POLITICAL INFORMATION CYCLES:WHEN DO VOTERS SANCTION INCUMBENT PARTIES

FOR HIGH HOMICIDE RATES?∗

JOHN MARSHALL†

APRIL 2018

I argue that voter news consumption tracks electoral cycles, and thus electoral ac-countability is often shaped by noisy performance indicators revealed before elections.Examining local homicides prior to Mexican municipal elections, I test this theory’scentral implications using three sources of plausibly exogenous variation. First, I showthat voters indeed consume more news and are better informed before local elections.Second, pre-election homicide spikes substantially decrease the incumbent party’s re-election prospects. Sanctioning is limited to mayors controlling local police forces,and does not reflect longer-term homicide rates. Third, the punishment of pre-electionhomicide shocks increases with access to local broadcast media stations. Such polit-ical information cycles appear to reflect poorly-informed Bayesian voters demandingmore information during election campaigns, rather than naive responses to exogenousadverse events. By highlighting the importance of when voters consume news, thesefindings can help explain variation in the media’s political influence and the mixedevidence that information supports electoral accountability.

JEL codes: D72, D78, O17.

∗I thank Jim Alt, Eric Arias, Abhijit Banerjee, Allyson Benton, David Broockman, Bruce Bueno de Mesquita,Melani Cammett, Francisco Cantu, Ana De La O, Oeindrila Dube, Raissa Fabregas, Nilesh Fernando, Hernan Flom,Alex Fouirnaies, Jeff Frieden, Thomas Fujiwara, Saad Gulzar, Sebastian Garrido de Sierra, Andy Hall, Torben Iversen,Joy Langston, Horacio Larreguy, Chappell Lawson, Sandra Ley, Rakeen Mabud, Noah Nathan, Brian Phillips, CarloPrato, Otis Reed, James Robinson, Rob Schub, Gilles Serra, Jim Snyder, Chiara Superti, Edoardo Teso, Giancarlo Vis-conti, Julie Weaver, and workshop or conference participants at presentations at APSA, Berkeley, Caltech, Columbia,Harvard, LSE, MIT, MPSA, NYU, NEWEPS, Stanford, Stanford GSB, and Yale for illuminating discussions anduseful comments. I thank Horacio Larreguy for kindly sharing municipal election data.†Department of Political Science, Columbia University. Email: jm4401@columbia.edu.

1

1 Introduction

Electoral accountability ideally supports effective governance that increases voter welfare by en-abling informed voters to select suitable politicians to represent them. Given the difficulty ofdirectly observing an incumbent’s competence or alignment with voter interests, performance onsalient issues like economic growth, redistribution, public security, or corruption can serve as valu-able signals of an incumbent’s continuing suitability for office (Fearon 1999; Rogoff 1990). How-ever, particularly in developing contexts, voters are often poorly informed about politics and cur-rent affairs (e.g. Keefer 2007; Pande 2011), or do not receive relevant information pertaining totheir incumbent party (Larreguy, Marshall and Snyder 2018; Snyder and Stromberg 2010).1 Thislack of politically-relevant information may explain why voters often fail to hold incumbents toaccount for their performance in office (e.g. Dunning et al. forthcoming).2

In response, scholars and policymakers have argued that increasing access to media can facil-itate electoral accountability. This rests on media outlets deciding to report relevant incumbentperformance indicators that voters would not otherwise obtain, and receives some empirical sup-port (see Ashworth 2012; Pande 2011).3 Although this may be a precondition for accountability,the focus on access neglects the reality that the availability of media coverage does not necessarilyimply that voters actually consume politically-relevant news.

I argue that how voters hold incumbent parties to account for their performance in office also re-flects the timing of their news consumption. News consumption is likely to track electoral cycles if,before elections, voters actively seek out politically-relevant information (e.g. Feddersen and San-droni 2006; Hamilton 2004; Marshall forthcoming) or media outlets increase their supply of suchinformation. Consequently, incumbent performance indicators reported by media outlets beforeelections—when voters consume most news—have the potential to disproportionately influenceelection outcomes, especially in low-information environments where voters possess impreciseprior beliefs. Such political information cycles could help account for the apparently large impactof the revelation shortly before the 2016 U.S. presidential election that the FBI was continuing toinvestigate Hillary Clinton.

1Ill-informed electorates also pose similar dilemmas in established democracies (Delli Carpini and Keeter 1996),but the risks of electing incompetent politicians are likely greater under weak democratic institutions.

2Evidence that voters in consolidating democracies punish or reward politicians for economic performance (An-derson 2007; Remmer 1991; Roberts and Wibbels 1999), levels of public security (Vivanco et al. 2015), or malfeasancein office (Arias et al. 2018; Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013; Ferraz and Finan 2008;Larreguy, Marshall and Snyder 2018) is mixed.

3Well-identified studies show that media programming affects voting behavior (e.g. DellaVigna and Kaplan 2007;Enikolopov, Petrova and Zhuravskaya 2011; Larreguy, Marshall and Snyder 2018; Martin and Yurukoglu 2017; Snyderand Stromberg 2010).

2

While voting behavior is likely to be sensitive to informative signals of incumbent type (e.g.Rogoff 1990), even pre-election performance signals that are uncorrelated with desirable char-acteristics could affect incumbent electoral prospects. Responses to the latter type of signal—thispaper’s empirical focus—could reflect rational or irrational belief updating. On one hand, Bayesianvoters with limited political information may rely on signals that reflect both performance-relevantcharacteristics and random shocks (e.g. Ashworth 2005), or learn about the incumbent’s type fromtheir response to the shock (Ashworth, Bueno de Mesquita and Friedenberg forthcoming; Besleyand Burgess 2002). On the other, voters could engage in a naive response to adverse pre-electionshocks (e.g. Achen and Bartels 2017; Wolfers 2007; Zaller 1992).

This paper examines the extent to which political information cycles drive voting behavior inthe context of Mexican voters holding municipal incumbent parties to account for local homicides.Following significant democratizing and decentralizing reforms in the 1990s, Mexican electionshave become competitive at national and sub-national levels, while responsibility for public ser-vice delivery such as policing is increasingly assigned to municipal, rather than state or national,governments. Public security is a major concern for Mexican voters, as in other Latin Americannations, and thus local homicides represent a salient measure of incumbent party performance.4

Moreover, since many voters are poorly informed about politics and a significant portion of theelectorate lacks strong partisan ties (Chong et al. 2015; Greene 2011; McCann and Lawson 2003),there is scope for crime reports to influence voting behavior.

Leveraging plausibly exogenous variation in temporal proximity to local elections, municipal-level homicide spikes occurring before elections, and then access to local broadcast media stations,I test the key features of the political information cycles argument at the individual and aggregatelevels. First, to demonstrate the existence of cycles in political news consumption, I exploit theirregular timing of survey waves with respect to Mexican states’ staggered electoral cycles to iso-late variation in the proximity of local—municipal and state legislative—elections (see also Eifert,Miguel and Posner 2010). Rather than separate supply and demand explanations for informationconsumption, this analysis identifies consumption cycles in equilibrium. I indeed find that votersreport consuming more political news through radio and television in the months preceding lo-cal elections. For some less-informed voters, they effectively only consume political news duringelections campaigns. These changes are reflected in a 0.2 standard deviation increase in topicalpolitical knowledge.

Second, using electoral returns from between 1999 to 2013, I show that spikes in the inci-

4Although municipal mayors could not seek re-election until 2018, voters generally hold the incumbent’s partyresponsible for performance in Mexico’s party-centric system (e.g. Arias et al. 2018; Chong et al. 2015; Larreguy,Marshall and Snyder 2018).

3

dence of homicides just before elections—when voters consume most news—substantially harmthe municipal incumbent party’s electoral performance. I estimate the effects of these pre-electionhomicide shocks by leveraging plausibly exogenous month-to-month volatility in homicide countswithin municipalities to capture idiosyncratic short-term deviations from longer-term trends. Specif-ically, I compare “shocked” municipal elections that experienced more homicides in the twomonths preceding the election to “control” elections from within the same municipality that expe-rienced at least as many homicides in the two months after the election.5 I support the design bydemonstrating that the distribution of homicide shocks is consistent with random sampling vari-ability, is balanced over 99 predetermined covariates, and inconsistent with manipulation by politi-cians or drug trafficking organizations (DTOs). Despite being uncorrelated with broader homicidelevels and trends, pre-election homicide shocks reduce the incumbent party’s probability of win-ning by 12 percentage points. Such punishment is consistent with Bayesian and naive responses tonoisy signals of incumbent competence.

Furthermore, consistent with poorly-informed voters relying on the pre-election signals avail-able when they actually consume news, voters do not hold incumbents accountable for longer-runhomicide rates. Using a difference-in-differences design, I find no evidence to suggest that homi-cide rates over the prior year or electoral cycle—arguably, more desirable indicators of incumbentperformance (Healy and Lenz 2014), but which may not be observed by voters predominantlyconsuming information just before elections—affect incumbent electoral performance. Moreover,my analysis of concurrent state and federal elections, and comparisons between municipalitieswith and without their own police force, indicate that the parties of incumbent mayors, rather thanhigher levels of government, are held to account for local homicide shocks (Vivanco et al. 2015).

Third, I link the individual and aggregate-level findings by showing that local broadcast mediaplays a crucial role in linking news consumption to electoral sanctioning. Local broadcastersreport politically-relevant news that voters may not be able to access otherwise, and local crimereceives substantial media coverage (e.g. Osorio 2015; Trelles and Carreras 2012). Using detailedcoverage maps for every radio and television station in the country, I leverage within-neighboringprecinct variation in signal access to identify the effects of broadcast media in more urban areas(see also Larreguy, Marshall and Snyder 2018). On average, I find that each additional local mediastation based within a precinct’s own municipality reduces the incumbent party’s vote share by 0.2percentage points. However, the 21% of precincts covered by less than 8 local media stations donot significantly punish homicide shocks. Neither additional non-local media stations nor party adshares on accessible outlets can account for these findings.

5I include only municipalities experiencing at least one homicide over the two months before and after an election.

4

Turning to the mechanisms, the observed electoral sanctioning is consistent with voters de-manding more news before elections and updating from the events in the news at this time in aBayesian fashion. First, indicating that voters internalize news in the media just before elections,pre-election homicide shocks both increase concern about public security and reduce confidencein the municipal incumbent by around 10 percentage points. In contrast with a “recency bias” ex-planation, homicide shocks occurring before surveys outside of electoral campaigns—when votersconsume less news—do not significantly influence voter beliefs in the short term. Second, il-lustrating voters’ capacity to integrate new information with their prior beliefs, the sanctioning ofhomicide shocks is greatest where a different incumbent party at the previous election did not over-see a pre-election homicide shock, or the same incumbent party experiences consecutive shocks.Such sophistication suggests that naive responses to shocks do not drive the results. Third, I showthat although media coverage of violent crime closely tracks the occurrence of homicides, cover-age does not increase during election campaigns. In conjunction with the fact that some votersonly consume news before elections, this suggests that demand for politically-relevant news is theprimary force underpinning political information cycles.

The overarching finding that political information cycles induce voters to heavily weight recentperformance indicators makes several contributions. First, the findings extend the literature doc-umenting the persuasive effects of broadcast media content (e.g. Adena et al. 2015; DellaVignaand Kaplan 2007; DellaVigna et al. 2014; Enikolopov, Petrova and Zhuravskaya 2011; Larreguy,Marshall and Snyder forthcoming; Martin and Yurukoglu 2017; Spenkuch and Toniatti 2016;Yanagizawa-Drott 2014), and especially the importance of local news (Ferraz and Finan 2008;Larreguy, Marshall and Snyder 2018; Snyder and Stromberg 2010). In contrast with these studiesfocusing on access to media at particular points in time, I demonstrate that the effects of the mediavary systematically with the timing of voter news consumption. I thus highlight the importanceof consumption patterns for understanding the broader effects of editorial content and coverage ofparticular events.

Second, integrating cyclical news consumption into models of electoral accountability helpsaccount for the mixed evidence that incumbent performance information influences voting behav-ior. For example, the timing of information consumption may explain why pre-election corruptionrevelations have produced larger electoral impacts than similar revelations occurring earlier in aelectoral cycle (Brollo 2009; Chang, Golden and Hill 2010), voters tend to overweight recent eco-nomic performance in evaluating U.S. presidents at the end of their term (Achen and Bartels 2017;Healy and Lenz 2014), and short-term fluctuations have induced electoral volatility across LatinAmerica (Roberts and Wibbels 1999). Furthermore, the importance of local media coverage in my

5

findings reiterates the substantially larger effects associated with media dissemination of incum-bent performance information (Banerjee et al. 2011; Ferraz and Finan 2008; Keefer and Khemani2014; Larreguy, Marshall and Snyder 2018) relative to dissemination campaigns that provide in-formation to voters in person (e.g. Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013;Dunning et al. forthcoming; Humphreys and Weinstein 2012). Nevertheless, the previous find-ings that local media can induce and amplify electoral sanctioning focus on revelations occurringshortly before elections. This paper thus tempers the media’s pro-accountability potential by il-lustrating the importance of timing information dissemination campaigns to coincide with voterengagement.

Third, the evidence further illuminates the process through which voters process informationand update their beliefs. Rather than responding naively to exogenous adverse shocks, voter be-liefs and behavior—which reflect positive and negative pre-election homicide shocks, are differen-tiated across mayors with and without a police force, and vary with previous pre-election homicideshocks experiences—is consistent with Bayesian updating in a career concerns model or fromevaluating responses to random shocks. These findings complement the sophisticated updatingdocumented by field experiments in both developed (Kendall, Nannicini and Trebbi 2015) anddeveloping contexts (Arias et al. 2018; Banerjee et al. 2011), but do so in the context of violentcrime—a metric of incumbent performance that has received limited empirical attention, despiterepresenting a key valence issue in many developing contexts.6 Nevertheless, in line with Durante,Pinotti and Tesei (2017), the results also highlight the substantial degree to which the behavior ofvoters with limited information can be influenced by media reporting on noisy signals of incumbentperformance.

Finally, the importance of political information cycles contributes to a broader political econ-omy literature emphasizing cyclical political behavior and outcomes. Previous studies have ob-served that economic policies and output (e.g. Brender and Drazen 2005), the distribution of aid(Faye and Niehaus 2012), and corruption levels (Bobonis, Fuertes and Schwabe 2013) reflect an-ticipated electoral cycles. My findings similarly chime with the strategic decision of politiciansto act surreptitiously when the media is covering other major events (Durante and Zhuravskayaforthcoming; Eisensee and Stromberg 2007). In contrast with such cycles driven by the actions ofpoliticians, I show that cyclical behavior primarily driven by voters can also have important elec-toral implications. These news consumption cycles mirror, and could induce, the cycles in partisan

6Existing studies use less plausible identification strategies and do not address political information cycles (Vi-vanco et al. 2015). Other studies examine the effects of terrorism (Berrebi and Klor 2006) or the effects of crime solelyon beliefs and ratings (Arnold and Carnes 2012), political participation (Bateson 2012; Ley 2014), or trust (Blanco2013; Corbacho, Philipp and Ruiz-Vega 2015).

6

identification observed across the world (Michelitch and Utych 2018).The paper proceeds as follows. Section 2 considers sources of political information cycles,

and the implications for electoral accountability. Section 3 describes the Mexican context used totest the argument. Sections 4-6 examine three central components of political information cycles:the timing of information consumption, the effect of homicides occurring at different points onincumbent electoral performance, and the moderating role of media coverage. Section 7 evaluatesmechanisms. Section 8 concludes by discussing broader implications.

2 Theoretical framework

Electoral accountability is underpinned by voters re-electing competent or aligned incumbent par-ties on the basis of relevant performance indicators (Fearon 1999; Maskin and Tirole 2004; Rogoff1990). In light of the mixed evidence that voters use performance information to hold incumbentsto account (see e.g. Anderson 2007; Ashworth 2012; Healy and Malhotra 2013; Pande 2011),scholars have emphasized the importance of voter access to politically-relevant news in the media(Ferraz and Finan 2008; Larreguy, Marshall and Snyder 2018; Snyder and Stromberg 2010). How-ever, such studies assume that voters actually consume the available information. Furthermore, byfocusing on performance metrics released before elections, extant research has failed to capturehow the timing of when news is released may differentially affect voter beliefs and behavior.

I instead propose that the impact of news on electoral accountability depends upon the tim-ing of voter news consumption. Specifically, I suggest that news consumption follows a cyclewhere many voters only seriously encounter politically-relevant information—performance indi-cators pertaining to incumbents that a voter can sanction electorally—just before elections. Votersthus rely on the incumbent performance indicators in the news at this time when casting theirballot, which are often noisy signals of incumbent competence or alignment with voter interests.

2.1 Causes of political information cycles

Voters seeking to influence outcomes of mass elections generally face strong incentives to remain“rationally ignorant” (Downs 1957). However, there are a number of reasons to believe that con-sumption of politically-relevant news increases before elections.

First, to the extent that political news is a consumption good (e.g. Baum 2002; Hamilton 2004),electoral campaigns are likely to particularly engage voters. While coverage of political events outof election season is often more focused on specific policy issues, dramatic campaigns appeal to abroader audience (e.g. Farnsworth and Lichter 2011).

7

Second, voters may acquire political information before elections for strategic reasons beyondaffecting election outcomes. Marshall (forthcoming) shows that voters in social networks that col-lectively value knowledge about politics acquire information about election campaigns to cultivatea reputation as politically sophisticated. Such dynamics may be especially strong around electionswhen political discussion is more common (Baker, Ames and Renno 2006). Similarly, voters mayacquire information ahead of elections to improve decision-making in other domains (Larcinese2005), e.g. if investment choices depend upon political risks. The information consumed for suchsocial and vocational reasons may spill over to influence vote choices.

Third, voters may feel a civic duty to become informed around elections. Feddersen and San-droni (2006) argue that “ethical” independent voters are intrinsically motivated to cast an informedballot, where information acquisition is concentrated among voters facing the lowest acquisitioncosts. A large body of research suggests that a voter’s civic duty is correlated with political en-gagement and participation (e.g. Blais 2000). Relatedly, Degan (2006) suggests that voters acquireinformation to avoid the psychic costs of voting for non-preferred candidates.

Fourth, voters without instrumental or intrinsic motivations may still consume more informa-tion around elections simply because more time and space is devoted to political news, advertising,and specific election programming. Even for uninterested voters, such information becomes dif-ficult to avoid—especially where there are relatively few channels available (Prior 2007) or suchinformation is recast as entertainment (Baum 2002)—and performance indicators are increasinglyplaced in the context of appraising incumbent politicians and parties (Semetko and Valkenburg2000).

Regardless of whether consumption is driven principally by voter demand for information ormedia supply of information, consumption of politically-relevant news is thus likely to increaseprior to elections. Such political information cycles would cause many voters—especially in de-veloping contexts where baseline political knowledge is comparatively low (Keefer 2007; Pande2011)—to essentially engage with the news before the election for the first time in an electoral cy-cle. Relatively politically engaged voters are, in turn, likely to ratchet up their exposure to politicalnews.

2.2 Consequences for electoral accountability

Provided that voters can access credible news (e.g. Chiang and Knight 2011) and wish to useincumbent performance indicators to elect better politicians (e.g. Arias et al. 2018; Krosnick andKinder 1990), political information cycles may affect electoral accountability by inducing votersto respond to politically-relevant information when news consumption is greatest before elections.

8

For example, if there are many homicides in the local news just before an election, voters maybecome more concerned about public security and negatively update their beliefs about the abilityof responsible incumbents to effectively reduce crime once re-elected. Such a response couldreflect both naive and sophisticated responses to events in the news.

A large literature suggests that voters may respond to political information in a naive manner.Zaller (1992) argues that, among voters that actually consume news without engaging in moti-vated reasoning, new information has short-lived effects on voter beliefs and vote choices. This isbecause the political opinions of such voters reflect recent considerations that are neither internal-ized nor critically evaluated. Healy and Lenz (2014) further suggest that voters consciously relyon less accurate signals of economic performance, although others highlight voters’ propensity toover-respond to weak signals (e.g. Ortoleva and Snowberg 2015). These information processingfailures appear to influence voting behavior: the findings of Achen and Bartels (2017), Healy,Malhotra and Mo (2010), and Wolfers (2007) suggest that voters sanction and reward incumbentsfor pre-election events—including shark attacks, local sports team defeats, and macroeconomicfluctuations—that are beyond the incumbent’s control.7

Alternatively, voters whose news consumption patterns track electoral cycles may rationallyholds officials to account based on information in the news just before elections. As AppendixA.1.1 shows formally in a simple career concerns model (Ashworth 2005), Bayesian voters mayupdate from events both within and beyond the incumbent’s control. When incumbent perfor-mance indicators such as economic growth or crime rates reflect both unobservable incumbentcompetence and random fluctuations, voters use performance cutoff strategies to sanction incum-bents that perform below expectations because they attribute the observed outcome to a combi-nation of both underlying sources of performance outcomes. As voters receive more independentperformance signals, by consuming news throughout the electoral cycle, they are able to averageacross idiosyncratic shocks and more frequently re-elect the most competent incumbents. How-ever, voters that principally consume news before elections only observe a small number of perfor-mance indicators, and thus sanction incumbents for poor performance even when such performancepartly reflects events beyond the incumbent’s control. Political information cycles may thus induceBayesian voters to rely on noisy performance indicators revealed before elections.

A second channel through which pre-election shocks—even when beyond the incumbent’scontrol—could rationally induce electoral accountability is if the occurrence of such shocks pro-vides an opportunity for voters to learn about the incumbent’s suitability for office. For instance,

7The first two empirical findings have been challenged by recent re-assessments (Fowler and Hall forthcoming;Fowler and Montagnes 2015).

9

Ashworth, Bueno de Mesquita and Friedenberg (forthcoming) develop a model where voters sanc-tion incumbents after natural disasters when such disasters interact with unobserved incumbentcompetence, such as disaster preparation. Rather than acting naively, voters will sanction incum-bents (with an electoral advantage) on average following a random shock if the shock reveals lowcompetence by amplifying the extent to which competence drives observed performance outcomes.Besley and Burgess (2002) convey similar insights. Appendix A.1.2 also develops a simple signal-ing model where separating equilibria are characterized by competent incumbents distinguishingthemselves in response to a shock (Rogoff 1990), e.g. by appearing empathetic in the media orproposing credible policies to mitigate violence if re-elected, in their response to a homicide shockbeyond their control.8 In such a model, a negative shock reduces the incumbent’s vote share onaverage where there is a large density of weak incumbent supporters—as is often the case in com-petitive elections—because incumbent vote tallies are more responsive to bad than good news.Consistent with such theories, Cole, Healy and Werker (2012), show that, conditional on a shockoccurring, (in)effective policy responses can mitigate (accentuate) the negative consequences ofnatural disasters. However, even without conditioning on post-treatment responses, these modelspredict differential responses to shocks beyond the incumbent’s control on average.

The following empirical analysis tests the central predictions of political information cycles:whether political news consumption increases before elections; whether adverse events reported inthe news before elections influence electoral accountability, and are driven by access to the media;and ultimately the extent to voting behavior reflects naive or Bayesian responses to news. Thesehypotheses are examined in the context of local homicides in Mexico—an important performanceindicator on a salient topic widely reported by local media—ahead of municipal elections.

3 Violent crime and political accountability in Mexico

Despite holding regular elections, Mexico was effectively a one-party state until the late 1990s.9

Mexico has since democratized significantly, with competitive elections between three main po-litical parties—the right-wing National Action Party (PAN), left-wing Party of the DemocraticRevolution (PRD), and previously-hegemonic Institutional Revolutionary Party (PRI).10 Despite

8In contrast with the career concerns model and Ashworth, Bueno de Mesquita and Friedenberg (forthcoming),this model assumes that incumbents know their own types.

9The PRI had retained a stranglehold on power since 1929 by implementing populist policies, establishing strongclientelistic ties, effectively mobilizing voters, creating barriers to political entry, and manipulating electoral outcomes(e.g. Cornelius 1996; Greene 2007; Magaloni 2006).

10After attaining legislative pluralities in the 1990s, the long-time conservative opposition PAN fully broke PRIcontrol when Vicente Fox won the Presidency in 2000. In 2006, Felipe Calderon retained the Presidency for the

10

growing partisan alignment, often induced by clientelistic exchanges, many Mexicans have yet todevelop strong ties to specific political parties (Greene 2011; Lawson and McCann 2005).

Mexico’s federal system is divided between three administrative and elected layers of gov-ernment: the federal government, 31 states (and the Federal District of Mexico City), and ap-proximately 2,500 municipalities. Constitutional reforms in the mid-1990s substantially increasedmayoral autonomy over the provision of local public services (see Diaz-Cayeros, Gonzalez andRojas 2006), inducing municipal spending to rise to around 20% of total government spending.11

The median municipality contains 13,000 people. Large cities with the exception of the FederalDistrict—which is divided into delegations—represent singular municipal bodies.

Municipal mayors have typically been elected every three years to non-renewable terms,12 andentered office between three and seven months after election day. Municipal elections are almostalways held concurrently with state legislative elections, although gubernatorial elections are oftenheld separately. I refer to simultaneous municipal and state legislative elections as “local elections.”Such local elections have become increasingly competitive, with fewer than 50% of municipalincumbent parties being re-elected and 31% of elections being won by less than 5 percentagepoints. The majority of local elections do not coincide with federal elections.13 Many mayors,especially those in large municipalities, seek other elected offices at the end of their term.

Although incumbent mayors could not seek re-election, Mexico’s party-centric system ensuresthat voters are likely to hold incumbent parties accountable for the performance of individual politi-cians. First, voters primarily decide between party labels, not individual candidates. Arias et al.(2018) report that 80% of voters can identify the party of their municipal incumbent; this farexceeds the probability of correctly naming individual local incumbents (Chong et al. 2015; Lar-reguy, Marshall and Snyder forthcoming). Second, differences in candidate selection mechanismsacross parties at the state level ensure that candidate choices are highly correlated within parties(Langston 2003). Third, various studies find that voters hold parties to account for an incumbent’smalfeasance in office (Arias et al. 2018; Chong et al. 2015; Larreguy, Marshall and Snyder 2018).

PAN, narrowly defeating the PRD’s Andres Manuel Lopez Obrador. After regaining legislative majorities, but neverrelinquishing regional control (Langston 2003), the PRI candidate Enrique Pena Nieto won the Presidency in 2012.Since 2015, a fourth major competitor—MORENA—has emerged at the national level.

11This expansion of spending included policing, although most municipalities already had their own police forces.Presidents Fox and Calderon also increased federal transfers to municipalities for policing in the 2000s (Sabet 2010).

12From 2018, re-election will become possible for legislators, and mayors in most states.13House and Senate elections are held every three years, while the President is elected every six years.

11

3.1 Public security forces

Responsibility for public security is shared across levels of government, and like other federalsystems such as the United States, mayors play an important role in fighting crime. Both stateand federal laws can be used to prosecute criminals, and uniformed (preventive) and investigativepolice forces exist at both the federal level and in each of Mexico’s 31 states and the FederalDistrict of Mexico City. State and federal police, and increasingly the army, are responsible forinvestigating major crimes in different jurisdictions. State police investigate state crimes such ashomicides, while federal officers focus on organized crime (Reames 2003). However, around three-quarters of municipalities also possess their own police force. Municipal forces support higher-level operations and supply information, but their principal role is preventive: they primarily patrolthe streets and maintain public order, address administrative issues, and respond first to criminalincidents (Reames 2003; Sabet 2010).14 Municipal police are by far most numerous, accountingfor more than half of Mexican enforcement personnel (Sabet 2010).

In municipalities with their own police force, mayors can influence the medium-term incidenceof crime. Mayors choose the local police chief and set local policies. PAN mayors have played akey role in supporting Calderon’s crackdown on Mexico’s DTOs (Dell 2015), while political align-ment across neighboring municipalities has reduced rates of violent crime (Durante and Gutierrez2015). The widely-publicized case of Iguala represents a particularly egregious example of po-litical control of the local police, where the mayor and his wife exploited their links with lawenforcement to cover up and possibly instigate the murder of 43 student protesters in 2014. Asshown below, mayors do not manipulate short-term homicide rates.

Given the number of municipal police officers and their “on-the-streets” presence, it is notsurprising that municipal police are the foremost police force voters’ minds. When asked whichtypes of law enforcement authorities they are aware of, the 2010 National Insecurity Survey (ENSI)finds that while 75% of voters could identify municipal police, only 38% and 48% respectivelyrecognized state and federal forces. Focus groups that I conducted in several large municipalitiesin 2015 suggest that voters often blame local government for crime in their community.

3.2 Trends in homicide rates

According to the United Nations Office on Drugs and Crime (UNODC), Mexico suffers one ofthe world’s highest homicide rates. In 2011, 22.6 people per 100,000 were intentionally mur-dered. This represents the 20th highest homicide rate in the world—slightly less than South Africa,

14A central police force controlled by the mayor of Mexico City covers all delegations within the Federal District.

12

Calderon's War on Drugs.2

.4.6

.81

1.2

Ave

rage

num

ber o

f hom

icid

es

1999 2002 2005 2008 2011 2014

Month (January of year)

(a) Trends in monthly homicides (as defined byINEGI) in the average Mexican municipality,

1999-2013

0.1

.2.3

.4

Pro

port

ion

answ

erin

g th

at c

rime

is m

ost i

mpo

rtan

t

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011

(b) Voters listing crime as the most importantproblem that Mexico faces, 2001-2011

Figure 1: Crime in Mexico

Notes: The data in Figure 1b are from the 2001-2011 Latinobarometro surveys. Crime includes concernsabout crime/public security, drug trafficking, violence, or terrorism/political violence.

Colombia, and Brazil, and slightly more than Nigeria, Botswana, and Panama. After briefly de-clining since 2011, the 2017 rate will almost certainly be the highest yet recorded.

