Unem-ployment Gap, Overall
Non-White Young No College
λEU 0.35 0.776 0.616(0.005) (0.008) (0.004)
λEI 0.12 0.351 0.185(0.003) (0.005) (0.003)
λUE 0.436 -0.188 0.116(0.006) (0.007) (0.006)
λUI -0.094 -0.193 -0.079(0.003) (0.006) (0.003)
λIE -0.056 -0.6 0.178(0.004) (0.009) (0.002)
λIU 0.293 0.857 0.011(0.004) (0.01) (0.003)
ε -0.048 -0.002 -0.027(0.002) (0.001) (0.001)
Standard errors in parentheses.
substantially larger fraction of the unemployment rate gap than flows out. However, for
non-white workers outflows can explain about 1/3 of the gap, while for young workers and
workers with no college outflows explain almost none of the gap.
Because the source of unemployment gap may differ during recessions, as in Section 3,
we estimate the fractions rk and their confidence intervals for the four periods of recession
in Table C.7, Appendix C. We find that the results in Table 5 hold for all four recessions in
our sample periods. Thus, we conclude that the unemployment gap is acyclical for all three
demographic groups.
5 Mechanisms
We have shown that elevated separation rates play an important role in higher unem-
ployment rates for disadvantaged groups, while greater decreases in hiring are key to cyclical
variation in unemployment rates. However, these changes in separations and hiring could be
due to either firm or worker behavior. In this section we investigate the mechanisms behind
these changes in flows, beginning with separations and then moving to hiring.
Separations
There are a variety of reasons disadvantaged workers may face higher separation rates.
Employers may practice “last in, first out” policies, which would make younger workers
more likely to lose their jobs during downsizing. Disadvantaged workers may be employed
in more volatile industries or occupations, leading to higher separation rates. Disadvantaged
25
Table 6: Employment Exit: EU + EI
Model a Model b Model c Model d
No Fixed Effect Industry Occupation Industry & OccupationFixed Effect Fixed Effect Fixed Effect
Intercept 2.1886*** 86.6810*** 3.3182*** 3.3629***(0.0058) (0.0410) (0.0695) (0.0695)
Young 3.1509*** 1.7549*** 1.4154*** 1.3136***(0.0109) (0.0079) (0.0074) (0.0075)
Non-White 1.2974*** 0.6671*** 0.4890*** 0.5368***(0.0134) (0.0098) (0.0092) (0.0092)
No College 2.4437*** 1.3658*** 0.6893*** 0.5960***(0.0086) (0.0065) (0.0065) (0.0066)
R2 0.0084 0.4878 0.5508 0.5536N 23343940 23343940 23343940 23343940
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
groups may be less connected to the labor market or by more likely to have personal and
family obligations which interrupt their ability to work consistently, leading to higher rates
of voluntary separation. Finally, discrimination may cause separations if employers choose
to lay off workers from less-favored groups. This is consistent with the “last hired, first fired”
hypothesis.19
We investigate this directly by examining how much of the difference in separation rates
between demographic groups can be explained by industry and occupation fixed effects.
Further, for individuals who exit employment to unemployment, we can divide between vol-
untary and involuntary separations, allowing us to estimate whether disadvantaged workers
are choosing to exit employment more often or if these elevated separations are driven by
employer behavior.
In Table 6 we regress the total employment exit rates (i.e., EU+EI) on indicators for each
of the three demographic groups: young, non-white, and no college education. In Column
(1) we see that membership in each group is associated with an increase in separation rate.
In Columns (2) through (4) we add in successive controls for the type of job to see how
much each reduces the explanatory power of the group indicator variables. Column (2)
adds fixed effects for major occupation, Column (3) adds fixed effects for major industry,
and Column (4) includes both. For each demographic group industry and occupation fixed
effects substantially reduce the coefficients, with occupations explaining a larger fraction of
separation rates than industry. All told, industry and occupation can explain 58% of the
higher separation rate for young workers, 59% for non-white workers, and 76% for workers
19See Couch and Fairlie (2010), Couch et al. (2016) and Xu and Couch (2017).
26
Table 7: Voluntary Separations
Model a Model b Model c
Intercept 0.0427*** 0.0811*** 0.0609***(0.0012) (0.0032) (0.0042)
Young 0.2758*** 0.2766*** 0.2833***(0.0026) (0.0026) (0.0096)
Non-White 0.0305*** 0.0298*** 0.0491***(0.0027) (0.0027) (0.0098)
No College 0.1248*** 0.1261*** 0.1620***(0.0018) (0.0018) (0.0069)
Unemployment Rate -0.6269*** -0.2981***(0.0498) (0.0651)
Young × Unemp. -0.1100(0.1462)
Non-White × Unemp. -0.3131**(0.1517)
No College × Unemp. -0.5723***(0.1046)
R2 0.0010 0.0010 0.0010N 23343940 23343940 23343940
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
with no college.
We then consider whether these higher rates are due to voluntary or involuntary separa-
tions. In the CPS all unemployed workers are asked about the reason for their job loss. We
classify individuals who report leaving their last job voluntarily as voluntary separations,
and individuals who were laid off, had a temporary job conclude, or responded “other job
losing” as involuntary. Table 7 gives the rates of voluntary separations. In Column (1)
we see that all three disadvantaged groups have higher rates of voluntary separations when
compared with white, older, and college-educated individuals. In Column (2), we see that
an increase in the unemployment rate is correlated with a decrease in the rate of voluntary
exits. Further, Column (3) shows this decrease is faster for non-white and no-college individ-
uals when compared with the respective dominant groups. The coefficient for young workers
is also negative, but not statistically significant. Thus, although disadvantaged groups are
more likely to voluntarily leave jobs, they see a greater reduction in voluntary separations
during recessions when compared with dominant groups.
Next we focus on involuntary separations. Across groups, about 90% of separations
are involuntary. In Column (1) of Table 8 we see that involuntary separation rates are
substantially greater for disadvantaged groups. As opposed to voluntary separations, we
27
Table 8: Involuntary Separations
Model a Model b Model c
Intercept 0.4595*** -0.1027*** 0.2956***(0.0027) (0.0082) (0.0106)
Young 0.4026*** 0.3912*** -0.0635***(0.0049) (0.0049) (0.0198)
Non-White 0.3412*** 0.3515*** 0.1540***(0.0065) (0.0065) (0.0250)
No College 0.7150*** 0.6950*** 0.1222***(0.0041) (0.0041) (0.0170)
Unemployment Rate 9.1752*** 2.6991***(0.1292) (0.1726)
Young × Unemp. 7.2645***(0.3185)
Non-White × Unemp. 3.2218***(0.4095)
No College × Unemp. 9.1682***(0.2720)
R2 0.0018 0.0020 0.0021N 23343940 23343940 23343940
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
see in Column (2) that involuntary separations increase with the state unemployment rate.
Column (3) shows these increases are larger for disadvantaged groups.
Thus, the fact that disadvantaged workers experience higher separation rates appears to
be primarily due to the types of jobs these workers hold. Although they are somewhat more
likely to voluntarily separate from their jobs (to unemployment), this effect is small compared
with involuntary separations. Between 25 and 40% of the difference in separation rates is
not explained by major industry and occupational groups. This could be due to differences
in performance or to active discrimination. We also find that the increase in separation to
unemployment among disadvantaged workers is entirely driven by firm-initiated separations.
Hiring
Disadvantaged workers may see lower hiring rates for a variety of reasons. They may be
more likely to lack skills that employers demand, for instance, as the share of jobs requiring
a college degree rises, less-educated individuals will find a smaller set of relevant vacancies.
Similarly, many job postings require a certain number of years of relevant work experience
as a way to screen for on-the-job acquired skills, making it harder for a new labor market
entrant to find work. Employers increase these skill requirements during recessions, making
28
it even more difficult for disadvantaged workers to find employment (Hershbein and Kahn
(2018), Modestino, Shoag, and Ballance (2015)). Forsythe (2020) finds that this is a larger
issue for young workers than for less-educated workers.
There is also well-documented discrimination in hiring, with non-white individuals espe-
cially facing discrimination in the labor market (e.g. Bertrand & Mullainathan, 2004). If
tight labor markets serve to constrain discriminatory preferences, this may lead to worsening
discrimination.
There may also be differences between groups in search behavior and search efficacy. If
disadvantaged workers spend less time searching or use less effective methods of search, they
may have a harder time finding a job. Moreover, many jobs are found informally via social
networks. If these workers have fewer social connections with individuals who may have
leads on job openings, they may also find job searching more difficult.
To investigate differences in hiring, we use data on search behavior from the American
Time Use Survey and the CPS. This allow us to investigate whether search effort differs
between groups and whether these differences vary cyclically. We cannot directly test other
mechanisms, such as efficacy of search and discrimination.
The ATUS data is available from 2003 through 2017 and records the amount of time the
individual spent on any given activity in the previous 24-hour period. We use two measures
to capture search behavior. First, we construct an indicator for whether the individual spent
any time searching for work to measure the extensive margin of search. Second, we measure
the number of minutes the individual spent searching, conditional on spending at least one
minute, to capture the intensive margin of search. In addition, the CPS measures the number
of methods of search used by unemployed individuals.
Although we are unable to link search behavior to job-search outcomes, we construct a
measure of the fraction of search time the individual spent interviewing as a measure for the
return to search effort. A higher fraction indicates that the individual is interviewing, and
hence getting more response from employers, for each minute of search. Summary statistics
are given in Appendix Table A.2.
Each coefficient in Table 9 comes from a separate regression, which includes year fixed
effects and is weighted using sampling weights. In the top panel, we see young and non-white
individuals are about one percentage point more likely to be searching, while individuals with
no college are about 1/4 of a percentage point less likely to be searching. When we separate
individuals by labor market status we see that individuals with no college are less likely
to spend any time searching in each category, while non-white workers are more likely to
spend time searching in each category. However, while young workers are more likely to be
searching from employment and NILF, they are 10 percentage points less likely to spend any
time searching while unemployed.
Interestingly, we find young and non-white workers are more likely to be searching on
29
the job than experienced and white workers, respectively, while workers with no college
are less likely to search on the job than college-educated workers. While it is outside the
scope of this paper, R. J. Faberman, Mueller, Sahin, and Topa (2017) found that workers
searching on-the-job receive more and better job offers and Bradley and Gottfries (2018)
find this is consistent with models of job-ladders. Arbex, O’Dea, and Wiczer (2019) find
that networks are an important source of these job-to-job moves, with better networked
individuals climbing job ladders faster. These career progressions are an important source of
earnings growth, so differences in the extent of on-the-job search may contribute to earnings
gaps between workers with and without any college.
In the second panel of Table 9 we investigate the intensive margin of search by assessing
the log of the number of minutes spent searching, conditional on spending at least one
minute searching. Since most workers do not spend time searching on any given day, this
dramatically reduces the sample sizes and power. We find that young workers spend 22% less
time searching than experienced workers, while workers with no college spend 24% less time
searching than those with any college. In contrast, there does not seem to be any difference
in search time between non-white and white workers overall, though NILF non-white workers
search 40% more than NILF white workers.
Finally, in the bottom panel of Table 9 we measure the percent of search time spent
interviewing and find that young workers spend about 6 percent more of their search time
interviewing than do more-experienced workers. Workers with no college are not statistically
significantly different from those with college. Non-white workers spend about 5 percent less
of their search time interviewing.
We can now reinterpret our results on transition dynamics. Although we showed that
young workers see faster transition rates from unemployment to employment, this does not
appear to be driven by search effort, as young unemployed workers are half as likely to be
searching as more-experienced workers (10 percentage points vs. 20 percentage points) and
spend 25% less time searching than do more-experienced workers. Nonetheless, they find
more interview success and faster hiring rates than more-experienced unemployed workers.
On the other hand, less-educated individuals’ slower transitions from unemployment to em-
ployment could be due to their less intensive search methods. Non-white workers search
more intensely on average, but see much lower hiring rates and spend less time interviewing
per minute of search. This indicates that these workers find less success from search effort,
consistent with resume audit studies showing that non-white workers receive lower callback
rates than their white peers (Bertrand & Mullainathan, 2004).
In Table 10 we examine whether search effort varies cyclically. We again use the ATUS
extensive margin measure for any search and intensive margin measure of the log of minutes
spent searching. Due to small sample sizes we also use the CPS measure of the log number
of methods of search. This measure is only available for unemployed individuals. The ATUS
30
Table 9: ATUS Search Behavior
All Employed Unemployed NILFATUS % With Any Search
Young 0.98*** 0.65*** -9.53*** 0.47**(0.12) (0.11) (1.07) (0.14)
No College -0.26** -0.19** -10.84*** -0.36***(0.08) (0.07) (1.16) (0.09)
Nonwhite 1.03*** 0.34** 0.64 0.49***(0.13) (0.12) (1.19) (0.12)
N 201,151 125,002 9,313 66,836ATUS Search conditional on Any Search
All Employed Unemployed NILFYoung -0.22** -0.10 -0.25** -0.34
(0.07) (0.12) (0.09) (0.22)No College -0.24*** -0.05 -0.39*** 0.01
(0.07) (0.12) (0.08) (0.20)Nonwhite -0.00 0.08 -0.09 0.40*
(0.07) (0.12) (0.09) (0.20)N 2,194 603 1,387 204
Interview RateYoung 5.55** 5.70 3.17 9.21
(1.82) (3.87) (1.87) (5.75)No College -1.56 -6.82 1.97 -4.96
(1.63) (3.79) (1.65) (4.81)Nonwhite -5.10*** -3.36 -4.53*** -8.88
(1.42) (4.10) (1.36) (4.90)N 2,194 603 1,387 204
Data from ATUS, 2003–2017. Robust standard errors in parentheses.* p < 0.1, ** p < 0.05, ***p < 0.01. Each coefficient is from a separateregression that includes year fixed effects.
31
data includes 5-year fixed effects (2003–2007, 2008–2012, 2013–2018), and the CPS data
includes year and month fixed effects, from 1994 through 2017. The unemployment rate
is the national non-seasonally-adjusted percentage. All specifications are weighted using
sampling weights and robust standard errors are shown.
In the first column of Table 10 we see that, on average, the share of workers spending
time searching for new work increases with the unemployment rate. However, in Column
(5) we see little evidence that the number of minutes spent searching increases with the
unemployment rate, with the caveat that the ATUS data is underpowered. In Column (9)
we see that the log of the number of search methods used increases by 2% for each one
percentage point increase in the unemployment rate. Thus, on net we see that search effort
increases with the unemployment rate, which is consistent with the literature (J. Faberman,
Kudlyak, et al. (2016), Mukoyama, Patterson, and Sahin (2018)).
When we consider heterogeneity by demographic status, we see little difference by race or
by education. Young workers’ rate of search appears to be less cyclical than more-experienced
workers, driven by a fall in search among unemployed young workers, but there is no similar
effect observed in the search intensity measures from either the ATUS or CPS. In the CPS
measure, we see that search effort for no-college workers may increase by less than for those
with any college education. In net, although there is some weak evidence that young and
less-educated job seekers may increase their search effort by less, all job seekers increase
search effort during periods of high unemployment. There is therefore little evidence that
worse labor market outcomes for disadvantaged workers during recessions is due to search
behavior.
Thus, while the flow-decomposition results indicate that hiring can explain the cyclical
unemployment gap between groups, we can weakly say that it does not appear that this is
due to worker search behavior. We have shown that separations are the primary explanation
for these differences in unemployment. In this section we find that between 60 and 75%
of the separation gap can be explained by demographic sorting between occupations and
industries.
For non-white workers, we found that hiring was almost as important as separations in
explaining the gap in the unemployment rate. We found that these workers see a particularly
low interview response rate per time spent actively searching in the time use data. Coupled
with audit evidence on low call-back rates for non-white workers, this suggests that hiring
discrimination is partially responsible for non-white workers’ elevated unemployment rates.
6 Conclusions and Policy Implications
In this paper, we examine the sources of elevated unemployment rates for three demo-
graphic subgroups: young workers, non-white workers, and workers with no college educa-
32
Tab
le10
:C
ycl
ical
Sea
rch
Beh
avio
r
AT
US
Any
Sea
rch
AT
US
Log
Sea
rch
CP
SL
ogSea
rch
All
Em
plo
yed
Unem
plo
yed
NIL
FA
llE
mplo
yed
Unem
plo
yed
NIL
FU
nem
plo
yed
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Pan
elA
:E
xp
erie
nce
You
ng
2.13
***
0.94
*4.
