Marshall-Dissertation-2016

Information Consumption and Electoral Accountability in Mexico

The Harvard community has made this article openly available. Please share how this access benefits you. Your story matters

Citation Marshall, John Louis. 2016. Information Consumption and Electoral Accountability in Mexico. Doctoral dissertation, Harvard University, Graduate School of Arts & Sciences.

Citable link http://nrs.harvard.edu/urn-3:HUL.InstRepos:33493495

Terms of Use This article was downloaded from Harvard University’s DASH repository, and is made available under the terms and conditions applicable to Other Posted Material, as set forth at http:// nrs.harvard.edu/urn-3:HUL.InstRepos:dash.current.terms-of- use#LAA Information Consumption and Electoral Accountability in Mexico

A dissertation presented

John Louis Marshall

The Department of Government

in partial fulﬁllment of the requirements for the degree of Doctor of Philosophy in the subject of Political Science

Harvard University Cambridge, Massachusetts

May 2016 c 2016 — John Louis Marshall

Information Consumption and Electoral Accountability in Mexico

Abstract

Electoral accountability rests on voters re-electing high-performing and removing low- performing incumbents. However, voters in many developing contexts are poorly informed about incumbent performance, particularly of local politicians. This dissertation asks: how do voters in low-information environments hold local governments to account for their performance in office? I seek to explain when Mexican voters obtain performance information pertaining to their municipal incumbents, and ultimately how it impacts their beliefs and voting behavior. I argue that voters are able and willing to sanction local governments upon receiving incumbent performance indicators. However, electoral accountability requires incentives for voters and media outlets to respectively acquire and supply politically-relevant news. Information in the news just before elections, when these incentives align, thus strongly influences electoral accountability. I test these propositions by examining in detail voter responses to two key issues in Mexican politics—malfeasance in office and violent crime. The first chapter, coauthored with Eric Arias, Horacio Larreguy and Pablo Querubín, uses a large-scale field experiment to establish that voters indeed update from and act on malfeasance revelations. Reflecting voters’ negative priors, the distribution of leaflets documenting mayoral malfeasance increases the incumbent party’s vote share on average. However, consistent with Bayesian learning, these rewards decrease with positive prior iii beliefs, the strength of such priors, the severity of malfeasance revelations, and the extent of negative updating. Moreover, surprising information mobilizes turnout, while relatively unsurprising information reduces turnout. The second chapter then explores when and why voters choose to become informed. I argue that voters strategically acquire costly political information to cultivate a reputation among their peers as politically sophisticated. Leveraging a field experiment and observational variation, I demonstrate that social incentives increase political knowledge among voters nested in groups that collectively value political knowledge. This effect is most pronounced among relatively unsophisticated voters seeking to reach a minimum standard within their group, but is also evident among more sophisticated voters seeking to differentiate themselves from less-informed peers. The third chapter, coauthored with Horacio Larreguy and James Snyder, shows how broadcast media regulates access to relevant incumbent performance indicators. Supporting our argument that a media station’s potential audience shapes their incentives to provide local news, voters only sanction malfeasant parties in precincts covered by media stations whose principal audience resides in the precinct’s municipality. Conversely, media outlets based outside the municipality do not aid, and in fact crowd out, electoral accountability. The final chapter combines these insights to explain how voters hold incumbents to account for homicides in their municipality. I argue that even short-term performance indicators in the news prior to elections shape the voting behavior of poorly informed citizens. I show that voters consume most news before elections and update about incumbent performance from pre-election homicide shocks reported at that time. Unlike longer-term homicide trends, pre-election homicides substantially reduce the incumbent party’s probability of re-election. Sanctioning again requires, and increases with, access to local media, and is concentrated where voter priors are weakest.