Using data from the National Institute of Statistics, Geography, and Information (INEGI), Mex-ico’s autonomous statistical agency, Figure 1a plots the number of homicides per month occurringin the average municipality between 1999 and 2013. INEGI defines an intentional homicide as anunnatural death, as determined by a coroner’s report.15 The monthly homicide rate has been sub-stantial throughout this period, but increased dramatically after President Calderon entered officein December 2006 and began Mexico’s “War on Drugs.” The homicide clearance rate is onlyapproximately 20% (Mexico Evalua 2012), and drug-related homicides—which are regionallyconcentrated—represent around half of the homicides over this period (Heinle, Rodrıguez Fer-reira and Shirk 2014). However, many municipalities only rarely experience a homicide; in fact,only one homicide occurs in the median municipality each year.

15Month and location of death are also based on the coroner’s report. Although there is no official measure ofdrug-related homicides enshrined in Mexican law, the federal government has sporadically released monthly data.This data suffers from various problems beyond its short time-series (see Heinle, Rodrıguez Ferreira and Shirk 2014),while Dell’s (2015) estimates show that total homicides measured by INEGI rose substantially more than drug-relatedhomicides following Calderon’s war on drugs. Furthermore, voter fears about public security are not only linked toorganized crime, but also reflect equally-prevalent non-drug related homicides (of which only a tiny fraction occurwithin families). Table A15 shows broadly similar results for the small sample (December 2006-October 2011) whenhomicides were classified as drug-related by a committee composed of representatives from the ministries constitutingthe National Council of Public Security.

13

These figures could understate the true homicide rate if unnatural causes cannot be detected(Mexico Evalua 2012). However, it is unlikely that the coroners reports are falsified becauseindividual-level reports are only released publicly several years after the event. Moreover, 94% ofhomicides were registered in the same month that they occurred.

Unsurprisingly, voters are concerned about Mexico’s high rates of violent crime. Like manyother Latin American countries (Blanco 2013), Figure 1b shows that the number of Mexican Lati-nobarometro respondents citing public security as the most important problem facing the countryincreased broadly in line with homicide rates. For most of the 2000s, public security registered asthe most salient issue for voters, ahead of the economy.

3.3 Media coverage and voter political knowledge

As in most developing countries, broadcast media is voters’ primary source of news. Accord-ing to the 2009 Latinobarometer, 83% of Mexicans receive political information from television,41% from radio, 30% from newspapers, and 41% from family, friends, and colleagues. In 2008,around 10% listened to radio every day, while the average person watched 4 hours of television(Ibope/AGB Mexico 2009). In contrast, only 21% have access to the internet at home in 2010,while 3G coverage only starting to expand rapidly in 2011.16 Broadcast media coverage thus playsa central role in determining whether voters receive politically-relevant information.

Mexico contains 852 AM radio stations, 1,097 FM stations, and 1,255 television stations, whichgenerally cover relatively small geographic areas (see Figure 3 below). Most stations form part ofbroader regional and national radio or television networks—such as Grupo ACIR, Radiorama,Televisa, and TV Azteca—where affiliates share branding or are owned-and-operated subsidiaries(Larreguy, Marshall and Snyder 2018). Among radios, 83% report news more than once a day, andtypically report on municipal rather than national issues (Larreguy, Lucas and Marshall 2016). Fortelevision stations, identical entertainment content is generally bought from or relayed by networkproviders. However, while national news is typically centrally provided, affiliates and regional sub-divisions emitting from major cities within each state also provide significant local news content.Of the 52 distinct television channels (excluding Mexico City’s 24-hour news channels) for whichschedules were available in 2015, the average channel broadcast 3.6 hours of news coverage eachweekday (both before and after the June elections). Slightly less than half of this news is devotedto state- or city-specific programming. Private radio and television outlets rely on advertising rev-enues to support themselves, and thus face strong incentives to tailor their programming to localaudiences (Larreguy, Marshall and Snyder 2018). The pro-PRI coverage biases that characterized

16Even in 2010 and 2011, 60% of Latinobarometro respondents reported never using the internet.

14

elections before the late 1990s have somewhat dissipated since (e.g. Hallin 2000; Lawson 2002).Homicides are not always reported in great depth, but are regularly covered in print (Ley 2014;

Osorio 2015) and are “omnipresent” in the local broadcast media (Trelles and Carreras 2012).17

Based on a survey of all local radio stations conducted in 2016 and 2017 as part of another study(Larreguy, Lucas and Marshall 2016), 96% report regularly covering local security stories, whileTable 8 below demonstrates that newspaper reporting of violent crime indeed tracks the homiciderate. The 2010 ENSI survey also reports that 87% of respondents learn about public security in thecountry and in their state from television news programs, while 34%, 29%, and 9% respectivelylearn from radio news programs, periodicals or newspapers, and the internet.18 Furthermore, 82%of voters believe that television news has the most important influence on public opinion.

Despite claiming reasonable levels of attention to political news, voters’ knowledge of publicaffairs is limited. I show below that only half of Mexicans can answer basic questions aboutpolitics. Castaneda Sabido (2011) and Chong et al. (2015) similarly find that voters are generallyunaware of mayoral responsibilities and performance in office. This lack of knowledge suggeststhat many voters may possess weak prior beliefs about their incumbent party’s suitability for office,and may thus internalize information when they actually follow the news.

4 The existence of political news consumption cycles

The political information cycles argument implies that voters consume more political informationjust prior to elections. While some voters only start consuming politically-relevant informationduring election campaigns, others will consume more than before. This section first estimates theeffect of upcoming local elections on such consumption patterns using Mexican survey data.

4.1 Data

I use four cross-sectional waves of the National Survey of Political Culture and Civil Practices(ENCUP) conducted over several weeks in November 2001, February 2003, December 2005, andAugust 2012.19 The surveys were commissioned by the Interior Ministry and designed to be

17For example, see this local news report of an gangland-style murder in Monterrey. The same news station alsoreports on domestic murders, e.g. this case of a woman strangling her partner. News programs may continue to reporton arrests and the rare cases that go to trial, especially when the defendants are found guilty.

18Although social networks may represent a key source of information on more explicitly political issues (Baker,Ames and Renno 2006), comparatively few learn from work colleagues (5%) or family, friends, and neighbors (13%).

19The study was implemented by INEGI in 2001 and 2003, and private firms in 2005 and 2012. A 2008 wave wasalso conducted, but did not provide a respondent’s municipality and asked different media consumption questions.

15

nationally representative, and focus on the country’s political culture rather than more contentiousquestions about elections.20 Each round draws stratified random samples of around 4,500 Mexicanvoters for face-to-face interviews from pre-selected electoral precincts within urban and rural stratadefined by the electoral register.21 The pooled sample includes up to 17,213 respondents across539 municipalities (including delegations within the Federal District).

Political news consumption is measured in two ways. First, I examine the frequency with whichvoters report watching or listening to the news, programs about politics, or programs about publicaffairs.22 To understand the margins at which consumption changes, I distinguish 5 consumptionintensities: never, at some point, at least monthly, at least weekly, and daily. I code indicators forat least each level above never and a 5-point scale (from 0 to 4). Second, I assess whether self-reported such news consumption translates into greater political knowledge. I focus specificallyon topical political knowledge that voters are more likely to encounter through the media (Barabaset al. 2014). Topical political knowledge is measured as the first (standardized) factor from aset of indicators coding correct responses to simple factual questions regarding recent politicalevents and the partisanship of a respondent’s state governor.23 The average respondent answeredaround half the questions correctly. An advantage of this measure is that voters cannot over-reporttheir knowledge. Because these consumption measures do not differentiate changes in demand forinformation among voters from changes in the supply of political news provided by media outlets,the results capture equilibrium political information consumption.

4.2 Identification strategy

The identification strategy leverages the irregular timing of survey waves with respect to state-specific electoral calendars (akin to Eifert, Miguel and Posner 2010). The timing of ENCUPsurveys varies across both the number of years between survey rounds (and thus does not trackfederal elections) and the month within the year when the survey was conducted, while Mexicanstates have historically followed different electoral cycles that vary in both the month and year of

20The specific objectives of the study, which does not address elections at all, are enumerated here.21In 2012, 5 broad strata were identified, and electoral precincts and then voters were randomly selected from

within such strata to match the strata’s rural-urban, gender, and age distribution. In 2005 and 2012, 10 voters weresurveyed from each precinct according to specific directions (see the 2012 methodological manual here). Althoughsuch detailed sampling information is not available for the earlier surveys, the overall design is similar.

22I focus on radio and television, which are by far the most prevalent sources of political information in Mexico.Comparable questions from 2001 were not available.

23In 2001, 2003, 2005, and 2012 respectively, respondents were asked 3, 2, 2, and 2 topical questions (see Ap-pendix). Questions regarding basic knowledge of political institutions were excluded because they are unlikely to becovered directly in the news.

16

their elections.24 Of Mexico’s 32 states (including the Federal District), 27 register an election inone of the survey years, of which 14 only hold an election in a survey year but after the surveyoccurred. Provided that the surveys were not timed to coincide (or not) with particular elections,whether a survey was administered just before a state’s local elections is plausibly exogenous.

Campaigns generally last around 5 months, although there is some variation across states.25

Accordingly, I code an indicator for respondents facing an upcoming municipal election, and typi-cally a simultaneous state legislative election, within 5 months of the survey.26 I then estimate theeffect of an upcoming local election using the following regression:

Yimt = βUpcoming local electionmt + µt + εimt , (1)

where Yimt is a measure of political news consumption for respondent i in municipality m at sur-vey year t. Survey fixed effects, µt , control for differences in the difficulty of political knowledgequestions across surveys and common period effects that could arise from concurrent federal elec-tions (in 2003 and 2012), Presidential elections (in 2012), or national trends in political behavior.Throughout, standard errors are clustered by municipality.

Crucially, there is no evidence to suggest that the ENCUP surveys were strategically timed withrespect to local elections. First, Table A2 shows that upcoming local elections are well-balancedover individual and municipal level characteristics: only 1 of 18 tests reports a significant differ-ence at the 10% level. As a robustness check, I control flexibly for the imbalance over respondenteducation. Second, conditional on a municipality’s inclusion in at least one survey wave, TableA5 shows that neither upcoming local elections nor violence predict whether a municipality isincluded in any given survey round.

Although upcoming local elections are plausibly exogenous to the timing of the survey, I alsoimplement a difference-in-differences design (by further including state fixed effects, λs) to demon-strate robustness. This design instead relies on the parallel trends assumption that states facing andnot facing upcoming elections at the time of the survey follow similar trends in political news con-

24See Table A1 in the Appendix for a full list of municipal elections by month. In Chiapas, Coahuila, Guerrero,Michoacan, Quintana Roo, Veracruz, and Yucatan, the typical 3-year cycle was adjusted over the sample period byswitching to a 2 or 4-year term for a single electoral cycle. Moreover, following a constitutional amendment in 2007,states were subsequently mandated to hold local elections on the same day as federal elections when the state cyclecoincides. Consequently, states also changed the month of their elections. To reduce constant electoral competition,some states holding off-cycle elections also homogenized elections after the reform.

25Political advertising slots are specifically allocated for these purposes five months prior to federal elections (Lar-reguy, Marshall and Snyder forthcoming).

26Table A12 shows that the results are robust to defining upcoming elections by any number of months between 1and 12.

17

sumption. While this assumption is plausible in a context where all states repeatedly hold elections,this design substantially reducing the identifying variation: because half the states in the datasetnever experience an election within 5 months of an election, there is no within-state variation inthe treatment for half the sample.

4.3 Upcoming elections increase news consumption and political knowledge

The results in panel A-C of Table 1 demonstrate that local elections increase both the likelihoodthat a voter consumes any political information at all as well as the level of information consumed.Column (1) of panel A shows that an upcoming local election increases the probability that a voterlistens to or watches news at all by 5 percentage points. Only consuming news before electionsis likely to represent a conscious acquisition choice by voters that had previously avoided the rel-atively extensive news coverage also available outside election campaigns. Column (1) in panelB also reports a large increase in weekly news and political program consumption. The positiveeffect on the five-point scale in column (1) of panel C is also statistically significant. These find-ings suggest that voters with both lower and higher baseline levels of engagement consume morepolitical information prior to statewide elections.

Panel D shows that the cycle of increased information consumption prior to elections mirrorschanges in political knowledge. Contrary to potential concerns about social desirability bias, col-umn (1) demonstrates that voters facing an upcoming local election are more than a quarter ofa standard deviation—or 9 percentage points—more likely to correctly answer a question testingtheir political knowledge.27 This increase suggests that voters actively engage by internalizingpolitical information, a likely prerequisite for political information to influence electoral account-ability.

Column (2)-(8) demonstrate the robustness of such political information cycles at both theextensive and intensive margins of political news consumption. First, to address the imbalance overeducation, column (2) shows that the findings are qualitatively similar when including fixed effectsfor all five educational categories. This alters the sample because education was not measured in2005. Second, at notable efficiency cost, column (3) reports the difference-in-differences resultsincluding state fixed effects. Although this approach effectively drops observations from states thatdo not experience an upcoming election in any survey (more than half the sample), the results arerelatively stable and generally remain statistically significant in spite of the reduced precision inthis subset of states. Third, the results do not depend upon the particular definition of the upcoming

27Interacting upcoming local elections with baseline covariates for both consumption and knowledge outcomessuggests that the same voters consuming more news also become more knowledgeable about politics.

18

Table 1: The effect of upcoming local elections on political news consumption and topicalpolitical knowledge

(1) (2) (3) (4) (5) (6) (7) (8)

Panel A: Outcome: Watch and listen to news and political programs everUpcoming local election 0.049*** 0.033* 0.006

(0.014) (0.017) (0.018)Upcoming local election 0.093*** 0.079***

(two-month) (0.015) (0.015)Months until next election -0.002*** -0.001* -0.002***

(0.001) (0.001) (0.001)

Observations 13,030 8,330 13,030 13,030 8,330 13,030 8,330 13,030Outcome range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1}Outcome mean 0.87 0.86 0.87 0.87 0.86 0.87 0.86 0.87Treatment mean 0.19 0.24 0.19 0.02 0.03 17.53 18.88 17.53

Panel B: Outcome: Watch and listen to news and political programs weeklyUpcoming local election 0.082*** 0.053** 0.058*

(0.022) (0.024) (0.030)Upcoming local election 0.116*** 0.095***

(two-month) (0.038) (0.030)Months until next election -0.005*** -0.005*** -0.004***

(0.001) (0.001) (0.001)

Observations 13,030 8,330 13,030 13,030 8,330 13,030 8,330 13,030Outcome range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1}Outcome mean 0.63 0.63 0.63 0.63 0.63 0.63 0.63 0.63Treatment mean 0.19 0.24 0.19 0.02 0.03 17.53 18.88 17.53

Panel C: Outcome: Watch and listen to news and political programs scaleUpcoming local election 0.276*** 0.174** 0.158

(0.075) (0.079) (0.098)Upcoming local election 0.416*** 0.343***

(two-month) (0.122) (0.093)Months until next election -0.016*** -0.014*** -0.012***

(0.002) (0.003) (0.003)

Observations 13,030 8,330 13,030 13,030 8,330 13,030 8,330 13,030Outcome range {0,1,2,3,4} {0,1,2,3,4} {0,1,2,3,4} {0,1,2,3,4} {0,1,2,3,4} {0,1,2,3,4} {0,1,2,3,4} {0,1,2,3,4}Outcome mean 2.58 2.57 2.58 2.58 2.57 2.58 2.57 2.58Outcome standard deviation 1.47 1.49 1.47 1.47 1.49 1.47 1.49 1.47Treatment mean 0.19 0.24 0.19 0.02 0.03 17.53 18.88 17.53

Panel D: Outcome: Topical political knowledgeUpcoming local election 0.312*** 0.152*** 0.119***

(0.051) (0.036) (0.044)Upcoming local election 0.218** 0.188**

(two-month) (0.102) (0.080)Months until next election -0.006*** -0.003** -0.002

(0.002) (0.001) (0.001)

Observations 17,213 12,513 17,213 17,213 12,513 17,213 12,513 17,213Outcome range [-2.11,1.56] [-2.11,1.56] [-2.11,1.56] [-2.11,1.56] [-2.11,1.56] [-2.11,1.56] [-2.11,1.56] [-2.11,1.56]Outcome mean 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00Outcome std. dev. 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00Treatment mean 0.16 0.18 0.16 0.03 0.04 18.58 19.88 18.58

Municipalities 539 386 539 539 386 539 386 539Education fixed effects X X XState fixed effects X X

Notes: All specifications include survey-year fixed effects, and are estimated using OLS. The outcomes in panels A-C were not collected in

the 2001 survey, and education was not collected in 2005. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, **

denotes p < 0.05, *** denotes p < 0.01.

19

election indicator. Columns (4)-(5) instead define an upcoming local election as a survey withintwo months before election day. Although only 2% of observations (from 4 states) experience suchan election,28 the larger point estimates provide further confidence that voters indeed consumemost political information before elections. Adopting a linear approach, columns (6)-(8) similarlyconfirm that voters exhibit greater news consumption and political knowledge as the number ofmonths until the next local election decreases.

5 Local violence and electoral accountability

Given that voters consume significantly more news just before elections, signal of incumbent qual-ity covered in the news at this time may influence voting behavior in contexts where voters possessrelatively low baseline levels of political information. To test this, I exploit plausibly exogenousspikes in the occurrence of homicides—a widely-covered news event—before elections to identifythe electoral effects of short-term performance indicators coinciding with voter news consump-tion cycles. I contrast the effect of such pre-election homicide shocks with longer-term homiciderates—likely to be more accurate signals of performance, but less frequently observed due to votercycles in politically-relevant information consumption.

5.1 Data

To examine the electoral effects of violence in the run-up to elections I utilize two main sources ofdata. First, electoral returns for municipal, state, and federal elections covering Mexico’s c.67,000electoral precincts were assembled from the Federal Electoral Institute (IFE),29 state electoral in-stitutes, and freedom of information requests. I focus primarily on municipal elections between1999 and 2013, for which data is widely available across all states, while state and federal electoralreturns are used to parse accountability channels. Since municipal elections generally occur everythree years, the full dataset contains around four elections per municipality. Second, I combine theelectoral data with the INEGI’s monthly municipal homicide data (described in section 3.2).

The empirical analysis focuses on two measures of incumbent electoral performance: a municipal-level indicator for whether the incumbent party won the election, and the precinct-level change inthe incumbent party’s vote share.30 Reflecting their persisting state-level power (Langston 2003),

28Since only 4 states have within-state variation, state fixed effects are not added.29IFE recently became the National Electoral Institute (INE).30Table A13 shows similar results the incumbent vote share level as a robustness check. Given that the analysis

weights by the number of registered voters in each precinct and standard errors are clustered by municipality, theprecinct-level vote share estimates are almost identical to municipal-level estimates. Because some precincts are

20

53% of incumbent mayors in the sample are from the PRI; 29% are PAN and 12% are PRD. Asnoted above, voters generally sanction parties for the performance of their politicians in Mexico’sparty-centric system.31 The average incumbent party experiences a 5.7 percentage point decline intheir vote share, but still wins 54% of races. Turnout rates are typically around 60%. Municipali-ties without their own police forces, including the delegations of the Federal District, are excludedfrom the analysis because mayors have limited capacity to affect homicide rates.32

5.2 Identification strategy

While the national homicide rate has changed relatively smoothly over time, this masks consider-able inter-month volatility within municipalities. At the height of the War on Drugs in 2010, themonthly homicide count in Ciudad Juarez, Chihuahua, oscillated dramatically, e.g. from 394 inSeptember to 477 in October before falling to 242 in November and again rising to 309 in Decem-ber. Conversely, the median municipality experienced only one homicide over a year. More gen-erally, municipal homicides counts appear to fluctuate fairly random in the short-run: the month-to-month count is almost exactly as likely to increase as decrease—both in the run-up to electionsand in the middle of a municipal government’s tenure.33

To plausibly identify the electoral effects of homicides occurring when voters consume mostpolitically-relevant news, I exploit short-term deviations from trend in the number of homicidesaround local elections.34 Specifically, I compare municipalities experiencing a short-term spikein the number of homicides just before a local election, relative to just after that election, to oth-erwise similar municipalities experiencing at least as many homicides just after the election as

missing within municipalities, a precinct-level analysis allows more municipal-election observations.31In the 32% of cases where the incumbent won as part of a coalition formed by several parties, and given such

coalitions can vary across elections, if the coalition changes I define the incumbent as the party with the largest voteshare at the following election. In general, coalitions are dominated by large parties; the three main parties representmore than 90% of incumbents. Table 2 shows that the results are robust to restricting attention to incumbents containingthe three largest parties and single-party incumbents.

32Using the descriptions of municipal public security forces provided by the municipal governments in the 2000,2002, 2004, 2011, and 2013 National Census of Municipal Governments (ENGM) surveys, and imputing data formissing years, I exclude municipalities without a police force or relying on state, federal, community, private, or othersecurity forces. Given the difficulty of assigning responsibility, the few municipalities with inter-municipal and civilassociation-run police are excluded. Appendix A.3.1 describes the imputation procedure. Table A14 shows that themain findings are robust to including the approximately 100 municipalities excluded for lacking a police force.

33Across municipalities since 1999, the monthly homicide change was positive in 11.86% of cases and negative in11.79% of cases. In election months, these shares are 11.90% and 11.87%. In the month preceding an election, theshares are 11.87% and 11.86%. Most changes are zero because most municipalities do not experience a homicide.This ratio is identical when conditioning on a homicide in either the last month or the current month.

34A benefit of the unavailability of the number of homicides reported by broadcast media stations is that theempirical strategy is not susceptible to differential media reporting biases. Table A6 shows that delayed homicideregistration, a potential indicator of strategic registration and missing homicides, is balanced over homicide shocks.

21

before. Akin to Ferraz and Finan’s (2008) comparison of Brazilian municipalities, where federalaudit reports were randomly released just before and just after municipal elections, I ensure thatthe comparison group captures similar municipalities (in homicide levels as well as other charac-teristics) by focusing on municipalities experiencing at least one homicide around the election.35

Formally, a municipality m is defined as experiencing a pre-election homicide shock if, con-ditional on at least one homicide occurring over the four months spanning the penultimate monthbefore an election and the second month after an election, m experiences more homicides in thetwo months before an election in month τ than the two months after the election:

Homicide shockmτ :=

1 if Homicidesmτ−2 +Homicidesmτ−1 > Homicidesmτ +Homicidesmτ+1

andτ+1∑

τ ′=τ−2Homicidesmτ ′ > 0,

0 if Homicidesmτ−2 +Homicidesmτ−1 ≤ Homicidesmτ +Homicidesmτ+1

andτ+1∑

τ ′=τ−2Homicidesmτ ′ > 0.,

. ifτ+1∑

τ ′=τ−2Homicidesmτ ′ = 0.

(2)

Since most elections are held on the first Sunday of the month, this four-month window capturesthe final two months of the campaign and the post-election period before the winner enters office,while covering a sufficiently short span that month-to-month changes in the homicide count areplausibly random.36 On average, Figure 2a shows that a homicide shock equates to a statisticallysignificant increase of around 5 homicides per month in shocked municipalities relative to controlmunicipalities. After dropping the 30% of municipalities where no homicide is registered in eitherthe two months before or after the election, Figure A1b in the Appendix maps the final sample of1,335 mostly larger municipalities with their own police force. Of these, 906 experience at leastone election with and one election without a homicide shock between 1999 and 2013.

I estimate the effects of homicide shocks just prior to the election in precinct p of municipalitym in election year t using regressions of the form:

Ypmt = βHomicide shockmt +ηm + µt + εpmt , (3)

35Municipalities with no homicides around local elections also differ significantly from even municipalities ex-periencing only a single homicide over the 4-month window I focus on. Municipalities without a homicide are lessdeveloped, less politically competitive, and less violent.

36For the 9% of elections held on the 16th or later of a given month, I use months τ−1 and τ for the pre-electionhomicide count and τ + 1 and τ + 2 for the post-election count. Table 2 shows that the results are robust to droppingelections not held during the first half of the month.

22

where Ypmt is either a municipal-level indicator for the incumbent party winning the election orthe precinct-level change in the incumbent party’s vote share relative to the previous election.Although they are not required for identification, municipality and year fixed effects—respectively,ηm and µt—are included to increase precision and ensure that the estimates do not capture time-invariant municipality characteristics of common shocks. I show below that similar results obtainwithout these fixed effects. All observations are weighted by the number of registered voters.

This strategy for isolating exogenous shocks relative to longer-run trends is preferred to a de-sign examining changes in pre-election homicides rates across consecutive electionsThis is becauseit is hard to purge the correlation between short-term increases in the homicide rate and the longer-run homicide trends that I seek to differentiate pre-election shocks from. Nevertheless, Table A16reports broadly similar results using a difference-in-differences design to examine deviations be-tween pre-election homicide rate and the homicide rate over the entire preceding electoral cycle.

5.3 Evidence supporting the identification strategy

The key identifying assumption is that the timing of homicides around elections is effectively ran-dom. Although month-to-month changes in homicide rates appear idiosyncratic at first glance, thisdoes not necessarily imply that differences in homicide shocks across elections in a given munici-pality occurred by chance. This subsection validates the plausibility of the identifying assumption.

I start by showing that pre-election homicide shocks are uncorrelated with a wide variety ofobservable pre-treatment covariates. First, Appendix Table A6 examines 99 time-varying andtime-invariant covariates capturing demographic, socioeconomic, media coverage, and politicalfeatures of each municipality.37 Only 11 differences are statistically significant at the 10% level.38

Most importantly, homicide shocks are equally distributed across political variables including theincumbent’s previous vote share, the competitiveness of the previous election, and the incumbent’sparty identity, but also with respect to the number of media stations, the share that own a televisionor radio, and educational attainment.

Second, shocked municipalities experience indistinguishable homicide levels and trends fromnon-shocked municipalities before elections. Figure 2a shows that the difference in monthly homi-cides between treated and untreated precincts is both relatively constant over the 10 months preced-ing the period used to define homicide shocks and never statistically significant.39 In fact, almostall the difference reflects a single observation—the extremely high homicide rate in Ciudad Juarez

37Municipality fixed effects are excluded for the time-invariant 2010 Census variables. Table A7 shows analogoustests for precinct-level outcomes.

38The same number of significant imbalances arise if municipality and year fixed effects are excluded.39Fewer homicides in treated municipalities would downwardly bias estimates if voters sanction greater violence.

23

in 2010. The balance tests in Table A6 also confirm that the average number of homicides in the12, 6, and 3 months preceding the period defining the treatment do not significantly vary withhomicide shocks. The balance tests further show that, for the 2011-2013 period when disaggre-gated crime reports coincide with my sample, homicide shocks are not significantly correlated withpre-election spikes in other types of major crime. Contrary to concerns about strategic reporting, amismatch between the month in which a homicide occurred and was reported is equally rare acrossin shocked elections. Furthermore, I show below that a placebo shock defined 6 months prior tothe election does not influence electoral outcomes, and that the results are robust to controlling forpre-election monthly homicide counts.

Third, consistent with sampling variability, Figure 2b shows that the distributions of two-monthhomicide counts prior to the period defining a shock, just before elections, and just after electionsare similar. If homicide counts were to instead reflect strategic manipulation, we might insteadexpect to observe a spike in the density around 0 or spikes in the tail of the distributions just beforethe election. However, comparing the distribution of homicides 2 months prior to the electionand over prior 2-month periods, Kolmogorov-Smirnov tests fail to reject equality of distribution ineach case. Appendix Figure A2 further shows—both graphically and using Kolmogorov-Smirnovtests—that these distributions do not differ by homicide shock status.

I also find no evidence to support the more specific concern that DTOs alter the number ofhomicides around elections to signal their preferred electoral outcome (e.g. Alesina, Piccolo andPinotti 2016), potentially by colluding with local governments. Table A6 shows that pre-electionhomicide shocks are not significantly correlated with the number of drug-related homicides in theprior year, for the 2006(Dec.)-2011 period when monthly data was made publicly available bythe Mexican National Security System. Homicide shocks are also balanced across municipalitiesregistering more than one drug-related homicide in any pre-election year over this period and whereDTOs are present. At the precinct-level, Table A7 shows that shocks are no more likely to occur inthe 5% of precincts designated by the IFE as high-risk (typically locations with high DTO activity).In addition, Table 2 shows that the results are robust to removing municipalities with high levelsof drug-related homicides and states with high DTO presence.

Nevertheless, the types of homicide could change without affecting overall levels. Ganglandkillings are typically concentrated among young and uneducated men (Dell 2015), and are oftencommitted using firearms or more gruesome methods—particularly if intended as signals. Usingthe International Classification of Diseases codes in INEGI’s coroner reports, Table A6 also exam-ines the causes of death and victim characteristics associated with homicides occurring in the twomonths before an election. In particular, homicide shocks are balanced with respect to the share

24

Afterelection

Monthly before - afterdifference: 5.3***

Monthly before - afterdifference: 4.2***

-20

-15

-10

-50

510

Diff

eren

ce in

mon

thly

hom

icid

es

-10123456789101112Months until election

Full sample Excluding Ciudad Juárez 2010

Beforeelection

(a) Difference in the number of homicides betweenmunicipalities with and without a pre-election

homicide shock, by month until the election (95%confidence intervals)

0.0

5.1

.15

.2D

ensi

ty

0 5 10 15 20 25Average number of homicides per month

Months 9 and 10 before election Months 7 and 8 before election

Months 5 and 6 before election Months 3 and 4 before election

Months 1 and 2 before election Months 1 and 2 after election

(b) Distribution of homicides by relation to election

Figure 2: Distribution of homicides by homicide shock and over time

Notes: Each bar in Figure 2a denotes the difference in the number of homicides in treated (positive shock) andcontrol (negative shock) municipalities in the ten months prior to the period defining the homicide shock aroundthe election. Similarly, the before-after difference is the difference between in the monthly average number ofhomicides over the two month before and after the election. Each estimate is from a regression akin to the mainestimating equation (3). The few cases of more than 25 homicides per month are not shown in Figure 2b tofacilitate graphical representation.

of pre-election homicides caused by a firearm, hanging, or chemical substances. Of relatively fre-quent types of homicide, only cutting-related homicides were greater in homicide shock electionsbefore the election. Furthermore, these homicides did not disproportionately afflict young, male,unmarried, or uneducated individuals that are most likely to be involved with organized crime inshocked municipalities.