730.
50-0
.10
-0.0
1-0
.18
0.09
-0.1
1***
(0.4
4)(0
.41)
(4.0
0)(0
.59)
(0.2
6)(0
.48)
(0.3
4)(0
.75)
(0.0
1)U
nem
p.
Rat
e0.
18**
*0.
010.
930.
06*
-0.0
10.
02-0
.03
-0.0
20.
02**
*(0
.03)
(0.0
3)(0
.56)
(0.0
2)(0
.03)
(0.0
6)(0
.03)
(0.0
7)(0
.00)
You
ng×
Unem
p.
-0.1
9**
-0.0
5-2
.15*
**-0
.00
-0.0
2-0
.02
-0.0
1-0
.05
0.00
(0.0
7)(0
.06)
(0.5
6)(0
.10)
(0.0
4)(0
.07)
(0.0
5)(0
.10)
(0.0
0)P
anel
B:
Educa
tion
No
Col
lege
-0.0
9-0
.34
-6.0
60.
16-0
.33
-0.4
1-0
.44
-0.5
10.
19**
*(0
.30)
(0.2
5)(4
.33)
(0.3
5)(0
.25)
(0.4
6)(0
.29)
(0.7
2)(0
.01)
Unem
p.
Rat
e0.
14**
-0.0
10.
490.
11-0
.01
0.01
-0.0
3-0
.07
0.02
***
(0.0
4)(0
.04)
(0.6
3)(0
.06)
(0.0
3)(0
.05)
(0.0
3)(0
.09)
(0.0
0)N
oC
olle
ge×
Unem
p.
-0.0
30.
02-0
.72
-0.0
80.
010.
060.
010.
07-0
.01*
**(0
.05)
(0.0
4)(0
.61)
(0.0
6)(0
.03)
(0.0
6)(0
.04)
(0.1
0)(0
.00)
Pan
elC
:R
ace
Non
whit
e0.
690.
730.
880.
430.
060.
29-0
.09
-0.0
8-0
.01
(0.4
7)(0
.47)
(4.4
9)(0
.44)
(0.2
7)(0
.63)
(0.3
2)(0
.68)
(0.0
1)U
nem
p.
Rat
e0.
12**
*0.
010.
000.
05-0
.01
0.03
-0.0
2-0
.07
0.02
***
(0.0
4)(0
.03)
(0.5
1)(0
.03)
(0.0
3)(0
.05)
(0.0
3)(0
.09)
(0.0
0)N
onw
hit
e×
Unem
p.
0.06
-0.0
6-0
.04
0.01
-0.0
1-0
.03
-0.0
00.
060.
00(0
.07)
(0.0
7)(0
.63)
(0.0
7)(0
.04)
(0.0
9)(0
.04)
(0.0
9)(0
.00)
N20
1,15
112
5,00
29,
313
66,8
362,
194
603
1,38
766
,836
641,
787
Dat
afr
omA
TU
S,
2003
–201
7an
dC
PS
1994
–2017.
Rob
ust
stan
dard
erro
rsin
par
enth
eses
.*p<
0.1,
**p<
0.05,
***p
<0.0
1.
AT
US
spec
ifica
tion
sin
clu
de
5-ye
arfi
xed
effec
ts;
CP
Ssp
ecifi
cati
on
incl
ud
esm
onth
and
year
fixed
effec
ts.
33
tion. We decompose the unemployment rate into flows between employment, unemployment,
and out-of-the-labor-force. For all three groups, as well as for the labor market as a whole,
reduced hiring is the primary source of cyclical fluctuations in the unemployment rate. This
indicates that policies that seek to boost labor demand during recessions will be beneficial
for all workers.
We find important differences in the sources of the persistent gap in the unemployment
rate between groups. Separations explain the entire gap for young workers, while for workers
with no college education separations explain most of the gap but hiring plays a small role,
and for non-white workers hiring is nearly as important as separations. For workers with no
college education 75% of the separation rate can be explained by industry and occupation,
indicating that much of this gap is due to sorting to more volatile jobs. However, we do find
that workers without any college spend less time searching than college educated workers
which could explain the role that hiring plays for these workers. For younger workers industry
and occupation only can account for 60% of the elevated separation rates, suggesting lifecycle
differences in job-shopping play a role. For non-white workers hiring can explain almost half
of the gap, although non-white workers spend more time searching for employment than
other groups. In light of evidence that shows non-white applicants face additional barriers
to being hired, this suggests that discrimination can likely explain part of non-white workers’
persistently high unemployment rates. Policies to address separation rates are thus unlikely
to be sufficient to close the racial unemployment gap.
34
References
Albanesi, S., & Sahin, A. (2018). The gender unemployment gap. Review of EconomicDynamics , 30 , 47 - 67. Retrieved from http://www.sciencedirect.com/science/
article/pii/S1094202517301229 doi: https://doi.org/10.1016/j.red.2017.12.005Arbex, M., O’Dea, D., & Wiczer, D. (2019). Network search: Climbing the job ladder faster.
International Economic Review , 60 (2), 693–720.Barrett, N. S., & Morgenstern, R. D. (1974). Why do blacks and women have high unem-
ployment rates? Journal of Human Resources , 452–464.Bertrand, M., & Mullainathan, S. (2004, September). Are emily and greg more employable
than lakisha and jamal? a field experiment on labor market discrimination. Ameri-can Economic Review , 94 (4), 991-1013. Retrieved from https://www.aeaweb.org/
articles?id=10.1257/0002828042002561 doi: 10.1257/0002828042002561Borowczyk-Martins, D., Bradley, J., & Tarasonis, L. (2017). Racial discrimination in the
u.s. labor market: Employment and wage differentials by skill. Labour Economics , 49 ,106 - 127. Retrieved from http://www.sciencedirect.com/science/article/pii/
S092753711730307X doi: https://doi.org/10.1016/j.labeco.2017.09.007Bradley, J., & Gottfries, A. (2018). A job ladder model with stochastic employment oppor-
tunities.Cairo, I., & Cajner, T. (2018). Human capital and unemployment dynamics: Why more ed-
ucated workers enjoy greater employment stability. The Economic Journal , 128 (609),652–682.
Cairo, I., & Cajner, T. (2018). Human capital and unemployment dynamics: Whymore educated workers enjoy greater employment stability. The Economic Journal ,128 (609), 652-682. Retrieved from https://onlinelibrary.wiley.com/doi/abs/
10.1111/ecoj.12441 doi: 10.1111/ecoj.12441Choi, S., Janiak, A., & Villena-Roldan, B. (2015). Unemployment, participation and worker
flows over the life-cycle. The Economic Journal , 125 (589), 1705–1733. Retrieved fromhttp://dx.doi.org/10.1111/ecoj.12176 doi: 10.1111/ecoj.12176
Clark, K. B., & Summers, L. H. (1982). The dynamics of youth unemployment. In The youthlabor market problem: Its nature, causes, and consequences (pp. 199–234). Universityof Chicago Press.
Couch, K. A., & Fairlie, R. (2010). Last hired, first fired? black-white unemployment andthe business cycle. Demography , 47 (1), 227–247. Retrieved from http://dx.doi.org/
10.1353/dem.0.0086 doi: 10.1353/dem.0.0086Couch, K. A., Fairlie, R., & Xu, H. (2016, Feburary). Racial differences in labor market
transitions and the great recession. Working Paper .Elsby, M., Hobijn, B., & Sahin, A. (2015, May). On the importance of the participation
margin for labor market fluctuations. Journal of Monetary Economics , 72 , 64-82.Elsby, M., Michaels, R., & Solon, G. (2009). The Ins and Outs of Cyclical Unemployment.
American Economic Journal(November 2006), 84–110.Elsby, M. W., Hobijn, B., & Sahin, A. (2010). The labor market in the great recession.
brookings papers on economic activity, spring 2010. Brookings Institution.Elsby, M. W. L., Michaels, R., & Ratner, D. (2015, September). The beveridge curve:
A survey. Journal of Economic Literature, 53 (3), 571-630. Retrieved from http://
35
www.aeaweb.org/articles?id=10.1257/jel.53.3.571 doi: 10.1257/jel.53.3.571Faberman, J., Kudlyak, M., et al. (2016). What does online job search tell us about the
labor market? FRB Chicago Economic Perspectives , 40 (1).Faberman, R. J., Mueller, A. I., Sahin, A., & Topa, G. (2017). Job search behavior among
the employed and non-employed (Tech. Rep.). National Bureau of Economic Research.Forsythe, E. (2020). Why Don’t Firms Hire Young Workers During Recessions?Freeman, R. B. (1973). Decline of labor market discrimination and economic analysis. The
American Economic Review , 63 (2), 280–286. Retrieved from http://www.jstor.org/
stable/1817087
Fujita, S., & Ramey, G. (2009). The cyclicality of separation and job finding rates. In-ternational Economic Review , 50 (2), 415–430. Retrieved from http://dx.doi.org/
10.1111/j.1468-2354.2009.00535.x doi: 10.1111/j.1468-2354.2009.00535.xGobillon, L., Rupert, P., & Wasmer, E. (2014). Ethnic unemployment rates and fric-
tional markets. Journal of Urban Economics , 79 , 108 - 120. Retrieved from http://
www.sciencedirect.com/science/article/pii/S0094119013000533 (Spatial Di-mensions of Labor Markets) doi: https://doi.org/10.1016/j.jue.2013.06.001
Gomes, P. (2015, Mar 26). The importance of frequency in estimating labour markettransition rates. IZA Journal of Labor Economics , 4 (1), 6. Retrieved from https://
doi.org/10.1186/s40172-015-0021-9 doi: 10.1186/s40172-015-0021-9Guo, N. (2018, March). The effect of an early career recessionon schooling and lifetime
welfare. International Economic Review . (Forthcoming)Hershbein, B. B., & Kahn, L. B. (2018). Do Recessions Accelerate Routine-Biased Tech-
nological Change? Evidence from Vacancy Postings. American Economic Review ,108 (7), 1737–1772.
Hoynes, H., Miller, D. L., & Schaller, J. (2012). Who Suffers During Recessions? Journalof Economic Perspectives , 26 (3), 27–48.
Lang, K., & Lehmann, J.-Y. K. (2012, December). Racial discrimination in the labor market:Theory and empirics. Journal of Economic Literature, 50 (4), 959-1006. Retrievedfrom http://www.aeaweb.org/articles?id=10.1257/jel.50.4.959 doi: 10.1257/jel.50.4.959
Modestino, A. S., Shoag, D., & Ballance, J. (2015). Upskilling: do employers demand greaterskill when workers are plentiful? Review of Economics and Statistics , 1–46.
Mukoyama, T., Patterson, C., & Sahin, A. (2018, January). Job search behavior overthe business cycle. American Economic Journal: Macroeconomics , 10 (1), 190-215.Retrieved from http://www.aeaweb.org/articles?id=10.1257/mac.20160202 doi:10.1257/mac.20160202
Nickell, S. (1979). Education and lifetime patterns of unemployment. Journal of PoliticalEconomy , 87 (5, Part 2), S117–S131.
Shimer, R. (2012, April). Reassessing the ins and outs of unemployment. Review of EconomicDynamics , 15 (2), 127-148.
Xu, H., & Couch, K. A. (2017). The business cycle, labor market transitions by age, andthe great recession. Applied Economics , 49 (52), 5370-5396. Retrieved from https://
doi.org/10.1080/00036846.2017.1307932 doi: 10.1080/00036846.2017.1307932
36
Appendix A Supplemental Tables and Figures
This appendix contains tables of summary data.
Table A.1 shows the shares of the eight types of workers in the sample. Table A.3 gives
unemployment rates and dates of minimum and maximum rates for subgroups during the
NBER recessions periods used. In Table 1 we use dates and unemployment rates based on all
workers, as opposed to the subgroup values given here because the periods that unemploy-
ment rate increases from minimum to maximum around recessions may differ for different
demographic groups.
In the paper we examined transition rates from employment to unemployment and from
unemployment to employment. Figures A.1 and A.2 give the transition rates between em-
ployment and out-of-the-labor-force (i.e., λIE and λEI) and between unemployment and out-
of-the-labor-force (i.e., λIU and λUI). Figure A.3 is a counterpart to Figure 5, showing
that our log-linear approximation is an accurate approximation for dominant worker groups’
unemployment fluctuations as well as for disadvantaged groups.
Table A.1: Data Summary
Non-White White Young Experienced
Non Weighted 14.65% 85.34% 25.19% 74.80%Weighted 16.66% 83.33% 26.02% 73.97%
No College With College Female Male
Non Weighted 54.69% 45.30% 52.74% 47.25%Weighted 53.23% 46.76% 52.11% 47.88%
Sample composition by worker type. Weighted numbers are computed accordingto CPS weight.
37
Table A.2: Search Intensity Summary Statistics
All Employed Unemployed NILFMean SD Mean SD Mean SD Mean SD
% of Group Searching:College 1.5 12.3 0.7 8.2 23.5 42.4 0.6 7.7No College 1.3 11.2 0.5 7.0 12.7 33.3 0.2 5.0Older 1.1 10.6 0.4 6.5 21.7 41.2 0.3 5.4Younger 2.1 14.4 1.1 10.3 12.1 32.6 0.8 8.7White 1.2 11.0 0.6 7.4 16.6 37.2 0.3 5.4Nonwhite 2.2 14.8 0.9 9.4 17.1 37.7 0.8 8.9Minutes of Search (conditional on any search):College 145.7 130.2 119.9 121.6 161.3 134.1 138.6 120.0No College 124.2 122.8 109.3 98.9 128.7 129.3 120.8 114.2Older 145.0 128.8 125.8 116.2 152.8 133.7 134.8 114.7Younger 124.8 124.9 107.5 113.5 134.2 130.6 127.6 122.7White 134.8 124.8 109.8 102.5 148.0 132.8 118.8 116.4Nonwhite 141.1 134.1 137.0 145.0 140.5 132.9 152.9 117.8% of Time Spent Interviewing:College 9.7 28.2 16.4 35.7 5.8 21.4 11.4 31.1No College 8.2 25.7 10.4 29.0 7.8 25.0 6.3 23.8Older 6.5 23.0 10.9 29.5 5.2 20.3 4.9 21.7Younger 12.7 31.8 18.2 37.4 9.2 27.2 15.6 35.2White 10.6 29.6 15.7 35.1 8.1 25.9 12.9 33.4Nonwhite 5.3 19.4 11.2 29.5 3.4 14.3 3.7 16.7Methods of Search (CPS):College 2.00 1.13No College 2.34 1.33Older 2.25 1.28Younger 2.01 1.15White 2.14 1.24Nonwhite 2.13 1.21
38
Table A.3: Dates of Minimum and Maximum Unemployment
Non-White White Young Experienced
1980s Recession 9.03% 4.59% 7.09% 4.39%1990s Recession 4.24% 2.00% 3.72% 2.02%2000s Recession 4.03% 1.73% 3.41% 1.77%Great Recession 7.79% 4.50% 7.31% 4.56%
No College With College Female Male
1980s Recession 6.97% 3.24% 4.47% 5.48%1990s Recession 3.54% 2.14% 2.30% 2.56%2000s Recession 3.30% 1.54% 2.20% 1.92%Great Recession 8.29% 4.09% 4.68% 5.76%
1980s Recession 1990s Recession 2000s Recession Great Recession
Non-White 1979/07–1983/01 1989/07–1992/12 2000/12–2003/03 2007/02–2010/01Young 1979/05–1982/11 1990/02–1992/06 2000/06–2003/09 2007/02–2009/08No College 1979/07–1982/11 1989/03–1992/01 2000/02–2003/06 2007/03–2009/12Female 1979/07–1982/10 1989/08–1992/09 2000/10–2003/08 2007/02–2009/08All 1979/07–1982/11 1989/03–1992/01 2000/06–2003/06 2007/03–2010/01
Values and dates of minimum and maximum unemployment rates for each worker subgroup.