iv | Contents

1 Introduction1 1.1 Dissertation overview...... 4 1.2 Implications and future research...... 9

2 Priors rule: When do malfeasance revelations help and hurt incumbent parties? 15 2.1 Introduction...... 15 2.2 Malfeasance, audits and elections in Mexican municipalities...... 22 2.3 Prior beliefs and voting behavior...... 26 2.4 Experimental design...... 35 2.5 Aggregate election results...... 52 2.6 How information inﬂuences voting behavior...... 62 2.7 Conclusion...... 75

3 Signaling sophistication: How social expectations can increase political information acquisition 77 3.1 Introduction...... 77 3.2 Voter information acquisition within social groups...... 83 3.3 Experimental evidence...... 91 3.4 Observational evidence...... 105 3.5 Conclusion...... 115

4 Publicizing malfeasance: When media facilitates electoral accountability in Mexico 118 4.1 Introduction...... 118 4.2 Why the intensity of media coverage matters...... 125 4.3 Political accountability in Mexico...... 128 4.4 Data...... 135 4.5 Empirical strategy...... 143 4.6 Results...... 154 4.7 Conclusion...... 172

v 5 Political information cycles: When do voters sanction incumbent parties for high homicide rates? 174 5.1 Introduction...... 174 5.2 Theoretical argument...... 180 5.3 Violent crime and political accountability in Mexico...... 187 5.4 Local elections, news consumption, and political perceptions...... 193 5.5 Local violence and electoral accountability...... 204 5.6 The moderating role of local media...... 224 5.7 Conclusion...... 233

A Appendix to Chapter 2 236

B Appendix to Chapter 3 254 B.1 Proofs...... 254 B.2 Variable deﬁnitions...... 257 B.3 Additional experimental results...... 262 B.4 Additional observational results...... 265

C Appendix to Chapter 4 270 C.1 Variable deﬁnitions...... 270 C.2 Audit reports...... 274 C.3 Additional local media stations and news consumption...... 274 C.4 Lack of balance across media stations...... 278 C.5 Additional results...... 278

D Appendix to Chapter 5 283 D.1 Formal model illustrating the voter updating process...... 283 D.2 Months and years of municipal elections...... 289 D.3 Data description...... 289 D.4 Map of municipalities included in different samples...... 296 D.5 Additional analyses...... 296

Bibliography 318

vi | Acknowledgments

This manuscript could not have been completed without the support and advice of my committee. I cannot imagine a better chair than Torben Iversen. His theoretical engagement, attention to the big picture and remarkable clarity of thought have inspired and disci- plined my work. While I have enjoyed our long meetings, the greatest testament to Torben is the energized feeling I’ve always felt when leaving his ofﬁce. I have learned tremen- dously from Jim Snyder’s incisive and thoughtful advice and passion for academic study. I can only hope to be as respected as Jim on both an intellectual and personal level. Jim Alt has taught me to be an academic, created research opportunities and supported me throughout my graduate studies. It has been an honor to be one of his last students. Most of all, I am indebted to Horacio Larreguy. His rigor and amazing desire to uncover truth have pushed me one step further at every turn. Horacio has inﬂuenced the location, scope and methods of my research and his commitment to my cause has exceeded any reasonable expectation. It is hard to imagine what this dissertation would look like without him. I hope that I can repay my committee’s investments in me by producing research that will satisfy the lofty standards that they have instilled. I also owe a great deal to the outstanding scholarly community that has supported me over the past six years. My research has been enriched by many conversations with and insights from Charlotte Cavaille, Adi Dasgupta, Julie Faller, Mauricio Fernández Duque, Nilesh Fernando, Noam Gidron, Michael Gill, Shelby Grossman, Andy Hall, Alex Hertel-