Another potential concern is that homicide counts change before elections because high-capacitygovernments can crack down on crime prior to elections to win votes. Although no expert onmunicipal politics that I interviewed believed that municipal governments crack down on localhomicides in the short-term, I examine this concern more systematically. First, proxies for statecapacity—including municipality size and budget, police per voter, alignment with governorsor the president, and neighbor homicide shocks—are uncorrelated with homicide shocks. Thissuggests no differential ability to engage in such crackdowns. Second, using Osorio’s (2015)newspaper-based measures of DTOs crackdowns, Table A6 further demonstrates that homicideshocks are not significantly correlated with violent enforcement, drug-related arrests, asset seizures,drug seizures, and gun seizures in the two months before the election.

25

Finally, defining homicide shocks using post-election homicides could introduce bias if electionoutcomes themselves affect post-election homicides counts. However, new mayors do not enteroffice within two months of the election. Furthermore, Table A8 shows that neither pre-electionpolitical variables nor election outcome variables predict homicide levels in the two months afteran election.40 Post-election homicides are also uncorrelated with interactions between electionoutcomes and background covariates. Nevertheless, as noted above, Table A16 reports similarresults examining pre-election deviations from trend in the homicide count based only on pre-election data.

5.4 Pre-election homicide shocks harm municipal incumbent parties

Table 2 shows that pre-election homicide shocks substantially harm the municipal incumbentparty’s electoral prospects. Column (1) of panel A reports that such homicide shocks reduce theincumbent party’s probability of being re-elected by 12 percentage points. This represents a 21%reduction in a mayor’s re-election probability relative to municipalities not experiencing a pre-election homicide shock. Column (1) of panel B reports that this decline reflects a 2.2 percentagepoint reduction in the incumbent party’s vote share. This equates to 0.45 percentage points for eachadditional homicide per month relative to the baseline levels in the treatment and control group.Table A11 indicates that the decline in incumbent vote share principally reflects greater turnout fornon-incumbent parties.

These sharp differences are robust to various checks. First, the results are not sensitive topossible violations of the identifying assumption. Column (2) shows that the results do not dependupon including municipality and year fixed effects, which also adds 429 elections to the samplefrom municipalities where a homicide shock is only defined for a single election. However, thegreatest identification threat reflects time-varying unobservables. To address possible concerns,column (3) of Table 2 first demonstrates robustness to including 32 time-varying municipal-levelcovariates.41 Column (4) next includes state-year fixed effects to control for election-specific state-level shocks. Columns (5) and (6) further include incumbent-specific (for the three largest parties)and municipality-specific linear trends to ensure that the results are respectively not driven bydifferential trends across different types of incumbent parties or municipalities. Finally, I conducta placebo test where a homicide shock is defined six months earlier, according to equation (3). Theresults in column (7) show that such pre-campaign shocks, which could be indicative of pre-trends,

40Dell (2015) similarly finds that homicide rates increased where PAN mayors were elected during the Calderonadministration, but not during the “lame duck” quarter before such mayors entered office.

41To preserve sample size, I the few exclude variables with greater than 95% missingness from this control set.

26

Tabl

e2:

Eff

ects

ofpr

e-el

ectio

nho

mic

ide

shoc

kson

mun

icip

alin

cum

bent

elec

tora

lout

com

es

Bas

elin

eN

ofix

edC

ontr

ols

Stat

e-In

cum

bent

Mun

icip

ality

Plac

ebo

Vio

lenc

eFe

wer

than

Few

Non

-O

ne-

Thr

ee-

Four

-Fi

rst

Mai

nSi

ngle

-sp

ec.

effe

cts

year

part

y-s

peci

fic6

mon

ths

bin

25ho

mic

ides

drug

DTO

mon

thm

onth

mon

thha

lfof

part

ypa

rty

effe

cts

tren

dstr

ends

earl

ier

effe

cts

perm

onth

hom

icid

esst

ates

shoc

ksh

ock

shoc

km

onth

incu

mbe

ntin

cum

bent

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

(13)

(14)

(15)

(16)

(17)

Pane

lA:O

utco

me:

Incu

mbe

ntpa

rty

win

Hom

icid

esh

ock

-0.1

22**

*-0

.088

**-0

.075

**-0

.100

***

-0.1

10**

*-0

.124

**-0

.129

***

-0.1

05**

*-0

.122

***

-0.1

11**

-0.1

16**

*-0

.123

***

-0.1

40**

(0.0

39)

(0.0

35)

(0.0

36)

(0.0

31)

(0.0

40)

(0.0

63)

(0.0

35)

(0.0

40)

(0.0

39)

(0.0

50)

(0.0

42)

(0.0

39)

(0.0

57)

Plac

ebo

hom

icid

e-0

.017

shoc

k(0

.053

)H

omic

ide

shoc

k-0

.047

(one

mon

th)

(0.0

47)

Hom

icid

esh

ock

-0.1

11**

*(t

hree

mon

th)

(0.0

42)

Hom

icid

esh

ock

-0.0

85**

(fou

rmon

th)

(0.0

41)

Obs

erva

tions

2,85

23,

281

2,68

92,

852

2,54

42,

852

2,40

82,

843

2,83

12,

852

1,89

51,

891

3,49

63,

977

2,62

22,

544

1,76

2U

niqu

em

unic

ipal

ities

906

1,33

586

590

682

390

682

090

690

590

661

663

71,

064

1,17

285

482

364

1O

utco

me

rang

e{0

,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

Out

com

em

ean

0.55

0.55

0.55

0.55

0.55

0.55

0.55

0.55

0.54

0.55

0.54

0.55

0.54

0.54

0.55

0.55

0.55

Hom

icid

esh

ock

mea

n0.

410.

420.

410.

410.

420.

450.

410.

410.

420.

410.

410.

480.

430.

420.

410.

420.

39

Pane

lB:O

utco

me:

Cha

nge

inin

cum

bent

part

yvo

tesh

are

Hom

icid

esh

ock

-0.0

22**

*-0

.018

**-0

.015

**-0

.018

***

-0.0

15*

-0.0

29**

*-0

.027

***

-0.0

22**

-0.0

22**

*-0

.026

**-0

.019

**-0

.022

***

-0.0

34**

*(0

.009

)(0

.007

)(0

.006

)(0

.007

)(0

.008

)(0

.011

)(0

.009

)(0

.009

)(0

.009

)(0

.010

)(0

.009

)(0

.008

)(0

.012

)Pl

aceb

oho

mic

ide

-0.0

05sh

ock

(0.0

12)

Hom

icid

esh

ock

-0.0

15(o

nem

onth

)(0

.011

)H

omic

ide

shoc

k-0

.024

***

(thr

eem

onth

)(0

.009

)H

omic

ide

shoc

k-0

.026

***

(fou

rmon

th)

(0.0

09)

Obs

erva

tions

167,

116

173,

115

162,

774

167,

116

157,

639

167,

116

145,

511

167,

116

155,

267

167,

116

102,

126

143,

840

180,

302

188,

067

152,

539

157,

639

104,

220

Uni

que

mun

icip

aliti

es90

61,

335

895

906

823

906

820

906

905

906

616

637

1,06

41,

172

854

823

641

Out

com

era

nge

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.87]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

Out

com

em

ean

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

6-0

.05

-0.0

5-0

.06

-0.0

5-0

.05

-0.0

4O

utco

me

std.

dev.

0.14

0.14

0.14

0.14

0.14

0.14

0.14

0.14

0.15

0.14

0.14

0.14

0.15

0.15

0.14

0.14

0.15

Hom

icid

esh

ock

mea

n0.

410.

420.

410.

410.

420.

470.

410.

410.

420.

410.

410.

480.

430.

420.

410.

420.

39

Not

es:

Col

umns

(2)i

nclu

des

the

cont

rols

inTa

ble

A6,

with

the

exce

ptio

nof

vari

able

sw

ithgr

eate

rtha

n5%

mis

sing

ness

.C

olum

n(3

)exc

lude

sm

unic

ipal

ityan

dye

arfix

edef

fect

s,an

dth

us

incl

udes

mun

icip

aliti

es42

9m

unic

ipal

ities

whe

rea

hom

icid

esh

ock

ison

lyde

fined

for

one

elec

tion

whe

reel

ecto

rald

ata

isav

aila

ble.

Col

umn

(4)

incl

udes

stat

e-ye

arfix

edef

fect

s.C

olum

n

(5)i

nclu

des

part

y-sp

ecifi

cin

cum

bent

year

tren

ds.C

olum

n(6

)inc

lude

sm

unic

ipal

ity-s

peci

ficye

artr

ends

.Col

umn

(7)u

ses

apl

aceb

oho

mic

ide

shoc

kde

fined

six

mon

ths

befo

reth

eel

ectio

n.

Col

umn

(8)i

nclu

des

fixed

effe

cts

fort

heto

taln

umbe

rofh

omic

ides

over

the

four

-mon

thw

indo

win

bins

ofsi

zete

n.C

olum

n(9

)exc

lude

sm

unic

ipal

ities

aver

agin

gm

ore

than

25ho

mic

ides

perm

onth

over

the

two

mon

ths

eith

erbe

fore

oraf

tert

heel

ectio

n.C

olum

n(1

0)ex

clud

esm

unic

ipal

ities

that

aver

age

mor

eth

anon

edr

ug-r

elat

edho

mic

ide

am

onth

over

the

12m

onth

sbe

fore

anel

ectio

nbe

twee

n20

06an

d20

11.C

olum

n(1

1)ex

clud

esst

ates

with

high

-lev

elof

DTO

activ

ity(s

eefo

otno

te43

).C

olum

ns(1

2)-(

14)r

espe

ctiv

ely

defin

eho

mic

ide

shoc

ksov

eron

e-,t

hree

-,

and

four

-mon

thw

indo

ws.

Col

umn

(15)

incl

udes

only

elec

tions

that

take

plac

eon

orbe

fore

the

15th

ofth

em

onth

.Col

umn

(16)

incl

udes

only

elec

tions

whe

reth

ePA

N,P

RD

,orP

RIa

rean

incu

mbe

nt.C

olum

n(1

7)in

clud

eson

lyob

serv

atio

nsw

here

the

incu

mbe

ntis

nota

coal

ition

.All

spec

ifica

tions

are

estim

ated

usin

gO

LS,

and

allo

bser

vatio

nsar

ew

eigh

ted

byth

enu

mbe

rof

regi

ster

edvo

ters

inth

eel

ecto

ralp

reci

nct.

Stan

dard

erro

rscl

uste

red

bym

unic

ipal

ityar

ein

pare

nthe

ses.

*de

note

sp<

0.1,

**de

note

sp<

0.05

,***

deno

tes

p<

0.01

.

27

do not affect incumbent electoral performance.Second, the results do not reflect municipalities experiencing particular levels or types of homi-

cide. By including fixed effects for the total number of homicides over the four-month windowused to define homicide shocks (in 10-homicide bins), column (8) provides further evidence thatthe results are not driven by differences in homicide levels across shocked and non-shocked munic-ipalities.42 Moreover, column (9) excludes the 21 municipality-elections experiencing more than25 homicides per month before or after the election to demonstrate that the results do not simplyreflect a few large municipalities with many homicides. Although homicide shocks are uncorre-lated with indicators of DTO presence, to further address the concern that the results are drivenby strategic political violence in areas with high drug-related criminal activity, I exclude the elec-tions most vulnerable to strategic political violence by DTOs. Specifically, column (10) excludesmunicipalities that average more than one drug-related murder a month over any pre-election yearbetween 2006 and 2011 (when such data were registered), while column (11) excludes 9 stateswith high DTO-related drug crime.43 In neither case do the estimates substantially change.

Third, the results are not driven by particular measures of homicide shock or electoral outcome.First, columns (12)-(14) show that homicide shocks defined by one-, three-, or four-month compar-isons also reduce the incumbent’s probability of winning by 5-11 percentage points and vote shareby around 2 percentage points. The smaller effects associated with the one-month comparison mayreflect the weaker signal imparted by shocks defined by fewer homicides. Column (15) shows thatthe results are robust to removing elections that occurred in the second half of the month. Theseelections are more challenging to classify with respect to homicides occurring before the electionbecause they are more scattered within the month. Column (16) restricts attention to PAN, PRD,and PRI incumbents, demonstrating that the results are not driven by the few mayors from smallerparties. Consistent with a clearer assignment of responsibility, column (17) reports slightly largereffects, especially on vote share, when restricting attention to single-party incumbents.

Together, these results robustly show that a temporary increase in local homicides can havesubstantial electoral implications for an incumbent when the spike coincides with the pre-electionperiod when voters consume more political information.

42Bins of size 1 and 5 similarly yield significant effects, at the cost of fully-explaining municipality-elections witha unique large number of homicides.

43Baja California, Chihuahua, Durango, Guerrero, Michoacan, Nuevo Leon, Sinaloa, Sonora, and Tamaulipas.

28

5.5 No effect of longer-run homicide rates

The political information cycles argument implies that voters are particularly likely to respond tohomicides that occur just before elections. In contrast, longer-run homicide metrics may not affectvoters’ responsibility attribution if, as already shown, many voters are insufficiently engaged atother times in the political cycle to learn about more systematic trends in homicide levels.

To examine whether homicide shocks before elections indeed induce greater incumbent sanc-tioning than long-run homicide rates, I estimate the effects of homicides over the prior year andprior (three-year) electoral cycle on the incumbent party’s electoral prospects in a difference-in-differences framework. I thus compare changes in support for incumbent parties in municipalitiesthat experienced large increases in their longer-run homicide rate between elections to changes insupport for incumbent parties in municipalities that did not using the following specification:

Ypmt = βAverage monthly homicide ratemt +ηm + µt + εpmt . (4)

As a robustness check, municipality-specific time trends are included to account for differentialtrends in incumbent support across municipalities experiencing different homicide rates.

The results in Table 3 indicate that long-run homicide trends have had limited impact on elec-toral outcomes. The point estimates in columns (1), (2), (5), and (6) of panel A show that homicidesover the year before the election are not significantly correlated with either the incumbent’s prob-ability of being re-elected or their vote share, regardless of the inclusion of municipality-specifictime trends. In fact, the estimates in columns (5) imply that 5 more homicides a month overthe year before an election—almost one a third of a standard deviation increase—only translatesinto a 0.15 percentage point decline in the municipal incumbent party’s vote share. This effect isaround 15 times smaller than the five-homicide monthly average differential between shocked andnon-shocked municipalities just before elections. Columns (3), (4), (7), and (8) include a quadraticterm, but also find little suggestion that any effect is non-linear. Panel B similarly shows that votersdo not sanction homicide rates averaged across the full electoral cycle. The evidence thus suggeststhat longer-run homicide trends play comparatively little role in informing electoral accountability.

5.6 Voters distinguish responsibility across levels of government

Although mayors play a role in local public security, state and federal governments are also impor-tant players. The punishment of municipal mayors could then reflect broader punishment of theparty controlling higher office. Conversely, if voters believe that mayors are primarily responsible

29

Tabl

e3:

Eff

ects

oflo

nger

-run

mun

icip

alho

mic

ide

rate

son

mun

icip

alin

cum

bent

elec

tora

lout

com

es

Incu

mbe

ntpa

rty

win

Cha

nge

inin

cum

bent

part

yvo

tesh

are

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Pane

lA:E

ffec

tofh

omic

ides

12m

onth

sbef

ore

elec

tion

Ave

rage

mon

thly

hom

icid

era

te0.

0009

-0.0

031

-0.0

009

-0.0

005

-0.0

003

-0.0

006

-0.0

014

-0.0

023

(12

mon

ths

befo

reel

ectio

n)(0

.001

6)(0

.001

9)(0

.006

2)(0

.008

0)(0

.000

3)(0

.000

6)(0

.001

2)(0

.002

3)A

vera

gem

onth

lyho

mic

ide

rate

0.00

00-0

.000

00.

0000

0.00

00(1

2m

onth

sbe

fore

elec

tion)

squa

red

(0.0

000)

(0.0

000)

(0.0

000)

(0.0

000)

Obs

erva

tions

2,85

22,

852

2,85

22,

852

167,

116

167,

116

167,

116

167,

116

Uni

que

mun

icip

aliti

es90

690

690

690

690

690

690

690

6O

utco

me

rang

e{0

,1}

{0,1}

{0,1}

{0,1}

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

Out

com

em

ean

0.55

0.55

0.55

0.55

-0.0

5-0

.05

-0.0

5-0

.05

Out

com

est

anda

rdde

viat

ion

0.50

0.50

0.50

0.50

0.14

0.14

0.14

0.14

Hom

icid

era

tem

ean

8.42

8.42

8.42

8.42

8.42

8.42

8.42

8.42

Hom

icid

era

test

anda

rdde

viat

ion

22.2

822

.28

22.2

822

.28

22.3

922

.39

22.3

922

.39

Mun

icip

ality

-spe

cific

time

tren

dsX

XX

X

Pane

lB:E

ffec

tofh

omic

ides

sinc

ela

stel

ectio

nA

vera

gem

onth

lyho

mic

ide

rate

0.00

27-0

.002

80.

0039

0.00

75-0

.000

2-0

.000

0-0

.000

4-0

.000

1(3

year

sbe

fore

elec

tion)

(0.0

019)

(0.0

040)

(0.0

084)

(0.0

147)

(0.0

003)

(0.0

008)

(0.0

014)

(0.0

033)

Ave

rage

mon

thly

hom

icid

era

te-0

.000

0-0

.000

10.

0000

0.00

00(3

year

sbe

fore

elec

tion)

squa

red

(0.0

000)

(0.0

001)

(0.0

000)

(0.0

000)

Obs

erva

tions

2,85

22,

852

2,85

22,

852

167,

116

167,

116

167,

116

167,

116

Uni

que

mun

icip

aliti

es90

690

690

690

690

690

690

690

6O

utco

me

rang

e{0

,1}

{0,1}

{0,1}

{0,1}

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

[-0.

96,0

.89]

Out

com

em

ean

0.55

0.55

0.55

0.55

-0.0

5-0

.05

-0.0

5-0

.05

Out

com

est

anda

rdde

viat

ion

0.50

0.50

0.50

0.50

0.14

0.14

0.14

0.14

Hom

icid

era

tem

ean

7.20

7.20

7.20

7.20

7.18

7.18

7.18

7.18

Hom

icid

era

test

anda

rdde

viat

ion

15.2

915

.29

15.2

915

.29

15.3

215

.32

15.3

215

.32

Mun

icip

ality

-spe

cific

time

tren

dsX

XX

X

Not

es:

All

spec

ifica

tions

are

estim

ated

usin

gO

LS,

and

incl

ude

mun

icip

ality

and

year

fixed

effe

cts.

All

obse

rvat

ions

are

wei

ghte

dby

the

num

ber

ofre

gist

ered

vote

rsfo

rth

ere

leva

ntel

ecto

ralu

nit.

Stan

dard

erro

rscl

uste

red

bym

unic

ipal

ityar

ein

pare

nthe

ses.

*de

note

sp<

0.1,

**de

note

sp<

0.05

,**

*de

note

sp<

0.01

.

30

for local crime (or that their actions are only weakly correlated with co-partisans at higher levels),they may update less about national parties following local homicide shocks. I next disentanglethese chains of accountability and responsibility assignment.

If voters believe that the mayor is responsible for local crime because they control the localpolice, pre-election homicide shocks should only be punished in municipalities with their ownpolice force. To test this, I add municipalities (and delegations in the Federal District) withouttheir own police force to the sample. Consistent with voters recognizing that elevated homicidesrates may be beyond the control of municipal mayors without local police forces, panel A of Table4 shows that homicide shocks only significantly harm the electoral prospects of incumbent partiescommanding a local police force. Columns (2) and (4) address the concern that the effect attributedto police forces reflects the lack of sanctioning in smaller and less developed municipalities bycontrolling for the interaction of a homicide shock with previous incumbent party vote share andthe average monthly homicide rate over the last year.44 Contrary to the naive perspective on votingbehavior, this finding suggests that voters are not indiscriminately punishing adverse events.

Nevertheless, if the parties of municipal mayors, state deputies, state Governors, or the Pres-ident are correlated, the substantial electoral penalties documented at the municipal level couldinstead reflect punishment of higher political actors. The results in columns (1)-(3) of Table 5suggest that, if anything, homicide shocks increase the municipal vote share of the state deputy’s,governor’s, and president’s parties’, and thus provide no support for punishment at higher levels ofgovernment spilling over to municipal elections.45 Alternatively, voters could restrict punishmentfor state and federal politicians in their corresponding races. However, columns (4) and (5) indi-cate that a homicide shock reduces neither the vote share of state nor federal deputies in concurrentstate and federal elections.

Although there is no evidence to suggest that state and federal incumbent parties are heldresponsible for homicide shocks by voters, the punishment of the local party could also spill upto the national level. However, when municipal and higher-level elections are held concurrently,columns (6) and (7) of Table 5 show that the state and federal deputy vote shares of the municipalincumbent’s party are not significantly affected by a homicide shock. Table A17 further shows thatvoters do not differentially punish particular parties for homicide shocks.

44Table A20 provides mixed evidence that voters account for homicide shocks affecting neighboring municipalities.45Consistent with Table A17 in the Appendix, the positive effect at the federal level reflects increased support for

Calderon’s PAN government.

31

Table 4: Electoral effects of homicide shocks, by presence of municipal police force

Incumbent Change in incumbentparty win party vote share

(1) (2) (3) (4)

Homicide shock -0.122*** -0.130*** -0.023*** -0.024***(0.037) (0.038) (0.008) (0.007)

No municipal police force 0.002 0.023 0.006 -0.005(0.099) (0.098) (0.016) (0.014)

Homicide shock 0.236* 0.234** 0.065*** 0.050**× No municipal police force (0.124) (0.116) (0.023) (0.022)

Observations 3,398 3,360 195,119 195,119Unique municipalities 1,026 1,019 1,026 1,026Outcome range {0,1} {0,1} [-0.96,0.89] [-0.96,0.89]Outcome mean 0.58 0.58 -0.04 -0.04Homicide shock mean 0.41 0.41 0.41 0.41Interaction mean 0.14 0.14 0.14 0.14Test: Homicide shock + Homicide shock × 0.32 0.33 0.05 0.21

No municipal police force = 0 (p value)Interactive controls X X

Notes: All specifications include municipal and year fixed effects, weight by the number of registered voters, andare estimated using OLS. The interactive controls are lagged incumbent party vote share and average monthlyhomicides over the last year (both standardized, and × Homicide shock). Standard errors clustered by municipal-ity are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

6 The amplifying effect of local media

The individual- and aggregate-level findings suggest that the electoral sanctioning of local homi-cides principally reflects increased voter news consumption before elections. This implication ofpolitical information cycles relies on voters having access to locally-based media likely to report onlocal events (e.g. Snyder and Stromberg 2010). Given that media sources may bias their reportingof news in different ways (e.g. Besley and Prat 2006; Gentzkow and Shapiro 2006; Mullainathanand Shleifer 2005) and market segmentation may mean that not all types of station reach all voters(Barabas and Jerit 2009; Prat and Stromberg 2005), the likelihood that voters consume relevant lo-cal news—and thus hold municipal incumbents to account—is likely to increase with the numberof local media outlets voters have access to. To examine the extent to which sanctioning of homi-cide shocks reflects such media coverage, I combine pre-election homicide shocks with plausiblyexogenous precinct-level variation in media signal coverage to estimate the effect of an additionallocal radio or television station likely to report on such shocks.

32

Table 5: Electoral effects of homicide shocks, by level of government

Change in municipal vote share of... Change in Change in Change in Change in...state ...state ...President’s state deputy federal state vote of federal vote ofdistrict Governor’s party incumbent incumbent municipal municipalparty party vote share vote share incumbent incumbent(1) (2) (3) (4) (5) (6) (7)

Homicide shock 0.001 0.029** 0.032* -0.004 -0.005 0.008 -0.017(0.009) (0.015) (0.018) (0.009) (0.015) (0.016) (0.017)

Observations 143,226 135,849 167,116 135,153 52,661 128,109 62,450Unique municipalites 892 845 906 895 391 875 413Outcome mean 0.00 0.03 0.03 -0.04 -0.05 0.45 0.38Outcome std. dev. 0.13 0.17 0.24 0.14 0.13 0.17 0.13Homicide shock mean 0.42 0.42 0.41 0.41 0.42 0.42 0.43

Notes: All specifications include municipal and year fixed effects, weight by the number of registered voters, and are estimated using OLS.

Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

6.1 Data

As part of Mexico’s major media reform in 2007 (see Larreguy, Marshall and Snyder forthcoming;Serra 2012), the IFE required that all radio and television stations in the country submit detailedcoverage and technological data. This included the location and power of their antennae and a mapdetailing their commercial quality coverage—the level of coverage that media stations may baseadvertising sales on and which the IFE deemed relevant for minimizing cross-state advertisingspillovers. The signal inside the commercial quality coverage area is very strong, and shouldcover virtually all households. However, signal quality declines quickly as the distance from thecommercial quality coverage boundary increases.46

Figure 3 maps the commercial quality coverage of each type of media station. While FM radio,and especially television, stations have limited and primarily urban coverage, AM radio stationscan travel considerable distances—particularly when aided by stretches of sea water with highelectrical ground conductivity. Virtually all of the population is covered by at least one media sta-tion, but the number of outlets—both emitting from within a municipality and without—providingcommercial quality signals to any given electoral precinct varies considerably. Following Larreguy,Marshall and Snyder (2018), I define an electoral precinct as covered by a given media station if atleast 20% of voters fall within the commercial coverage boundary.47

46The IFE defines the boundary of the coverage area using a 60 dBµ threshold for signal strength. According to theU.S.-based National Communications and Information Administration, this “60 dBµ level is recognized as the area inwhich a reliable signal can be received using an ordinary radio receiver and antenna.” This is the threshold commonlyused to determine a radio station’s audience and sell advertising space commercially in the US.

47The INEGI provides detailed Census population counts in 2010, for both rural localities and urban blocks. This

33

(a) AM radio stations (b) FM radio stations (c) Television stations

Figure 3: Media station commercial quality signal coverage areas

6.2 Identification strategy

To estimate the electoral effects of an additional media station covering local homicides, I leveragegeographic variation in coverage on the intensive margin (see also Larreguy, Marshall and Snyder2018; Spenkuch and Toniatti 2016). I compare neighboring electoral precincts within the samemunicipality that differ in the total number of local AM, FM, and television stations—defined asoutlets whose antennae are located within the electoral precinct’s municipality and that providelocal content48—by which they are covered.49 Since signal quality rapidly, but not entirely, de-clines across the commercial quality boundary, this strategy identifies an “intent to treat” effect ofincreasing the probability of exposure to an additional local media station.50 Figure 4 illustratesthe identification strategy graphically in case of a local television station, using the example ofelectoral precincts 1571 and 1583 from the same municipality in Oaxaca. The identifying assump-tion is that these neighboring precincts differ only because precinct 1583 receives a commercialquality signal from a local television station that does not cover precinct 1571.

More generally, for each electoral precinct, I identify the set of neighboring precincts fromwithin the same municipality that differ in the number of local media stations that they are coveredby. Each such grouping n is defined by a single “treated” precinct and the set of “control” neigh-boring precincts that receive different local media exposure. Focusing on municipalities includedin the homicide shock sample produced 2,689 neighboring groups, containing an average of 2.3

data is used to identify the proportion of voters within a precinct covered by a commercial quality signal.48Local content excludes television stations solely retransmitting Televisa and TV Azteca national broadcasts.49Although data does not exist to adjust for news consumption “non-compliance,” any effect would be larger among

compliers that only receive news because they were exposed to an additional commercial quality local signal.50Exposure to commercial quality coverage does not constitute a geographic regression discontinuity design be-

cause the treatment is not binary (some neighbors differ by more one media station) and where neighbors differ bymore than one media station it is not clear how to coherently define the running variable.

34

Figure 4: Identification strategy example

Notes: Both precincts are located in the municipality of Villa de Tututepec de Melchor Ocampo in thestate of Oaxaca. Precinct 1583 is covered by the local television emitting from within the municipality,but precinct 1571 is not.

comparison units per election. The average precinct in this sample is covered by 6.3, 8.9, and3.2 local AM, FM, and television stations respectively. The total number of local media stationscovering a precinct ranges from 0 to 40.51

There are good reasons to believe that, among neighboring electoral precincts, local media cov-erage at the commercial quality boundary is exogenous. First, neighboring precincts often differ incoverage because of physical characteristics such as water, geographic contours, and large objectsthat aid or impede ground conductivity (in the case of AM radio) and “line of sight” radio waves(in the case of FM radio and television) between an antenna and precincts at the coverage bound-ary (Larreguy, Marshall and Snyder 2018; Olken 2009). Second, given that media stations lack thetechnology to differentiate media markets at fine-grained levels,52 and voters who locate accordingto the availability of local media are unlikely to choose to live close to a coverage boundary (pre-ferring a location guaranteeing high-quality coverage), it is unlikely that coverage reflects strategicsorting by either media stations or voters around coverage boundaries. Finally, I restrict attentionto neighboring precincts with an area of less than 2km2. This removes larger precincts where me-dia outlets could more plausibly target their signals, and prevents stark comparisons between urbanand suburban, or suburban and rural, precincts that may respond differently to homicide shocks forother reasons. The final sample is shown in Figure A1c.

51Given that data from the Secretariat of Communications and Transportation show that the number of mediastations has barely changed since 2003, I continue to pool the years 1999 to 2013.

52The IFE data show that the power of signal transmitters are fairly discrete. The power output in watts for AM,FM, and television stations is almost exclusively round thousands that are divisible by five.

35

Combining this within-neighbor variation in local media coverage with the homicide shocksleveraged above, I use the following specification to estimate the interaction between pre-electionhomicide spikes and local media coverage:

Ypnmt = β1Homicide shockmt +β2Local mediap

+β3

(Homicide shockmt×Local mediap

)+ ζn + µt + εpnmt , (5)

where ζn is a fixed effect for each set of neighboring precincts.53 To weight comparisons withineach neighbor-year set equally, each “treated” precinct is weighted by electorate size and “control”precincts are weighted by electorate size divided by the number of controls per neighbor-year set.