39
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.025
0.030
0.035
0.040
0.045 λEI: Non-WhiteλEI: White
Non-White vs. White
1978 1983 1988 1993 1998 2003 2008 2013 20180.035
0.040
0.045
0.050
0.055
0.060
0.065λ IE: Non-Whiteλ IE: White
Non-White vs. White
1978 1983 1988 1993 1998 2003 2008 2013 20180.020
0.025
0.030
0.035
0.040
0.045
0.050
0.055
λEI: YoungλEI: Experienced
Young vs. Experienced
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.04
0.06
0.08
0.10
0.12
0.14λ IE: Youngλ IE: Experienced
Young vs. Experienced
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.020
0.025
0.030
0.035
0.040
0.045 λEI: No College EducationλEI: College Education
No College vs. Any College
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.03
0.04
0.05
0.06
0.07λ IE: No College Educationλ IE: College Education
No College vs. Any College
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.020
0.025
0.030
0.035
0.040
0.045
0.050
0.055λEI: FemaleλEI: Male
Female vs. Male
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.035
0.040
0.045
0.050
0.055
0.060
0.065
0.070 λ IE: Femaleλ IE: Male
Female vs. Male
Figure A.1: λEI and λIE
Left panels give λEI and right panels λIE. Gray-shaded areas indicate NBER Recession periods. Data source:
FRED & IPUMS-CPS.
40
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35
0.40λUI: Non-WhiteλUI: White
Non-White vs. White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.02
0.03
0.04
0.05
0.06λ IU: Non-Whiteλ IU: White
Non-White vs. White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35
0.40λUI: YoungλUI: Experienced
Young vs. Experienced
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.02
0.04
0.06
0.08
λ IU: Youngλ IU: Experienced
Young vs. Experienced
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
λUI: No College EducationλUI: College Education
No College vs. Any College
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.0150
0.0175
0.0200
0.0225
0.0250
0.0275
0.0300
λ IU: No College Educationλ IU: College Education
No College vs. Any College
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.10
0.15
0.20
0.25
0.30
0.35 λUI: FemaleλUI: Male
Female vs. Male
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.015
0.020
0.025
0.030
0.035
λ IU: Femaleλ IU: Male
Female vs. Male
Figure A.2: λUI and λIU
Left panels give λUI and right panels λIU. Gray-shaded areas indicate NBER Recession periods. Data source:
FRED & IPUMS-CPS.
41
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4
0.6R 2 = 1.0
White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4
0.6 R 2 = 1.0
Experienced
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4
0.6R 2 = 1.0
With Some College Education
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4
0.6
R 2 = 1.0
All
Figure A.3: Observed Unemployment Fluctuation vs. Approximation
Gray-shaded areas indicate NBER Recession periods. The black solid line shows observed unemployment
fluctuations and the green dashed line approximated first-order log-linearized unemployment fluctuations.
Data source: FRED & IPUMS-CPS.
42
Appendix B Gender Differences in Unemployment
The difference between the unemployment rates for female and male workers (i.e., the
gender unemployment gap) is not persistently positive, with the gap existing before the 1980
and vanishing after 1980. The increase in male workers’ unemployment during recessions is
much larger than that for female workers. In this appendix we examine the sources of the
gender unemployment gap before the 1980 and why this gap disappeared after the 1980s. In
addition, we see which transition rates cause cyclical fluctuations in unemployment rates for
both male and female workers.
Before we conduct our quantitative analysis, we review the dynamics of the observed
transition rates. Figure B.1 plots the transitions rates between employment, unemployment
and out-of-the-labor-force for female and male workers. There are several features of note.
First, we would expect the sharp increases in the separation rates λEU (especially for male
workers) and the significant decline in the job-finding rates λUE to be the main sources
of the increase in unemployment rates during recessions. Second, from 1978 to 1993, we
find a declining trend in female workers’ transition rates from employment to out-of-the-
labor-force (λEI). After the 1980, the difference between male and female workers’ λEU
enlarged significantly, particularly during recessions. The decline in in female workers’ λEI
means that female workers became more likely to stay in the labor force. Because female
workers have lower λEU and similar λUE to male workers, the decline in λEI would reduce the
unemployment rate for female workers and therefore cause the gender unemployment gap to
vanish. Moreover, male workers’ higher separation rates, λEU, would further cause that the
gender unemployment gap not just to vanish, but to become negative.
To better understand the impact of these transition events on female and male workers’
unemployment fluctuations, we first repeat the analysis of Section 3. Figure B.2 shows
the approximate unemployment fluctuation based on Equation (11) as well as the observed
rates for both female and male workers. The correlation between the approximated and
observed fluctuations is 0.99. Therefore, our analytic approach to unemployment fluctuations
in Section 3 is reliable for female and male workers.
We use Equation (14) to measure the contribution of each transition flow in explaining
unemployment fluctuations. The left part of Table B.1 gives the β coefficients, measuring the
importance of each transition rate in accounting for unemployment fluctuations. As in Table
4, we also give β coefficients for all workers (i.e., fluctuations in the aggregate unemployment).
As expected, the variation in the job-finding rate λUE is the major source of unemployment
rate fluctuations, accounting for sixty percentage of total fluctuations in the female workers’
unemployment rates and half of total fluctuations in male workers’ rates. In contrast, the
change in the separation rates λEU explains thirty percent in the total fluctuations in male
workers’ unemployment rates, but only 11 percent for for female workers. In comparison
43
with the impact of job finding and separation rates on aggregate unemployment rates, we
find the separation margin plays a relatively important role in explaining male workers’
unemployment rate fluctuations, while hiring margin accounts for more of the fluctuations
in female workers’ rates.
Next, we begin to analyze the causes of the gender unemployment gap, the level difference
between female and male workers’ unemployment rates. Figure 1 shows that there has not
always been a gender unemployment gap, so we proceed as follows. First, we examine the
gender unemployment gap during the period from July 1978 to May 1979, during which time
women were more likely to be unemployed than men. Second, we replace each transition rate
(e.g., separation rate) during July 1978 to May 1979 with the average of this transition rate
during periods from January 1993 to December 1993, during which gender unemployment
gap disappeared and the average unemployment rate (6%) is similar to that during July
1978 to May 1979 (6.1%). Then, using Equation (10), we compute the counterfactual female
unemployment rate from July 1978 to May 1979 using the transition rates from the period
without a gender unemployment gap (January 1993 to December 1993) and construct a
counterfactual gender unemployment gap.
We use the percentage differences between the observed and counterfactual gender unem-
ployment gaps to understand the causes of the gender unemployment gap, as well as what
caused it to disappear. The right part of Table B.1 shows that using average separation
rates from January 1993 to December 1993 eliminates forty percentage of the gender unem-
ployment gap in the July 1978 to May 1979 period. The decline in female transition rates
from employment to out-of-the-labor-force, λEI, reduces the gender unemployment gap by
65 percent.
Our findings on the gender unemployment gap based on Equation (10) are similar to
those of Albanesi and Sahin (2018), who argued that the decline in λEI showed the labor
force attachment of female workers converging with that of male workers. Moreover, Albanesi
and Sahin (2018) indicated that separation rates differed because female and male workers
favored different industries.
44
1978 1983 1988 1993 1998 2003 2008 2013 20180.0125
0.0150
0.0175
0.0200
0.0225
0.0250
0.0275
0.0300λEU: FemaleλEU: Male
λEU
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55λUE: FemaleλUE: Male
λUE
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.020
0.025
0.030
0.035
0.040
0.045
0.050
0.055λEI: FemaleλEI: Male
λEI
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.035
0.040
0.045
0.050
0.055
0.060
0.065
0.070 λ IE: Femaleλ IE: Male
λIE
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.10
0.15
0.20
0.25
0.30
0.35 λUI: FemaleλUI: Male
λUI
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.015
0.020
0.025
0.030
0.035
λ IU: Femaleλ IU: Male
λIU
Figure B.1: Transition Rates: Female and Male Workers
Gray-shaded areas indicate NBER Recession periods. Data source: FRED & IPUMS-CPS.
45
1978 1983 1988 1993 1998 2003 2008 2013 20180.4
0.2
0.0
0.2
0.4
Female Workers
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.4
0.2
0.0
0.2
0.4
0.6
Male Workers
Figure B.2: Approximated vs. Observed Unemployment Fluctuations
Gray-shaded areas indicate NBER Recession periods. Data source: FRED & IPUMS-CPS. The left figure
compares the approximation and observed unemployment fluctuations for female workers, while the right
compares these for male workers.
Table B.1: Quantitative Analysis: Fluctuations and Gap
β Coefficients Female Male All % Change in Gap
λEU 0.107 0.28 0.206 λEU 39.37(0.013) (0.012) (0.01)
λEI -0.056 -0.024 -0.037 λEI 64.99(0.01) (0.005) (0.006)
λUE 0.561 0.483 0.512 λUE -67.38(0.014) (0.011) (0.01)
λUI 0.135 0.124 0.138 λUI -24.23(0.01) (0.006) (0.007)
λIE 0.109 0.054 0.075 λIE -16.34(0.008) (0.004) (0.005)
λIU 0.149 0.089 0.111 λIU -5.64(0.011) (0.007) (0.008)
ε -0.004 -0.006 -0.005(0.001) (0.001) (0.001)
95% Confidence interval in brackets.
46
Appendix C Supplemental Tables: Recessions
This appendix contains tables providing more detailed data regarding recessions. Table
2 uses the unemployment rate to examine the relationship between unemployment the tran-
sition rates. To examine the dynamics of the transition rates during recessions we repeat
the estimates but replace the date dummy with periods with the largest increase in the
unemployment rates during a recession.
There are four recession periods in our sample: ones in the 1980s, 1990s, and 2000s, as
well as the Great Recession. However, to start, we do not distinguish between recession pe-
riods. Table C.1 gives the base results. Because the date dummy indicated the periods that
unemployment rates increased from the minimum to the maximum during a recession, Table
C.1 shows that the average unemployment rates this period are not significantly different
from that of sample average. This is the reason that we use the observed aggregate unem-
ployment rate rather than such periods in the main text because we can clearly determine
the relationship between the unemployment rates and the transition rates.
Next, we repeat the estimate with dummies for the four specific recessions. Tables C.2,
C.3, C.4, and C.5 show the estimation results respectively for non-white, young workers,
those with no college education, and female workers. The main difference we find is that non-
white workers and young workers suffer more during the 1980s recession in the comparison
with their situation in the Great Recession. In contrast, the situation for those with no
college education is similar in both the 1980s and the Great Recession. Moreover, during the
recession period where the average unemployment is larger than the sample average, we see
similar changes in the transition rates. For example, separation rates increased while job-
finding rates declined for non-white workers during the Great Recession. This is consistent
with the findings in Table 2.
Tables 4 and 5 show the contributions of each transition event to unemployment fluctu-
ations and gaps, respectively. To understand whether the relative importance of transition
events change during recessions, we repeat the analysis for recessions only. The recessions
periods are given in Table A.3, based on the unemployment rates for all workers. Table
C.6 lists results for unemployment fluctuations and Table C.7 for unemployment gaps. In
general, recessions do not alter the conclusions of Tables 4 and 5.
47
Table C.1: Estimated Transitions: Disadvantaged Workers
Non-White u λEU λEI λUE λUI λIE λIU
Intercept 5.740*** 1.292*** 2.552*** 27.542*** 20.964*** 4.281*** 2.196***(0.254) (0.038) (0.092) (0.628) (0.359) (0.167) (0.076)
Recession 0.245 0.141*** -0.184 -0.249 -1.959*** -0.206 0.013(0.151) (0.028) (0.120) (0.551) (0.284) (0.221) (0.041)
Non-White 4.928*** 0.695*** 0.723*** -8.048*** 5.781*** 0.449 2.325***(0.168) (0.031) (0.168) (0.473) (0.274) (0.277) (0.063)
Recession × Non-White -0.023 -0.002 0.036 -0.402 0.966** -0.010 -0.044(0.329) (0.061) (0.177) (0.749) (0.466) (0.292) (0.106)
R2 0.55 0.48 0.13 0.44 0.52 0.08 0.71N 980 980 980 980 980 980 980
Young u λEU λEI λUE λUI λIE λIU
Intercept 5.320*** 1.117*** 2.109*** 24.400*** 19.629*** 2.253*** 1.310***(0.204) (0.032) (0.104) (0.649) (0.376) (0.269) (0.090)
Recession 0.076 0.104*** -0.248 -0.122 -2.066*** -0.247 -0.058(0.141) (0.025) (0.155) (0.533) (0.322) (0.260) (0.092)
Young 4.234*** 1.017*** 1.759*** 2.541*** 6.243*** 7.107*** 5.155***(0.128) (0.027) (0.153) (0.454) (0.298) (0.562) (0.101)
Recession × Young 0.310 0.087* 0.216 -0.242 0.528 -0.208 0.146(0.257) (0.050) (0.179) (0.756) (0.508) (0.619) (0.168)
R2 0.62 0.72 0.33 0.29 0.50 0.31 0.83N 980 980 980 980 980 980 980
No College u λEU λEI λUE λUI λIE λIU
Intercept 4.504*** 0.978*** 2.006*** 28.765*** 18.860*** 5.929*** 2.473***(0.196) (0.026) (0.085) (0.754) (0.375) (0.258) (0.047)
Recession -0.033 0.042** -0.125 -0.114 -1.390*** -0.230 -0.017(0.114) (0.021) (0.077) (0.643) (0.298) (0.386) (0.050)
No College 4.303*** 0.963*** 1.463*** -3.890*** 5.844*** -2.423*** 0.164***(0.124) (0.021) (0.200) (0.483) (0.277) (0.389) (0.036)
Recession × No College 0.213 0.136*** -0.224 0.056 -0.816* 0.126 0.017(0.244) (0.042) (0.206) (0.808) (0.461) (0.405) (0.064)