To assess the claim that differences in local media coverage between neighboring precinctsreflect plausibly exogenous topological features, I use equation (5) to examine balance across de-mographic, socioeconomic, homicide, and political municipal and precinct characteristics.54 TableA9 shows that only 8 of these 92 variables are significantly correlated with the number of local me-dia stations at the 10% level. In particular, there are no significant differences in key indicators ofrural-urban geography (such as precinct area, electorate size, or distance to the municipality head),the number of non-local media stations, homicide indicators, or previous electoral outcomes. Ishow below that the results are robust to controlling interactively for the significant imbalances.

6.3 Local media increase punishment of homicide shocks

Column (2) of Table 6 shows that access to local media stations plays a significant role in support-ing the electoral sanctioning of homicide shocks in this sample of predominantly urban precincts.55

Indicating that local media are necessary for Mexican voters to punish the incumbent party forpre-election homicides, the null lower-order effect of a homicide shock shows that pre-electionhomicide spikes do not affect electoral outcomes in precincts without access to local media. How-ever, the significant negative interaction coefficient implies that each additional local media stationreduces the vote share of incumbents experiencing a pre-election homicide shock by 0.2 percent-age points. For more than eight local media stations (79% of the sample), Figure 5 shows thatthe overall effect of a homicide shock is to substantially reduce the municipal incumbent party’svote share.56 The positive coefficient on the lower-order local media term suggests that voters may

53Since I use within-municipality neighbors, neighbor fixed effects incorporate municipality fixed effects.54Table A10 shows that homicide shocks remain well-balanced across pre-treatment variables in this subsample.55Column (1) shows that the effect of homicide shocks is larger in the local media subsample.56Since the identification strategy relies on within-neighbor variation, this represents a linear extrapolation of the

average marginal effect.

36

-.15

-.1-.0

50

.05

Mar

gina

l effe

ct o

f a h

omic

ide

shoc

k

0 10 20 30 40Number of local media stations

Figure 5: Electoral effect of a homicide shock, conditional on the number of local media stations(95% confidence interval)

Notes: Calculated using the estimates from column (1) of Table 6. The gray density above the x axisplots the distribution of the local media variable in the sample.

also reward incumbents that do not experience a homicide shock before the election. The naturallogarithm of the number of local media stations used in column (3) indicates that each additionallocal media station alters the incumbent party’s vote share by approximately 3%.

The findings are robust to various potential threats to identification. First, column (4) interac-tively controls for the number of non-local media outlets covering a precinct to show that precinctswith greater access to local media do not simply respond more to homicide shocks because theypossess greater access to media in general. Second, column (5) similarly demonstrates that theresults are robust to controlling interactively for standardized version of the 8 variables with sig-nificant imbalances over local media. Third, to ensure that distance from radio antennae—and thusthe location of precincts within municipalities more generally—is not driving the results, column(6) restricts the sample to comparison precincts that are no more than 50 meters nearer to, or furtherfrom, the average local media station of each neighboring group’s “treated” precinct. The resultsof this Olken (2009)-type approach provide similar estimates in this smaller sample.

Table 6 also shows that the results do not depend on particular design choices. Columns (7)-(9)first demonstrate that the estimates are not sensitive to the 20% cutoff used to define whether a localmedia station covers a precinct. Columns (10)-(12) show that the results are robust to consideringtighter (precincts with an area of 1km or less) or more generous restrictions (5km or less or no arearestriction) on the maximum precinct area permitted.

Moreover, if the results indeed reflect local media covering homicide shocks that occur just

37

Table 6: Effect of homicide shocks on precinct-level change in incumbent party vote share,moderated by access to local and non-local media

Outcome: Change in incumbent party vote share

No local Baseline Logarithmic Interactive Interactive Average 10% 50% 100%media local transformation non-local imbalanced control media precinct precinct precinct

media media control controls within 50m coverage coverage coverage(1) (2) (3) (4) (5) (6) (7) (8) (9)

Homicide shock -0.0563*** -0.0148 0.0220 -0.0114 -0.0143 -0.0143 -0.0222 -0.0095 -0.0009(0.0153) (0.0249) (0.0364) (0.0300) (0.0234) (0.0240) (0.0248) (0.0249) (0.0176)

Local media 0.0017* 0.0017* 0.0016* -0.0001 0.0013 0.0007 0.0017**(0.0009) (0.0009) (0.0009) (0.0011) (0.0008) (0.0012) (0.0009)

Homicide shock × Local media -0.0023** -0.0024** -0.0024** -0.0023** -0.0020** -0.0023** -0.0027***(0.0010) (0.0010) (0.0009) (0.0010) (0.0009) (0.0009) (0.0007)

Local media (log) 0.0249***(0.0089)

Homicide shock × Local media (log) -0.0291**(0.0114)

Non-local media -0.0002(0.0009)

Homicide shock × Non-local media -0.0001(0.0008)

Observations 29,736 29,736 29,736 29,736 29,736 11,606 29,533 33,056 51,288Unique municipalities 118 118 118 118 118 115 116 123 198Outcome range [-0.63,0.50] [-0.63,0.50] [-0.63,0.50] [-0.63,0.50] [-0.63,0.50] [-0.63,0.50] [-0.63,0.50] [-0.63,0.50] [-0.64,0.61]Outcome mean -0.04 -0.04 -0.04 -0.04 -0.04 -0.04 -0.04 -0.04 -0.05Outcome standard deviation 0.13 0.13 0.13 0.13 0.13 0.13 0.13 0.13 0.13Homicide shock mean 0.40 0.40 0.40 0.40 0.40 0.40 0.41 0.39 0.42Media measure media mean 18.69 18.69 18.69 18.69 18.76 19.93 20.49 18.49Media measure standard deviation 10.30 10.30 10.30 10.30 10.32 11.14 10.84 11.25

1km area 5km area No area Longer-run Non-local AM ad FM TV adrestriction restriction restriction homicides media incumbent incumbent incumbent

share share share(10) (11) (12) (13) (14) (15) (16) (17)

Homicide shock -0.0169 -0.0101 0.0126 -0.0095 -0.0054 -0.0174 -0.0080(0.0299) (0.0210) (0.0142) (0.0184) (0.0826) (0.0521) (0.0441)

Local media 0.0016 0.0021*** 0.0012* 0.0010(0.0010) (0.0007) (0.0007) (0.0009)

Homicide shock × Local media -0.0023* -0.0026*** -0.0027***(0.0012) (0.0009) (0.0008)

Average monthly homicide rate 0.0024(3 years before election) (0.0018)

Average monthly homicide rate (3 years -0.0001before election) × Local media (0.0001)

Non-local media 0.0000(0.0004)

Homicide shock × Non-local media -0.0005(0.0004)

Incumbent ad share -0.1320 -0.0168 -0.0843(0.2326) (0.1341) (0.1130)

Homicide shock × Incumbent ad share -0.0637 -0.0168 -0.0527(0.3018) (0.1923) (0.1594)

Observations 25,297 35,323 101,713 29,736 92,596 33,643 33,643 33,643Unique municipalities 91 139 375 118 383 379 379 379Outcome range [-0.63,0.47] [-0.63,0.53] [-0.90,0.89] [-0.63,0.50] [-0.75,0.73] [-0.77,0.87] [-0.77,0.87] [-0.77,0.87]Outcome mean -0.04 -0.04 -0.05 -0.04 -0.05 -0.07 -0.07 -0.07Outcome standard deviation 0.13 0.13 0.15 0.13 0.14 0.13 0.13 0.13Homicide shock mean 0.39 0.41 0.45 15.31 0.46 0.43 0.43 0.43Media measure media mean 19.69 17.90 13.23 18.69 8.89 0.27 0.27 0.26Media measure standard deviation 9.99 10.48 10.78 10.30 9.56 0.06 0.08 0.08

Notes: All specifications include neighbor group and year fixed effects, and are estimated using OLS. All observations are weighted by the

number of registered voters divided by the number of comparison units within each neighbor group. The (standardized) interactive controls in

column (4) are omitted to save space. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05,

*** denotes p < 0.01.

38

before elections, then content that media outlets have weak incentives to report should not altervote choices. In particular, news that is not recent or not relevant to a media outlet’s audience islikely to be broadcast less. In line with these expectations, columns (13) and (14) find no significantinteraction between longer-run homicide levels and access to local media or between pre-electionhomicide shocks and non-local media. Nevertheless, voters could still be influenced by contentthat also references homicide shocks, beyond the news and political programming of local mediaoutlets themselves. Mexico’s election campaign ads are the most likely source, although theymostly contain national content (Larreguy, Marshall and Snyder forthcoming). However, columns(15)-(17) report no evidence that the share of broadcast media ads allocated to the party of themunicipal incumbent moderates the effects of homicide shocks. These falsification tests suggestthat electoral sanctioning occurs where media outlets are likely to report on local homicides at thetime when voters consume most news, and thus emphasize the importance for voting behavior ofthe intersection between the consumption and supply of politically-relevant news.

7 Mechanisms and interpretation

The preceding analysis showed that voters consume most political information just before elec-tions, and are responsive to homicides that occur at this time, but only when covered by localmedia stations likely to report such news. This section explores the mechanisms driving thesepolitical information cycles. I find that voters update their beliefs in a Bayesian manner, beforeshowing that political information cycles are more likely driven by changes in voter demand fornews than changes in media coverage around elections.

7.1 Voter belief updating reflects political information cycles

By examining voter beliefs directly, I test whether the preceding results reflect political infor-mation cycles. If news consumption reflects political information cycles, voters will principallyupdate their beliefs—potentially both about the incumbent’s performance and the importance ofaddressing security concerns—in response to events occurring just before elections.

However, the observed aggregate voting behavior is also consistent with several alternative ex-planations. First, pre-election homicide spikes could induce emotional electoral responses that donot rely on voters updating—whether in a naive or sophisticated manner—about incumbent qual-ity (Achen and Bartels 2017; Healy, Malhotra and Mo 2010). Second, cycles of information con-sumption may not be driving the results if voters’ beliefs only reflect recent events (Zaller 1992).In this case, pre-election homicides would drive election outcomes because voter beliefs did not

39

permanently incorporate news consumed earlier. Third, voters may only respond to pre-electionhomicide shocks because incumbents have time to win back disgruntled voters when homicidespikes occur earlier in their terms (Brollo 2009). In contrast with the political information cyclesargument that voters are unlikely to react to events during periods where they consume little news,these alternative explanations imply that voters should also update in the short-term outside ofelectoral campaigns as well (i.e. while homicides remain recent or before incumbents respond).

To distinguish between these accounts, I use the ENCUP surveys to examine how voter con-cerns about public security and confidence in elected officials vary with homicide shocks duringand outside local election campaigns.57 I define an indicator for respondents citing crime and inse-curity, drug trafficking, violence, or vandalism as the most important problem for their communityto solve.58 I also create an indicator of low confidence in the municipal mayor, defined by re-spondents ranking their confidence at 5 or below on a scale from 0 to 10 (for 2012) or expressingno, or almost no, confidence (in 2001, when a ten-point scale was not provided).59 The resultsin Table 7 extend equation (1) by interacting upcoming local elections with pre-survey homicideshocks (defined analogously to pre-election homicide shocks) and longer-term monthly homiciderates that are unlikely to be reported.60 The plausible exogeneity of both upcoming local electionsand pre-survey homicide shocks credibly identifies the lower-order and interaction effects.61

The results in Table 7 demonstrate that homicide shocks just before an election increase fearsregarding public security and reduce confidence in the mayor. The insignificant effects of thelower-order homicide shock terms in columns (1) and (2) of panel A, combined with the largepositive interactions, imply that homicide shocks only increase concerns about security amongvoters in municipalities experiencing an upcoming local election. Relative to the sample mean, ahomicide shock increases security concerns by almost 50% during an election campaign. Columns(1) and (2) of panel B report similar results for confidence in the mayor, showing that only pre-election homicide shocks induce voters to reduce such confidence, by up to one third of the samplemean. For both outcomes, the lower-order upcoming local election coefficients suggest that while

57The Federal District, where policing is not administered at the delegation-level, is excluded.58Along with all other political concerns, I include as zeroes respondents that do not cite a problem.59The cutoff on the 0-10 scale was chosen to broadly match the distribution of responses in 2001. However, the

results do not depend upon the choice of cutoff. Vote intentions were not elicited in any survey wave.60Since the day of the survey varies, but homicide data is monthly, a homicide in the month of the survey could

occur after the survey was carried out. I address this by defining a homicide shock as occurring if, relative to the twomonths after the survey was conducted, there were more homicides either in the two months preceding the surveymonth or the preceding month and survey month itself. In 2005, when day of month is available, I use the twopreceding months for the 15th or earlier, and the preceding month and survey month for the 16th or later.

61Supporting this design, Tables A3 and A4 show that short-run homicide shocks, and their interaction with up-coming local elections, are not systematically correlated with individual-level and municipal covariates.

40

Table 7: Heterogeneous effects of upcoming local elections on concern for public security andconfidence in the mayor, by short-run and long-run municipal homicide measures

Homicide measure:Pre-survey Homicides per Homicides per

homicide shock month last year month last 3 years(1) (2) (3) (4) (5) (6)

Panel A: Outcome: Public insecurity the major problem in the communityUpcoming local election 0.040** -0.028 0.039** 0.002 0.039** 0.003

(0.020) (0.022) (0.016) (0.016) (0.016) (0.016)Homicide measure 0.002 -0.005 0.011*** 0.008*** 0.011*** 0.008***

(0.014) (0.010) (0.002) (0.001) (0.002) (0.001)Upcoming local election 0.060** 0.071*** 0.001 0.001 0.001 0.001× Homicide measure (0.029) (0.025) (0.003) (0.002) (0.003) (0.002)

Observations 10,632 10,632 13,428 13,428 13,428 13,428Clusters 302 302 423 423 423 423Outcome range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1}Outcome mean 0.14 0.14 0.12 0.12 0.12 0.12Local election mean 0.22 0.22 0.20 0.20 0.20 0.20Homicide measure mean 0.47 0.47 3.27 3.27 3.35 3.35Survey year without data 2012 2012 2012 2012 2012 2012

Panel B: Outcome: Low confidence in the municipal mayorUpcoming local election -0.029 0.169 -0.006 0.091 -0.006 0.093

(0.041) (0.109) (0.035) (0.056) (0.038) (0.058)Homicide measure -0.049** -0.024 0.001 0.001 0.001*** 0.001**

(0.024) (0.020) (0.001) (0.001) (0.000) (0.000)Upcoming local election 0.114** 0.077 -0.001 0.010 -0.002 0.012× Homicide measure (0.047) (0.145) (0.009) (0.011) (0.015) (0.021)

Observations 6,453 6,453 7,764 7,764 7,764 7,764Clusters 293 293 384 384 384 384Outcome range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1}Outcome mean 0.35 0.35 0.35 0.35 0.35 0.35Local election mean 0.02 0.02 0.03 0.03 0.03 0.03Homicide measure mean 0.42 0.42 6.48 6.48 7.08 7.08Survey years without data 2003, 2005 2003, 2005 2003, 2005 2003, 2005 2003, 2005 2003, 2005

State fixed effects X X X

Notes: All specifications include survey year fixed effects, and are estimated using OLS. The number of observa-tions in columns (1) and (2) is lower because voters in municipalities failing to experience any homicide over thetwo months either side of the survey were dropped. Standard errors clustered by municipality are in parentheses.* denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

41

voters may become somewhat more positive when a homicide shock does not occur, any positivechanges are far smaller than the negative changes associated with homicide shocks. Furthermore,and reinforcing voters’ perception that municipal governments are responsible for local homicides(implied by the precinct-level results in Table 5), Appendix Table A18 shows that pre-electionhomicide shocks do not cause voters to lose confidence in the president or their state governor.

Given that some voters do consume news outside of election campaigns, it is perhaps surprisingthat homicide shocks do not affect beliefs at other times as well. This could be because changesin beliefs and voting behavior are concentrated among the voters that shift from consuming verylittle news outside election campaigns to more regular consumption before elections. More fre-quent consumers of news—who are, on average, significantly more educated—may be harder toinfluence because their greater news consumption engenders more precise prior beliefs that areharder to shift. Table A21 supports this claim: voters without higher education update their beliefsabout public security being a major problem and reduce their confidence in the mayor more inresponse to pre-election homicide shocks, while the effect of local media covering such shocks onincumbent vote share is lower in precincts with below-median levels of higher education.

It remains possible that voters are actually responding to longer-run homicide rates. The lower-order terms unsurprisingly show that public security is a greater concern in more violent municipal-ities. However, the placebo-like tests in columns (3)-(6), which interact upcoming local electionswith longer-term homicide measures, do not support this possibility.

These results suggest that voter responses are not simply emotional, but entail belief updating.Moreover, highlighting the importance of information consumption cycles, the insignificant effectsof a homicide shock outside local election campaigns imply that voters do not consume informationand react similarly outside of election campaigns, just without durably updating their beliefs. Theresults are instead consistent with voters only internalizing events likely to have been reported inthe news before elections. Nevertheless, although beliefs change, this analysis cannot establish theextent to which belief updating is sophisticated or actually influences voting behavior.

7.2 Prior beliefs moderate voter responses to homicide shocks

To better assess whether voting behavior reflects sophisticated belief updating, I consider morenuanced responses to homicide shocks. If voters respond in a Bayesian manner, their reactionto homicide shocks will interact with their prior beliefs about incumbent and challenger parties.I test for such learning by examining whether homicide shocks before the previous election—aplausible proxy for voters’ prior beliefs about the competence of the current incumbent party (ifstill in office) or challenger parties (if a different party was previously in office), given voters’

42

-.1-.0

50

.05

.1M

argi

nal e

ffect

of h

omic

ide

shoc

k

Different incumbent,no previous shock

Same incumbent,previous shock

Same incumbent,no previous shock

Different incumbent,previous shock

Figure 6: The effect of a homicide shock on the change in incumbent party vote share, byprevious homicide shocks and incumbent identity (95% confidence intervals)

Notes: Same incumbent is an indicator for the current incumbent party winning the previous election.Previous shock is an indicator for a homicide shock occurring before the previous election.

cyclical news consumption—moderate the precinct-level electoral effects of pre-election homicideshocks. Conversely, naive Zallerian voters are unlikely to respond in such a nuanced manner.

Figure 6 shows how the effect of a homicide shock on the change in the incumbent party’s voteshare varies with the occurrence of a homicide shock before the previous election and whetherthe current incumbent party differs from the previous incumbent party.62 The first bar shows thatsanctioning is greatest among voters in municipalities experiencing a homicide shock before thecurrent election and which did not experience such a shock before the previous election undera different incumbent party. Similarly, the negative effect of a homicide shock is also large if thesame incumbent party was in office for consecutive pre-election homicide shocks. These effects aresomewhat larger in magnitude than for the third bar, where the current incumbent was in office atthe previous election but did not experience a shock. However, the starkest contrast is between thefirst two cases—where homicide shocks entail substantial sanctioning—and the fourth bar, where aprevious incumbent from a different party presided over a homicide shock at the previous electionand experienced no voter sanctioning. These findings reinforce the survey results suggesting thatvoters update their beliefs in a sophisticated manner, and thus lend greater credibility to the careerconcerns and signaling models predicting that Bayesian voters will sanction high homicide rates.

62The associated regression estimates are provided in Table A19.

43

7.3 Media coverage of violent crime is not more common during local elec-tion campaigns

So far, this article has treated political information cycles as an equilibrium outcome. However,from both a theoretical and policy perspective, it is important to understand whether these cyclesreflect changes in voter demand for politically-relevant news and/or changes in the media’s supplyof such news around elections. The increased political news consumption at the extensive margindocumented in Table 1 suggests that tuning in for the news is a conscious choice, given that radioand television schedules do not change around elections. However, for the many voters increasingconsumption along the intensive margin, it is hard to separate increased voter interest in followingnews before elections from news providers broadcasting more politically-relevant news beforeelections.

For news relating to local homicides, a first step in separating these effects is to examinechanges in coverage around elections. In the absence of comprehensive broadcast media tran-scripts, I use newspaper data from Osorio (2015). He uses machine-learning techniques to collatereports relating to drug violence from 105 government agencies and national and local newspapersbetween 2000 and 2010. The resulting report counts capture the frequency of coverage, but not itstone. Broadcast media coverage is likely to be similar, given that newspapers serve as an impor-tant information source for broadcast media outlets in developing contexts (Keefer and Khemani2016). In a nationwide survey of local newspapers and radio stations, 64% report sourcing newsfrom other local media outlets (Larreguy, Lucas and Marshall 2016).

To test whether reporting of violent events increases before elections in the municipalitiesused for the main analysis, I examine whether the correlation between the occurrence and report-ing of violent events increases within the five months before local elections using the followingdifference-in-differences specification:

Reportsmt = β1Homicidesmt +β2Upcoming local electionmt

+β3

(Upcoming local electionmt×Homicidesmt

)+ηm + µt + εmt , (6)

where Reportsmt is a month-year municipal-level count of the number of violent events (includinghomicides as well as shootings, kidnappings, torture etc.) between DTOs reported for that month,63

and µm and µt are municipality and month-year fixed effects.The results in Table 8 show a strong positive correlation between homicides and violent events,

63The outcome was aggregated up from Osorio’s (2015) weekly count. Similar results obtain when using reportsof violent enforcement, arrests, drug seizures, asset seizures, and gun seizures.

44

Table 8: Correlation between monthly municipal homicide counts and government agency andnewspaper reports on violence between gangs, by upcoming local elections

Number of inter-DTO violence reports per month-year(1) (2) (3) (4) (5)

Homicides per month 0.236*** 0.225*** 0.237*** 0.237*** 0.226***(0.026) (0.024) (0.026) (0.026) (0.024)

Upcoming local election 0.005(0.018)

Homicides per month -0.005 -0.006 -0.004× Upcoming local election (0.017) (0.017) (0.014)

Observations 117,780 117,780 117,780 117,780 117,780Clusters 906 906 906 906 906Outcome range [0,102] [0,102] [0,102] [0,102] [0,102]Outcome mean 0.28 0.28 0.28 0.28 0.28Outcome std. dev. 2.12 2.12 2.12 2.12 2.12Homicides per month mean 0.93 0.93 0.93 0.93 0.93Homicides per month std. dev. 5.24 5.24 5.24 5.24 5.24Upcoming local election mean 0.15 0.15 0.15 0.15 0.15Additional fixed effects State × Election × Election × state

month-year month-year × month-year

Notes: All specifications include municipality and year-month fixed effects, and are estimated using OLS. Stan-dard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotesp < 0.01.

but that such coverage is not significantly greater before elections. Column (1) first demonstratesthat for each homicide that occurs in a municipality, the number of violent events reported increasesby 0.24. Since not all newspapers are studied and reports only pertain to drug-related violence, thisestimate likely understates the propensity of newspapers to report homicides. Turning to differen-tial coverage around elections, column (3) shows that reporting is not significantly greater in thefive months preceding local elections. This is broadly consistent with surveys of media outletsin 2017, which indicate that while politics in general is covered more before elections, only 18%(5%) of outlets report on public security issues more (less) than usual before elections (Larreguy,Lucas and Marshall 2016). Both sets of results are precisely estimated, and are robust to includingstate-month-year fixed effects (and their interaction with upcoming local elections). These findingsindicate that media coverage of homicides is not sensitive to the electoral cycle, and thus suggestthat the sanctioning of pre-election homicides is more likely driven by increased voter demand forpolitically-relevant news before elections than changes in the supply of such news.

45

8 Conclusion

This study demonstrates the importance of the timing of voter news consumption for understand-ing how voters hold governments to account. Leveraging survey and electoral returns data, andidentification strategies isolating three central and complementary components of the political in-formation cycles argument, I show that voters indeed consume more politically-relevant news be-fore elections, and that pre-election homicide shocks substantially reduce support for the municipalincumbent party among voters exposed to local media likely to report on such events. Conversely,there is little evidence that long-run performance homicide indicators affect electoral outcomes.

Although this paper’s evidence draws from Mexico, its findings are likely to apply morebroadly to contexts where voters do not regularly consume news outside of election campaigns andto issues that are similarly salient to voters. Political knowledge levels are low in many democ-racies, especially in the developing world. In the U.S., major news stories preceding electionsseem to frequently and substantially influence election results. Moreover, macroeconomic out-comes and corruption revelations may evoke similarly strong responses from voters, given the easewhich media outlets can explain—and potentially dramatize (Baum 2002)—its link to politicianperformance.

Consequently, understanding the origins and further substantiating mechanisms underpinningpolitical information cycles demands further research. First, this article did not attempt to separatebetween the factors—such as consumption benefits, social pressure, or civic duty—that induce vot-ers to demand more news before elections. Second, although the frequency of newspaper coverageof homicides does not change during election campaigns, it is important to establish the extent towhich such stability extends to the tone of coverage and to other more explicitly political newsstories. Third, this study addressed the issue of media bias by simply focusing on access to anaverage local media outlet. However, it may be important to move beyond access to examine theextent to which coverage is politically slanted (Gentzkow and Shapiro 2010; Gentzkow et al. 2015),and the agenda is strategically determined (Durante and Zhuravskaya forthcoming; Eisensee andStromberg 2007). Finally, exactly what voters infer from pre-election homicides warrants deeperexamination. While I observe updating consistent with Bayesian information processing, addi-tional data is required to distinguish whether voters learn from homicide spikes occurring or fromresponses in their aftermath.

Another open question regards the normative implications of political information cycles. Whethervoting on the basis of noisy pre-election performance signals, which add noise to the selectionprocess, reduces voter welfare depends on the factors that would have determined vote choices

46

otherwise. Although exogenous pre-election homicide shocks might cause voters to downweightsignals that are more informative about desirable politician characteristics, in contexts like Mexicoit is also possible that clientelistic relationships are broken down, which could increase welfareif clientelism perpetuates the rule of low-quality incumbents. Nevertheless, the possibility thatinformation reduces the efficiency of election mechanisms chimes with other theoretical resultsshowing that it cannot be assumed that greater political information is necessarily welfare en-hancing (Ashworth and Bueno de Mesquita 2014; Ashworth, Bueno de Mesquita and Friedenbergforthcoming). Ultimately, whether homicide shocks—and other captivating news stories—resultin the election of worse politicians and lower performance are empirical questions of significantinterest policy-makers designing informational interventions.

Regardless of the welfare implications, it is clear that political parties, interest groups, andNGOs should strategically target incumbent performance information toward the least informedvoters around elections—when voters consume political news. While such groups may use infor-mation to support their goals, it is essential that purely informative news should relate to horizonsbeyond proximate security or economic shocks, and thus provide the longer-term indicators thatvoters profess to value (Healy and Lenz 2014). Although NGO-supported interventions have at-tempted to achieve this, often by delivering leaflets or newspapers (e.g. Banerjee et al. 2011; Chonget al. 2015; Dunning et al. forthcoming), broadcast media appears to be the most effective tool fordissemination (Ferraz and Finan 2008; Larreguy, Marshall and Snyder 2018). Consequently, akey challenge is encouraging political actors and media stations to overcome sensationalist andpolitical biases in favor of longer-term performance indicators.

47

References

Achen, Christopher H. and Larry M. Bartels. 2017. Democracy for realists: Why elections do not

produce responsive government. Princeton University Press.

Adena, Maja, Ruben Enikolopov, Maria Petrova, Veronica Santarosa and Ekaterina Zhuravskaya.2015. “Radio and the Rise of the Nazis in Prewar Germany.” Quarterly Journal of Economics

130(4):1885–1939.

Alesina, Alberto, Salvatore Piccolo and Paolo Pinotti. 2016. “Organized Crime, Violence, andPolitics.” Working paper.

Anderson, Christopher J. 2007. “The end of economic voting? Contingency dilemmas and thelimits of democratic accountability.” Annual Review Political Science 10:271–296.

Arias, Eric, Horacio A. Larreguy, John Marshall and Pablo Querubın. 2018. “Priors rule: Whendo malfeasance revelations help and hurt incumbent parties?” Working paper.

Arnold, R. Douglas and Nicholas Carnes. 2012. “Holding mayors accountable: New York’s exec-utives from Koch to Bloomberg.” American Journal of Political Science 56(4):949–963.

Ashworth, Scott. 2005. “Reputational dynamics and political careers.” Journal of Law, Economics,

and Organization 21(2):441–466.

Ashworth, Scott. 2012. “Electoral accountability: recent theoretical and empirical work.” Annual

Review of Political Science 15:183–201.

Ashworth, Scott and Ethan Bueno de Mesquita. 2014. “Is Voter Competence Good for Voters?Information, Rationality, and Democratic Performance.” American Political Science Review

108(3):565–587.

Ashworth, Scott, Ethan Bueno de Mesquita and Amanda Friedenberg. 2017. “Accountability andinformation in elections.” American Economic Journal: Microeconomics 9(2):95–138.

Ashworth, Scott, Ethan Bueno de Mesquita and Amanda Friedenberg. forthcoming. “LearningAbout Voter Rationality.” American Journal of Political Science .

Baker, Andy, Barry Ames and Lucio R. Renno. 2006. “Social context and campaign volatility innew democracies: networks and neighborhoods in Brazil’s 2002 Elections.” American Journal

of Political Science 50(2):382–399.

48

Banerjee, Abhijit V., Selvan Kumar, Rohini Pande and Felix Su. 2011. “Do Informed Voters MakeBetter Choices? Experimental Evidence from Urban India.” Working paper.

Barabas, Jason and Jennifer Jerit. 2009. “Estimating the Causal Effects of Media Coverage onPolicy-Specific Knowledge.” American Journal of Political Science 53(1):73–89.

Barabas, Jason, Jennifer Jerit, William Pollock and Carlisle Rainey. 2014. “The Question (s) ofPolitical Knowledge.” American Political Science Review 108(4):840–855.

Bateson, Regina. 2012. “Crime victimization and political participation.” American Political Sci-

ence Review 106(3):570–587.

Baum, Matthew A. 2002. “Sex, lies, and war: How soft news brings foreign policy to the inattentivepublic.” American Political Science Review 96(1):91–109.

Berrebi, Claude and Esteban F. Klor. 2006. “On terrorism and electoral outcomes theory andevidence from the Israeli-Palestinian conflict.” Journal of Conflict Resolution 50(6):899–925.

Besley, Timothy and Andrea Prat. 2006. “Handcuffs for the grabbing hand? Media capture andgovernment accountability.” American Economic Review 96(3):720–736.

Besley, Timothy and Robin Burgess. 2002. “The political economy of government responsiveness:Theory and evidence from India.” Quarterly Journal of Economics 117(4):1415–1451.

Blais, Andre. 2000. To vote or not to vote? The merits and limits of rational choice theory.University of Pittsburgh Pre.

Blanco, Luisa R. 2013. “The impact of crime on trust in institutions in Mexico.” European Journal

of Political Economy 32:38–55.

Bobonis, Gustavo J., Luis R. Camara Fuertes and Rainer Schwabe. 2013. “Monitoring CorruptiblePoliticians.” Working paper.

Brender, Adi and Allan Drazen. 2005. “Political budget cycles in new versus established democ-racies.” Journal of Monetary Economics 52(7):1271–1295.

Brollo, Fernanda. 2009. “Who Is Punishing Corrupt Politicians—Voters or the Central Govern-ment? Evidence from the Brazilian Anti-Corruption Program.” Working paper.

49

Castaneda Sabido, Fernando. 2011. “Investigacion sobre la percepcion ciudadana acerca de latransparencia, rendicion de cuentas y fiscalizacin en el uso de los recursos publicos en Mexico.”Reporte para la Auditorıa Superior de la Federacion .

Chang, Eric C.C., Miriam A. Golden and Seth J. Hill. 2010. “Legislative malfeasance and politicalaccountability.” World Politics 62(2):177–220.

Chiang, Chun-Fang and Brian Knight. 2011. “Media bias and influence: Evidence from newspaperendorsements.” Review of Economic Studies 78(3):795–820.

Chong, Alberto, Ana De La O, Dean Karlan and Leonard Wantchekon. 2015. “Does CorruptionInformation Inspire the Fight or Quash the Hope? A Field Experiment in Mexico on VoterTurnout, Choice and Party Identification.” Journal of Politics 77(1):55–71.

Cole, Shawn, Andrew Healy and Eric Werker. 2012. “Do voters demand responsive governments?Evidence from Indian disaster relief.” Journal of Development Economics 97(2):167–181.

Corbacho, Ana, Julia Philipp and Mauricio Ruiz-Vega. 2015. “Crime and erosion of trust: Evi-dence for Latin America.” World Development 70:400–415.

Cornelius, Wayne A. 1996. Mexican politics in transition: The breakdown of a one-party-dominant

regime. La Jolla: Center for US-Mexican Studies, University of California, San Diego.

Da Silveira, Bernardo S. and Joao M.P. De Mello. 2011. “Campaign advertising and electionoutcomes: Quasi-natural experiment evidence from gubernatorial elections in Brazil.” Review of

Economic Studies 78(2):590–612.

de Figueiredo, Miguel F.P., F. Daniel Hidalgo and Yuri Kasahara. 2013. “When Do Voters PunishCorrupt Politicians? Experimental Evidence from Brazil.” Working paper.

Degan, Arianna. 2006. “Policy Positions, Information Acquisition and Turnout.” Scandinavian

Journal of Economics 108(4):669–682.

Dell, Melissa. 2015. “Trafficking Networks and the Mexican Drug War.” American Economic

Review 105(6):1738–1779.

DellaVigna, Stefano and Ethan Kaplan. 2007. “The Fox News effect: Media bias and voting.”Quarterly Journal of Economics 122(3):1187–1234.

50

DellaVigna, Stefano, Ruben Enikolopov, Vera Mironova, Maria Petrova and Ekaterina Zhu-ravskaya. 2014. “Cross-border media and nationalism: Evidence from Serbian radio in Croatia.”American Economic Journal: Applied Economics 6(3):103–32.

Delli Carpini, Michael X. and Scott Keeter. 1996. What Americans Know about Politics and Why

It Matters.

Diaz-Cayeros, Alberto, Jose Antonio Gonzalez and Fernando Rojas. 2006. Mexico’s Decentraliza-tion at a Crossroads. In Federal and Economic Reform: International Perspectives, ed. Jessica S.Wallack and T. N. Srinivasan. Cambridge University Press pp. 364—406.

Downs, Anthony. 1957. An Economic Theory of Democracy. Addison Wesley.

Dunning, Thad, Guy Grossman, Macartan Humphreys, Susan Hyde and Craig McIntosh. forth-coming. Metaketa I: The Limits of Electoral Accountability. Cambridge University Press.

Durante, Ruben and Ekaterina Zhuravskaya. forthcoming. “Attack When the World Is Not Watch-ing? U.S. News and the Israeli-Palestinian Conflict.” Journal of Political Economy .

Durante, Ruben and Emilio Gutierrez. 2015. “Fighting crime with a little help from my friends:Party affiliation, inter-jurisdictional cooperation and crime in Mexico.” Working paper.

Durante, Ruben, Paolo Pinotti and Andrea Tesei. 2017. “The political legacy of entertainmentTV.”.

Eifert, Benn, Edward Miguel and Daniel N. Posner. 2010. “Political Competition and EthnicIdentification in Africa.” American Journal of Political Science 54(2):494–510.

Eisensee, Thomas and David Stromberg. 2007. “News droughts, news floods, and US disasterrelief.” Quarterly Journal of Economics pp. 693–728.

Enikolopov, Ruben, Maria Petrova and Ekaterina Zhuravskaya. 2011. “Media and political per-suasion: Evidence from Russia.” American Economic Review 101(7):3253–3285.

Farnsworth, Stephen J. and S. Robert Lichter. 2011. The Nightly News Nightmare: Media Coverage

of US Presidential Elections, 1988-2008. Rowman and Littlefield.

Faye, Michael and Paul Niehaus. 2012. “Political aid cycles.” American Economic Review

102(7):3516–30.

51

Fearon, James D. 1999. Electoral Accountability and the Control of Politicians: Selecting GoodTypes versus Sanctioning Poor Performance. In Democracy, Accountability, and Representation,ed. Adam Przeworski, Susan Stokes and Bernard Manin. Cambridge University Press.

Feddersen, Timothy and Alvaro Sandroni. 2006. “Ethical voters and costly information acquisi-tion.” Quarterly Journal of Political Science 1(3):287–311.

Ferraz, Claudio and Frederico Finan. 2008. “Exposing Corrupt Politicians: The Effects of Brazil’sPublicly Released Audits on Electoral Outcomes.” Quarterly Journal of Economics 123(2):703–745.

Fowler, Anthony and Andrew B. Hall. forthcoming. “Do shark attacks influence presidential elec-tions? Reassessing a prominent finding on voter competence.” Journal of Politics .

Fowler, Anthony and B. Pablo Montagnes. 2015. “College football, elections, and false-positive results in observational research.” Proceedings of the National Academy of Sciences

112(45):13800–13804.

Fox, Jonathan. 1994. “The difficult transition from clientelism to citizenship: Lessons from Mex-ico.” World Politics 46(2).

Gentzkow, Matthew and Jesse M. Shapiro. 2006. “Media bias and reputation.” Journal of Political

Economy 114(2):280–316.

Gentzkow, Matthew and Jesse M. Shapiro. 2010. “What drives media slant? Evidence from USdaily newspapers.” Econometrica 78(1):35–71.

Gentzkow, Matthew, Nathan Petek, Jesse M. Shapiro and Michael Sinkinson. 2015. “Do news-papers serve the state? Incumbent party influence on the US press, 1869–1928.” Journal of the

European Economic Association 13(1):29–61.

Greene, Kenneth F. 2007. Why Dominant Parties Lose: Mexico’s Democratization in Comparative

Perspective. Cambridge University Press.

Greene, Kenneth F. 2011. “Campaign Persuasion and Nascent Partisanship in Mexico’s NewDemocracy.” American Journal of Political Science 55(2):398–416.

Hallin, Daniel. 2000. Media, political power, and democratization in Mexico. In De-Westernizing

Media Studies, ed. James Currant and Myung-Jin Park. London: Routledge pp. 97–110.

52

Hamilton, James. 2004. All the news that’s fit to sell: How the market transforms information into

news. Princeton University Press.

Healy, Andrew and Gabriel S. Lenz. 2014. “Substituting the End for the Whole: Why Vot-ers Respond Primarily to the Election-Year Economy.” American Journal of Political Science

58(1):31–47.

Healy, Andrew J., Neil Malhotra and Cecilia Hyunjung Mo. 2010. “Irrelevant events affect vot-ers’ evaluations of government performance.” Proceedings of the National Academy of Sciences

107(29):12804–12809.

Healy, Andrew and Neil Malhotra. 2013. “Retrospective Voting Reconsidered.” Annual Review of

Political Science 16:285–306.

Heinle, Kimberly, Octavio Rodrıguez Ferreira and David A. Shirk. 2014. “Drug Violence in Mex-ico: Data and Analysis Through 2013.” Justice in Mexico Project Special Report.

Hirano, Shigeo, Gabriel S. Lenz, Maksim Pinkovskiy and James M. Snyder, Jr. 2015. “VoterLearning in State Primary Elections.” American Journal of Political Science 59(1):91–108.

Holmstrom, Bengt et al. 1999. “Managerial Incentive Problems: A Dynamic Perspective.” Review

of Economic Studies 66(1):169–182.

Humphreys, Macartan and Jeremy Weinstein. 2012. “Policing Politicians: Citizen Empowermentand Political Accountability in Uganda Preliminary Analysis.” Working paper.

Ibope/AGB Mexico. 2009. “Audiencias y Medios en Mexico.”.

Keefer, Philip and Stuti Khemani. 2014. “Mass media and public education: The effects of accessto community radio in Benin.” Journal of Development Economics 57:57–72.

Keefer, Philip and Stuti Khemani. 2016. “The Government Response to Informed Citizens: NewEvidence on Media Access and the Distribution of Public Health Benefits in Africa.” World Bank

Economic Review 30(2):233–267.

Keefer, Phillip. 2007. Seeing and Believing: Political Obstacles to Better Service Delivery. InThe Politics of Service Delivery in Democracies: Better Access for the Poor, ed. ShantayananDevarajan and Ingrid Widlund. Stockholm: Expert Group on Development Issues pp. 42–55.

53

Kendall, Chad, Tommaso Nannicini and Francesco Trebbi. 2015. “How do voters respond toinformation? Evidence from a randomized campaign.” American Economic Review 105(1):322–53.

Krosnick, Jon A. and Donald R. Kinder. 1990. “Altering the foundations of support for the presi-dent through priming.” American Political Science Review 84(2):497–512.

Langston, Joy. 2003. “Rising from the ashes? Reorganizing and unifying the PRI’s state partyorganizations after electoral defeat.” Comparative Political Studies 36(3):293–318.

Larcinese, Valentino. 2005. “Electoral Competition and Redistribution with Rationally InformedVoters.” BE Journal of Economic Analysis and Policy 4(1):1–28.

Larreguy, Horacio A., Christopher Lucas and John Marshall. 2016. “When do media stationssupport political accountability? A field experiment in Mexico.” AEA RCT Registry 0001186.

Larreguy, Horacio A., John Marshall and James M. Snyder, Jr. 2018. “Publicizing malfeasance:How local media facilitates electoral sanctioning of Mayors in Mexico.” Working Paper.

Larreguy, Horacio A., John Marshall and James M. Snyder, Jr. forthcoming. “Leveling the play-ing field: How campaign advertising can help non-dominant parties.” Journal of the European

Economic Association .

Lawson, Chappell H. 2002. Building the fourth estate: Democratization and the rise of a free press

in Mexico. University of California Press.

Lawson, Chappell and James A. McCann. 2005. “Television News, Mexico’s 2000 Elections andMedia Effects in Emerging Democracies.” British Journal of Political Science 35(1):1–30.

Ley, Sandra. 2014. “Citizens in Fear: Participation and Voting Behavior in the Midst of Violence.”Dissertation.

Magaloni, Beatriz. 2006. Voting for Autocracy: Hegemonic Party Survival and its Demise in

Mexico. New York: Cambridge University Press.

Marshall, John. forthcoming. “Signaling sophistication: How social expectations can increasepolitical information acquisition.” Journal of Politics .

Martin, Gregory J and Ali Yurukoglu. 2017. “Bias in Cable News: Real Effects and Polarization.”American Economic Review 107(9):2565–2599.

54

Maskin, Eric and Jean Tirole. 2004. “The Politician and the Judge: Accountability in Government.”American Economic Review 94(4):1034–1054.

McCann, James A. and Chappell H. Lawson. 2003. “An Electorate Adrift? Public Opinion and theQuality of Democracy in Mexico.” Latin American Research Review 38(3):60–81.

Mexico Evalua. 2012. “Seguridad y Justicia Penal en los estados.” Report.

Michelitch, Kristin and Stephen Utych. 2018. “Electoral Cycle Fluctuations in Partisanship:Global Evidence from Eighty-Six Countries.” Journal of Politics 80(2):000–000.

Mullainathan, Sendhil and Andrei Shleifer. 2005. “The Market for News.” American Economic

Review 95(4):1031–1053.

Olken, Benjamin A. 2009. “Do television and radio destroy social capital? Evidence from Indone-sian villages.” American Economic Journal: Applied Economics 1(4):1–33.

Ortoleva, Pietro and Erik Snowberg. 2015. “Overconfidence in Political Behavior.” American

Economic Review 105(2):504–535.

Osorio, Javier. 2015. “The Contagion of Drug Violence: Spatiotemporal Dynamics of the MexicanWar on Drugs.” Journal of Conflict Resolution 59(8):1403–1432.

Pande, Rohini. 2011. “Can informed voters enforce better governance? Experiments in low-income democracies.” Annual Review of Economics 3(1):215–237.

Prat, Andrea and David Stromberg. 2005. “Commercial Television and Voter Information.”. CEPRDiscussion Paper No. 4989.

Prior, Markus. 2007. Post-broadcast democracy: How media choice increases inequality in polit-

ical involvement and polarizes elections. Cambridge University Press.

Reames, Benjamin. 2003. “Police forces in Mexico: A profile.” Center for US-Mexican Studies.

Remmer, Karen L. 1991. “The political impact of economic crisis in Latin America in the 1980s.”American Political Science Review 85(3):777–800.

Roberts, Kenneth M. and Erik Wibbels. 1999. “Party systems and electoral volatility in LatinAmerica: a test of economic, institutional, and structural explanations.” American Political Sci-

ence Review 93(3):575–590.

55

Rogoff, Kenneth. 1990. “Equilibrium Political Budget Cycles.” American Economic Review

80(1):21–36.

Sabet, Daniel. 2010. Police Reform in Mexico: Advances and Persistent Obstacles. In Shared

Responsibility: U.S.-Mexico Policy Options for Confronting Organized Crime, ed. Eric L. Olson,David A. Shirk and Andrew Selee. Woodrow Wilson International Center for Scholars MexicoInstitute.

Semetko, Holli A. and Patti M. Valkenburg. 2000. “Framing European politics: A content analysisof press and television news.” Journal of Communication 50(2):93–109.

Serra, Gilles. 2012. “The Risk of Partyarchy and Democratic Backsliding in Mexico.” Taiwan

Journal of Democracy 8(1):93–118.

Snyder, Jr, James M. and David Stromberg. 2010. “Press Coverage and Political Accountability.”Journal of Political Economy 118(2):355–408.

Spenkuch, Jorg L. and David Toniatti. 2016. “Political Advertising and Election Outcomes.” Work-ing Paper.

Trelles, Alejandro and Miguel Carreras. 2012. “Bullets and votes: Violence and electoral partici-pation in Mexico.” Journal of Politics in Latin America 4(2):89–123.

Vivanco, Edgar F., Jorge Olarte, Alberto Dıaz-Cayeros and Beatriz Magaloni. 2015. Drugs, Bullets,and Ballots: The Impact of Violence on the 2012 Presidential Election. In Mexico’s Evolving

Democracy: A Comparative Study of the 2012 Elections, ed. Jorge I. Domınguez, Kenneth F.Greene, Chappell H. Lawson and Alejandro Moreno. Johns Hopkins University Press pp. 153–181.

Wolfers, Justin. 2007. “Are Voters Rational? Evidence from Gubernatorial Elections.” Workingpaper.

Yanagizawa-Drott, David. 2014. “Propaganda and conflict: Evidence from the Rwandan geno-cide.” Quarterly Journal of Economics 129(4):1947–1994.

Zaller, John. 1992. The Nature and Origins of Mass Opinion. Cambridge University Press.

56

A Appendix

Contents

A.1 Random shocks and electoral accountability with Bayesian voters . . . . . . . . . A2A.1.1 A career concerns model . . . . . . . . . . . . . . . . . . . . . . . . . . . A2A.1.2 An incumbent signaling model . . . . . . . . . . . . . . . . . . . . . . . A5

A.2 Months and years of municipal elections . . . . . . . . . . . . . . . . . . . . . . . A7A.3 Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A9

A.3.1 ENCUP survey data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A9A.3.2 Precinct and municipal homicide and electoral data . . . . . . . . . . . . . A11A.3.3 Precinct local media coverage data . . . . . . . . . . . . . . . . . . . . . . A12

A.4 Map of municipalities included in different samples . . . . . . . . . . . . . . . . . A12A.5 Additional analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A12

A.5.1 Assessing the identification assumptions . . . . . . . . . . . . . . . . . . . A12A.5.2 Effects on turnout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A14A.5.3 Additional robustness checks for the aggregate electoral results . . . . . . A22A.5.4 Additional robustness checks for the aggregate electoral results . . . . . . A22A.5.5 Electoral effects of drug-related homicides . . . . . . . . . . . . . . . . . A22A.5.6 Difference-in-differences approach to identifying the effects of pre-election

homicides . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A29A.5.7 Differential effects of homicide shocks across parties . . . . . . . . . . . . A30A.5.8 Belief updating about presidents and state governors . . . . . . . . . . . . A33A.5.9 Differential updating across elections . . . . . . . . . . . . . . . . . . . . A33A.5.10 Learning from the experiences of neighboring municipalities . . . . . . . . A33A.5.11 Less-educated voters respond most to pre-election homicide shocks . . . . A35

A1

A.1 Random shocks and electoral accountability with Bayesian voters

Bayesian voters will generally update from credible signals of incumbent competence or alignmentwith voters’ interests. For example, voters may use long-run economic performance or crime ratesto infer incumbent type, on the basis that incumbent policy choices and implementation qualityreflect underlying competence of alignment with voters’ interests. However, voters may also up-date from exogenous shocks, like the pre-election homicide shocks studied in the main paper. Thissection presents two simple models that rationalize why voters may alter their beliefs and votingbehavior in response to shock beyond the incumbent’s control.

A.1.1 A career concerns model

Consider a career concerns model where an incumbent party seeks re-election in a given munici-pality. For simplicity, there are two electoral terms τ ∈ {1,2}, each containing T discrete periodsduring which a noisy signal of the incumbent’s underlying competence θ is observed—in this ap-plication, the number of homicides that occurred in the municipality during period t. The termsare split by an election, where the signal helps a unit mass of voters to decide whether to re-electincumbent I or instead elect challenger C.

The incumbent seeks to retain office. For each term in office, I receives “ego” rents R > 0.Each incumbent has a fixed level of competence, which is drawn from θI ∼N (θ ,1/hθ ), wherethe realization is known neither to I nor the representative voter.64 At the beginning of each term,I can also exert costly effort aτ ≥ 0 to reduce the homicide rate, in an attempt to convince votersthat they are sufficiently competent to merit re-election. In this context, aτ could involve allocatingmore resources to policing, selecting a competent chief of police with few ties to organized crime,or requiring municipal police to work with police forces at higher levels of government. The costof incumbent effort is c(aτ), where c′(·)> 0, c′′(·), and c(0) = 0. Incumbents are forward-lookingand risk-neutral; they maximize p(aτ)R− c(aτ), where p(aτ) is the (endogenously-determined)probability of re-election.

Following Holmstrom et al.’s (1999) simple model, the public signal of incumbent competence—the number of homicides in the municipality in period t of term τ—is given by:

Htτ = b−θI−aτ + et , (A1)

where et ∼N (0,1/he) is a period-specific normally-distributed shock, and b > 0 is a constant

64Signaling models, like the one considered below, instead assume that the incumbent’s type is only known to theincumbent (e.g. Rogoff 1990).

A2

baseline homicide level. This shock is independent of incumbent competence and across periods,and its distribution is common knowledge. The homicide count is thus greater for less competentincumbents and incumbents that work less hard, but also reflects idiosyncratic factors beyond theincumbent’s control.

Finally, voters elect their preferred party for the second term, based on the sum of the party’scompetence and their “ideological” bias. All voters value competence equally, either directly tomaintain security, or to address other issues that the ability to reduce homicides may be corre-lated with. In addition, the ideological bias δi of voter toward Iis realized after aτ is selected,and is uniformly-distributed such that δi ∼ U (b− 1

2ψ,b+ 1

2ψ) with mean bias B ∈ R. Voter i

compares I to challenger C, who is a random draw from the same pool as incumbents, such thatθC ∼N (θ ,1/hθ ). Their vote choice is vi ∈ {I,C}.

The game’s timing can be summarized as:

1. Nature draw incumbent competence θI and challenger competence θC from N (θ ,1/hθ ).This is unknown to all politicians and voters.

2. At the beginning of τ = 1, the incumbent selects a1.

3. Each voter observes T independent signals, Ht1, and their ideological bias δi is realized.

4. Each voter decides whether to re-elect the incumbent or not, vi ∈ {I,C}. The party with themajority of votes wins.

5. At the beginning of τ = 2, the incumbent (election winner) selects a2.

Working backwards, any incumbent party in the second term (τ = 2) exerts effort a∗2 = 0, giventhat they lack re-election incentives.65 Voter i votes for I (i.e. vi = I) when E[θI|H11, ...,HT 1, a1]+

δi ≥ E[θC], where a1 is the voter’s conjecture about incumbent effort that needs to be filteredout. Upon observing the homicide count in each period, voters’ posterior belief about incumbentcompetence is normally distributed with expectation:

E[θI|H11, ...,HT 1, a1] = λ (b−H1− a1)+ (1−λ )θ , (A2)

where λ := T hehθ+T he

is the weight attached to the mean observed signal Hτ := T−1∑

Tt=1 Htτ . Re-

arranging, voter i then votes to re-elect the incumbent if the number of homicides observed is

65This reflects the simplifying two-period assumption, although the results generalize to infinitely-repeated games.

A3

sufficiently low: H1 ≤ δiλ+ b− θ − a1. Aggregating across voters yields the incumbent party’s

vote share:

VI =12+B+λψ(b−H1− θ − a1) (A3)

Intuitively, the incumbent’s vote share reflects average ideological biases and voter inferencesabout incumbent competence—the only thing that matters in the second term where there are no re-election incentives to work hard—based on the number of homicides (after filtering out incumbenteffort).

At the beginning of the first term, the incumbent selects their effort to reduce the number ofhomicides in anticipation of this voting rule. Given the ex ante re-election probability (i.e. beforethe number of homicides is observed), the incumbent thus solves:

maxa1≥0

[1−Φ

(a1−a1− B

ψλ

h

)]R− c(a1), (A4)

where h :=√

1hθ+ T

heand Φ(·) is the Normal cumulative distribution function. In equilibrium, vot-

ers’ expectations of effort are correct (i.e. a1 = a∗1), implying that effort (in an interior equilibrium)is:66

φ

(− B

hψλ

)R

h= c′(a∗1), (A5)

where φ (·) is the Normal probability distribution function. As previous career concern modelshave shown, effort is thus increasing in rewards (R), the relative responsiveness of voters to signalsof competence (ψ), and the precision of the signal (λ or he).

The preceding analysis generates the following empirical predictions:

Proposition 1. The incumbent is re-elected if H1 ≤ Bψλ

+ b− θ + a∗1, and receives vote share

VI =12 + B + ψλ (b−H1− θ − a∗1). Conditional on the signal, incumbent electoral prospects

decrease in H1.

Proof : follows straightforwardly from the preceding analysis. �The result that incumbents overseeing more homicides are, on average, less likely to be re-

elected reflects two sources of variation in homicide counts. First, as intended by voters, thesanctioning of high homicide counts in part captures cases where the incumbent is relatively less

66As in Ashworth (2005), the second-order condition holds for B≤ c′′(0)(

1hθ

+ The

)√πe2 .

A4

competence (i.e. θI < θ ). Second, because voters cannot fully solve the signal extraction prob-lem, even rational voters willingly in part sanction the incumbent for a high average idiosyncraticshock (e1). In contrast, because voters anticipate incumbent efforts to convince them that they arecompetent, such efforts are both inefficient and ineffective in equilibrium.

While voters respond to homicides induced by both incumbent competence and chance, theseresponses are moderated differently by the number of signals, T . When T is large, the idiosyncraticshocks average out (i.e. limT→∞ e1 → 0), and thus voters extract a clear signal of competence,which also causes the weight attached to the signal to also increase (i.e. ∂λ

∂T > 0). When T is small,voters still update from, and sanction on the basis of, the observed number of homicides, but placerelatively less weight on the less precise signal in casting their ballot. Consequently, variation inhomicides due to et only significantly influences voters that consume a small number of signals(i.e. low T ). In contrast, voters that consume many signals (i.e. high T ) are able to achieve theirgoal of sanctioning principally on the basis of variation in homicides due to θI .

This logic makes clear how political information cycles can influence electoral accountabilityin a career concerns model. To the extent that such cycles induce voters to primarily consume newsjust before the election, voters will rely on a small number of signals. Since random shocks willoften not average out with small T , voters with relatively imprecise prior beliefs will rationallysanction incumbents for the pre-election homicide shocks examined empirically in the main paper.If votes instead receive many signals, homicide shocks—which are shown to be uncorrelated withproxies for incumbent competence—should not influence voting behavior. In this case, longer-termhomicide rates (i.e. the sum of many signals) would heavily influence vote choice.

A.1.2 An incumbent signaling model

I next develop a simple signaling model where the incumbent party responds to an exogenoushomicide shock and voters update their beliefs based on their observation of this response. Themodel demonstrates that where politicians are generally low-quality, circumstances that allow com-petence politicians to reveal their type will reduce the incumbent’s vote share on average. Thismodel contrasts with Ashworth, Bueno de Mesquita and Friedenberg (forthcoming), who gener-ate similar results in a (non-signaling) selection model where random shocks amplify or mute therole of unobserved incumbent competence in the performance outcomes voters use to evaluateincumbents, but incumbents do not act in response to shocks.67

Consider an election campaign where a unit mass of voters individually decide whether to vote

67Ashworth, Bueno de Mesquita and Friedenberg (2017) show that such insights would generalize to a careerconcerns-type model where politicians do not know their types.

A5

to retain an incumbent party I or vote for a challenger party C. The competence θ ∈ {θ ,θ} of bothI and C is only known to the party themselves, where parties are assumed to be unitary actors andθ > θ > 0. From the perspective of voters, Pr(θ ) = q is the common prior belief that any politicianis the more competent type (θ ). Each voter has a net “ideological” bias bi ∈ R toward I (whichis unknown to parties), with the distribution of shocks given by F and expectation E[bi] = b.Voters vote expressively for their preferred candidate, by maximizing the weighted sum of thepolitician’s expected competence and their idiosyncratic bias; i.e. they vote for the incumbent iffE[θI|·]+ωbi≥E[θC|·], where ω > 0 is a common relative weight attached to the ideological bias.

Voters learn about the incumbent’s competence if, with probability p∈ (0,1), a homicide shockoccurs just before the election. In particular, if a homicide shock occurs, I chooses (continuous)action a ≥ 0 to mitigate the security concern at cost a/θ , in order to maximize the differencebetween their vote share and their costs. For example, a could involve speaking to reporters onthe radio or television in an effort to calm voter concerns about the homicide spike, proposingpolicies that the party would implement address security concerns, or convincing voters that thegovernment was not responsible for the violence. For simplicity, if no homicide shock occurs, I

cannot credibly signal their type to vote.When no homicide shock occurs (state N), the expected competence of both parties is θ =

qθ +(1−q)θ . A voter i thus votes for I if bi ≥ 0, and I’s vote share is then V NI = 1−F(0).

When a homicide shock does occur (state S), the incumbent chooses a to maximize their voteshare, given the inferences voters draw from their choice of a. In any pooling equilibrium, in whicha(θ ) = a(θ ), voters will not update and thus behave identically to the case where no homicideshock occurs. In a separately equilibrium, however, voters can learn about I’s competence. Apply-ing the Cho-Kreps intuitive criterion, a θ incumbent will always select a∗(θ ) = 0 in a separatingequilibrium. A θ incumbent will select a > 0 provided that the following individual rationality andincentive compatibility constraints are satisfied:

1−F(

θ −θ

ω

)− a

θ≥ 1−F(0), (A6)

1−F(

θ −θ

ω

)≥ 1−F(0)− a

θ, (A7)

where the first inequality ensures that θ types prefer to separate than pool, and the second inequal-ity ensures that θ types find it too costly to exert the level of a required to mimic θ types. The

A6

incentive constraint will bind in the separating equilibrium, implying that:

a∗(θ ) = θ

[F(

θ −θ

ω

)−F(0)

], (A8)

subject to the rationality constraint holding, such that a∗(θ ) ≤ θ

[F(0)−F

(θ−θ

ω

)]. The space

in which separating equilibria exist is thus defined by θ

θ≥ −

F(0)−F(

θ−θ

ω

)F(0)−F

(θ−θ

ω

) . In such an equilib-

rium, the more competent incumbent’s vote share is V IS (θ ) = 1−F

(θ−θ

ω

)and the less competent

incumbent’s vote share is V IS (θ ) = 1−F

(θ−θ

ω

).