R2 0.64 0.78 0.16 0.29 0.48 0.12 0.34N 980 980 980 980 980 980 980
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
48
Table C.2: Estimated Transitions: Non-White Workers
u λEU λEI λUE λUI λIE λIU
Const. 5.789*** 1.301*** 2.555*** 27.483*** 20.927*** 4.274*** 2.207***(0.235) (0.035) (0.092) (0.600) (0.355) (0.164) (0.072)
80s 1.379*** 0.436*** 0.099 -1.361* -3.730*** -0.179 0.208***(0.209) (0.045) (0.132) (0.731) (0.371) (0.236) (0.050)
90s -0.244 0.122*** -0.309** 1.524* -3.663*** -0.102 -0.129**(0.161) (0.034) (0.133) (0.790) (0.371) (0.249) (0.056)
00s -1.018*** -0.097*** -0.207 3.038*** 0.614 -0.011 -0.227***(0.144) (0.028) (0.127) (0.849) (0.374) (0.238) (0.056)
GR 0.908** 0.104** -0.357*** -4.542*** -1.294*** -0.567** 0.210**(0.359) (0.052) (0.129) (0.964) (0.470) (0.239) (0.097)
Non-White 4.928*** 0.695*** 0.723*** -8.048*** 5.781*** 0.449 2.325***(0.168) (0.031) (0.168) (0.473) (0.274) (0.277) (0.063)
80s× Non-White 2.453*** 0.376*** 0.010 -1.714* 2.153*** -0.457 0.768***(0.455) (0.096) (0.211) (0.920) (0.681) (0.318) (0.153)
90s× Non-White 0.165 0.090 0.021 -1.271 1.618*** -0.160 0.232**(0.278) (0.058) (0.196) (0.984) (0.627) (0.329) (0.115)
00s× Non-White -1.402*** -0.175*** 0.092 -0.596 0.422 0.486 -0.407***(0.295) (0.056) (0.195) (1.191) (0.768) (0.331) (0.108)
GR× Non-White -1.418** -0.320*** 0.014 2.161 -0.388 0.069 -0.815***(0.613) (0.081) (0.193) (1.344) (0.802) (0.326) (0.186)
N 980 980 980 980 980 980 980R2 0.64 0.59 0.14 0.48 0.55 0.09 0.75
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
49
Table C.3: Estimated Transitions: Young Workers
u λEU λEI λUE λUI λIE λIU
Const. 5.351*** 1.124*** 2.109*** 24.345*** 19.575*** 2.245*** 1.319***(0.191) (0.030) (0.104) (0.621) (0.368) (0.266) (0.088)
80s 0.620*** 0.236*** -0.108 -1.140* -3.240*** -0.368 -0.083(0.208) (0.046) (0.177) (0.693) (0.454) (0.369) (0.148)
90s -0.189 0.104*** -0.370* 0.794 -3.299*** -0.288 -0.122(0.151) (0.034) (0.189) (0.724) (0.486) (0.401) (0.168)
00s -0.930*** -0.062** -0.206 2.868*** 0.013 -0.085 -0.214(0.147) (0.028) (0.180) (0.870) (0.539) (0.389) (0.158)
GR 0.889** 0.147*** -0.341* -3.342*** -1.961*** -0.259 0.214(0.361) (0.054) (0.188) (0.903) (0.583) (0.398) (0.184)
Young 4.234*** 1.017*** 1.759*** 2.541*** 6.243*** 7.107*** 5.155***(0.128) (0.027) (0.153) (0.454) (0.298) (0.562) (0.101)
80s× Young 1.770*** 0.477*** 0.160 -0.970 -1.702*** -0.139 1.142***(0.365) (0.081) (0.265) (1.013) (0.639) (0.704) (0.277)
90s× Young -0.342 0.077 0.134 1.205 -0.209 0.863 0.323(0.263) (0.062) (0.264) (1.035) (0.654) (0.853) (0.302)
00s× Young -0.482* -0.079 0.316 0.059 1.956** 0.412 -0.238(0.255) (0.052) (0.214) (1.188) (0.776) (0.737) (0.226)
GR× Young 0.204 -0.153* 0.242 -1.169 2.120*** -2.050*** -0.713***(0.582) (0.081) (0.220) (1.347) (0.788) (0.712) (0.274)
N 980 980 980 980 980 980 980R2 0.67 0.77 0.33 0.34 0.55 0.31 0.84
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
50
Table C.4: Estimated Transitions: Workers with No College Education
u λEU λEI λUE λUI λIE λIU
Const. 4.525*** 0.982*** 2.007*** 28.727*** 18.816*** 5.932*** 2.479***(0.186) (0.025) (0.085) (0.725) (0.366) (0.258) (0.048)
80s 0.206 0.122*** 0.084 0.802 -1.940*** 0.446 0.281***(0.143) (0.037) (0.104) (0.924) (0.449) (0.399) (0.079)
90s -0.490*** -0.011 -0.285*** 1.530 -3.489*** 0.414 0.015(0.124) (0.029) (0.092) (0.961) (0.394) (0.429) (0.063)
00s -0.634*** -0.040 -0.131 2.252** 0.293 -0.524 -0.374***(0.139) (0.028) (0.091) (1.093) (0.476) (0.395) (0.058)
GR 0.841*** 0.098** -0.199** -5.515*** -0.690 -1.280*** 0.030(0.293) (0.042) (0.093) (0.940) (0.498) (0.393) (0.108)
No College 4.303*** 0.963*** 1.463*** -3.890*** 5.844*** -2.423*** 0.164***(0.124) (0.021) (0.200) (0.482) (0.276) (0.389) (0.036)
80s× No College 1.066*** 0.322*** -0.424* -2.242** -3.047*** -0.550 -0.096(0.345) (0.074) (0.225) (1.118) (0.614) (0.428) (0.096)
90s× No College -0.159 0.143** -0.312 -0.003 -0.520 -0.531 -0.206***(0.252) (0.056) (0.222) (1.158) (0.532) (0.455) (0.078)
00s× No College -0.800*** -0.067 -0.049 0.645 0.971 0.714* 0.183**(0.240) (0.049) (0.218) (1.332) (0.640) (0.423) (0.075)
GR× No College 0.784 0.155* -0.117 2.041 -0.647 0.848** 0.170(0.605) (0.087) (0.221) (1.344) (0.761) (0.428) (0.146)
N 980 980 980 980 980 980 980R2 0.68 0.80 0.16 0.34 0.54 0.13 0.39
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
51
Table C.5: Estimated Transitions: Female Workers
u λEU λEI λUE λUI λIE λIU
Const. 6.560*** 1.561*** 1.936*** 27.068*** 18.357*** 5.143*** 3.257***(0.194) (0.032) (0.089) (0.631) (0.362) (0.193) (0.054)
80s 1.339*** 0.517*** -0.224* -1.665** -4.783*** 0.071 0.524***(0.282) (0.061) (0.122) (0.761) (0.405) (0.303) (0.096)
90s -0.133 0.232*** -0.358*** 1.440* -4.387*** -0.230 -0.188**(0.215) (0.057) (0.117) (0.789) (0.384) (0.313) (0.079)
00s -1.079*** -0.127*** -0.028 2.739*** 1.680*** -0.043 -0.298***(0.187) (0.040) (0.113) (0.938) (0.455) (0.290) (0.072)
GR 1.344*** 0.189** -0.083 -4.250*** -0.311 -0.932*** 0.162(0.470) (0.078) (0.113) (1.052) (0.529) (0.289) (0.136)
Female -0.041 -0.309*** 1.601*** -3.256*** 9.363*** -1.352*** -0.972***(0.120) (0.023) (0.177) (0.469) (0.280) (0.316) (0.041)
80s× Female 0.414 -0.155** 0.769*** 0.373 3.320*** -0.357 -0.331***(0.341) (0.071) (0.212) (1.014) (0.610) (0.373) (0.109)
90s× Female -0.342 -0.229*** 0.056 -0.422 2.778*** 0.148 0.129(0.250) (0.063) (0.207) (1.065) (0.545) (0.381) (0.092)
00s× Female -0.216 0.036 -0.413** 0.210 -2.062*** 0.090 0.022(0.236) (0.050) (0.196) (1.239) (0.631) (0.358) (0.086)
GR× Female -1.208** -0.243*** -0.622*** 0.528 -1.928** 0.528 -0.120(0.554) (0.086) (0.199) (1.361) (0.762) (0.361) (0.159)
N 980 980 980 980 980 980 980R2 0.18 0.50 0.22 0.34 0.72 0.11 0.61
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
52
Table C.6: β Coefficient: Unemployment Fluctuations,Recessions
Non-White 1980s 1990s 2000s Great Recession
λEU 0.315 0.159 0.115 0.207(0.055) (0.073) (0.047) (0.03)
λEI -0.116 -0.044 -0.019 -0.066(0.024) (0.038) (0.039) (0.016)
λUE 0.445 0.468 0.524 0.476(0.041) (0.101) (0.052) (0.03)
λUI 0.164 0.123 0.198 0.136(0.018) (0.041) (0.026) (0.019)
λIE 0.153 0.151 0.101 0.109(0.024) (0.036) (0.027) (0.014)
λIU 0.047 0.139 0.068 0.149(0.022) (0.068) (0.036) (0.024)
ε -0.008 0.005 0.015 -0.011(0.003) (0.008) (0.003) (0.004)
Young 1980s 1990s 2000s Great Recession
λEU 0.267 0.157 0.101 0.168(0.03) (0.046) (0.062) (0.037)
λEI -0.077 -0.089 0.023 -0.095(0.022) (0.039) (0.027) (0.02)
λUE 0.562 0.579 0.599 0.597(0.042) (0.056) (0.087) (0.033)
λUI 0.126 0.131 0.145 0.113(0.014) (0.034) (0.025) (0.017)
λIE 0.092 0.167 0.114 0.12(0.014) (0.022) (0.027) (0.012)
λIU 0.039 0.047 0.014 0.112(0.02) (0.036) (0.042) (0.02)
ε -0.01 0.009 0.005 -0.015(0.003) (0.004) (0.004) (0.004)
No College 1980s 1990s 2000s Great Recession
λEU 0.324 0.206 0.153 0.215(0.029) (0.049) (0.053) (0.035)
λEI -0.046 -0.087 -0.015 -0.06(0.014) (0.018) (0.025) (0.016)
λUE 0.454 0.521 0.569 0.507(0.035) (0.049) (0.048) (0.033)
λUI 0.129 0.138 0.152 0.149(0.014) (0.021) (0.023) (0.015)
λIE 0.095 0.131 0.053 0.076(0.016) (0.016) (0.021) (0.013)
λIU 0.053 0.086 0.073 0.126(0.012) (0.027) (0.022) (0.014)
ε -0.009 0.005 0.015 -0.013(0.003) (0.003) (0.002) (0.004)
Standard error in parentheses.
53
Table C.7: Compositions of Unemployment Gaps, Re-cessions
Non-White 1980s 1990s 2000s Great Recession
λEU 0.361 0.374 0.328 0.302(0.013) (0.014) (0.021) (0.023)
λEI 0.069 0.091 0.143 0.18(0.007) (0.007) (0.01) (0.015)
λUE 0.472 0.441 0.467 0.519(0.012) (0.022) (0.02) (0.028)
λUI -0.102 -0.104 -0.094 -0.147(0.007) (0.011) (0.012) (0.022)
λIE 0.009 -0.017 -0.102 -0.101(0.008) (0.009) (0.01) (0.018)
λIU 0.263 0.27 0.3 0.301(0.009) (0.012) (0.017) (0.022)
ε -0.072 -0.055 -0.042 -0.054(0.004) (0.005) (0.006) (0.01)
Young 1980s 1990s 2000s Great Recession
λEU 0.724 0.88 0.778 0.753(0.017) (0.027) (0.021) (0.027)
λEI 0.234 0.341 0.35 0.406(0.007) (0.012) (0.011) (0.023)
λUE -0.08 -0.27 -0.179 -0.136(0.019) (0.029) (0.021) (0.027)
λUI -0.112 -0.192 -0.205 -0.27(0.007) (0.015) (0.016) (0.026)
λIE -0.474 -0.63 -0.563 -0.57(0.015) (0.027) (0.019) (0.033)
λIU 0.726 0.872 0.813 0.817(0.018) (0.029) (0.022) (0.034)
ε -0.019 -0.001 0.007 0(0.003) (0.005) (0.005) (0.006)
No College 1980s 1990s 2000s Great Recession
λEU 0.568 0.65 0.648 0.651(0.014) (0.013) (0.017) (0.013)
λEI 0.119 0.151 0.204 0.197(0.007) (0.006) (0.007) (0.006)
λUE 0.194 0.122 0.079 0.069(0.02) (0.019) (0.02) (0.015)
λUI -0.022 -0.068 -0.103 -0.086(0.007) (0.007) (0.009) (0.007)
λIE 0.19 0.201 0.146 0.158(0.007) (0.009) (0.006) (0.007)
λIU -0.018 -0.028 0.048 0.039(0.008) ( 0.01) (0.009) (0.009)
ε -0.03 -0.027 -0.024 -0.028(0.002) (0.002) (0.003) (0.002)
Standard error in parentheses.
54
Appendix D Data Errors
In this appendix we discuss two data errors that exist in the CPS data. First, we consider
time aggregation error. The frequency of CPS observations is monthly, but transition flows
in the labor market may occur within a single month, so time-aggregation biases exist.
Second, we consider misclassification error, meaning that a worker’s labor force status may
be incorrectly recorded. We discuss how to correct these errors and report relevant figures
here.
D.1 Time Aggregation Error
We correct for time aggregation error following Shimer (2012). From the CPS data, we
can construct the monthly transition matrix Pmt for month t. Due to the multiple transition
in a month, we have
Pmt = (Pt)
n,
where Pt is the transition matrix for each sub-period (e.g., week) in a month. Here n is the
number of sub-periods in a month. The goal is to derive Pt based on the observed Pmt . For
simplicity, we skip the subscript t in the following derivation.
By eigen decomposition, we can write Pm as QMQ−1. Therefore, we can have
P = QM1/nQ−1.
For example, if n is equal to 4, we can derive the weekly transition matrix P based on
monthly transition matrix Pm. If time is continuous, then n → ∞. As shown in Gomes
(2015), we can have continuous time transition matrix
P = limx→0
QMxQ−1 − Ix
where x equals 1/n. In the main text, to avoid time aggregation we use a continuous time
transition matrix, rather than the observed monthly transition matrix, for all analysis.
D.2 Misclassification Errors
We correct for this error by following the deNUNified approach proposed by M. Elsby et
al. (2015) to remove the high frequency movement between unemployment and out-of-the-
labor-force. We compare the raw transition rates λIU and λUI with the adjusted ones after
correction for misclassification. As in M. Elsby et al. (2015), the level of both transition rates
decline. Because other transition rates are not affected based on this adjustment approach,
we do not give them here.
55
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.20
0.25
0.30
0.35
0.40
0.45
0.50 AdjustedRaw
Non-White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.20
0.25
0.30
0.35
0.40
0.45
0.50AdjustedRaw
Young
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35
0.40
0.45 AdjustedRaw
No College Education
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.20
0.25
0.30
0.35
0.40
0.45AdjustedRaw
Female
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35
0.40 AdjustedRaw
White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35
AdjustedRaw
Experienced
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35 AdjustedRaw
Any College Education
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.10
0.15
0.20
0.25
0.30
0.35 AdjustedRaw
Male
Figure D.1: λUI: Adjusted vs. Raw
Gray-shaded areas indicate NBER Recession periods. Data source: FRED & IPUMS-CPS.
56
1978 1983 1988 1993 1998 2003 2008 2013 20180.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09 AdjustedRaw
Non-White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.04
0.06
0.08
0.10
0.12AdjustedRaw
Young
1978 1983 1988 1993 1998 2003 2008 2013 20180.015
0.020
0.025
0.030
0.035
0.040
0.045
AdjustedRaw
No College Education
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.015
0.020
0.025
0.030
0.035
0.040
AdjustedRaw
Female
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.015
0.020
0.025
0.030
0.035
0.040 AdjustedRaw
White
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.010
0.015
0.020
0.025
0.030 AdjustedRaw
Experienced
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.015
0.020
0.025
0.030
0.035
0.040
0.045
AdjustedRaw
Any College Education
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.02
0.03
0.04
0.05
0.06
AdjustedRaw
Male
Figure D.2: λIU: Adjusted vs. Raw
Gray-shaded areas indicate NBER Recession periods. Data source: FRED & IPUMS-CPS.
57
Appendix E Deriving the Decomposition of the Un-
employment Rate
By the Taylor Theorem, we can rewrite the log difference in unemployment rate ut =
lnut− lnu∗t = lnu(Λt)−u(Λ∗t ) as a log-difference in transition rates λxt = lnλxt − lnλ∗xt , where
x ∈ X = {EU,EI,UE,UI, IE, IU}. To understand workers’ unemployment fluctuations,
lnut − ln ut, λ∗ijt will be the trend component in workers’ transition rate. When we discuss
a group of disadvantaged workers’ unemployment gap, lnugt − lnugt , we need to replace λ∗ijtwith the counterpart group of workers’ transition rate λg,ijt . The decomposition of the log
difference in unemployment rate, lnut − lnu∗t = lnu(Λt) − u(Λ∗t ), can be explicitly written
as
lnut − lnu∗t =∑x∈X
∂ lnu(Λt)
∂λxt
∣∣∣∣Λt=Λ∗
t
× λ∗xt · (lnλxt − lnλ∗xt ) + εt
= aUt (λ∗IEt + λ∗IUt )λ∗EUt · (lnλEU
t − lnλ∗EUt )
+ aUt λ∗IUt λ∗EI
t · (lnλEIt − lnλ∗EI
t )
+ aEt (λ∗IEt + λ∗IUt )λ∗UE · (lnλUEt − lnλ∗UE
t )
+ aEt λ∗IEt λ∗UI
t · (lnλUIt − lnλ∗UI
t )
+[aUt λ
∗EUt + aEt (λ∗UE
t + λ∗UIt )]λ∗IEt · (lnλIE
t − lnλ∗IEt )
+[aUt (λ∗EU
t + λ∗EIt ) + aEt λ
∗UEt
]λ∗IUt · (lnλIU
t − lnλ∗IUt ) + εt
=FEUt + FEI
t + FUEt + FUI
t + F IEt + F IU
t + εt,
(17)
where the coefficient aUt is equal to (1− u∗t )/U∗t and aE is equal to −u∗t/U∗t . Moreover,
U∗ = λ∗EUλ∗IU + λ∗IEλ∗EU + λ∗EIλ∗IU represents the total inflows to unemployment given
transition rates vector Λ∗t . Equation (17) gives the total log difference in unemployment,
lnut − lnu∗t , as consisting of six factors (F ) that depend on the log difference in transition
rate lnλt − lnλ∗xt . For example, FUE accounts for the part of lnut − lnu∗t that is driven
by the job-finding rate (i.e., lnλUEt − lnλ∗UE
t ). During recessions, FUEt will increase because
of the decline in job-finding rate, lnλUEt − ln λUE
t , and the negative component aEt . In the
calculations of unemployment gaps, FUEt would be positive because of the lower job-finding
rate of disadvantaged workers and aEt < 0. The error term εt accounts for the residual
between lnut − lnu∗t and the sum of these six factors.