In municipalities where a separating equilibrium would exist if a homicide shock occurred, theincumbent party’s expected vote share declines when:

E[V IS ]−E[V I

N ] < 0⇐⇒ q <F(

q(θ−θ )ω

)−F(0)

F(

q(θ−θ )ω

)−F

(q(θ−θ )−(θ−θ )

ω

) . (A9)

This condition clearly shows that the homicides reducing the incumbent party’s electoral prospectsdepends on the distribution of ideological biases. In particular, homicide shocks reduce the in-cumbent party’s vote share when a large fraction of voters slightly favor the incumbent (relativeto those slightly favoring the challenger) when no shock occurs (i.e. F

(q(θ−θ )

ω

)−F(0) is large).

This is because the electoral effect of better-than-expected news (i.e. when the incumbent demon-strates competence) is greater than the electoral effect of worse-than-expected news (i.e. whenthe incumbent demonstrates incompetence). For this reason, a uniform distribution of ideologicalbiases implies that E[V I

S ]−E[V IN ] = 0.

Similar results could be obtained using variants of this signaling model that are united by theoccurrence of a homicide shock that reveals the incumbent’s type. For example, incumbents couldinvest in their capacity to respond to homicide shocks in advance, enabling voters to separatecompetent from less competent parties when a homicide shock reveals the investment.

A.2 Months and years of municipal elections

Table A1 reports the municipal elections potentially covered by the survey and aggregate electionsin the main analysis.

A7

Table A1: Municipal elections, 1999-2013, by state

State Election dates

Aguascalientes August 2001, August 2004, August 2007, July 2010, July 2013.Baja California June 2001, August 2004, August 2007, July 2010, July 2013.Baja California Sur February 1999, February 2002, February 2005, February 2008, February 2011.Campeche July 2000, July 2003, July 2006, July 2009, July 2012.Chiapas October 2001, October 2004, October 2007, July 2010, July 2012.Chihuahua July 2001, July 2004, July 2007, July 2010, July 2013.Coahuila September 1999, September 2002, September 2005, October 2008, July 2010, July 2013.Colima July 2000, July 2003, July 2006, July 2009, July 2012.Distrito Federal July 2000, July 2003, July 2006, July 2009, July 2012.Durango July 2001, July 2004, July 2007, July 2010, July 2013.Guanajuato July 2000, July 2003, July 2006, July 2009, July 2012.Guerrero October 1999, October 2002, October 2005, October 2008, July 2012.Hidalgo November 1999, November 2002, November 2005, November 2008, July 2011.Jalisco November 2000, July 2003, July 2006, July 2009, July 2012.Estado de Mexico July 2000, March 2003, March 2006, July 2009, July 2012.Michoacan November 2001, November 2004, October 2007, October 2011.Morelos July 2000, July 2003, July 2006, July 2009, July 2012.Nayarit July 1999, July 2002, July 2005, July 2008, July 2011.Nuevo Leon July 2000, July 2003, July 2006, July 2009, July 2012.Oaxaca August 2001, August 2004, August 2007, July 2010, July 2013.Puebla November 2001, November 2004, November 2007, July 2010, July 2013.Queretaro July 2000, July 2003, July 2006, July 2009, July 2012.Quintana Roo February 1999, February 2002, February 2005, February 2008, July 2010, July 2013.San Luis Potosı July 2000, July 2003, July 2006, July 2009, July 2012.Sinaloa November 2001, November 2004, October 2007, July 2010, July 2013.Sonora July 2000, July 2003, July 2006, July 2009, July 2012.Tabasco October 2000, October 2003, October 2006, October 2009, July 2012.Tamaulipas October 2001, November 2004, November 2007, July 2010, July 2013.Tlaxcala November 2001, November 2004, October 2007, July 2010, July 2013.Veracruz September 2000, September 2004, September 2007, July 2010, July 2013.Yucatan May 2001, May 2004, May 2007, May 2010, July 2012.Zacatecas July 2001, July 2004, July 2007, July 2010, July 2013.

Notes: Emboldened election are counted as upcoming local elections in the survey analysis. State-level electionswere held in Hidalgo in February 2002 without concurrent municipal elections, and are counted as upcominglocal elections. Italicized elections are not included in the sample for the homicide shocks analysis due to dataunavailability (or exclusion in the case of the Federal District). Except in the cases of Baja California 2001 and2004 and Oaxaca 2013, missingness reflects the fact that data from the preceding election required to define thechange in vote share was not available. For Baja California 2001 and 2004, the precinct numbering changed acrosselections and thus cannot be matched. For Oaxaca 2013, precinct level data was unavailable.

A8

A.3 Data description

A.3.1 ENCUP survey data

Upcoming local election. Indicator coded 1 for respondents living in a state/municipality with anupcoming local election occurring within the year of the survey. States/municipalities where anelection has already occurred within the year of the survey are coded 0.

Watch and listen to news and political programs ever/monthly/weekly/daily. Indicator coded 1for a respondent that answers that they watch political programs or listen to news at least ever/oncea month/at least once a week/daily. (“¿Que tan seguido escucha noticias o ve programas sobrepolıtica?”)

Watch and listen to news and political programs scale. 5-point scale from 0 to 4, with valuescorresponding to levels of watching and listening to new and political programs (in ascendingorder).

Topical political knowledge. First factor from a factor analysis containing the following ques-tions: What is the name of the youth movement that recently started in Mexico? (2012) Wherewas the plan to build an airport that was subsequently abandoned due to local pressure? (2003,2005) Which political party intends to charge VAT on medicines, food, and tuition? (2001) Whichparty holds your state governorship? (2001, 2003, 2005, 2012) What is the name of your stateGovernor? (2001)

Education. 5-point variable, where 0 is less than completed primary education, 1 is a maximumeducation level of completing high school, 2 is a maximum level of completing lower secondaryeducation, 3 is a maximum level of complete secondary education (preparatoria), and 4 is at leasta university degree.

Homicide shock. This indicator is coded 1 if either the number of homicides in the two monthsprior to the month of the survey (including the survey month) or the two months prior to thesurvey month exceed those in the two months immediately after the month of the survey, basedon the INEGI monthly homicide statistics for the occurrence of homicides among a municipality’sresidents. In 2005, the indicator is coded using the current month if the day of the month is greaterthan the 16th. (Note that homicides figures are subject to reclassification across time.)

Homicides last year. Average number of residents suffering a homicide per month within themunicipality in the preceding 12 months (excluding the current month). (Note that homicidesfigures are subject to reclassification across time.)

Homicides last 3 years. Average number of residents suffering a homicide per month withinthe municipality in the preceding 36 months (excluding the current month). (Note that homicides

A9

figures are subject to reclassification across time.)Public insecurity the major problem in the community. Indicator coded 1 if, in an open re-

sponse, a respondent lists violence, crime or public security as the main problem facing theircommunity (including as 0s respondents that listed no problem).

Low confidence in mayor. Indicator coded 1 for respondents answering that their confidence inthe municipal mayor is 5 or below on a scale from 0 to 10 (for the 2003, 2005, and 2012 surveys) orexpressing no or almost no confidence in the municipal mayor (in 2001), in response to a questionasking about the level of confidence that respondents have in the listed institutions.

Number of organizations. The number of organizations that a respondent reports being a mem-ber of, or previously being a member of. The number of possible organizations slightly variesacross survey.

Organizations talk about politics. Indicator coded 1 for respondents that answer that politics isdiscussed at the organizations they are a member of.

Number of group meetings. The number of political organizations at which an individual hasattended a meeting during the last year.

Discuss community problems. A scale measuring the regularity with which respondents dis-cuss problems in the community with friends and neighbors, ranging through never (coded 0),occasionally (coded 1) and frequently (coded 2).

Incumbent win margin (lag). The difference in vote share between the incumbent and second-placed finisher at the previous municipal mayoral election (or an election held later in the year ofthe survey). In Usos y Custombres in Oaxaca, the incumbent win margin is set to the maximum of1.

ENPV (lag). The effective number of political parties (by vote share) at the previous municipalmayoral election (or an election held later in the year of the survey). In Usos y Custombres inOaxaca, ENPV is set to the maximum of 1.

Incumbent won (lag). Indicator coded 1 for municipalities where the incumbent party wasre-elected at the most recent election (or an election held later in the year of the survey).

Incumbent vote share (lag). The municipal vote share of the incumbent party at the most recentelection (or an election held later in the year of the survey).

Police per voter (lag). Total number of municipal security employees in the previous year (inthousands), divided by the total number of registered voters.

A10

A.3.2 Precinct and municipal homicide and electoral data

Change in incumbent party vote share. Change in the precinct-level share of all votes cast forthe incumbent between the current municipal election and the prior municipal election (3 or 4years earlier). When multiple parties form an incumbent coalition, the incumbent vote share isdetermined by the vote share of the largest party/coalition containing an incumbent party at thenext election in terms of vote share.

Incumbent party win. Indicator coded 1 if the incumbent party wins the municipal election. Inthe case of coalitions, is defined similarly to the above.

Change in turnout. Change in the precinct-level turnout rate between the current municipalelection and the prior municipal election (3 or 4 years earlier).

Change in incumbent vote share (registered). Change in the precinct-level share of votes, as ashare of all registered voters, cast for the incumbent between the current municipal election andthe prior municipal election (3 or 4 years earlier).

Homicide shock. Defined in equation (2) of the main paper, using INEGI homicide statisticsfor intentional homicides that occurred in each month to residents of a given municipality. One-and three-month versions are similarly defined. (Note that homicides figures are subject to reclas-sification across time.)

Placebo homicide shock (6 months earlier). Defined as in equation (2) of the main paper, withthe exception that all months are shifted 6 months forward in time. (Note that homicides figuresare subject to reclassification across time.)

Average monthly homicide rate (12 months/3 years before election). Average number of resi-dents suffering a homicide per month within the municipality in the 12 months/3 years prior to themunicipal election, again based on INEGI homicide data. (Note that homicides figures are subjectto reclassification across time.)

Post-2006. Indicator coded 1 for elections held since December 2006.No municipal police force. Indicator coded 1 for municipalities without a municipal police

force under its direct control. This category includes municipalities that work solely with statepolice or federal police, work with the community, run security using a private or other service,or have no service at all. Municipalities that share police forces or use civil associations wereexcluded because channels of accountability are hard to discern. These categorizations were ho-mogenized across the 2000, 2002, 2004, 2011 and 2013 ENGM surveys. Missing years wereimputed according to the following rules: I first used the last available data, and if no previouscoding was available took the nearest year in the future.

Calderon Presidency. Indicator coded one for elections in the years 2007-2012.

A11

PAN/PRD. Indicator coded 1 for PAN/PRD municipal incumbents.

A.3.3 Precinct local media coverage data

Local media. Number of AM radio, FM radio or television stations, with an emitter based in theprecinct’s municipality, covering at least 20% of the precinct population (as defined by detailedpopulation data—block-level population in urban areas, and rural locality locations).

Local AM radio/FM radio/television. Number of AM radio/FM radio/television stations, withan emitter based in the precinct’s municipal, covering at least 20% of the precinct population(as defined by detailed population data—block-level population in urban areas, and rural localitylocations).

Non-local media. Number of AM radio, FM radio or television stations, with an emitter basedoutside the precinct’s municipality, covering at least 20% of the precinct population (as defined bydetailed population data—block-level population in urban areas, and rural locality locations).

High higher education. Indicator coded 1 for electoral precincts where more than 40% ofresidents had experienced higher education in 2010, according to the 2010 Census.

A.4 Map of municipalities included in different samples

In separate figures, Figure A1 shades in red the municipalities that appear at least once in each ofthe main empirical analyses—the survey sample, the homicide shock sample, and the local mediasample.

A.5 Additional analyses

The following subsections present the results of additional analyses cited in the main paper.

A.5.1 Assessing the identification assumptions

Tables A2-A9 show the results of balance tests for the three main sets of empirical findings in thepaper: Table A2 shows that upcoming local elections are well balanced across individual and mu-nicipal characteristics in the ENCUP survey data; Tables A6 and A7 shows that homicide shocksare well balanced across a wide variety of covariates, where municipality fixed effects are excludedfor time-invariant Census and geographic variables; and Table A9 shows that the number of localmedia stations is well-balanced across these same covariates.

In addition, Tables A3 and A4 show that homicides and their interaction with upcoming localelections are well-balanced in the ENCUP surveys. Moreover, Table A5 demonstrates that neither

A12

(a) Municipalities in the ENCUP survey samples

(b) Municipalities in the homicide shock sample

(c) Municipalities in the local media neighbor sample

Figure A1: Municipalities included in each empirical analysis (shaded in red)

A13

changes in upcoming local elections nor changes in measures of local violence predict whether amunicipality is included in any given year. In each panel the outcome is an indicator for whethera municipality is included in that survey wave, conditional on a municipality appearing in at leastone of the four ENCUP surveys.

I also provide several additional tests to support the exogeneity of homicide shocks. First,Figure A2 shows that the distribution of homicides in months prior to the window used to de-fine homicide shocks does not vary by whether a municipality ultimately experienced a homicideshock. For each set of months, Kolmogorov-Smirnov tests fail to reject the null of equality ofdistribution: respectively, the p-values associated with the combined test for 9 and 10 months be-fore the election, 7 and 8 months before the election, 5 and 6 months before the election, and 3and 4 months before the election are 0.998, 1.000, 0.335, and 0.856. Second, Table A8 showsthat election outcomes are uncorrelated with homicide counts in the two months after elections. Inparticular, columns (3) and (4) find no correlation between the identity of the winning party andpost-election homicides either throughout the sample or during the Calderon administration. Thisdoes not conflict with Dell (2015), who focuses on drug-related homicides over the subsequent

mayor’s term, or as many months as possible of that mayor’s term.

A.5.2 Effects on turnout

Table A11 examines how homicide shocks affect turnout and incumbent vote share (as a proportionof registered voters). The results show that a homicide shock significantly increases turnout atthe 90% level, but barely reduces incumbent vote shares, as a proportion of registered voters.Together, these results suggest that the decline in incumbent vote share (as a proportion of thosethat turned out) principally reflects a rise in turnout for non-incumbent parties. The sample sizes areslightly smaller than reported in the main text because turnout at the previous election is not alwaysavailable, and thus the outcome cannot be constructed for some cases. Column (1) demonstratesthat the overall effect on vote share continues to hold in this subsample.

The results in Table A15 report the effect of drug-related homicide shocks just before an elec-tion, defined according to equation (2) but instead using drug-related homicides. Given that onlyfive years of data exist, and thus many municipalities only appear once, I use state fixed effects in-stead of municipality fixed effects. Although the large drop in sample size unsurprisingly reducesprecision substantially, the point estimates are similar to those reported in Table 2. These findingsreinforce the claim in the main text that voters respond similarly to different types of homicide.

A14

Tabl

eA

2:B

alan

ceof

upco

min

glo

cale

lect

ions

inth

eE

NC

UP

surv

eys

over

18in

divi

dual

and

mun

icip

alle

velv

aria

bles

Fem

ale

Cat

holic

Age

Edu

catio

nE

mpl

oyed

Ow

nec

on.

Num

ber

Org

.sN

umbe

rla

stye

arin

mon

thof

talk

abou

tof

grou

por

g.s

polit

ics

mee

tings

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Upc

omin

glo

cale

lect

ion

-0.0

07-0

.011

-0.1

000.

202*

**-0

.015

-0.0

29-0

.066

-0.0

13-0

.084

(0.0

11)

(0.0

20)

(0.3

94)

(0.0

75)

(0.0

16)

(0.0

21)

(0.0

77)

(0.0

21)

(0.1

03)

Obs

erva

tions

15,9

7611

,983

15,9

7611

,756

11,7

5612

,322

15,9

7612

,576

15,9

76O

utco

me

mea

n0.

550.

8140

.76

1.70

0.47

1.57

1.06

0.26

1.52

Loc

alel

ectio

nm

ean

0.16

0.20

0.16

0.18

0.18

0.20

0.16

0.20

0.16

Dis

cuss

Hom

icid

eH

omic

ide

Hom

icid

eIn

cum

bent

EN

PVIn

cum

bent

Incu

mbe

ntPo

lice

com

mun

itysh

ock

last

year

last

3ye

ars

win

mar

gin

(lag

)w

onvo

tesh

are

perv

oter

prob

lem

s(l

ag)

(lag

)(l

ag)

(las

tyea

r)(1

0)(1

1)(1

2)(1

3)(1

4)(1

5)(1

6)(1

7)(1

8)

Upc

omin

glo

cale

lect

ion

0.00

5-0

.022

1.60

71.

607

0.02

20.

011

0.09

40.

012

-0.0

63(0

.029

)(0

.081

)(1

.080

)(1

.090

)(0

.016

)(0

.066

)(0

.064

)(0

.011

)(0

.178

)

Obs

erva

tions

12,5

7612

,664

15,9

4115

,941

15,9

7615

,976

15,9

7615

,778

13,6

66O

utco

me

mea

n0.

700.

444.

504.

820.

152.

570.

610.

482.

30L

ocal

elec

tion

mea

n0.

200.

170.

160.

160.

160.

160.

160.

170.

16

Not

es:

All

spec

ifica

tions

incl

ude

surv

eyye

arfix

edef

fect

s,an

dar

ees

timat

edus

ing

OL

S.St

anda

rder

rors

clus

tere

dby

mun

icip

ality

inpa

rent

hese

s.*

deno

tes

p<

0.1,

**de

note

sp<

0.05

,

***

deno

tes

p<

0.01

.

A15

Tabl

eA

3:B

alan

ceof

hom

icid

esh

ock

inth

eE

NC

UP

surv

eys

over

18in

divi

dual

and

mun

icip

alle

velv

aria

bles

Fem

ale

Cat

holic

Age

Edu

catio

nE

mpl

oyed

Ow

nec

on.

Num

ber

Org

.sN

umbe

rla

stye

arin

mon

thof

talk

abou

tof

grou

por

g.s

polit

ics

mee

tings

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Hom

icid

esh

ock

-0.0

02-0

.008

-0.2

67-0

.010

-0.0

060.

024

0.02

6-0

.010

0.02

9(0

.007

)(0

.012

)(0

.273

)(0

.053

)(0

.012

)(0

.016

)(0

.064

)(0

.014

)(0

.062

)

Obs

erva

tions

12,6

649,

522

12,6

649,

464

9,46

49,

591

12,6

649,

764

12,6

64O

utco

me

mea

n0.

550.

8140

.58

1.79

0.47

1.57

1.05

0.26

1.45

Hom

icid

esh

ock

mea

n0.

440.

440.

440.

450.

450.

460.

440.

460.

44

Dis

cuss

Upc

omin

gH

omic

ide

Hom

icid

eIn

cum

bent

EN

PVIn

cum

bent

Incu

mbe

ntPo

lice

com

mun

itylo

cal

last

year

last

3ye

ars

win

mar

gin

(lag

)w

onvo

tesh

are

perv

oter

prob

lem

sel

ectio

n(l

ag)

(lag

)(l

ag)

(las

tyea

r)(1

0)(1

1)(1

2)(1

3)(1

4)(1

5)(1

6)(1

7)(1

8)

Hom

icid

esh

ock

-0.0

15-0

.010

-2.3

57*

-3.1

43*

0.00

2-0

.032

-0.0

000.

004

-0.1

39(0

.020

)(0

.036

)(1

.267

)(1

.860

)(0

.010

)(0

.042

)(0

.041

)(0

.006

)(0

.106

)

Obs

erva

tions

9,76

412

,664

12,6

6412

,664

12,6

6412

,664

12,5

2112

,606

10,5

84O

utco

me

mea

n0.

690.

175.

626.

030.

152.

560.

540.

482.

17H

omic

ide

shoc

km

ean

0.46

0.44

0.44

0.44

0.44

0.44

0.44

0.44

0.45

Not

es:

All

spec

ifica

tions

incl

ude

surv

eyye

arfix

edef

fect

s,an

dar

ees

timat

edus

ing

OL

S.St

anda

rder

rors

clus

tere

dby

mun

icip

ality

inpa

rent

hese

s.*

deno

tes

p<

0.1,

**de

note

sp<

0.05

,

***

deno

tes

p<

0.01

.

A16

Tabl

eA

4:In

tera

ctiv

eba

lanc

eof

upco

min

glo

cale

lect

ion

and

hom

icid

esh

ock

inth

eE

NC

UP

surv

eys

over

17in

divi

dual

and

mun

icip

alle

velv

aria

bles

Fem

ale

Cat

holic

Age

Edu

catio

nE

mpl

oyed

Ow

nec

on.

Num

ber

Org

.sN

umbe

rla

stye

arin

mon

thof

talk

abou

tof

grou

por

g.s

polit

ics

mee

tings

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Upc

omin

glo

cale

lect

ion

-0.0

07-0

.028

-0.4

580.

156

-0.0

07-0

.012

-0.0

76-0

.039

-0.0

92(0

.018

)(0

.031

)(0

.595

)(0

.116

)(0

.025

)(0

.034

)(0

.108

)(0

.028

)(0

.123

)H

omic

ide

shoc

k-0

.005

-0.0

17-0

.307

-0.0

15-0

.003

0.03

3*0.

060

-0.0

080.

068

(0.0

08)

(0.0

13)

(0.3

19)

(0.0

61)

(0.0

14)

(0.0

20)

(0.0

68)

(0.0

15)

(0.0

60)

Hom

icid

esh

ock

0.02

20.

042

0.21

80.

023

-0.0

14-0

.043

-0.2

15*

-0.0

14-0

.249

×U

pcom

ing

loca

lele

ctio

n(0

.023

)(0

.032

)(0

.859

)(0

.160

)(0

.032

)(0

.047

)(0

.128

)(0

.032

)(0

.188

)

Obs

erva

tions

12,6

649,

522

12,6

649,

464

9,46

49,

591

12,6

649,

764

12,6

64O

utco

me

mea

n0.

550.

8140

.58

1.79

0.47

1.57

1.05

0.26

1.45

Upc

omin

glo

cale

lect

ion

mea

nH

omic

ide

shoc

km

ean

Dis

cuss

Upc

omin

gH

omic

ide

Hom

icid

eIn

cum

bent

EN

PVIn

cum

bent

Incu

mbe

ntPo

lice

com

mun

itylo

cal

last

year

last

3ye

ars

win

mar

gin

(lag

)w

onvo

tesh

are

perv

oter

prob

lem

sel

ectio

n(l

ag)

(lag

)(l

ag)

(las

tyea

r)(1

0)(1

1)(1

2)(1

3)(1

4)(1

5)(1

6)(1

7)

Upc

omin

glo

cale

lect

ion

-0.0

31-0

.429

-1.0

970.

056*

*0.

071

0.15

9*0.

008

-0.1

57(0

.047

)(1

.504

)(1

.748

)(0

.023

)(0

.095

)(0

.091

)(0

.017

)(0

.257

)H

omic

ide

shoc

k-0

.009

-3.1

10*

-4.1

39*

0.00

9-0

.018

0.02

00.

003

-0.1

76(0

.025

)(1

.585

)(2

.267

)(0

.012

)(0

.053

)(0

.048

)(0

.008

)(0

.123

)H

omic

ide

shoc

k-0

.031

4.61

0**

6.06

7**

-0.0

36-0

.085

-0.1

100.

004

0.21

0×

Upc

omin

glo

cale

lect

ion

(0.0

56)

(2.2

56)

(2.8

16)

(0.0

35)

(0.1

26)

(0.1

39)

(0.0

25)

(0.3

32)

Obs

erva

tions

9,76

412

,664

12,6

6412

,664

12,6

6412

,521

12,6

0610

,584

Out

com

em

ean

0.69

5.62

6.03

0.15

2.56

0.54

0.48

2.17

Upc

omin

glo

cale

lect

ion

mea

nH

omic

ide

shoc

km

ean

Not

es:

All

spec

ifica

tions

incl

ude

surv

eyye

arfix

edef

fect

s,an

dar

ees

timat

edus

ing

OL

S.St

anda

rder

rors

clus

tere

dby

mun

icip

ality

inpa

rent

hese

s.*

deno

tes

p<

0.1,

**de

note

sp<

0.05

,

***

deno

tes

p<

0.01

.

A17

Table A5: Predictors of municipalities included in each survey wave

Panel A: Linear predictors Surveyed municipality indicator(1) (2) (3) (4) (5)

Local election 0.016(0.037)

Homicide within last month -0.020(0.039)

Homicide shock 0.036(0.039)

Homicides last year -0.001(0.001)

Homicides last 3 years -0.000(0.001)

Observations 2,092 2,088 1,351 2,088 2,088Outcome mean 0.46 0.46 0.52 0.46 0.46

Panel B: Election-homicide interactions Surveyed municipality indicator(1) (2) (3) (4)

Local election 0.032 0.050 0.021 0.020(0.055) (0.075) (0.042) (0.042)

Homicide within last month -0.017(0.040)

Local election × Homicide within last month -0.029(0.070)

Homicide shock 0.052(0.043)

Local election × Homicide shock -0.124(0.102)

Homicides last year -0.001(0.001)

Local election × Homicides last year -0.003(0.004)

Homicides last 3 years -0.000(0.001)

Local election × Homicides last 3 years -0.002(0.004)

Observations 2,088 1,351 2,088 2,088Outcome mean 0.46 0.52 0.46 0.46

Notes: All specifications include municipality and survey-year fixed effects, and are estimated using OLS. Stan-dard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotesp < 0.01.

A18

Table A6: Municipal-level balance on 99 variables over pre-election homicide shocks

Outcome Effect of homicide shock Outcome Effect of homicide shockCoef. SE Obs. Coef. SE Obs.

Electorate 5,003.795 (4,685.159) 2,852 Victims with complete primary schooling before election 0.874 (6.664) 1,658Electorate density (log) -0.010 (0.010) 2,837 Victims with incomplete secondary schooling before election 1.270 (4.129) 1,658Incumbent vote share (lag) 0.000 (0.007) 2,815 Victims with complete secondary schooling before election 0.943 (4.001) 1,658Incumbent win (lag) 0.040 (0.042) 2,332 Single victims before election -0.037 (3.442) 1,698Win margin (lag) 0.004 (0.011) 2,815 Job-related homicides before election 0.934 (1.004) 1,217ENPV (lag) -0.021 (0.037) 2,852 Organs examined after homicide before election 2.247 (6.832) 1,562Turnout (lag) -0.017** (0.008) 2,644 Family-incident homicides before election 0.050 (0.061) 1,737PRI incumbent -0.076* (0.039) 2,852 Urban homicide victims before election 1.245 (7.528) 1,385PAN incumbent 0.053 (0.040) 2,852 Non-Mexican homicide victims before election -0.031 (0.114) 1,792PRD incumbent 0.024 (0.030) 2,852 Neighbor average homicide shock 0.004 (0.027) 2,739Mayor’s party aligned with President 0.070* (0.037) 2,852 Average non-homicide deaths (prior 3 years) -2.700 (2.020) 2,852Mayor’s party aligned with Governor -0.082** (0.041) 2,852 Average non-homicide deaths (prior year) -4.073* (2.248) 2,852Mayor’s party aligned with President and Governor 0.008 (0.023) 2,852 Average child mortalities (prior 3 years) -0.051 (0.258) 2,124Total municipal spending 23.572 (66.800) 1,759 Average child mortalities (prior year) 0.180 (0.028) 2,124Police per voter 0.054 (0.070) 2,852 Area (km2) 417.555* (233.087) 2,852Homicides 12 months before election -2.130 (2.310) 2,852 Local media -1.205 (1.217) 2,849Homicides 11 months before election -1.628 (2.822) 2,852 Non-local media 0.005 (1.350) 2,849Homicides 10 months before election -2.336 (2.644) 2,852 Occupants per room 0.004 (0.013) 2,826Homicides 9 months before election -2.197 (2.546) 2,852 Share with 2 bedrooms -0.009 (0.008) 2,826Homicides 8 months before election -2.988 (2.031) 2,852 Share 3+ bedrooms -0.008 (0.008) 2,826Homicides 7 months before election -2.085 (2.210) 2,852 Share female 0.000 (0.001) 2,826Homicides 6 months before election -2.004 (2.400) 2,852 Share working age 0.000 (0.002) 2,826Homicides 5 months before election -1.542 (1.721) 2,852 Children per woman 0.000 (0.019) 2,826Homicides 4 months before election -0.775 (2.176) 2,852 Female years of schooling 0.003 (0.086) 2,826Homicides 3 months before election -1.066 (2.272) 2,852 Male years of schooling -0.010 (0.092) 2,826Average homicides (prior 3 years) -0.763 (1.466) 2,852 Share illiterate 0.001 (0.003) 2,826Average homicides (prior year) -1.875 (2.258) 2,852 Share economically active -0.003 (0.003) 2,826Top homicide quartile (prior year) 0.007 (0.033) 2,852 Share born out of state -0.003 (0.015) 2,826Top homicide decile (prior year) 0.006 (0.021) 2,852 Share Catholic 0.001 (0.005) 2,826Presence of DTO 0.006 (0.031) 2,017 Share indigenous speakers 0.002 (0.003) 2,826Average drug-related homicides (prior year) 0.006 (0.011) 2,852 Share without health care 0.000 (0.006) 2,826Reports of inter-DTO violence before election -0.494 (0.817) 1,924 Share with electricity, running water, and drainage -0.010 (0.009) 2,826Reports of violent enforcement before election 0.026 (0.108) 1,924 Share washing machine -0.012 (0.010) 2,826Reports of arrests before election 0.083 (0.772) 1,924 Share landline telephone -0.012 (0.015) 2,826Reports of drug seizures before election 0.575 (1.073) 1,924 Share radio -0.004 (0.007) 2,826Reports of asset seizures before election -0.340 (0.262) 1,924 Share fridge -0.006 (0.008) 2,826Reports of gun seizures before election 0.464 (0.450) 1,924 Share cell phone -0.004 (0.010) 2,826Gun-related homicides before election -0.261 (5.383) 2,852 Share television -0.002 (0.003) 2,826Chemical substance-related homicides before election -0.028 (0.017) 2,852 Share car or truck -0.005 (0.010) 2,826Hanging-related homicides before election 0.137 (0.198) 2,852 Share computer -0.006 (0.011) 2,826Drowning-related homicides before election 0.111*** (0.039) 2,852 Share internet -0.005 (0.010) 2,826Explosives-related homicides before election 0.005 (0.005) 2,852 Nonhomicide crimes shock 0.126 (0.086) 591Smoke/fire-related homicides before election 0.033 (0.032) 2,852 Sexual assault shock 0.140 (0.101) 338Cutting object-related homicides before election 0.648*** (0.221) 2,852 Kidnapping shock -0.191 (0.150) 104Blunt object-related homicides before election 0.154* (0.080) 2,852 Robbery shock 0.127 (0.082) 556Delayed homicide registrations before election 1.755 (5.769) 2,852 Carjacking shock 0.168 (0.124) 140Male homicide victims before election 1.663 (6.323) 1,792 Criminal injury shock 0.073 (0.087) 517Age of homicide victims before election 1.529*** (0.584) 1,749 Property theft shock -0.055 (0.090) 531Victims with no schooling before election 0.322 (0.288) 1,658 Other crime shock 0.229*** (0.081) 576Victims with incomplete primary schooling before election 1.322 (7.061) 1,658

Notes: Specifications include municipality and year fixed effects (unless otherwise stated), and are estimated using OLS. The time-

invariant variables—area, media coverage, and the 2010 Census variables (occupants per dwelling-share internet)—and pre-election shocks

regarding other types of crime (data are only available from 2011 onwards) exclude municipality fixed effects. Standard errors clustered

by municipality in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A19

Table A7: Precinct-level balance over 105 variables by pre-election homicide shocks

Outcome Effect of homicide shock Outcome Effect of homicide shockCoef. SE Obs. Coef. SE Obs.

Electorate 81.500 (53.762) 167,116 Age of homicide victims before election 1.4822** (0.581) 144,229Municipal electorate 5,088.790 (4,737.408) 167,116 Victims with no schooling before election 0.320 (0.288) 141,789Electorate density (log) -0.005 (0.007) 164,722 Victims with incomplete primary schooling before election 1.282 (7.135) 141,789Incumbent vote share (lag) 0.000 (0.007) 167,116 Victims with complete primary schooling before election 0.834 (6.732) 141,789Municipal-level incumbent vote share (lag) 0.000 (0.007) 166,238 Victims with incomplete secondary schooling before election 1.253 (4.169) 141,789Incumbent win (lag) 0.042 (0.042) 139,407 Victims with complete secondary schooling before election 0.921 (4.040) 141,789Win margin (lag) 0.001 (0.009) 167,116 Single victims before election -0.063 (3.483) 142,988Municipal-level win margin (lag) 0.004 (0.011) 166,238 Job-related homicides before election 0.935 (1.005) 126,754Municipal-level ENPV (lag) -0.019 (0.032) 167,113 Organs examined after homicide before election 2.216 (6.894) 138,468ENPV (lag) -0.020 (0.036) 167,116 Family-incident homicides before election 0.050 (0.061) 142,454Turnout (lag) -0.0178** (0.007) 157,893 Urban homicide victims before election 1.178 (7.617) 123,125Municipal-level turnout (lag) -0.0175** (0.007) 157,290 Non-Mexican homicide victims before election -0.034 (0.115) 145,210PRI incumbent -0.0741* (0.039) 167,116 Neighbor average homicide shock 0.003 (0.027) 163,862PAN incumbent 0.053 (0.039) 167,116 Average non-homicide deaths (prior 3 years) -2.746 (2.025) 167,116PRD incumbent 0.023 (0.030) 167,116 Average non-homicide deaths (prior year) -4.1205* (2.249) 167,116Mayor’s party aligned with President 0.0695* (0.037) 167,116 Average child mortalities (prior 3 years) -0.047 (0.258) 134,190Mayor’s party aligned with Governor -0.0848** (0.040) 167,116 Average child mortalities (prior year) 0.183 (0.282) 134,190Mayor’s party aligned with President and Governor 0.006 (0.023) 167,116 Area (km2) 4.6625* (2.539) 164,722Total municipal spending 23.217 (66.622) 111,360 Local media -1.250 (1.217) 165,079Police per voter 0.058 (0.070) 167,116 Non-local media 0.103 (1.368) 165,079Homicides 12 months before election -2.117 (2.330) 167,116 Occupants per room 0.004 (0.013) 166,783Homicides 11 months before election -1.632 (2.854) 167,116 Share with 2 bedrooms -0.007 (0.008) 166,653Homicides 10 months before election -2.338 (2.672) 167,116 Share 3+ bedrooms -0.007 (0.008) 166,653Homicides 9 months before election -2.217 (2.577) 167,116 Share female 0.000 (0.001) 166,664Homicides 8 months before election -2.996 (2.052) 167,116 Share working age 0.000 (0.002) 166,664Homicides 7 months before election -2.100 (2.229) 167,116 Children per woman 0.002 (0.018) 166,783Homicides 6 months before election -2.013 (2.418) 167,116 Female years of schooling -0.001 (0.090) 166,783Homicides 5 months before election -1.535 (1.737) 167,116 Male years of schooling -0.015 (0.097) 166,783Homicides 4 months before election -0.766 (2.199) 167,116 Share illiterate 0.001 (0.003) 166,653Homicides 3 months before election -1.050 (2.294) 167,116 Share economically active -0.003 (0.003) 166,664Average homicides (prior 3 years) -0.763 (1.480) 167,116 Share born out of state -0.004 (0.013) 166,664Average homicides (prior year) -1.876 (2.282) 167,116 Share Catholic 0.002 (0.005) 166,664Top homicide quartile (prior year) 0.007 (0.033) 167,116 Share indigenous speakers 0.002 (0.003) 166,653Top homicide decile (prior year) 0.006 (0.021) 167,116 Share without health care -0.001 (0.006) 166,664Presence of DTO 0.006 (0.031) 131,039 Share with electricity, running water, and drainage -0.010 (0.009) 166,653High-risk electoral precinct 0.002 (0.003) 50,392 Share washing machine -0.011 (0.010) 166,653Average drug-related homicides (prior year) 0.006 (0.011) 167,116 Share landline telephone -0.012 (0.014) 166,653Reports of inter-DTO violence before election -0.481 (0.827) 126,996 Share radio -0.004 (0.006) 166,653Reports of violent enforcement before election 0.031 (0.109) 126,996 Share fridge -0.006 (0.008) 166,653Reports of arrests before election 0.111 (0.786) 126,996 Share cell phone -0.004 (0.010) 166,653Reports of drug seizures before election 0.619 (10.900) 126,996 Share television -0.002 (0.003) 166,653Reports of asset seizures before election -0.329 (0.264) 126,996 Share car or truck -0.005 (0.010) 166,653Reports of gun seizures before election 0.488 (0.458) 126,996 Share computer -0.007 (0.011) 166,653Gun-related homicides before election -0.257 (5.449) 167,116 Share internet -0.006 (0.010) 166,653Chemical substance-related homicides before election -0.0286* (0.017) 167,116 Nonhomicide crimes shock 0.126 (0.088) 32,770Hanging-related homicides before election 0.130 (0.198) 167,116 Sexual assault shock 0.134 (0.103) 25,513Drowning-related homicides before election 0.1107*** (0.039) 167,116 Kidnapping shock -0.176 (0.148) 13,005Explosives-related homicides before election 0.005 (0.005) 167,116 Robbery shock 0.142* (0.085) 32,558Smoke/fire-related homicides before election 0.032 (0.032) 167,116 Carjacking shock 0.180 (0.128) 7,450Cutting object-related homicides before election 0.6499*** (0.223) 167,116 Criminal injury shock 0.071 (0.089) 31,818Blunt object-related homicides before election 0.1557* (0.080) 167,116 Property theft shock -0.075 (0.091) 31,608Delayed homicide registrations before election 1.752 (5.840) 167,116 Other crime shock 0.245*** (0.082) 32,738Male homicide victims before election 1.626 (6.395) 145,210

Notes: Specifications include municipality and year fixed effects, and are estimated using OLS. The time-invariant variables—area, media

coverage, and the 2010 Census variables (occupants per dwelling-share internet)—exclude municipality fixed effects. Standard errors

clustered by municipality in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A20

0.0

5.1

.15

.2D

ensi

ty

0 5 10 15 20 25Average number of homicides per month

Months 9 and 10 before election (shocked) Months 9 and 10 before election (not shocked)

Months 7 and 8 before election (shocked) Months 7 and 8 before election (not shocked)

Months 5 and 6 before election (shocked) Months 5 and 6 after election (not shocked)

Months 3 and 4 before election (shocked) Months 3 and 4 after election (not shocked)

Figure A2: Distribution of pre-election homicides, by homicide shock status

A21

Table A8: Correlation between election outcomes and post-election homicides

Average homicides in the twomonths after an election

(1) (2) (3) (4)

Incumbent party win 4.309(7.287)

Change in incumbent party vote share -7.643(6.249)

PAN win 0.134 -19.256(4.931) (14.469)

PRI win 1.151 1.867(3.805) (10.776)

PRD win -8.705 -13.488(8.283) (29.500)

Observations 2,852 173,389 2,852 975Unique municipalities 906 1,329 906 484Outcome mean 19.54 18.99 19.54 30.33Election outcome mean 0.55 -0.05

Notes: All specifications include municipality and year fixed effects, and are estimated using OLS. Standard errorsclustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A.5.3 Additional robustness checks for the aggregate electoral results

Table A12 shows that the information consumption results are robust to defining upcoming localelections by any number of months between 1 and 12 before the election.

A.5.4 Additional robustness checks for the aggregate electoral results

Table A13 reports the results from Table 2 using incumbent vote share instead of the difference inincumbent vote share with respect to the previous election. The results do not meaningfully differ.Table A14 similarly shows that the main results are robust to including municipalities without apolice force.

A.5.5 Electoral effects of drug-related homicides

As noted in the main text, the Calderon government released monthly municipal data on the numberof drug-related deaths between 2007 and 2011. However, there are data issues with such data.First, these numbers are contentious (see Heinle, Rodrıguez Ferreira and Shirk 2014), and do not

A22

Table A9: Precinct-level balance over 92 variables by the number of local media stations

Outcome Effect of local media Outcome Effect of local mediaCoef. SE Obs. Coef. SE Obs.

Homicide shock 0.000 (0.000) 29,736 Gun-related homicides before election 0.009 (0.009) 29,736Area (km2) -0.003 (0.006) 29,736 Chemical substance-related homicides before election 0.000 (0.000) 29,736Electorate -28.851* (15.100) 29,736 Hanging-related homicides before election 0.000 (0.001) 29,736Municipal electorate 30.425 (25.718) 29,736 Drowning-related homicides before election 0.000 (0.000) 29,736Electorate density (log) -0.012 (0.009) 29,701 Explosives-related homicides before election NA NA 29,736Distance from centroid to municipal head (log) -0.005 (0.003) 29,736 Smoke/fire-related homicides before election 0.000 (0.000) 29,736Non-local media -0.031 (0.239) 29,736 Cutting object-related homicides before election 0.000 (0.001) 29,736Incumbent vote share (lag) -0.000 (0.000) 29,736 Blunt object-related homicides before election 0.000 (0.000) 29,736Incumbent win (lag) -0.000 (0.000) 24,186 Delayed homicide registrations before election 0.010 (0.010) 29,736Win margin (lag) -0.000 (0.001) 29,736 Male homicide victims before election 0.009 (0.009) 28,181Municipality win margin (lag) 0.000 (0.000) 29,736 Age of homicide victims before election 0.246 (0.304) 27,930ENPV (lag) 0.001 (0.002) 29,736 Victims with no schooling before election -0.000 (0.001) 27,931Municipality ENPV (lag) 0.000 (0.000) 29,736 Victims with incomplete primary schooling before 0.007 (0.010) 27,931Turnout (lag) 0.001 (0.000) 28,779 Victims with complete primary schooling before election 0.006 (0.009) 27,931Municipality turnout (lag) -0.000** (0.000) 28,819 Victims with secondary schooling before election 0.004 (0.006) 27,931PRI incumbent 0.000 (0.000) 29,736 Single victims before election 0.004 (0.004) 28,010PAN incumbent -0.000 (0.000) 29,736 Job-related homicides before election 0.001 (0.001) 26,281PRD incumbent -0.000 (0.000) 29,736 Organs examined after homicide before election 0.009 (0.010) 27,631Mayor’s party aligned with President 0.000 (0.000) 29,736 Family-incident homicides before election 0.000 (0.000) 27,486Mayor’s party aligned with Governor -0.000 (0.000) 29,736 Urban homicide victims before election 0.003 (0.008) 24,715Mayor’s party aligned with President and Governor 0.000 (0.000) 29,736 Non-Mexican homicide victims before election 0.000 (0.000) 28,181Total municipal spending 0.236 (0.271) 19,842 Neighbor average homicide shock 0.000 (0.000) 28,608Police per voter 0.001** (0.000) 29,736 Average non-homicide deaths (prior 3 years) -0.027 (0.018) 29,736Homicides 12 months before election 0.009 (0.008) 29,736 Average non-homicide deaths (prior year) -0.026 (0.018) 29,736Homicides 11 months before election 0.009 (0.011) 29,736 Average child mortalities (prior 3 years) 0.000 (0.001) 23,080Homicides 10 months before election 0.002 (0.006) 29,736 Average child mortalities (prior year) -0.000 (0.001) 23,080Homicides 9 months before election 0.007 (0.007) 29,736 Occupants per room -0.004** (0.002) 29,736Homicides 8 months before election 0.005 (0.006) 29,736 Share with 2 bedrooms 0.000 (0.001) 29,736Homicides 7 months before election -0.004 (0.009) 29,736 Share 3+ bedrooms 0.001 (0.001) 29,736Homicides 6 months before election 0.000 (0.010) 29,736 Share female 0.000 (0.000) 29,736Homicides 5 months before election 0.003 (0.009) 29,736 Share working age -0.000 (0.000) 29,736Homicides 4 months before election 0.004 (0.007) 29,736 Children per woman -0.003 (0.002) 29,736Homicides 3 months before election 0.005 (0.006) 29,736 Female years of schooling 0.014 (0.011) 29,736Average homicides (prior 3 years) 0.005 (0.005) 29,736 Male years of schooling 0.023** (0.012) 29,736Average homicides (prior year) 0.004 (0.008) 29,736 Share illiterate -0.000** (0.000) 29,736Top homicide quartile (prior year) -0.000 (0.000) 29,736 Share economically active 0.000 (0.000) 29,736Top homicide decile (prior year) -0.000 (0.000) 29,736 Share born out of state 0.001 (0.001) 29,736Average drug-related homicides (prior year) 0.000 (0.000) 29,736 Share Catholic 0.000 (0.001) 29,736Presence of DTO -0.000 (0.000) 22,764 Share indigenous speakers 0.000 (0.000) 29,736High-risk electoral precinct 0.002 (0.002) 7,370 Share without health care -0.000 (0.000) 29,736Reports of inter-DTO violence before election -0.001 (0.002) 22,048 Share with electricity, running water, and drainage 0.002* (0.001) 29,736Reports of violent enforcement before election -0.000 (0.000) 22,048 Share washing machine 0.001 (0.001) 29,736Reports of arrests before election 0.002 (0.002) 22,048 Share fridge 0.001 (0.001) 29,736Reports of drug seizures before election 0.001 (0.002) 22,048 Share cell phone 0.000 (0.001) 29,736Reports of asset seizures before election -0.001 (0.001) 22,048 Share car or truck 0.001 (0.001) 29,736Reports of gun seizures before election 0.000 (0.001) 22,048 Share computer 0.002* (0.001) 29,736

Notes: All specifications include neighbor group and year fixed effects, and are estimated using OLS. All observations are weighted by the

number of registered voters divided by he number of comparison units within each neighbor group. There is no variation in the number of

explosive-related homicides or non-homicide pre-election crime shocks (for which overlapping data are only available for one election) across

local media in this sample. Standard errors clustered by municipality in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes

p < 0.01.

A23

Table A10: Precinct-level balance over 92 variables by pre-election homicide shocks (in the localmedia sample)

Outcome Effect of homicide shock Outcome Effect of homicide shockCoef. SE Obs. Coef. SE Obs.

Local media 0.005 (0.005) 29,736 Gun-related homicides before election -6.266 (14.269) 29,736Area (km2) 0.001 (0.002) 29,736 Chemical substance-related homicides before election -0.014 (0.022) 29,736Electorate 38.419 (47.667) 29,736 Hanging-related homicides before election -0.301 (0.542) 29,736Municipal electorate 17,743.371** (6,872.275) 29,736 Drowning-related homicides before election 0.239** (0.097) 29,736Electorate density (log) 0.014 (0.017) 29,701 Explosives-related homicides before election NA NA 29,736Distance from centroid to municipal head (log) 0.000 (0.001) 29,736 Smoke/fire-related homicides before election 0.046 (0.075) 29,736Non-local media 0.001 (0.002) 29,736 Cutting object-related homicides before election 1.378** (0.598) 29,736Incumbent vote share (lag) -0.004 (0.016) 29,736 Blunt object-related homicides before election 0.249 (0.205) 29,736Incumbent win (lag) -0.115 (0.101) 24,186 Delayed homicide registrations before election -3.705 (15.418) 29,736Win margin (lag) 0.012 (0.020) 29,736 Male homicide victims before election -4.803 (14.482) 28,181Municipality win margin (lag) 0.024 (0.021) 29,736 Age of homicide victims before election -88.070 (496.152) 27,930ENPV (lag) -0.022 (0.051) 29,736 Victims with no schooling before election 0.912 (0.598) 27,931Municipality ENPV (lag) -0.013 (0.053) 29,736 Victims with incomplete primary schooling before -4.573 (15.258) 27,931Turnout (lag) -0.002 (0.010) 28,779 Victims with complete primary schooling before election -5.087 (14.354) 27,931Municipality turnout (lag) -0.003 (0.010) 28,819 Victims with secondary schooling before election -2.706 (8.637) 27,931PRI incumbent 0.018 (0.064) 29,736 Single victims before election -3.689 (7.481) 28,010PAN incumbent -0.008 (0.060) 29,736 Job-related homicides before election 1.929 (2.873) 26,281PRD incumbent 0.000 (0.037) 29,736 Organs examined after homicide before election 0.391 (14.330) 27,631Mayor’s party aligned with President 0.046 (0.062) 29,736 Family-incident homicides before election 0.008 (0.155) 27,486Mayor’s party aligned with Governor -0.051 (0.065) 29,736 Urban homicide victims before election -3.839 (17.133) 24,715Mayor’s party aligned with President and Governor 0.003 (0.018) 29,736 Non-Mexican homicide victims before election -0.180 (0.313) 28,181Total municipal spending 253.374*** (79.149) 19,842 Neighbor average homicide shock -0.008 (0.063) 28,608Police per voter 0.279* (0.164) 29,736 Average non-homicide deaths (prior 3 years) -4.047 (4.420) 29,736Homicides 12 months before election -7.635 (6.711) 29,736 Average non-homicide deaths (prior year) -8.508 (5.308) 29,736Homicides 11 months before election -6.272 (8.528) 29,736 Average child mortalities (prior 3 years) -0.439 (0.544) 23,080Homicides 10 months before election -7.599 (7.421) 29,736 Average child mortalities (prior year) 0.389 (0.423) 23,080Homicides 9 months before election -6.437 (7.299) 29,736 Occupants per room 0.000* (0.000) 29,736Homicides 8 months before election -7.448 (5.702) 29,736 Share with 2 bedrooms -0.000** (0.000) 29,736Homicides 7 months before election -0.980 (7.790) 29,736 Share 3+ bedrooms -0.001*** (0.000) 29,736Homicides 6 months before election -3.042 (8.420) 29,736 Share female -0.000 (0.000) 29,736Homicides 5 months before election -7.166 (5.596) 29,736 Share working age -0.000** (0.000) 29,736Homicides 4 months before election -4.718 (6.179) 29,736 Children per woman -0.000 (0.001) 29,736Homicides 3 months before election -3.315 (6.356) 29,736 Female years of schooling -0.002 (0.003) 29,736Average homicides (prior 3 years) -3.308 (4.052) 29,736 Male years of schooling -0.001 (0.004) 29,736Average homicides (prior year) -5.461 (6.719) 29,736 Share illiterate 0.000 (0.000) 29,736Top homicide quartile (prior year) 0.035 (0.045) 29,736 Share economically active -0.000** (0.000) 29,736Top homicide decile (prior year) 0.035 (0.045) 29,736 Share born out of state -0.000 (0.000) 29,736Average drug-related homicides (prior year) 0.013 (0.017) 29,736 Share Catholic -0.000** (0.000) 29,736Presence of DTO -0.049 (0.061) 22,764 Share indigenous speakers -0.000 (0.000) 29,736High-risk electoral precinct -0.016 (0.011) 7,370 Share without health care 0.000 (0.000) 29,736Reports of inter-DTO violence before election -1.724 (1.653) 22,048 Share with electricity, running water, and drainage 0.000 (0.000) 29,736Reports of violent enforcement before election -0.075 (0.242) 22,048 Share washing machine -0.000 (0.000) 29,736Reports of arrests before election 0.370 (2.089) 22,048 Share fridge 0.000 (0.000) 29,736Reports of drug seizures before election 4.106 (4.027) 22,048 Share cell phone -0.000 (0.000) 29,736Reports of asset seizures before election -0.190 (0.441) 22,048 Share car or truck 0.000 (0.000) 29,736Reports of gun seizures before election -0.025 (0.602) 22,048 Share computer -0.000 (0.000) 29,736

Notes: All specifications include neighbor group and year fixed effects, and are estimated using OLS. All observations are weighted by the

number of registered voters divided by he number of comparison units within each neighbor group. There is no variation in the number of

explosive-related homicides or non-homicide pre-election crime shocks (for which overlapping data are only available for one election) across

local media in this sample. Standard errors clustered by municipality in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes

p < 0.01.

A24

Table A11: Effect of pre-election homicide shocks on precinct-level turnout

Change in Change in Change inincumbent turnout incumbent

party partyvote share vote share(share of (share ofturnout) registered voters)

(1) (2)

Homicide shock -0.022** 0.016* -0.004(0.009) (0.009) (0.005)

Observations 157,893 157,893 157,893Unique municipalities 905 905 905Outcome range [-0.96,0.89] [-0.76,0.67] [-0.93,0.81]Outcome mean -0.06 0.00 -0.03Homicide shock mean 0.42 0.42 0.42

Notes: All specifications are estimated using OLS, and include state and year fixed effects. All observations areweighted by the number of registered voters. Standard errors clustered by municipality are in parentheses. *denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

necessarily follow the homogeneous coroner’s report criteria used by INEGI. Second, the limitedavailability of this data, combined with the definition of a homicide shock that requires at least onedrug-related death over the four-month window around elections, substantially reduces the samplesize by around 75%. More generally, it is not clear theoretically whether voters should respondmore or less to drug-related homicides, as opposed to other homicides. In fact, voters might thinkthat these are less relevant to them than more arbitrary murders which constitute around 50% oftotals homicides even during the drug war (only a tiny fraction of such homicides as domestic).Nevertheless, I use the drug-related homicide data as a robustness check.

The results in Table A15 report the effect of drug-related homicide shocks just before an elec-tion, defined according to equation (2) but instead using drug-related homicides. Given that onlyfive years of data exist, and thus many municipalities only appear once, I use state fixed effects in-stead of municipality fixed effects. Although the large drop in sample size unsurprisingly reducesprecision substantially, the point estimates are similar to those reported in Table 2. These findingsreinforce the claim in the main text that voters respond similarly to different types of homicide.

A25

Tabl

eA

12:T

heef

fect

ofup

com

ing

loca

lele

ctio

nson

polit

ical

new

sco

nsum

ptio

nan

dto

pica

lpol

itica

lkno

wle

dge,

bynu

mbe

rof

mon

ths

befo

reth

eel

ectio

n

Num

bero

fmon

ths

befo

reel

ectio

nus

edto

defin

ean

upco

min

glo

cale

lect

ion

1m

onth

2m

onth

s3

mon

ths

4m

onth

s5

mon

ths

6m

onth

s7

mon

ths

8m

onth

s9

mon

ths

10m

onth

s11

mon

ths

12m

onth

s(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)(1

0)(1

1)(1

2)

Pane

lA:O

utco

me:

Wat

chan

dlis

ten

tone

wsa

ndpo

litic

alpr

ogra

mse

ver

Upc

omin

glo

cale

lect

ion

0.09

3***

0.09

3***

0.06

7***

0.06

7***

0.04

9***

0.04

9***

0.05

9***

0.06

1***

0.06

1***

0.06

2***

0.04

4***

0.04

4***

(0.0

15)

(0.0

15)

(0.0

13)

(0.0

13)

(0.0

14)

(0.0

14)

(0.0

13)

(0.0

13)

(0.0

13)

(0.0

13)

(0.0

10)

(0.0

10)

Obs

erva

tions

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

Mun

icip

aliti

es53

953

953

953

953

953

953

953

953

953

953

953

9O

utco

me

rang

e{0

,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

Out

com

em

ean

0.87

0.87

0.87

0.87

0.87

0.87

0.87

0.87

0.87

0.87

0.87

0.87

Loc

alel

ectio

nm

ean

0.02

0.02

0.06

0.06

0.19

0.19

0.31

0.33

0.33

0.33

0.44

0.44

Pane

lB:O

utco

me:

Wat

chan

dlis

ten

tone

wsa

ndpo

litic

alpr

ogra

msw

eekl

yU

pcom

ing

loca

lele

ctio

n0.

116*

**0.

116*

**0.

105*

**0.

105*

**0.

082*

**0.

082*

**0.

092*

**0.

088*

**0.

088*

**0.

092*

**0.

091*

**0.

091*

**(0

.038

)(0

.038

)(0

.032

)(0

.032

)(0

.022

)(0

.022

)(0

.019

)(0

.019

)(0

.019

)(0

.019

)(0

.016

)(0

.016

)

Obs

erva

tions

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

Mun

icip

aliti

es53

953

953

953

953

953

953

953

953

953

953

953

9O

utco

me

rang

e{0

,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

Out

com

em

ean

0.63

0.63

0.63

0.63

0.63

0.63

0.63

0.63

0.63

0.63

0.63

0.63

Loc

alel

ectio

nm

ean

0.02

0.02

0.06

0.06

0.19

0.19

0.31

0.33

0.33

0.33

0.44

0.44

Pane

lC:O

utco

me:

Wat

chan

dlis

ten

tone

wsa

ndpo

litic

alpr

ogra

mss

cale

Upc

omin

glo

cale

lect

ion

0.41

6***

0.41

6***

0.37

3***

0.37

3***

0.27

6***

0.27

6***

0.32

7***

0.31

9***

0.31

9***

0.32

7***

0.30

6***

0.30

6***

(0.1

22)

(0.1

22)

(0.1

06)

(0.1

06)

(0.0

75)

(0.0

75)

(0.0

62)

(0.0

61)

(0.0

61)

(0.0

62)

(0.0

50)

(0.0

50)

Obs

erva

tions

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

13,0

3013

,030

Mun

icip

aliti

es53

953

953

953

953

953

953

953

953

953

953

953

9O

utco

me

rang

e{0

,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

Out

com

em

ean

2.58

2.58

2.58

2.58

2.58

2.58

2.58

2.58

2.58

2.58

2.58

2.58

Loc

alel

ectio

nm

ean

0.02

0.02

0.06

0.06

0.19

0.19

0.31

0.33

0.33

0.33

0.44

0.44

Pane

lD:O

utco

me:

Topi

calp

oliti

calk

now

ledg

eU

pcom

ing

loca

lele

ctio

n0.

448*

**0.

218*

*0.

403*

**0.

403*

**0.

312*

**0.

312*

**0.

265*

**0.

245*

**0.

224*

**0.

212*

**0.

149*

**0.

149*

**(0

.119

)(0

.102

)(0

.085

)(0

.085

)(0

.051

)(0

.051

)(0

.045

)(0

.045

)(0

.043

)(0

.042

)(0

.032

)(0

.032

)

Obs

erva

tions

17,2

1317

,213

17,2

1317

,213

17,2

1317

,213

17,2

1317

,213

17,2

1317

,213

17,2

1317

,213

Mun

icip

aliti

es53

953

953

953

953

953

953

953

953

953

953

953

9O

utco

me

rang

e[-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6][-

2.11

,1.5

6]O

utco

me

mea

n0.

000.

000.

000.

000.

000.

000.

000.

000.

000.

000.

000.

00L

ocal

elec

tion

mea

n0.

010.

030.

060.

060.

160.

160.

250.

270.

280.

290.

370.

37

Not

es:

All

spec

ifica

tions

incl

ude

surv

ey-y

ear

fixed

effe

cts,

and

are

estim

ated

usin

gO

LS.

The

outc

omes

inpa

nels

A-C

wer

eno

tcol

lect

edin

the

2001

surv

ey.

Stan

dard

erro

rscl

uste

red

by

mun

icip

ality

are

inpa

rent

hese

s.*

deno

tes

p<

0.1,

**de

note

sp<

0.05

,***

deno

tes

p<

0.01

.