E.1 Different Transition Events
In a given month, an unemployed worker may directly become employed (U→ E) or may
give up on job searching, then participate in the labor force again and find a job (U→ I→ E).
Similarly, an employed worker may become unemployed but continue to job search (E→ U)
58
or may give up job searching for a time but then participate labor force again (E→ I→ U).
We therefore do not naively decompose the unemployment rate according to each transition
rate based on Equation (17). Recall from Equations (8) and (9) that transition rate λUI
only influences λUIt λ
IEt . However, the transition rate λIE
t influences the unemployment rate
through the inflows to unemployment together with λEU, and the inflows to employment
together with λUI and λUE. Therefore, we extract the component related to transition flows
λUIt from factor F IE
t as fUIEt . We combine this extracted part fUIE
t with FUIt , the component
that depends on lnλUIt − lnλ∗UI
t . We define fUIEt + FUI
t as FUIEt and define F IE
t − fUIEt as
F IEt , which do not contain any components related to flows U → I → E. Similarly, λIU
t
influence the unemployment rate through the inflows to unemployment together with λEI
and λEU, and the inflows to employment together with λUE. Therefore, we combine fEIUt ,
the component related to transition flows λEIt , in factor F IU
t together with FEIt to get FEIU
t .
Again, we define F IUt = F IU
t − fEIUt , which does not depend on flows E → I→ U. After we
combine the parts that influence U→ I→ E and E→ I→ U together respectively as FUIEt
and FEIUt , we rewrite Equation (17) as
Ft = FEUt + FEIU
t + FUEt + FUIE
t + F IEt + F IU
t + εt. (18)
Here, for simplicity, we use Ft to represent the “total” log difference in unemployment target,
lnut − lnu∗t . In the main text, we use F tot for total unemployment fluctuation and F gap for
total unemployment gap. This equation is the same as Equation (17), with rearranged
components related to U → I → E and E → I → U flows. The factor F k still describes the
component that influences Ft through transition flow k. Here we have six transition flows.
Two depend on separation margins, E → U and E → I → U, and two depend on hiring
margins, U→ E and U→ I→ E, while the last two, I→ E and I→ U, are related to labor
force participation. Because F IE and F IU now only contain factors related to the transitions
from I to E and U, we can clearly determine the “pure” impact of labor force participation
margin on unemployment fluctuations or on disadvantaged workers’ unemployment gaps.
The part that cannot be accounted for by these components is the residual εt, which is the
same as that in Equation (17). The components in Equation (18) can be explicitly written
as:
FEUt = aUt (λIE
t λIUt )λEU
t λEU,ct ,
FUEt = aEt (λIE
t + λIUt )λUE,c
t ,
FEIUt = aUt λ
IUt λ
EIt λ
EI,ct + aUt λ
EIt λ
IUt λ
IU,ct ,
FUIEt = aEt λ
IEt λ
UIt λ
UI,ct + aEt λ
UIt λ
IEt λ
IE,ct ,
F IEt =
(aUt λ
EUt + aEt λ
UEt
)λIEt λ
IEt , and
F IUt =
(aUt λ
EUt + aEt λ
UEt
)λIUt λ
IUt .
(19)
59
Table E.1: β Coefficient: Unemployment Fluctuations, Overall
Non-White Young No College All
E → U 0.167 0.158 0.213 0.206(0.014) (0.013) (0.012) (0.010)
E → I → U 0.058 0.022 0.056 0.079(0.014) (0.013) (0.010) (0.009)
U → E 0.490 0.570 0.503 0.512(0.013) (0.014) (0.012) (0.010)
U → I → E 0.289 0.242 0.232 0.204(0.012) (0.010) (0.009) (0.007)
I → E 0.005 0.021 0.008 0.008(0.000) (0.001) (0.001) (0.001)
I → U -0.003 -0.006 -0.004 -0.006(0.000) (0.001) (0.000) (0.000)
ε -0.006 -0.006 -0.007 -0.005(0.001) (0.001) (0.001) (0.001)
White Experienced With College All
E → U 0.218 0.246 0.216 0.206(0.011) (0.012) (0.012) (0.010)
E → I → U 0.085 0.097 0.110 0.079(0.009) (0.009) (0.010) (0.009)
U → E 0.522 0.464 0.523 0.512(0.011) (0.010) (0.013) (0.010)
U → I → E 0.177 0.197 0.155 0.204(0.007) (0.008) (0.008) (0.007)
I → E 0.008 0.004 0.007 0.008(0.001) (0.000) (0.001) (0.001)
I → U -0.006 -0.003 -0.007 -0.006(0.000) (0.000) (0.001) (0.000)
ε -0.005 -0.005 -0.004 -0.005(0.001) (0.001) (0.001) (0.001)
95% Confidence interval in brackets.
We give results (β and γ) based on the decomposition approach of Equation (18) as we did
in Tables 4 and 5.
Table E.1 gives estimated β based on Equation (18) for unemployment fluctuations. In
Table E.1, we also report standard errors so that importance of a specific transition rate
can be compared with other transition rates. The estimated β of the different worker types
share common features. First, the contribution of hiring margins to workers’ unemploy-
ment fluctuations (FUE + FUIE or FUE only) is significantly larger than that of separations
(FEU + FEIU or FEU only). Second, the role of direct inflows from U to E (FUE) is more
important than that of indirect inflows from U to I to E (FUIE). The same is true of the two
separation margins components FEU and FEIU. Third, direct inflows from I to E or from
I to U play trivial roles in unemployment fluctuations. Therefore, in the comparison with
separation or hiring margins, the pure impact of labor force participation (directly from I
60
Table E.2: Compositions of UnemploymentGap, Overall
Non-White Young No College
E → U 0.35 0.776 0.616(0.005) (0.008) (0.004)
E → I → U 0.433 1.237 0.196(0.006) (0.013) (0.004)
U → E 0.436 -0.188 0.116(0.006) (0.007) (0.006)
U → I → E -0.142 -0.746 0.064(0.005) (0.013) (0.003)
I → E -0.009 -0.048 0.035(0.001) (0.003) (0.001)
I → U -0.021 -0.029 0.001(0.001) (0.002) (0.000)
ε -0.048 -0.002 -0.027(0.002) (0.001) (0.001)
95% Confidence interval in brackets.
to U or E) in explaining unemployment fluctuations is very minor. Our analysis shows that
out-of-the-labor-force I is just a transient status when workers move between employment
and unemployment in a given month. Our results about I actually complement the findings
in M. Elsby et al. (2015) which emphasize the role of transition between I and U in unem-
ployment rate fluctuations. We show that IU or UI transition flows are important only when
I is a transient status between E and U.
Table E.2 shows the estimated fractions of the factors which depend on specific flows in
the total unemployment gap and the corresponding confidence interval. We can see that, for
all disadvantaged workers, the main factor in the unemployment gap is FEU. In particular,
the component FUE only accounts for 10 percent of the total unemployment gap for workers
with no college education, and its contribution is less than zero for young workers. This
finding for young workers is not surprising. It is consistent with the estimated parameter
in Table 2, which shows that young workers have higher job-finding rates compared with
experienced workers. However, for non-white workers 40 percent of the total unemployment
gap can be attributed to the contribution of FUE. Particularly, non-white workers’ ratio of
FUE to the total unemployment gap is significantly higher than that of FEU in total. We
also see that the impacts of FEIU and FUIE on the unemployment gap are different. For non-
white and young workers the unemployment gap is mainly caused by FEIU, a component
that belongs to the separation margin. However, FUIE, a component in the hiring margin,
does not contribute to the unemployment gap for all disadvantaged workers. As with the
estimated results from Table E.1, the pure labor force participation components F IE and F IU
do not account for any fraction of the unemployment gap. Transition flows related to not-in-
61
the-labor-force I are important only when I is a transient status for workers transiting from
E to U. Overall, hiring margins contribute to the unemployment gap for non-white workers,
while separation margins are the most crucial factors in explaining the unemployment gap
for all disadvantaged workers.
62
Appendix F Alternative Trend Measure
In this section we give estimation results based on the sample average as the trend
component, rather than the HP-filter trend. Based on Table F.1, the main results hold:
workers’ unemployment fluctuations can mainly be attributed to hiring margin rather than
separation margin. In addition, even though we use the sample average as the trend, the
small magnitude of βε shows again the accuracy of our decomposition approach.
Given the sample average as the trend, we show the separation margin for non-white
workers may contribute to their unemployment changes during normal times. When we use
the HP-filter trend in the main text, we find that the contribution of the separation margin
to unemployment fluctuations for non-white and young workers declines. Therefore, the
contribution of the separation margin to non-white workers’ unemployment fluctuations is
not significantly larger than that of other workers.
Moreover, based on the HP-filter trend, the contribution of the hiring margin to young
workers’ unemployment fluctuations is significantly larger than that for experienced workers.
Because using the HP-filter trend will increase the contribution of the hiring margin, based on
mean as the trend component, we have a lower bound for the contribution of the hiring margin
to non-white and young workers’ unemployment fluctuations. Fujita and Ramey (2009)
also addressed that the contribution of the hiring (separation) margin to unemployment
fluctuations is higher (lower) when we use the HP-filter trend. Our conclusion that hiring
influences disadvantaged workers’ unemployment fluctuations does not change even when we
use the HP-filter trend.
63
Table F.1: β Coefficient: Sample Average
Non-White Young No College All
λEU 0.31 0.209 0.207 0.245(0.014) (0.016) (0.012) (0.013)
λEI -0.004 -0.016 -0.029 -0.004(0.006) (0.009) (0.006) (0.006)
λUE 0.339 0.499 0.52 0.473(0.015) (0.02) (0.014) (0.016)
λUI 0.105 0.124 0.133 0.13(0.006) (0.007) (0.007) (0.005)
λIE 0.105 0.109 0.096 0.064(0.006) (0.009) (0.005) (0.004)
λIU 0.151 0.082 0.083 0.097(0.009) (0.012) (0.006) (0.006)
ε -0.006 -0.007 -0.01 -0.005(0.002) (0.002) (0.001) (0.001)
White Experienced With College All
λEU 0.248 0.246 0.2 0.245(0.014) (0.011) (0.012) (0.013)
λEI 0.003 -0.004 0.015 -0.004(0.006) (0.005) (0.006) (0.006)
λUE 0.49 0.475 0.562 0.473(0.016) (0.012) (0.015) (0.016)
λUI 0.12 0.138 0.089 0.13(0.005) (0.005) (0.006) (0.005)
λIE 0.05 0.052 0.059 0.064(0.004) (0.004) (0.007) (0.004)
λIU 0.094 0.099 0.087 0.097(0.005) (0.006) (0.008) (0.006)
ε -0.006 -0.006 -0.011 -0.005(0.001) (0.001) (0.001) (0.001)
Standard error in parentheses.
64
Appendix G β for Unemployment Gap
In the main text we use Equation (16) to derive the ratio of each factor to total unem-
ployment gap, so that we can determine the contribution of each factor to the unemployment
gap. The β coefficients in Equation (14) only reveal the importance of these factors to the
variation of unemployment gap, rather than the level of the gap. Therefore, we do not use
the β coefficient for analysis of the unemployment gap in the main text. This appendix gives
results for the β coefficients when we apply Equation (14) to the unemployment gap.
As we can see in Table G.1, variation in the unemployment gap is still mainly caused
by fluctuations in the separation factors FEU and FEIU for non-white and young workers.
For workers with no college education, 60 percent of the unemployment gap variation is ac-
counted for by the hiring margin, which only contributes around 20 percent to unemployment
gap. Table G.2 shows the β coefficients for the different recession periods. As in Table G.1,
variation in the unemployment gap is determined by the separation margin. Particularly,
the importance of the hiring margin for workers with no college education declines dur-
ing recessions (especially the Great Recession). Therefore, hiring margin only explains the
fluctuations in normal times for workers with no college education, rather than in recessions.
Although Figure 8 shows that the hiring margin contributes trivially to the unemployment
gap, the β coefficients for the hiring margin factors are even positive for young workers.
Moreover, the β coefficient for hiring margin is much larger than that for the separation
margin for those with no college education. Because β is the contribution of the fluctuations
of a factor to the variation, rather than the level, of the unemployment gap, β coefficients
are inconsistent with the features in Figure 8. For this reason we do not use β in the main
text for the unemployment gap.
65
Table G.1: β Coefficient: Unemployment Gap,Overall
Non-White Young No College
λEU 0.402 0.421 0.375[0.368, 0.436] [0.377, 0.465] [0.339, 0.411]
λEI 0.002 0.095 0.016[-0.017, 0.021] [0.069, 0.12] [-0.006, 0.038]
λUE 0.400 0.187 0.424[0.359, 0.441] [0.142, 0.232] [0.376, 0.471]
λUI 0.014 0.12 0.086[-0.003, 0.032] [0.098, 0.142] [0.066, 0.106]
λIE 0.100 -0.075 0.111[0.08, 0.119] [-0.107, -0.042] [0.088, 0.133]
λIU 0.156 0.27 0.023[0.134, 0.178] [0.234, 0.305] [-0.007, 0.054]
ε -0.074 -0.018 -0.035[-0.086, -0.062] [-0.029, -0.007] [-0.043, -0.028]
95% Confidence interval in brackets.
66
Table G.2: β Coefficient: Unemployment Gap, Recessions
Non-White 1980s 1990s 2000s Great Recession
λEU 0.365 0.292 0.512 0.199[0.178, 0.551] [0.168, 0.416] [0.273, 0.75] [0.039, 0.36]
λEI 0.165 0.065 0.009 0.099[0.078, 0.252] [-0.001, 0.13] [-0.117, 0.134] [0.012, 0.186]
λUE 0.374 0.466 0.346 0.3[0.207, 0.541] [0.258, 0.674] [0.107, 0.585] [0.142, 0.458]
λUI 0.094 0.106 0.121 0.13[0.041, 0.147] [0.044, 0.168] [0.012, 0.229] [0.028, 0.232]
λIE 0.051 0.026 0.089 0.079[-0.065, 0.167] [-0.055, 0.107] [-0.029, 0.207] [-0.008, 0.167]
λIU -0.024 0.097 0.008 0.134[-0.057, 0.104] [0.008, 0.185] [-0.161, 0.177] [0.027, 0.242]
ε -0.073 -0.051 -0.084 0.058[-0.127, -0.019] [-0.106, 0.004] [-0.148, -0.019] [-0.007, 0.124]
Young 1980s 1990s 2000s Great Recession
λEU 0.287 0.265 0.319 0.418[0.154, 0.42] [0.124, 0.406] [0.142, 0.497] [0.289, 0.548]
λEI 0.158 0.13 0.113 0.165[0.086, 0.23] [0.053, 0.207] [0.042, 0.184] [0.095, 0.236]
λUE 0.353 0.195 0.244 0.073[0.178, 0.528] [0.016, 0.374] [0.063, 0.425] [-0.079, 0.225]
λUI 0.066 0.162 0.142 0.141[0.014, 0.117] [0.067, 0.257] [0.048, 0.237] [0.079, 0.203]
λIE -0.139 0.024 -0.093 -0.138[-0.263, -0.015] [-0.115, 0.163] [-0.202, 0.016] [-0.215, -0.061]
λIU 0.301 0.195 0.279 0.327[0.158, 0.445] [0.04, 0.351] [0.152, 0.406] [0.234, 0.42]
ε -0.026 0.029 -0.004 0.013[-0.059, 0.007] [-0.017, 0.075] [-0.058, 0.051] [-0.029, 0.055]
No College 1980s 1990s 2000s Great Recession
λEU 0.412 0.37 0.245 0.496[0.295, 0.529] [0.252, 0.489] [0.108, 0.382] [0.327, 0.664]
λEI 0.053 0.094 0.043 0.138[-0.017, 0.124] [0.026, 0.162] [-0.022, 0.109] [0.059, 0.217]
λUE 0.399 0.41 0.474 0.217[0.232, 0.565] [0.215, 0.605] [0.287, 0.661] [0.028, 0.406]
λUI 0.004 0.07 0.174 0.047[-0.055, 0.063] [0.008, 0.132] [0.112, 0.237] [-0.036, 0.13]
λIE 0.135 0.034 0.026 0.098[0.075, 0.195] [-0.046, 0.114] [-0.028, 0.079] [0.003, 0.193]
λIU 0.036 0.049 0.053 0.027[-0.047, 0.119] [-0.064, 0.162] [-0.039, 0.145] [-0.098, 0.151]
ε -0.040 -0.028 -0.015 -0.022[-0.066, -0.013] [-0.055, -0.001] [-0.045, 0.015] [-0.051, 0.007]
95% Confidence interval in brackets.
67
Appendix H Definition of Young Workers
In the main text we define potential experience as age minus schooling years, defining
young workers as those with less than ten years of potential experience. However, given the
variation in years of schooling, the oldest of these “young” workers is 38. 20% of these young
workers are under 20, 60% are under 25, and more than 90% are under 30. In this appendix
we use age rather than potential experience. First, we use 18 years-old as the cut-off point,
defining workers who are 18 or younger as the young workers. We then use 20 as the cut-off
point to see if there is any difference.
Figure H.1 compares the unemployment rates of workers 18 (or 20) and younger with
older workers. We see there is still an unemployment gap, with workers 18 (or 20) and
younger showing a higher unemployment rate than older workers. Moreover, the dashed and
dotted lines show that the approximated unemployment rate based on Equation (10) still
tracks the observed unemployment rate well. Therefore, the unemployment gap is robust to
using age rather than potential experience to define young workers.
Figure H.2 compares the transition rates for young workers according to our definition
with the counterpart experienced workers. Figures H.3 and H.4 compare the transition rates
for workers aged 18 (and 20) and younger with the counterpart older workers. We find that
no matter whether using potential experience or age to define young workers, the dynamics
and the features of the transition rates do not change. For example, the entry rates to
unemployment (from the employment or the out-of-the-labor-force) are higher for younger
workers than for the counterpart older workers. Moreover, the level of job-finding rates
for younger workers are similar to those of older workers and decrease significantly during
recessions. Therefore, we confirm that transition rates have similar features whether using
potential experience or age to define young workers.
Because the unemployment rates and transition flows have the same features and dynam-
ics when using age to define young workers, the source of the fluctuations and unemployment
gap should also be similar to our analysis in the main text. Table H.1 gives results for workers
18 (or 20) and younger. For the source of fluctuations, we find that the decline in job-finding
rates still explain around 50% of the fluctuations in the unemployment rates. This is consis-
tent with the findings for young workers in the main text. In addition, for the unemployment
gap between younger workers and older workers we find that entry rates to unemployment
(i.e., λEU and λIU) and transition rates from employment to out-of-the-labor-force are the
main sources of the unemployment gap. Particularly, the separation rate λEU accounts for
most of the unemployment gap. These features are similar to those found in the main text,
with the only difference being that the contribution to the unemployment gap of λIU declines
and the contribution of λIE increases.
So, whether we use years of potential experience or solely age to define young work-
68
ers, we reach similar conclusions about the sources of unemployment fluctuations and the
unemployment gap.
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.05
0.10
0.15
0.20
0.25
0.30
Younger than 19
Older than 19
Unemployment Rate: 18 Years as Cut-Off
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.05
0.10
0.15
0.20
0.25
Younger than 21
Older than 21
Unemployment Rate: 20 Years as Cut-Off
Figure H.1: Unemployment Rate: Young Workers
Solid lines represent the observed unemployment rate while dashed and dotted lines are the approximated
unemployment rate based on Equation (10), i.e., Shimer’s approach.
1978 1983 1988 1993 1998 2003 2008 2013 20180.010
0.015
0.020
0.025
0.030
0.035
0.040λEU: YoungλEU: Experienced
λEU
1978 1983 1988 1993 1998 2003 2008 2013 20180.020
0.025
0.030
0.035
0.040
0.045
0.050
0.055
λEI: YoungλEI: Experienced
λEI
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55 λUE: YoungλUE: Experienced
λUE
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35
0.40λUI: YoungλUI: Experienced
λUI
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.04
0.06
0.08
0.10
0.12
0.14λ IE: Youngλ IE: Experienced
λIE
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.02
0.04
0.06
0.08
λ IU: Youngλ IU: Experienced
λIU
Figure H.2: Transition Rates: Young Workers
Black solid lines are the observed transition rates for young workers while green-dashed lines are those for
experienced workers.
69
1977 1982 1987 1992 1997 2002 2007 2012 20170.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09 Younger than 19: λEUOlder than 19: λEU
λEU
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.05
0.10
0.15
0.20
Younger than 19: λEIOlder than 19: λEI
λEI
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.2
0.3
0.4
0.5
0.6
0.7
0.8 Younger than 19: λUEOlder than 19: λUE
λUE
1977 1982 1987 1992 1997 2002 2007 2012 20170.1
0.2
0.3
0.4
0.5
0.6Younger than 19: λUIOlder than 19: λUI
λUI
1977 1982 1987 1992 1997 2002 2007 2012 20170.025
0.050
0.075
0.100
0.125
0.150
0.175
0.200 Younger than 19: λ IEOlder than 19: λ IE
λIE
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.02
0.04
0.06
0.08
0.10Younger than 19: λ IUOlder than 19: λ IU
λIU
Figure H.3: Transition Rates: Workers 18 and Younger
Black solid lines are the observed transition rates for workers age 18 or younger while green-dashed lines
those for older workers.
70
1977 1982 1987 1992 1997 2002 2007 2012 20170.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08 Younger than 21: λEUOlder than 21: λEU
λEU
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16Younger than 21: λEIOlder than 21: λEI
λEI
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.2
0.3
0.4
0.5
0.6
0.7
0.8Younger than 21: λUEOlder than 21: λUE
λUE
1977 1982 1987 1992 1997 2002 2007 2012 20170.1
0.2
0.3
0.4
0.5
0.6Younger than 21: λUIOlder than 21: λUI
λUI
1977 1982 1987 1992 1997 2002 2007 2012 20170.025
0.050
0.075
0.100
0.125
0.150
0.175
0.200 Younger than 21: λ IEOlder than 21: λ IE
λIE
1977 1982 1987 1992 1997 2002 2007 2012 2017
0.02
0.04
0.06
0.08
0.10Younger than 21: λ IUOlder than 21: λ IU
λIU
Figure H.4: Transition Rates: Workers 20 and Younger
Black solid lines are the observed transition rates for workers age 20 or younger while green-dashed lines are
those for older workers.
71
Table H.1: The β Coefficients and r Ratio
Fluctuation: β Coefficient 18 and Younger 20 and Younger
λEU 0.115 0.140(0.017) (0.014)
λEI 0.013 -0.020(0.014) (0.010)
λUE 0.501 0.533(0.021) (0.017)
λUI 0.049 0.072(0.013) (0.011)
λIE 0.271 0.213(0.012) (0.010)
λIU 0.062 0.074(0.018) (0.014)
ε -0.012 -0.012(0.003) (0.003)
Gap: r Ratio 18 and Younger 20 and Younger
λEU 0.615 0.690(0.006) (0.005)
λEI 0.489 0.453(0.004) (0.003)
λUE -0.058 -0.093(0.005) (0.004)
λUI -0.169 -0.153(0.003) (0.003)
λIE -0.182 -0.248(0.004) (0.003)
λIU 0.283 0.350(0.004) (0.003)
ε 0.022 0.001(0.003) (0.002)
Standard errors in parentheses.
72
Appendix I Heterogeneity within Demographic Groups
This appendix provides more detail on sources of heterogeneity within our six demo-
graphic groups. Disadvantaged workers may differ from their counterpart groups across a
wide variety of dimensions such as geography, occupation, and other demographic charac-
teristics. We investigate whether these other dimensions can explain the group differences
found in the main results.
We begin by investigating demographic variation. Table I.1 shows the percentages of
non-white workers, those with no college, female, and young workers in each demographic
subgroup. For example, among non-white workers 55 percent have no college experience.
The last column in Table I.1 gives the percentages of non-white, without college, female,
and young workers in the total sample. According to Table I.1, the percentages of workers
with these demographic characteristics are similar across the eight worker groups. Thus, it
is not the case that other sources of demographic heterogeneity are driving the differences
we observe between groups in the main results.
Next we turn to geography and occupations. If disadvantaged workers are clustered
in locations with worse labor markets or are more likely to be employed in jobs that are
more volatile, these characteristics could lead to worse labor market outcomes. To formally
investigate the role these characteristics play, we use an analysis similar to Albanesi and
Sahin (2018), who analyze how composition has affected female and male unemployment
rates. In particular, we construct counterfactual transition rates and unemployment rates,
using the counterpart groups’ distribution across a characteristic to construct a re-weighted
flow or unemployment rate. This tells us how much of the difference in the flow is due
to differences in transition rates within groups versus differences in composition between
groups.
First, we compute the group share in each cell (e.g., occupation or state), which we then
use to weight the labor market flows and the unemployment rate within the cell. Specifically,
we calculate
λxt =∑i∈I
witλxit, (20)
where x ∈ {EU,EI,UE,UI, IE, IU}, i represents the characteristics (occupation, location,
etc) and t represents the period. For example, i may represent different occupations. Here,
wit means the percentage of workers with characteristic i and λit means the transition flows
for workers with characteristic i. Similarly, the formula for weighted unemployment rate is
ut =∑i∈I
wituit, (21)
where uit is the unemployment rate for workers with characteristic i.
Next, we construct the weights for each disadvantaged group’s counterpart group (wCit ),
73
which we then use to construct the counterfactual flow and unemployment rates. We then
compute the following measure:
errx =
∑Tt=0(ln λx,Ct − ln λxt )
2
T× 100
erru =
∑Tt=0(ln ux,Ct − ln uxt )
2
T× 100
(22)
which are the logarithm root mean square errors (RMSE) between the counterfactual tran-
sition flows ln λx,Ct and unemployment rates ln ux,Ct compared with those observed (ln λxt and
ln uxt ). We use the logarithms of these variables to avoid the issue that transition flow rates
and unemployment rates differ in magnitudes. For example, λEU is only one-quarter of λUE,
so if we did not use the logarithm of λEU and merely compute the RMSE, the value would
be small and could not accurately measure the difference between the observed and counter-
factual λEU. Because all variables are now in logarithm, errx and erru represent the relative
changes between the observed flow rates (or unemployment rates) and counterfactual ones.
I.1 Analysis
In addition to the four characteristics (non-white, young, female, and no college educa-
tion), we also consider the decomposition by age, occupation and location in the analysis.
For the decomposition by age, we follow Albanesi and Sahin (2018) and use three groups:
below 24, 24 to 55, and greater than 55. Table I.2 shows the relative differences (in percent-
ages) for transition flows and unemployment comparing the observed and counterfactual.
Recall the counterfactual transition flow and unemployment rates are constructed by re-
placing the compositions with their counterpart advantaged workers. Table I.2 shows that
when non-white workers have the same compositions as white workers, transition rates and
unemployment rates are similar to those originally observed. Therefore, the impact of being
non-white on the transition and unemployment rates is quite significant.
Table I.2 shows that the relative differences are large for λIU considering compositions and
λUE considering occupation compositions. Figure I.1 shows these two series and compares
the original and the counterfactual. Based on Figure I.1, we find they are close to each other.
In the following discussion, we therefore use 1.5%, which is the relative difference for λIU
considering age composition as the criterion to determine the significance of the effect of the
decomposition.
Table I.3 gives the relative differences for workers with no college. Unlike for non-white
workers, decomposition by age and occupation significantly affects the transition flow rates
and unemployment rates for the workers with no college. Figure I.2 plots those flows or
unemployment rate with relative differences greater than 1.5%.
We find that when the workers with no college have the same age composition as workers
74
with any college, transition rates from employment to out-of-the-labor-force and from out-of-
the-labor-force into the labor force both declined. Despite the age decomposition affecting
transition rates, these difference did not cause significant changes in unemployment rates
(the relative difference is only 0.23%). In addition to age composition, Figure I.2 shows
that when the workers with no college have the same occupational composition as that of
workers with any college, exit rates from employment become lower while entry rates into
employment become higher. The influence of occupational decomposition on transition flows
also causes the unemployment rate for workers with no college to decline.
Table I.4 gives the relative differences for young workers. Decomposition by age and oc-
cupation significantly affect transition flow rates and unemployment rates for young workers.
It is not surprising that age composition causes difference in transition flow and unemploy-
ment rates as we define young workers by workers’ age and schooling. Therefore, this finding
confirms that our definition of young workers can reveal workers’ age. Figure I.3 plots the
flows or unemployment rate with a relative difference greater than 1.5% after we consider
the occupational composition.
When young workers have the same occupational composition as experienced workers,
the exit rate from employment increases while the entry rate into employment decreases.
Therefore, the occupational composition causes young workers’ unemployment to increase,
showing that young workers and experienced worker work in different occupations.
These results show that occupational segregation by education and age can explain part
of the differences in transition rates between groups. This could be due to searching in
different labor markets, or to differences in qualifications or search ability. However, for
non-white workers, differences in occupational segregation do not appear to affect transition
rates.
I.2 Controlling for Occupation & Geography
In the previous analysis we analyzed how occupation, location, and age decomposition
affected transition flows and unemployment rates for each group of disadvantaged workers.
As we showed show that occupational composition does affect transition flows and unem-
ployment rates, so one may wonder whether the features found in Table 2 still hold.
We therefore redo the regression, controlling for occupation together with location (by
state of residence). However we do not consider age decomposition as it can be captured by
the regression for young workers. The estimation equation is
ln yt =β0 + βd1d + βr lnut + βdr (lnut × 1d) +∑m
am1m
+ βocc1occ + βocc,d(1occ × 1d) + βs1State + βs,d(1State × 1d) + εt.
(23)
where 1d is an indicator for the disadvantaged demographic group, and 1m are month-by-year
75
fixed effects. Here, we add 1occ as the indicator for occupation and 1s as the indicator for
the workers’ state of residence. Again, we use yt to represent either the unemployment rates
or transition rates.
Table I.5 gives regression results after controlling for occupation. We find that the results
are mainly similar to those reported in Table 2 for minority workers. This shows that the
impact of race on transition flows and unemployment rates is quite universal and will not
change, even controlling for occupation or state. For workers with no college, after controlling
for occupation, the transition flows between unemployment and out-of-the-labor-force (λUI
and λIU) become more volatile than for counterpart workers with college experience. In
Table 2 these two transitions flows for workers with no college are higher than those for
workers with any college, but they become higher after we control for occupation. Because
education level limits the occupations in which workers can search for jobs, we believe this
can be attributed to the fact that workers with no college experience search and work in
different occupations than do workers with college degrees. Therefore, after we control for
occupation, the magnitude of the impact of attending college or not on λUI and λIU changes.
There do exist some differences for young workers. Except for the separation rate λEU,
the signs for the other transition rates are different from those in Table 2. For example,
in Table 2 young workers have a higher entry rate to employment (λUE and λIE), while in
Table I.6 young workers’ entry rate to employment is lower than that of experienced workers
after controlling for occupation. When we do not control for occupation, young workers
see higher employment entry rates. This difference indicated that the labor market young
workers search for jobs is segregated and different from those for experienced workers.
Additionally, after controlling for occupation, the entry rates to unemployment from out-
of-the-labor-force (λIU) and the employment entry rate (λEU + λEI) both become higher for
young workers. We believe this can be attributed to the labor markets for young workers’
search being different from those for experienced workers. However, we cannot conclude that
the impact of age and potential experience on transition flows changes after controlling for
occupation. As years of work experience can limit the occupations in which a worker can
search, occupational difference therefore matters here, and the different labor markets for
young workers’ search will depend on their years of experience.