A26

Tabl

eA

13:E

ffec

tsof

pre-

elec

tion

hom

icid

esh

ocks

onm

unic

ipal

incu

mbe

ntvo

tesh

are

Bas

elin

eN

ofix

edC

ontr

ols

Stat

e-In

cum

bent

Mun

icip

ality

Plac

ebo

Vio

lenc

eFe

wer

than

Few

Non

-O

ne-

Thr

ee-

Four

-Fi

rst

Mai

nSi

ngle

-sp

ec.

effe

cts

year

part

y-s

peci

fic6

mon

ths

bin

25ho

mic

ides

drug

DTO

mon

thm

onth

mon

thha

lfof

part

ypa

rty

effe

cts

tren

dstr

ends

earl

ier

effe

cts

perm

onth

hom

icid

esst

ates

shoc

ksh

ock

shoc

km

onth

incu

mbe

ntin

cum

bent

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

(13)

(14)

(15)

(16)

(17)

Out

com

e:In

cum

bent

part

yvo

tesh

are

Hom

icid

esh

ock

-0.0

22**

*-0

.015

**-0

.015

**-0

.017

***

-0.0

16**

-0.0

26**

*-0

.025

***

-0.0

20**

*-0

.022

***

-0.0

18**

-0.0

22**

*-0

.021

***

-0.0

25**

*(0

.007

)(0

.007

)(0

.006

)(0

.006

)(0

.007

)(0

.009

)(0

.007

)(0

.007

)(0

.007

)(0

.009

)(0

.008

)(0

.007

)(0

.009

)Pl

aceb

oho

mic

ide

0.00

6sh

ock

(0.0

10)

Hom

icid

esh

ock

-0.0

09(o

nem

onth

)(0

.009

)H

omic

ide

shoc

k-0

.022

***

(thr

eem

onth

)(0

.007

)H

omic

ide

shoc

k-0

.025

***

(fou

rmon

th)

(0.0

08)

Obs

erva

tions

167,

116

173,

115

162,

774

167,

116

157,

639

167,

116

145,

511

167,

116

157,

555

167,

116

102,

126

143,

840

180,

302

188,

067

152,

539

157,

639

104,

220

Uni

que

mun

icip

aliti

es90

61,

335

895

906

823

906

820

906

905

906

616

637

1,06

41,

172

854

823

641

Out

com

em

ean

0.42

0.42

0.42

0.42

0.42

0.42

0.42

0.42

0.42

0.42

0.40

0.42

0.42

0.41

0.42

0.42

0.42

Out

com

est

d.de

v.0.

150.

150.

140.

150.

140.

150.

140.

150.

150.

150.

140.

140.

150.

150.

140.

140.

15H

omic

ide

shoc

km

ean

0.41

0.42

0.41

0.41

0.42

0.47

0.41

0.41

0.42

0.41

0.41

0.48

0.43

0.42

0.41

0.42

0.39

Not

es:C

olum

n(2

)exc

lude

sm

unic

ipal

ityan

dye

arfix

edef

fect

s,an

dth

usin

clud

esm

unic

ipal

ities

429

mun

icip

aliti

esw

here

aho

mic

ide

shoc

kis

only

defin

edfo

rone

elec

tion

whe

reel

ecto

ral

data

isav

aila

ble.

Col

umns

(3)

incl

udes

the

cont

rols

inTa

ble

A6,

with

the

exce

ptio

nof

vari

able

sw

ithgr

eate

rth

an5%

mis

sing

ness

.C

olum

n(4

)in

clud

esst

ate-

year

fixed

effe

cts.

Col

umn

(5)i

nclu

des

part

y-sp

ecifi

cin

cum

bent

year

tren

ds.C

olum

n(6

)inc

lude

sm

unic

ipal

ity-s

peci

ficye

artr

ends

.Col

umn

(7)u

ses

apl

aceb

oho

mic

ide

shoc

kde

fined

six

mon

ths

befo

reth

eel

ectio

n.

Col

umn

(8)i

nclu

des

fixed

effe

cts

fort

heto

taln

umbe

rofh

omic

ides

over

the

four

-mon

thw

indo

win

bins

ofsi

zete

n.C

olum

n(9

)exc

lude

sm

unic

ipal

ities

aver

agin

gm

ore

than

25ho

mic

ides

perm

onth

over

the

two

mon

ths

eith

erbe

fore

oraf

tert

heel

ectio

n.C

olum

n(1

0)ex

clud

esm

unic

ipal

ities

that

aver

age

mor

eth

anon

edr

ug-r

elat

edho

mic

ide

am

onth

over

the

12m

onth

sbe

fore

anel

ectio

nbe

twee

n20

06an

d20

11.C

olum

n(1

1)ex

clud

esst

ates

with

high

-lev

elof

DTO

activ

ity(s

eefo

otno

te43

).C

olum

ns(1

2)-(

14)r

espe

ctiv

ely

defin

eho

mic

ide

shoc

ksov

eron

e-,t

hree

-,

and

four

-mon

thw

indo

ws.

Col

umn

(15)

incl

udes

only

elec

tions

that

take

plac

eon

orbe

fore

the

15th

ofth

em

onth

.Col

umn

(16)

incl

udes

only

elec

tions

whe

reth

ePA

N,P

RD

,orP

RIa

rean

incu

mbe

nt.C

olum

n(1

7)in

clud

eson

lyob

serv

atio

nsw

here

the

incu

mbe

ntis

nota

coal

ition

.All

spec

ifica

tions

are

estim

ated

usin

gO

LS,

and

allo

bser

vatio

nsar

ew

eigh

ted

byth

enu

mbe

rof

regi

ster

edvo

ters

inth

eel

ecto

ralp

reci

nct.

Stan

dard

erro

rscl

uste

red

bym

unic

ipal

ityar

ein

pare

nthe

ses.

*de

note

sp<

0.1,

**de

note

sp<

0.05

,***

deno

tes

p<

0.01

.

A27

Tabl

eA

14:E

ffec

tsof

pre-

elec

tion

hom

icid

esh

ocks

onm

unic

ipal

incu

mbe

ntel

ecto

ralo

utco

mes

,inc

ludi

ngm

unic

ipal

ities

with

outa

polic

efo

rce

Bas

elin

eN

ofix

edC

ontr

ols

Stat

e-In

cum

bent

Mun

icip

ality

Plac

ebo

Vio

lenc

eFe

wer

than

Few

Non

-O

ne-

Thr

ee-

Four

-Fi

rst

Mai

nSi

ngle

-sp

ec.

effe

cts

year

part

y-s

peci

fic6

mon

ths

bin

25ho

mic

ides

drug

DTO

mon

thm

onth

mon

thha

lfof

part

ypa

rty

effe

cts

tren

dstr

ends

earl

ier

effe

cts

perm

onth

hom

icid

esst

ates

shoc

ksh

ock

shoc

km

onth

incu

mbe

ntin

cum

bent

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

(13)

(14)

(15)

(16)

(17)

Pane

lA:O

utco

me:

Incu

mbe

ntpa

rty

win

Hom

icid

esh

ock

-0.1

16**

*-0

.078

**-0

.080

**-0

.101

***

-0.1

05**

*-0

.126

**-0

.127

***

-0.1

03**

*-0

.116

***

-0.1

12**

-0.1

06**

*-0

.117

***

-0.0

94*

(0.0

34)

(0.0

31)

(0.0

34)

(0.0

28)

(0.0

37)

(0.0

54)

(0.0

32)

(0.0

36)

(0.0

34)

(0.0

44)

(0.0

37)

(0.0

35)

(0.0

52)

Plac

ebo

hom

icid

e-0

.021

shoc

k(0

.046

)H

omic

ide

shoc

k-0

.037

(one

mon

th)

(0.0

41)

Hom

icid

esh

ock

-0.1

02**

*(t

hree

mon

th)

(0.0

37)

Hom

icid

esh

ock

-0.0

91**

(fou

rmon

th)

(0.0

37)

Obs

erva

tions

3,36

63,

773

3,15

73,

366

3,03

33,

366

2,90

23,

366

3,35

23,

366

2,26

12,

247

4,11

44,

685

3,10

53,

033

2,11

0U

niqu

em

unic

ipal

ities

1,01

81,

425

966

1,01

893

41,

018

948

1,01

81,

017

1,01

869

972

81,

186

1,31

297

393

474

5O

utco

me

rang

e{0

,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

{0,1}

Out

com

em

ean

0.56

0.56

0.57

0.56

0.57

0.56

0.57

0.56

0.56

0.56

0.55

0.57

0.56

0.55

0.57

0.57

0.56

Hom

icid

esh

ock

mea

n0.

410.

410.

410.

410.

410.

450.

410.

410.

410.

410.

410.

480.

430.

420.

410.

410.

38

Pane

lB:O

utco

me:

Cha

nge

inin

cum

bent

part

yvo

tesh

are

Hom

icid

esh

ock

-0.0

22**

*-0

.018

***

-0.0

13**

-0.0

17**

*-0

.016

**-0

.027

***

-0.0

25**

*-0

.022

***

-0.0

22**

*-0

.022

**-0

.018

**-0

.021

***

-0.0

29**

*(0

.008

)(0

.007

)(0

.006

)(0

.006

)(0

.008

)(0

.009

)(0

.008

)(0

.008

)(0

.008

)(0

.010

)(0

.008

)(0

.008

)(0

.011

)Pl

aceb

oho

mic

ide

-0.0

07sh

ock

(0.0

10)

Hom

icid

esh

ock

-0.0

12(o

nem

onth

)(0

.010

)H

omic

ide

shoc

k-0

.025

***

(thr

eem

onth

)(0

.008

)H

omic

ide

shoc

k-0

.027

***

(fou

rmon

th)

(0.0

08)

Obs

erva

tions

194,

378

198,

903

188,

681

194,

378

184,

161

194,

378

172,

291

194,

378

183,

882

194,

378

121,

534

167,

297

209,

741

219,

061

176,

231

184,

161

121,

809

Uni

que

mun

icip

aliti

es1,

018

1,42

41,

005

1,01

893

41,

018

948

1,01

81,

017

1,01

869

972

81,

186

1,31

297

393

474

5O

utco

me

rang

e[-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

7][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9][-

0.96

,0.8

9]O

utco

me

mea

n-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.05

-0.0

5-0

.04

Out

com

est

d.de

v.0.

140.

150.

140.

140.

140.

140.

140.

140.

150.

140.

140.

140.

150.

150.

150.

140.

15H

omic

ide

shoc

km

ean

0.41

0.41

0.41

0.41

0.41

0.47

0.41

0.41

0.41

0.41

0.41

0.48

0.43

0.42

0.41

0.41

0.38

Not

es:C

olum

n(2

)exc

lude

sm

unic

ipal

ityan

dye

arfix

edef

fect

s,an

dth

usin

clud

esm

unic

ipal

ities

429

mun

icip

aliti

esw

here

aho

mic

ide

shoc

kis

only

defin

edfo

rone

elec

tion

whe

reel

ecto

ral

data

isav

aila

ble.

Col

umns

(3)

incl

udes

the

cont

rols

inTa

ble

A6,

with

the

exce

ptio

nof

vari

able

sw

ithgr

eate

rth

an5%

mis

sing

ness

.C

olum

n(4

)in

clud

esst

ate-

year

fixed

effe

cts.

Col

umn

(5)i

nclu

des

part

y-sp

ecifi

cin

cum

bent

year

tren

ds.C

olum

n(6

)inc

lude

sm

unic

ipal

ity-s

peci

ficye

artr

ends

.Col

umn

(7)u

ses

apl

aceb

oho

mic

ide

shoc

kde

fined

six

mon

ths

befo

reth

eel

ectio

n.

Col

umn

(8)i

nclu

des

fixed

effe

cts

fort

heto

taln

umbe

rofh

omic

ides

over

the

four

-mon

thw

indo

win

bins

ofsi

zete

n.C

olum

n(9

)exc

lude

sm

unic

ipal

ities

aver

agin

gm

ore

than

25ho

mic

ides

perm

onth

over

the

two

mon

ths

eith

erbe

fore

oraf

tert

heel

ectio

n.C

olum

n(1

0)ex

clud

esm

unic

ipal

ities

that

aver

age

mor

eth

anon

edr

ug-r

elat

edho

mic

ide

am

onth

over

the

12m

onth

sbe

fore

anel

ectio

nbe

twee

n20

06an

d20

11.C

olum

n(1

1)ex

clud

esst

ates

with

high

-lev

elof

DTO

activ

ity(s

eefo

otno

te43

).C

olum

ns(1

2)-(

14)r

espe

ctiv

ely

defin

eho

mic

ide

shoc

ksov

eron

e-,t

hree

-,

and

four

-mon

thw

indo

ws.

Col

umn

(15)

incl

udes

only

elec

tions

that

take

plac

eon

orbe

fore

the

15th

ofth

em

onth

.Col

umn

(16)

incl

udes

only

elec

tions

whe

reth

ePA

N,P

RD

,orP

RIa

rean

incu

mbe

nt.C

olum

n(1

7)in

clud

eson

lyob

serv

atio

nsw

here

the

incu

mbe

ntis

nota

coal

ition

.All

spec

ifica

tions

are

estim

ated

usin

gO

LS,

and

allo

bser

vatio

nsar

ew

eigh

ted

byth

enu

mbe

rof

regi

ster

edvo

ters

inth

eel

ecto

ralp

reci

nct.

Stan

dard

erro

rscl

uste

red

bym

unic

ipal

ityar

ein

pare

nthe

ses.

*de

note

sp<

0.1,

**de

note

sp<

0.05

,***

deno

tes

p<

0.01

.

A28

Table A15: Effect of pre-election drug-related homicide shocks on municipal incumbent electoraloutcomes

Incumbent Change inparty incumbentwin party

vote share(1) (2)

Drug-related homicide shock -0.068 -0.020(0.074) (0.017)

Observations 471 37,943Unique municipalities 416 416Outcome mean 0.53 -0.05Homicide shock mean 0.44 0.44

Notes: All specifications are estimated using OLS, and include state and year fixed effects. All observations areweighted by the number of registered voters. Standard errors clustered by municipality are in parentheses. *denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A.5.6 Difference-in-differences approach to identifying the effects of pre-election homicides

The results in the main text focus on pre-election homicide shocks coded as a binary variable. Thisapproach is theoretically appealing because its short-term nature both captures difference relativeto the municipality’s general baseline as well as closely maps to cycles in political informationconsumption. Moreover it is empirically appealing because month-to-month variation in homicideshocks is both plausibly exogenous to a wide variety of possible confounds and uncorrelated withthe broader trends in homicide rates that this article seeks to differentiate the effects of short-termspikes around elections from. However, I also now consider an alternative approach to capturingthe effects of short-run shocks around local elections that does not rely on homicide data after theelection.

Since the homicide level, as opposed to the short-term deviation exploited in the main paper,just before an election is highly correlated with the general homicide level, a meaningful testof increased homicides around elections requires a subtler design.68 I thus use a difference-in-differences design to identify how deviations in the number of homicides just before elections,relative to the number of homicides occurring earlier in the electoral cycle, affect vote choice.Specifically, I first use the raw difference between the average number of homicides in the two

68The results of such an approach are thus similar to the small long-run estimates reported in Table 3.

A29

months prior to an election and the average number of homicides over the prior electoral cycle (i.e.preceding 34 months). A second measure then takes the proportional difference by normalizingthese deviations by the average number of homicides over the prior electoral cycle. Finally, athird approach codes indicators for above- and below-median municipalities according to the twopreceding variables. These measures seek to capture differences in the magnitude of pre-electionhomicide deviations across elections, and thus exploit variation in the intensity of these shocksacross municipalities. The identifying assumption is that municipalities experiencing differenthomicide deviations from election cycle trends before elections otherwise follow parallel trends inincumbent support.

Table A16 reports estimates from the analog of equation (4), and demonstrates that this alterna-tive approach to capturing pre-election homicide deviations also decreases the incumbent party’svote share. Although the estimates are often borderline statistically significant, columns (1) and(2) first show that an increase in the raw pre-election deviation decreases the probability of theincumbent winning and the incumbent’s vote share. This estimates are more precisely estimatedwhen separating above and below median deviations in columns (3) and (4). Columns (5)-(8)report similar results when considering the proportional deviation instead.

A.5.7 Differential effects of homicide shocks across parties

I also examine differential punishment across parties. Due to the PRI’s more extensive clientelis-tic ties (e.g. Cornelius 1996; Fox 1994; Greene 2007; Magaloni 2006), PRI voters may be lesssusceptible to performance information, and thus less inclined to punish PRI incumbents for homi-cide shocks. Table A17 provides tentative evidence consistent with this expectation. Althoughthe interaction terms are not statistically significant, column (1) shows that the PAN and PRD arepunished relatively more than the PRI. Although punishment of PRI incumbents is not statisti-cally significant, columns (2)-(4) register significant punishment of PAN and PRD mayors. Thisis potentially because the PAN and especially PRD control more urban areas where the effects ofhomicide shocks are larger.

Finally, who do voters turn to in order to reduce local violence? If reducing violence is a majorconcern, and voters believe that Calderon’s tough stance on drug-related crime may help theirmunicipality (see Dell 2015), even PAN mayors may benefit from a homicide shock. Column (5)of panel C finds suggestive evidence for this: although a homicide shock increases the PAN’s voteshare by 4.6 percentage points during Calderon’s presidency, this interaction is not statisticallysignificant.

A30

Tabl

eA

16:D

iffer

ence

-in-

diff

eren

ces

appr

oach

toes

timat

ing

the

effe

ctof

pre-

elec

tion

devi

atio

nsin

hom

icid

era

tes

onm

unic

ipal

incu

mbe

ntel

ecto

ralo

utco

mes

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Pane

lA:O

utco

me:

Incu

mbe

ntpa

rty

win

Dev

iatio

nfr

omav

erag

em

onth

lyho

mic

ide

-0.0

016

-0.0

077*

**(0

.003

3)(0

.002

0)A

bove

med

ian

devi

atio

nfr

omav

erag

em

onth

lyho

mic

ide

-0.0

932*

*-0

.025

3(0

.043

7)(0

.061

3)Pr

opor

tiona

ldev

iatio

nfr

omav

erag

em

onth

lyho

mic

ide

-0.0

176*

-0.0

250

(0.0

094)

(0.0

168)

Abo

vem

edia

npr

opor

tiona

ldev

iatio

nfr

omav

erag

em

onth

lyho

mic

ide

-0.1

038*

*-0

.047

1(0

.043

1)(0

.059

5)

Obs

erva

tions

2,85

22,

852

2,85

22,

852

2,80

82,

808

2,80

82,

808

Uni

que

mun

icip

aliti

es90

690

690

690

688

788

788

788

7O

utco

me

mea

n0.

550.

550.

550.

550.

550.

550.

550.

55H

omic

ide

mea

sure

mea

n1.

671.

670.

500.

500.

320.

320.

500.

50H

omic

ide

mea

sure

stan

dard

devi

atio

n14

.31

14.3

10.

500.

501.

521.

520.

500.

50

Pane

lB:O

utco

me:

Cha

nge

inin

cum

bent

part

yvo

tesh

are

Dev

iatio

nfr

omav

erag

em

onth

lyho

mic

ide

-0.0

006

-0.0

008

(0.0

004)

(0.0

005)

Abo

vem

edia

nde

viat

ion

from

aver

age

mon

thly

hom

icid

e-0

.030

5***

-0.0

311*

*(0

.010

3)(0

.013

7)Pr

opor

tiona

ldev

iatio

nfr

omav

erag

em

onth

lyho

mic

ide

-0.0

024

-0.0

041

(0.0

020)

(0.0

035)

Abo

vem

edia

npr

opor

tiona

ldev

iatio

nfr

omav

erag

em

onth

lyho

mic

ide

-0.0

336*

**-0

.033

9***

(0.0

095)

(0.0

128)

Obs

erva

tions

167,

116

167,

116

167,

116

167,

116

166,

892

166,

892

166,

892

166,

892

Uni

que

mun

icip

aliti

es90

690

690

690

690

590

590

590

5O

utco

me

mea

n0.

550.

550.

550.

550.

550.

550.

550.

55H

omic

ide

mea

sure

mea

n1.

701.

700.

500.

500.

330.

330.

500.

50H

omic

ide

mea

sure

stan

dard

devi

atio

n14

.39

14.3

90.

500.

501.

541.

540.

500.

50

Mun

icip

ality

-spe

cific

time

tren

dsX

XX

X

Not

es:A

llsp

ecifi

catio

nsar

ees

timat

edus

ing

OL

S,an

din

clud

em

unic

ipal

ityan

dye

arfix

edef

fect

s.A

llob

serv

atio

nsar

ew

eigh

ted

byth

enu

mbe

rofr

egis

tere

dvo

ters

.Sta

ndar

der

rors

clus

tere

d

bym

unic

ipal

ityar

ein

pare

nthe

ses.

*de

note

sp<

0.1,

**de

note

sp<

0.05

,***

deno

tes

p<

0.01

.

A31

Table A17: Effects of pre-election homicide shocks on municipal incumbent electoral outcomes,by incumbent party

Change in Change in Change in Change in Change inincumbent PAN PRD PRI PAN

party incumbent incumbent incumbent incumbentvote share vote share vote share vote share vote share

(1) (2) (3) (4) (5)

Homicide shock -0.007 -0.037*** -0.049* -0.009 0.026(0.013) (0.012) (0.026) (0.013) (0.022)

Homicide shock × PAN -0.016(0.023)

Homicide shock × PRD -0.046(0.035)

Homicide shock × Calderon 0.046Presidency (0.044)

Observations 158,954 56,597 19,767 82,590 167,116Unique municipalities 898 432 286 769 906Outcome mean -0.05 -0.06 -0.09 -0.04 -0.05Outcome standard deviation 0.14 0.14 0.17 0.13 0.14Homicide shock mean 0.42 0.43 0.44 0.40 0.41

Notes: All specifications are estimated using OLS. All observations are weighted by the number of registeredvoters. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, ***denotes p < 0.01.

A32

Table A18: Heterogeneous effects of upcoming local elections on institutional confidence athigher levels of government, by short-run homicide shocks

Low confidence Low confidencein the President in state Governor

(1) (2) (3) (4)

Upcoming local election 0.064* 0.051* -0.059 0.064(0.035) (0.030) (0.040) (0.089)

Homicide shock -0.009 -0.010 -0.054** -0.032(0.014) (0.012) (0.025) (0.020)

Upcoming local election × Homicide shock 0.033 0.015 -0.001 -0.073(0.040) (0.038) (0.061) (0.158)

Observations 13,551 13,551 6,445 6,445Clusters 390 390 293 293Outcome range {0,1} {0,1} {0,1} {0,1}Outcome mean 0.27 0.27 0.33 0.33Local election mean 0.16 0.16 0.02 0.02Homicide measure mean 0.45 0.45 0.42 0.42Survey years without data 2003, 2005 2003, 2005State fixed effects X X

Notes: All specifications include survey year fixed effects, and are estimated using OLS. Standard errors clusteredby municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A.5.8 Belief updating about presidents and state governors

Table A18 extends the results of Table 7 to higher-level of government. As noted in the main text,the results indicate that—in contrast with confidence in their mayor—voters do not become lessconfident in the president or their state governor when they experience a pre-election homicideshock. This reinforces the responsibility attribution implied by Table 5.

A.5.9 Differential updating across elections

Table A19 reports the regression estimates underlying Figure 6.

A.5.10 Learning from the experiences of neighboring municipalities

To identify how comparative performance information influences voter behavior, I compute theaverage homicide shock across neighboring municipalities and the same average weighted by thesize of neighboring municipalities’ registered electorates. Consistent with voters benchmarking

A33

Table A19: The effect of a homicide shock on the change in incumbent party vote share, byprevious homicide shocks and incumbent identity

Change in incumbentparty vote share

(1)

Homicide shock -0.040**(0.018)

Homicide shock at last election -0.042*(0.025)

Homicide shock × Homicide shock at last election 0.063**(0.031)

Incumbent won previous election -0.046**(0.019)

Homicide shock × Incumbent won previous election 0.013(0.029)

Homicide shock at last election × Incumbent won previous election 0.035(0.031)

Homicide shock × Homicide shock at last election -0.070× Incumbent won previous election (0.045)

Observations 111,910Unique municipalities 730Outcome range [-0.90,0.89]Outcome mean -0.06Outcome standard deviation 0.14Homicide shock mean 0.43Homicide shock at last election mean 0.40Incumbent won previous election mean 0.55

Notes: All specifications include municipality and year fixed effects, and are estimated using OLS. Allobservations are weighted by the number of registered voters. Standard errors clustered by municipalityin parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A34

their incumbent’s performance against that of their neighbors, column (1) of Table A20 shows thatthe probability that an incumbent party is re-elected increases with the average shock experiencedby neighboring municipalities. However, the negative interaction between the homicide shock andthe neighbor average suggests that relative performance considerations do not apply when a voter’sown municipality is afflicted by a shock.69 Plausibly consistent with increased salience, mayoralpunishment instead increases with the proportion of shocks experienced by neighbors. The changein vote share follows a similar pattern but is not statistically significant, while columns (2) and (4)report similar results when an electorate-weighted neighboring municipal shock measure is usedinstead.

A.5.11 Less-educated voters respond most to pre-election homicide shocks

Voters consuming politically-relevant information for the first time in months or years are likely topossess imprecise prior beliefs, and thus be most susceptible to news because their beliefs are mostmalleable (e.g. Hirano et al. 2015; Zaller 1992). Conversely, voters consuming news throughoutthe electoral cycle possess stronger prior beliefs over the issues they regard as important and canbetter distinguish informative from uninformative signals. These arguments follow, for example,from the career concerns model discussed in Appendix A.1.1 above. They also receive empiricalsupport from Da Silveira and De Mello (2011) and Larreguy, Marshall and Snyder (forthcom-ing), who find that relatively uneducated voters in Brazil and Mexico respectively respond most topolitical advertising.

Education is widely believed to increase political knowledge by providing the ability, opportu-nity and motivation to acquire information (Barabas et al. 2014; Delli Carpini and Keeter 1996).Taking higher education as a proxy for the precision of prior beliefs, this subsection examines theextent to which it is the case—at both the individual and aggregate levels—that the least educatedvoters are most likely to respond to homicide shocks.

To test whether less educated voters respond more to performance indicators in the news aroundelections, I first examine how the effects of short-run homicide shocks on voter beliefs vary by ed-ucation. Using the ENCUP survey data, I define an education indicator for the 18% of respondentsthat have completed a university degree (“higher”) education. This is significantly positively corre-lated with news consumption and political knowledge. The marginal effects of an upcoming localelection reported in Table A21 suggest that increased concern about public insecurity and reducedconfidence in municipal incumbents are most pronounced among less educated voters. Although

69Unreported results separating neighboring incumbents from the same and different parties, but show no mean-ingful differences.

A35

Table A20: Effect of pre-election homicide shocks on municipal incumbent electoral outcomes,by neighbor average homicide shock

Incumbent Incumbent Change in Change inparty party incumbent incumbentwin win party party

vote share vote share(1) (2) (3) (4)

Homicide shock 0.007 -0.012 -0.000 -0.004(0.070) (0.069) (0.018) (0.017)

Neighbor average homicide shock 0.163** 0.027(0.082) (0.020)

Homicide shock × Neighbor -0.264** -0.045average homicide shock (0.118) (0.029)

Electorate-weighted neighbor 0.152* 0.027average homicide shock (0.080) (0.019)

Homicide shock × Electorate-weighted -0.230* -0.038neighbor average homicide shock (0.117) (0.028)

Observations 2,739 2,848 163,862 167,076Unique municipalities 876 904 899 904Outcome mean 0.55 0.55 -0.05 -0.05Outcome standard deviation 0.50 0.50 0.14 0.14Homicide shock mean 0.41 0.41 0.41 0.41Interaction mean 0.47 0.46 0.47 0.46

Notes: All specifications are estimated using OLS. All observations are weighted by the number of registeredvoters. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, ***denotes p < 0.01.

A36

Table A21: Effect of upcoming local elections on concern about public insecurity and lowconfidence in incumbent mayors, by homicide shock and completion of higher education

Public insecurity the Low confidence in Change inmajor problem in the municipal mayor incumbent partythe community vote share

(1) (2) (3) (4) (5)

Upcoming local election -0.021 -0.048* -0.036 0.163(0.022) (0.025) (0.049) (0.118)

Homicide shock -0.006 -0.005 -0.043* -0.013 0.0058(0.015) (0.013) (0.026) (0.022) (0.0307)

Higher education 0.082*** 0.064*** -0.001 0.008 0.0367**(0.021) (0.021) (0.029) (0.024) (0.0165)

Upcoming local election 0.107*** 0.096*** 0.143** 0.095× Homicide shock (0.037) (0.034) (0.063) (0.149)

Upcoming local election 0.037 0.045 0.196*** 0.182***× Higher education (0.033) (0.034) (0.051) (0.058)

Homicide shock × Higher education 0.000 0.010 -0.035 -0.037 -0.0443(0.026) (0.024) (0.040) (0.035) (0.0306)

Upcoming local election -0.057 -0.073 -0.494*** -0.487***× Homicide shock × Higher education (0.054) (0.053) (0.070) (0.075)

Local media 0.0027***(0.0010)

Homicide shock × Local media -0.0034**(0.0013)

Local media × Higher education -0.0019**(0.0009)

Homicide shock × Local media 0.0021× Higher education (0.0018)

Observations 6,564 6,564 5,916 5,916 29,736Clusters 196 196 293 293 118Outcome range {0,1} {0,1} {0,1} {0,1} [-0.63,0.50]Outcome mean 0.13 0.13 0.34 0.34 -0.04Homicide measure mean 0.47 0.47 0.42 0.42 0.40Higher education measure mean 0.20 0.20 0.16 0.16 0.50Upcoming local election mean 0.25 0.25 0.02 0.02Local media mean 18.69Survey years without data 2003, 2005 2003, 2005 2003, 2005 2003, 2005

Notes: The specifications in columns (1)-(4) are estimated using the ENCUP data and include survey year fixed effects, and are estimated

using OLS; the number of observations differ between columns due to non-responses. The specifications in columns (5) and (6) are estimated

using precinct-level electoral returns and includes neighbor group and year fixed effects, and are estimated using OLS; all observations are

weighted by the number of registered voters divided by the number of comparison units within each neighbor group. Standard errors clustered

by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A37

the triple interaction is not quite statistically significant in the case of public security concerns, theinteraction between upcoming local elections and homicide shocks indicate that—summing acrosscoefficients—a pre-election homicide shock only significantly increases the concerns of voters thatdid not complete upper secondary schooling.

I link these individual-level findings to precinct-level voting outcomes by examining whetherthe effect of local media covering a homicide shock also varies with higher education. I definean indicator for electoral precincts above-median levels of higher education, according to the 2010Census; i.e. 42% or more of voters with higher education in the relatively urban neighbors sample).Consistent with the survey results in columns (1)-(4), the triple interaction in column (5) of TableA21 suggests, albeit not statistically significantly so, that the electoral sanctioning induced by localmedia is greater in precincts with low fractions of highly educated voters.

A38