Table I.6 gives the results after we control for both occupation and workers’ state of
residence. We find results similar to those in Table I.5. Therefore, location does not cause
significant differences in transition flows and unemployment rate.
76
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.015
0.020
0.025
0.030
0.035
0.040
0.045
0.050
λIU: Age Composition
1978 1983 1988 1993 1998 2003 2008 2013 20180.12
0.14
0.16
0.18
0.20
0.22
0.24
0.26
0.28
λUE: Occupational Composition
Figure I.1: Transition Rates: Non-White Workers
Black solid lines are the original observed series while green-dashed lines are the counterfactual series given
the composition of counterpart advantaged workers.
77
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.0225
0.0250
0.0275
0.0300
0.0325
0.0350
0.0375
λEI: Age Composition
1978 1983 1988 1993 1998 2003 2008 2013 20180.030
0.035
0.040
0.045
0.050
0.055
λIE: Age Composition
1978 1983 1988 1993 1998 2003 2008 2013 20180.0100
0.0125
0.0150
0.0175
0.0200
0.0225
0.0250
0.0275
λIU: Age Composition
1978 1983 1988 1993 1998 2003 2008 2013 20180.50
0.55
0.60
0.65
0.70
0.75
0.80
λEE: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.010
0.015
0.020
0.025
0.030
0.035
0.040
λEU: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.20
0.25
0.30
0.35
0.40
0.45
λEI: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 20180.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
λUE: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.0015
0.0020
0.0025
0.0030
0.0035
0.0040
0.0045
λIE: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.10
0.15
0.20
0.25
0.30
0.35
0.40
u: Occupational Composition
Figure I.2: Transition Rates: No College Degree
Black solid lines are the original observed series while green-dashed lines are the counterfactual series given
the composition of counterpart advantaged workers.
78
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.55
0.60
0.65
0.70
0.75
λEE: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.01
0.02
0.03
0.04
0.05
0.06
λEU: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 20180.20
0.25
0.30
0.35
0.40
0.45
λEI: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 20180.02
0.04
0.06
0.08
0.10
0.12
λUE: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.005
0.010
0.015
0.020
0.025
λIE: Occupational Composition
1978 1983 1988 1993 1998 2003 2008 2013 2018
0.15
0.20
0.25
0.30
0.35
0.40
0.45
u: Occupational Composition
Figure I.3: Transition Rates: Young Workers
Black solid lines are the original observed series while green-dashed lines are the counterfactual series given
the composition of counterpart advantaged workers.
Table I.1: Compositions: Each Subgroup
White Non-White College Above College Below Experienced Young All
Non-White − − 16% 17% 16% 20% 17%College Below 53% 55% − − 54% 51% 53%Young 25% 31% 27% 25% − − 26%Female 52% 54% 51% 53% 53% 50% 52%
Percentages of non-white workers, those with no college, and young workers in each subgroup and intotal workers.
79
Table I.2: Relative Difference: Non-White
Age Occupation State No College Female Young
errEE 0.00% 0.04% 0.00% 0.00% 0.00% 0.00%errEU 0.08% 0.85% 0.15% 0.07% 0.01% 0.10%errEI 0.03% 0.17% 0.05% 0.03% 0.01% 0.10%errUE 0.01% 1.25% 0.10% 0.01% 0.00% 0.01%errUU 0.00% 0.02% 0.01% 0.00% 0.00% 0.00%errUI 0.02% 0.03% 0.09% 0.01% 0.01% 0.04%errIE 0.80% 0.83% 0.06% 0.03% 0.01% 0.10%errIU 1.54% 0.09% 0.13% 0.01% 0.01% 0.37%errII 0.01% 0.00% 0.00% 0.00% 0.00% 0.00%erru 0.09% 0.27% 0.05% 0.07% 0.00% 0.13%
Relative differences between counterfactual and observed flows andunemployment rates.
Table I.3: Relative Difference: No College Degree
Age Occupation State Gender Race Young
errEE 0.00% 3.24% 0.00% 0.00% 0.00% 0.00%errEU 0.12% 7.51% 0.05% 0.03% 0.00% 0.20%errEI 1.99% 14.57% 0.03% 0.06% 0.00% 0.11%errUE 0.03% 2.56% 0.03% 0.01% 0.00% 0.01%errUU 0.02% 0.08% 0.01% 0.00% 0.00% 0.00%errUI 0.31% 0.12% 0.03% 0.06% 0.00% 0.04%errIE 1.66% 17.39% 0.05% 0.02% 0.00% 0.26%errIU 2.39% 0.15% 0.07% 0.03% 0.00% 0.78%errII 0.01% 0.00% 0.00% 0.00% 0.00% 0.00%erru 0.23% 15.88% 0.03% 0.00% 0.00% 0.20%
Relative difference between counterfactual and observed flowsand unemployment rate.
Table I.4: Relative Difference: Young
Age Occupation State College Gender Race
errEE 0.17% 2.83% 0.00% 0.00% 0.00% 0.00%errEU 33.86% 3.77% 0.01% 0.31% 0.01% 0.01%errEI 83.23% 12.30% 0.00% 0.12% 0.01% 0.00%errUE 0.96% 39.32% 0.00% 0.02% 0.00% 0.01%errUU 0.84% 0.12% 0.00% 0.00% 0.00% 0.00%errUI 13.38% 1.26% 0.01% 0.02% 0.01% 0.00%errIE 3.33% 64.68% 0.00% 0.04% 0.01% 0.00%errIU 6.92% 1.00% 0.01% 0.02% 0.01% 0.00%errII 0.09% 0.01% 0.00% 0.00% 0.00% 0.00%erru 37.90% 8.30% 0.00% 0.31% 0.00% 0.01%
Relative difference between counterfactual and observed flows andunemployment rate.
80
Table I.5: Estimated Transitions: Disadvantaged Workers, Controlling for Occupation
Non-White u λEU λEI λUE λUI λIE λIU
Unemployment 0.406*** 0.084*** -0.036*** -0.349*** -0.336*** -0.338*** -0.150***(0.005) (0.010) (0.012) (0.004) (0.010) (0.003) (0.006)
Non-White 1.060*** 0.978*** 0.932*** -0.105*** -0.116*** -0.183*** -0.025(0.020) (0.041) (0.044) (0.017) (0.029) (0.012) (0.019)
Unemp × Non-White -0.125*** -0.024 0.024 0.087*** 0.132*** 0.164*** 0.142***(0.009) (0.020) (0.020) (0.009) (0.014) (0.006) (0.010)
R2 0.25 0.27 0.61 0.11 0.30 0.11 0.83N 696924 249245 157674 276804 83890 419661 232037
Young u λEU λEI λUE λUI λIE λIU
Unemployment 0.407*** 0.109*** 0.037*** -0.350*** -0.195*** -0.314*** -0.152***(0.006) (0.012) (0.014) (0.005) (0.010) (0.004) (0.007)
Young 1.030*** 1.131*** 1.257*** -0.045*** 0.367*** -0.059*** -0.002(0.018) (0.033) (0.038) (0.014) (0.028) (0.010) (0.018)
Unemp × Young -0.212*** -0.167*** -0.189*** 0.086*** -0.133*** 0.081*** 0.076***(0.009) (0.017) (0.019) (0.007) (0.013) (0.006) (0.009)
R2 0.22 0.25 0.60 0.12 0.28 0.09 0.79N 736511 258560 166067 289075 88436 443403 234618
No College u λEU λEI λUE λUI λIE λIU
Unemployment 0.387*** 0.155*** 0.078*** -0.228*** -0.149*** -0.175*** -0.025***(0.007) (0.014) (0.014) (0.005) (0.009) (0.004) (0.007)
No College 1.302*** 1.670*** 1.838*** 0.384*** 0.588*** 0.463*** 0.512***(0.018) (0.035) (0.039) (0.014) (0.028) (0.010) (0.019)
Unemp × No College -0.097*** -0.169*** -0.202*** -0.137*** -0.206*** -0.198*** -0.150***(0.009) (0.018) (0.019) (0.007) (0.013) (0.005) (0.010)
R2 0.20 0.23 0.60 0.12 0.32 0.11 0.83N 714529 253759 165607 282924 85958 438705 230654
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
81
Table I.6: Estimated Transitions: Disadvantaged Workers, Controlling for Occupation andState
Non-White u λEU λEI λUE λUI λIE λIU
Unemployment 0.403*** 0.084*** -0.020* -0.348*** -0.330*** -0.333*** -0.154***(0.005) (0.009) (0.012) (0.004) (0.009) (0.003) (0.006)
Non-White 0.911*** 0.597*** 0.596*** -0.217*** -0.360*** -0.256*** -0.148***(0.024) (0.048) (0.051) (0.022) (0.036) (0.016) (0.025)
Unemp × Non-White -0.129*** 0.005 0.016 0.092*** 0.126*** 0.164*** 0.145***(0.009) (0.018) (0.019) (0.009) (0.013) (0.006) (0.010)
R2 0.31 0.36 0.66 0.17 0.37 0.14 0.84N 696924 249245 157674 276804 83890 419661 232037
Young u λEU λEI λUE λUI λIE λIU
Unemployment 0.404*** 0.111*** 0.040*** -0.348*** -0.189*** -0.310*** -0.156***(0.006) (0.011) (0.013) (0.005) (0.009) (0.004) (0.006)
Young 1.105*** 1.143*** 1.227*** -0.106*** 0.275*** -0.049*** -0.003(0.023) (0.043) (0.047) (0.019) (0.034) (0.015) (0.025)
Unemp × Young -0.211*** -0.160*** -0.175*** 0.086*** -0.136*** 0.082*** 0.076***(0.008) (0.015) (0.017) (0.007) (0.013) (0.005) (0.009)
R2 0.27 0.35 0.65 0.18 0.35 0.13 0.81N 736511 258560 166067 289075 88436 443403 234618
No College u λEU λEI λUE λUI λIE λIU
Unemployment 0.385*** 0.168*** 0.084*** -0.226*** -0.143*** -0.173*** -0.032***(0.007) (0.013) (0.014) (0.005) (0.008) (0.003) (0.007)
No College 1.166*** 1.511*** 1.744*** 0.319*** 0.628*** 0.435*** 0.380***(0.025) (0.047) (0.049) (0.019) (0.034) (0.014) (0.025)
Unemp × No College -0.097*** -0.180*** -0.194*** -0.138*** -0.207*** -0.196*** -0.144***(0.009) (0.017) (0.018) (0.007) (0.013) (0.005) (0.009)
R2 0.25 0.33 0.65 0.19 0.39 0.15 0.84N 714529 253759 165607 282924 85958 438705 230654
Standard errors in parentheses. * p < 0.1, ** p < 0.05, ***p < 0.01.
82
Table I.7: Measures for Composition Effects
Log-Difference corr R2
λEU,w 11.4926% 0.9268 0.8725λEI,w 10.3985% 0.7162 0.5334λUE,w 1.3444% 0.9950 0.9917λUI,w 5.1792% 0.9711 0.9451λIE,w 1.0328% 0.9872 0.9758λIU,w 6.6231% 0.9694 0.9524uw 10.0580% 0.9754 0.9541
I.3 Effect of Demographic Composition on Transition Rates
In Figure 3 we showed that the unemployment rate would have been higher from the
1990s on if the demographic composition was held constant in 1978. In this appendix we
perform a similar exercise for transition rates. In particular, we construct weighted aggregate
transition rates as
λw,xt =16∑j=1
αjtxjt ; x ∈ {EU,EI,UE,UI, IE, IU}, (24)
where λw,xt represents group j’s weighted transition rate in period t. To determine the im-
portance of compositions of workers, we construct counterfactual transition rates by setting
the weight αj to the average level in the year 1978. Thus, the counterfactual transition rates
are
λw,xt =16∑j=1
αj1978xjt (25)
Figure I.4 compares observed transition rates (black solid lines) with the counterfac-
tual corresponding series (green dashed lines). We can see that the biggest differences in the
counterfactuals are for transition rates from employment to unemployment (λw,EU) and from
employment to out-of-the-labor-force (λw,EI). To quantify the differences, we compute the
log-deviation between the observed and counterfactual series, compute the correlation coef-
ficient, and find the R2 by regressing the observed series on counterfactual. Table I.7 gives
the results. We see that λw,EU, λw,EI and uwt would have been 10 percentage points larger if
the demographic composition of the labor force was kept at 1978 levels. Thus, the changing
demographic composition of the labor force is associated with a lower unemployment rate
and fewer job separations than we would have expected in the counterfactual.
83
1980 1985 1990 1995 2000 2005 2010 2015 20200.012
0.014
0.016
0.018
0.020
0.022
0.024
0.026
0.028
λEU,w
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.028
0.030
0.032
0.034
0.036
0.038
0.040
0.042
0.044
λEI,w
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55
λUE,w
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.16
0.18
0.20
0.22
0.24
0.26
0.28
0.30
λUI,w
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.020
0.025
0.030
0.035
0.040
λIU,w
1980 1985 1990 1995 2000 2005 2010 2015 20200.040
0.045
0.050
0.055
0.060
0.065
0.070
λIE,w
Figure I.4: Composition Effects
Black solid lines are the observed weighted transition rates while green dashed lines are the counterfactual
weighted transition rates. Data source: FRED & IPUMS-CPS.
84
I.4 Alternative Definitions of Disadvantaged Workers
In the main text we group individuals into pairs of groups based on potential experience,
education, and race. In this section we further divide groups to investigate whether our main
choice of groups is reasonable.
In particular, we divide individuals into five education groups: those with degree less
than high school, those with high school degree, those who did not have college degree but
have high school degree, those with college degree, and those with a degree beyond college.20
We divide workers into following age subgroups: less than 20, 20 to less than 30, 30 to less
than 40, 40 to less than 50, 50 to less than 60, and 60 and above. Finally, we divide workers
into three race categories: white workers, black workers, and all other races.
Table I.8 presents basic information about these workers. We see that workers with more
education have lower unemployment rates, lower exit rates (EU + EI), and lower unemploy-
ment entry rates (EU + IU). Moreover, workers with more education are less likely to leave
the labor force, those that have left the labor force are hired at higher rates, and job-finding
rates are typically higher for more-educated workers. Interestingly, job-finding rates decline
somewhat for individuals with further education post-college.
When we examine age groups we see that workers under 30 have the highest unemploy-
ment rates, employment exit rates, and unemployment entry rates. In addition, unemployed
younger workers are more likely to leave the labor force, although when those workers leave
labor force they are more likely to find a job than are other groups. In general, young workers
have the highest job-finding rates. When we examine race, we see that black workers have
the highest unemployment rate, lowest job-finding rate, highest employment exit rate, and
highest unemployment entry rate. Other non-white racial groups have unemployment and
transition rates in between those of white and black workers.
In Table I.8 we also show the correlation between the observed unemployment rate and
the approximated unemployment rate based on transition rates according to Equation (10).
For workers with high school degrees the approximation works well after 1992 (before: 0.227,
after: 0.973), while for the workers with above college degrees the approximation becomes
better after 1994 (before: 0.643, after: 0.883). Thus, when we compute the contribution of
each transition flow to the unemployment rate for workers with high school degrees we used
data from after 1992, while for workers with above college degrees we used data from after
1994.
Table I.9 gives results for the within-group analysis. The main results still hold: job-
finding rates are the primary source of cyclical unemployment fluctuations, although sepa-
ration rates make important contributions. We conclude that our results would not change
dramatically with different groupings of workers.
20The CPS did not introduce the graduate education category until 1992.
85
Table I.8: Basic Information
u EU EI UE UI IE IU ratio corr
Education< HS 0.108 0.034 0.053 0.334 0.283 0.037 0.026 0.327 0.952HS 0.069 0.022 0.032 0.340 0.223 0.052 0.030 0.215 0.227(HS, after 1992) 0.020 0.028 0.336 0.216 0.038 0.022 0.973>= HS & < College 0.048 0.015 0.028 0.379 0.206 0.062 0.028 0.267 0.978College 0.032 0.009 0.019 0.362 0.167 0.054 0.019 0.166 0.956> College 0.022 0.007 0.017 0.336 0.170 0.053 0.015 0.085 0.643> College (after 1994) 0.007 0.017 0.341 0.175 0.053 0.015 0.883Age< 20 0.169 0.051 0.135 0.379 0.393 0.094 0.068 0.074 0.90220–30 0.083 0.027 0.034 0.374 0.231 0.104 0.071 0.187 0.97530–40 0.051 0.016 0.018 0.339 0.187 0.075 0.044 0.191 0.96740–50 0.042 0.013 0.016 0.323 0.175 0.067 0.034 0.173 0.95650–60 0.040 0.011 0.020 0.296 0.175 0.045 0.018 0.148 0.940>= 60 0.036 0.010 0.066 0.280 0.287 0.015 0.003 0.228 0.917RaceWhite 0.052 0.018 0.028 0.378 0.210 0.044 0.021 0.833 0.981Black 0.115 0.029 0.037 0.246 0.261 0.047 0.045 0.114 0.962Other 0.065 0.018 0.033 0.316 0.264 0.054 0.032 0.053 0.903
86
Table I.9: In-Group Analysis
EU EI UE UI IE IU ε
< HS 0.261 0.001 0.420 0.094 0.128 0.103 -0.007HS 0.204 -0.048 0.513 0.128 0.072 0.138 -0.008>= HS & < College 0.200 -0.017 0.515 0.108 0.046 0.152 -0.005College 0.261 -0.013 0.469 0.126 0.032 0.130 -0.005> College 0.214 -0.007 0.494 0.119 0.036 0.157 -0.012< 20 0.111 0.010 0.476 0.071 0.258 0.085 -0.01120–30 0.219 -0.026 0.524 0.104 0.097 0.090 -0.00830–40 0.269 -0.030 0.472 0.129 0.065 0.100 -0.00440–50 0.282 -0.018 0.449 0.129 0.057 0.106 -0.00650–60 0.261 -0.025 0.450 0.137 0.050 0.132 -0.005>= 60 0.121 -0.054 0.406 0.218 0.037 0.278 -0.006White 0.217 -0.035 0.514 0.129 0.058 0.123 -0.006Black 0.171 -0.053 0.485 0.142 0.141 0.121 -0.007Other 0.201 -0.054 0.468 0.137 0.113 0.143 -0.008
87
Appendix J Aging Young Workers
Since young workers eventually age into the experienced workers category, this may af-
fect our estimates of labor market transitions. As we can see in the Figure 4, in comparison
with other subgroups there exists a small gap between young workers’ approximated unem-
ployment based on Equation (10) and their observed unemployment rate. In this appendix
we redo the analysis for young workers incorporating the transition flow of young workers
becoming experienced workers into our approximation method for the unemployment rate.
In our sample, around 2.7% of the respondents in the IPUMS-CPS have their status change
between young and experienced workers. We find that incorporating the flow between young
and experienced workers into Equation (10) does improve the performance of the steady-
state approach in explaining the observed unemployment rate. Although incorporating this
flow may slightly change the approximation formula and decomposition approach, our main
results for young workers’ unemployment fluctuations and unemployment gap still hold. The
analysis in our main results is quite robust, even though transition flows between young and
experienced workers are not incorporated.
J.1 New Approximation for Young Workers
Because young workers may become experienced workers while experienced workers will
never become young workers, the transition from young to experienced workers is one direc-
tion only. Although experienced workers will retire and leave the labor force, this outflow
will be measured by the transition rate between working and out-of-the-labor-force. Hence
we only have to incorporate the outflow from young workers to experienced workers into our
current approximation approach.
After we incorporate flows from young to experienced workers, young workers’ transitions
from employment, unemployment and out-of-the-labor-force can be rewritten as
Eyt = λEE
t Eyt−1 + λUE
t Uyt−1 + λIE
t Iyt−1 − αE
yt−1,
Uyt = λEU
t Eyt−1 + λUU
t Uyt−1 + λIU
t Iyt−1 − θU
yt−1,
Iyt = λEIt E
yt−1 + λUU
t Uyt−1 + λIE
t Iyt−1 − γI
yt−1.
(26)
Here, Ey, Uy and Iy respectively represent the stocks of employment, unemployment, and
out-of-the-labor-force for young workers, and λxzt represents young workers’ transition rate
from labor force state x to z. In addition to the transition rates between labor force states,
we have outflow rates from young to experienced workers. Here α is the transition probability
that an employed young worker becomes an experienced workers by the end of month t, while
θ and γ represent the similar transition probabilities for young workers who are unemployed
and not in the labor force, respectively. Figure J.5 shows these outflow rates. The magnitude
88
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.010
0.012
0.014
0.016
0.018
0.020
0.022α θ γ
Figure J.5: α, θ, and γ
Outflow rates from young workers to experienced workers (α, θ, γ).
of these outflow rates is around one-third as large as young workers’ employment exit rate
(λEU + λEI).
Thus, given the steady-state assumption, we have
(λEU + λEI + α)Ey = λUEUy + λIEIy,
(λUE + λUI + θ)Uy = λEUEy + λIUIy,
(λIE + λIU + γ)Iy = λUIUy + λEIEy.
(27)
Equation (27) shows that the inflow of each labor force state will equal outflow. Here, the
outflow rates from young to experienced workers are now incorporated into this formula. By
solving the Equation (27) we can derive the steady-state approximated unemployment stock
and employment stock respectively to be:
Uyt = ((λIE
t + λIUt + γ)λEU
t + λIUt λ
EIt )(λEI
t λUEt + λUI
t (λEUt + λEI
t + α)),
Eyt = ((λIE
t + λIUt + γ)λUE
t + λIEt λ
UIt )(λUI
t λEUt + λEI
t (λUEt + λUI
t + θ)),
so the steady-state approximated unemployment rate will be
uyt =Uy
Uy + Ey. (28)
89
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.06
0.08
0.10
0.12
0.14ut ug, ct
No α, θ, γ
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.06
0.08
0.10
0.12
0.14
0.16 ut uyt
Including α, θ, γ
Figure J.6: Observed vs. Constructed Unemployment Rate
The left graph compares the observed unemployment rate ut with the approximated unemployment rate
ug,ct that does not consider α, θ, γ. The right graph compares the observed unemployment rate ut with the
approximated unemployment rate uyt that incorporates α, θ, γ.
When α = θ = γ = 0, then uyt = will be
ug,ct =λEIt λ
IUt + λIE
t λEUt + λIU
t λEUt
(λEIt λ
IUt + λIE
t λEUt + λIU
t λEUt ) + (λUI
t λIEt + λIE
t λUEt + λIU
t λUEt )
, (29)
which is the original approximation formula of unemployment rate given in Equation (10).
Figure J.6 compares observed unemployment ut with approximated unemployment rates
uyt and ug,ct for young workers. Figure J.6 clearly shows that the approximated unemploy-
ment rate based on Equation (28) incorporating outflows from young to experienced workers
(α, θ, γ) fits the unemployment rate better than the original approximation based on Equa-
tion (29). Particularly, after incorporating α, θ, γ into the approximation formula, the gap
between the observed and approximated unemployment rates disappeared. Hence, including
outflows from young to experienced workers does improve the performance of steady-state
approximation formula in explaining young workers’ unemployment rate.
As α, β and γ represent outflow rates from young to experienced workers, they can also
be interpreted as inflow rates to experienced workers from young workers. In the main
text we use Equation (29) to compute the approximated unemployment rate for all types
of workers. Although Equation (29) does not include α, β and γ for experienced workers,
the approximated unemployment rate based on Equation (29) can capture well the dynamic
changes in the observed unemployment rates. Thus, it is not necessary or quantitatively
important for us to incorporate α, θ and γ as inflow component for experienced workers
This improvement in the performance of the approximated unemployment rate for young
workers can be attributed to the percentages of experienced and young workers in the sample.
90
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.4
0.2
0.0
0.2
0.4
0.6Observed Approximated
Observed vs. Approximated
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.4
0.2
0.0
0.2
0.4
λUE + λ IE λEU + λEI
Sources of Unemployment Fluctuations
Figure J.7: Results for Unemployment Fluctuations
The left graph compares observed (solid line) and approximated (dashed line) unemployment fluctuations.
The right graph compares observed unemployment fluctuations (solid line), contributions from separation
(dotted line), and contributions from hiring (dashed line).
The percentage of experienced workers in the sample is around 75%, compared with only
25% young workers (refer to Appendix A). Because the population of young workers is much
smaller than that of experienced workers, outflows from young to experienced workers are
therefore quantitatively important in determining young workers’ transition flows between
employment, unemployment, and out-of-the-labor-force. Thus incorporating the outflow
rates α, β and γ can help improve the performance of the young workers’ approximated
unemployment rate. In contrast, because the total population of experienced workers is
larger, this will not have a significant impact on experienced workers’ stocks and hence the
approximated unemployment rate for experienced workers from Equation (29) is still a good
approximation for their observed unemployment rates.
J.2 Unemployment Fluctuations Analysis
To discuss the sources of unemployment fluctuations, we can still use the Taylor Theorem
and log-linearize Equation (28) around the steady-state (trend). We can again decompose
the unemployment fluctuations into six components related to following transition flows:
λEU, λEI, λUE, λUI, λIU, λIE. Thus, we can have following decomposition:
F tott︷ ︸︸ ︷
ln uyt − lnuyt =
F approxt︷ ︸︸ ︷
FEUt + FEI
t + FUEt + FUI
t + F IEt + F IU
t +εt, (30)
where εt represents the error term that captured the contribution from α, θ and γ, and the
bias due to the log-linearization approximation. Moreover, lnuyt represents the trend of young
91
Table J.10: Contribution of Transition Flows
λEU λEI λUE λUI λIE λIU ε
Fluctuations (β) 0.165 −0.041 0.581 0.089 0.14 0.076 −0.012
λEU λEI λUE λUI λIE λIU ε
Gap (r) 0.667 0.325 −0.155 −0.182 −0.541 0.769 0.117
The contribution of each transition rate to unemployment fluctuations and the contributionof each transition rate to the unemployment gap for young workers.
workers’ unemployment rate and F tott represents the observed unemployment fluctuations.
Each component in the right-hand-side of Equation (30) is
F xt =
∂ lnuyt∂λxt
∣∣∣∣Λt=Λt,Γt=Γt
× λxt · (lnλxt − ln λxt ), x ∈ {EU,EI,UE,UI, IE, IU}. (31)
Here ln λxt represents the trend component of transition flows λx. We again use Λt to represent
the vector of the six transition rates between employment, unemployment, and not-in-the-
labor-force, while we use Γ to represent the vector αt, θt, γt. Therefore, Λt represents the
trend component for these six transition rates between labor force states while Γt is the trend
of the transition flows between young and experienced. We again use HP filter to derive the
trend. F approxt represents the approximation for the unemployment fluctuations. Thus, we
can measure the contribution of each transition rates by computing
βx =cov (F tot
t , F xt )
var (F tott )
.
Figure J.7 shows the related analysis for young workers’ unemployment fluctuations based
on the log-linearized approximation according to Equation (28). The left part compares the
observed unemployment fluctuations F tott and the approximation F approx
t . The correlation
between F tott and F approx
t is 0.998. In addition, we regress F tott on F approx
t and find the R2
to be 0.997. These measures show that the log-linearized approximation captures well the
fluctuations in unemployment. Moreover, the right side of Figure J.7 shows the contributions
of employment exit and entry to unemployment fluctuations. We can clearly see that our
primary results did not change: unemployment fluctuations for young workers are mainly
attributed to the job-finding margin.
The upper part of Table J.10 gives the β coefficients, which measure the contributions
of each transition rate to the unemployment fluctuations. Our main results still hold: the
most important source of young workers’ unemployment fluctuations is the hiring λUE. In
contrast, separations λEU are relatively unimportant. Although the β coefficient for the error
term εt after we consider αt, θt, and γt is larger than the value we found in our main results,
92
the impact of εt is still trivial.
J.3 Unemployment Gap Analysis
After discussing young workers’ unemployment fluctuations, we then discuss sources of
the unemployment gap between young and experienced workers. Because we incorporated
outflow from young to experienced workers, young workers’ approximated unemployment
rate is calculated based on Equation (28). In contrast, because we use no such outflow
for experienced workers, experienced workers’ approximated unemployment rate is com-
puted from Equation (28) based on their transition rates between labor force states with
αt, θt, γt as zero.21 We can again use the Taylor Theorem and log-linearize young workers’
approximated unemployment rates to decompose the unemployment fluctuations into six
components related to the different transition flows. The decomposition equation will be
F gapt︷ ︸︸ ︷
ln uyt − lnue,ct =
F approx,gapt︷ ︸︸ ︷
FEUt + FEI
t + FUEt + FUI
t + F IEt + F IU
t +εt, (32)
where lnue,ct represents the logarithm of the approximated unemployment rate for experi-
enced workers while lnuyt is the logarithm of young workers’ approximated unemployment
rate. Because uyt and ue,ct are both good approximations for the observed unemployment
rate, F gapt represents the observed unemployment gap between young workers and experi-
enced workers, while F xt in the right-hand side measures the contribution of transition flows
to the unemployment gap and x ∈ {EU,EI,UE,UI, IE, IU}. The error term εt captures the
contribution from α, θ, and γ, as well as the bias due to the log-linearization approximation.
Here each component in the right-hand side of Equation (32) is
F xt =
∂ lnuyt∂λxt
∣∣∣∣Λt=Λe
t ,Γt=Γet=0
× λx,et · (lnλxt − lnλx,et ), (33)
where lnλx,et represent the transition rates of experienced workers, Λet the vector of the six
transition rates between labor force states, and Γet the vector of αt, θt, γt for experienced
workers. Because there is no outflow from experienced workers to young workers, here Γet is
a zero vector. Thus, F approx,gapt represents the unemployment gap that can be attributed to
the transition rate between labor force states, and we can measure the contribution of each
transition rate to unemployment gap by computing
rx =1
T
T∑t=1
F gapt
F xt
.
21This is exactly the same as using Equation (29).
93
1980 1985 1990 1995 2000 2005 2010 2015 20200.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0Observed Approximated
Observed vs. Approximated
1980 1985 1990 1995 2000 2005 2010 2015 2020
0.6
0.4
0.2
0.0
0.2
0.4
0.6
0.8
1.0λUE + λ IE λEU + λEI
Sources of Unemployment Gap
Figure J.8: Results for Unemployment Gap
The left graph gives the observed (solid line) and approximated (dashed line) unemployment gaps. The right
shows the observed unemployment gap (solid line), the contributions from separation (dotted line), and the
contributions from hiring (dashed line).
Figure J.8 shows the related results for young workers’ unemployment gap. The left side
compares the observed unemployment gap F gapt and its approximation F approx
t . The corre-
lation between F gapt and F approx,gap
t is 0.99. In addition, we regress F gapt on F approx,gap
t and
find the R2 to be 0.98. Thus, these measures show that the log-linearized approximation
can capture well the changes in the unemployment gap. Moreover, the right side of Figure
J.8 shows the contributions of separation and hiring to the unemployment gap. Our main
results still hold: separation is the primary source of the unemployment gap between young
and experienced workers.
The lower part of Table J.10 gives the r coefficients which measure the contributions of
each transition rate to the unemployment gap. Our main results still hold: the unemployment
gap is due to the unemployment inflow rate (λEU and λIU) and the transition rate from
employment to out-of-the-labor-force (λEI). In contrast, hiring (λUE and λIE) does not
contribute to the unemployment gap. However, incorporating the outflow rates changes the
contribution of the error term εt to the unemployment gap. Table J.10 shows that around
12 percentage of the unemployment gap can be attributed to εt, while our analysis in the
main text shows the contributions of εt is trivial. Because we are now considering transition
flows between young and experienced workers, this difference in the contribution of εt to the
unemployment gap can be attributed to the impact of the outflow rates from young workers
to experienced workers (α, θ, and γ). In spite of this difference, our primary results do not
change.
94
J.4 Summary
Based our analysis, when we consider transition flows between young and experienced
workers the approximated unemployment rate fits young workers’ observed unemployment
rate better. In addition, though we incorporated these transition flows between young and
experienced workers, our main results regarding the sources of young workers’ unemployment
fluctuations and unemployment gap remain the same. Thus, based on this extension, we find
that our main results are reliable and robust.
95