vii Fernandez, Akos Lada, Chris Lucas, Rakeen Mabud, Noah Nathan, Max Palmer, Jonathan Phillips, Solé Artiz Prilliman, Rob Schub, Brandon Stewart, Chiara Superti and George Yin. I am proud to count you all as colleagues and friends. I am also grateful for the advice and critiques of Jorge Domínguez, James Robinson and especially Jeff Frieden. Many people that have made my life as a graduate student enjoyable. My office mates Mauricio Fernández-Duque, Leslie Finger, Jen Pan, Molly Roberts and Chiara Superti have been frequent sources of cheer, although I especially miss debating statistics and life with Brandon Stewart early in the morning. As friends and housemates, Nilesh Fernando and Nils Hagerdal have provided well-placed skepticism and quality food. Greg Conti and James Brandt have provided beers, basketball and political theory. Unfortunately, Greg remains undefeated in our occasionally epic table tennis matches. When not drilling back- hand winners down the line against me on the tennis court, Rob Schub has sampled with me the gamut of lunch and drinking spots within 500 meters of CGIS. My GSAS football team have provided fun, injuries, occasional trophies and, most of all, the opportunity to continue playing the sport I love. I would have never embarked on a Ph.D. without guidance from key formative figures. At university, Nigel Bowles and Stephen Fisher taught me how think and write like an academic. Turning further back still, Neil Commin’s enthusiasm and encouragement cultivated my interest in politics that led me away from mathematics. At school, Mr (Mick) Hickman instilled in me a rigor that I have never forgotten. My largest debt is to my wonderful fiancée, Rakeen Mabud. She has been my greatest supporter, tirelessly reading through each paper I write, removing “the claw” from my presentation style, and (just about) tolerating the fact that I am never quite done. Beyond my research, Rakeen has quite simply made me a better person. She has taught me to be silly, professional and worldly, among many other things. Her positive influences on me are innumerable, and I cannot wait to share my life with her. viii Finally, I want to thank my parents, Barbara and Chris, and my brother Stephen. I could not wish for a more supportive and loving family. They provided me with freedom and opportunities, cultivated my inquisitive and independent nature, and from a young age tolerated me incessantly asking “why?” about anything from why we were playing football in the park to why people believe in God. Without them, none of this would exist.

ix 1| Introduction

Electoral accountability is the idea that voters hold governments accountable for their performance in office, re-electing high-performing incumbent parties and removing low- performing incumbent parties (e.g. Fearon 1999; Ferejohn 1986). This is widely regarded as a cornerstone of representative democracy. As the means through which incumbents can be appraised, voter information—about both performance outcomes and who is responsible for them—thus plays an essential role in this process. Idealized models of voting behavior typically rest on the assumption that voters cast ballots based on well-formed expectations of what different parties offer (Manin, Prze- worski and Stokes 1999). However, in practice, voters across the world—and particularly in developing democracies—are often poorly informed about incumbent performance in office, especially of local politicians (Keefer 2007; Pande 2011). Beyond a lack of representation in the abstract, this lack of voter information can have important consequences. By failing to remove incompetent or corruption politicians (see Pande 2011) or incentiviz- ing incumbents to target public funds towards engaged voters (e.g. Besley and Burgess 2002; Casey 2015), a misinformed electorate may have important efficiency and distributive implications for political, policy and socioeconomic outcomes. These concerns motivate the overarching question of this dissertation: how do voters in low-information environments hold local governments to account for their performance in office? I focus on incumbent performance information pertaining to local incumbent par-

1 ties, and address this question at the heart of democratic theory by first asking whether such information indeed influences voter beliefs, and in turn how such beliefs affect choices in the voting booth. Delving deeper, I tackle from both a voter demand and media supply perspective the more foundational question of when, if ever, voters actually become sufficiently informed to effectively hold their local government to account. In this dissertation, I thus seek to provide answers to various questions: How do voters become informed? Is an active media necessary for electoral accountability? What types of information affect which types of voters, and by how much? Are news stories before elections more likely to affect voting behavior? I examine these questions through the lens of Mexican municipal politics. Following hegemonic rule by the Partido Revolucionario Institucional from 1929 until the 1990s, Mexico has taken important steps toward democratic consolidation. Incumbent turnover, major institutional reforms, increasing media independence and growing multi-party competition are manifestations of an increasingly well-functioning democracy. However, the country still faces major challenges including poverty reduction, corruption and drug violence. These challenges are perhaps most pronounced among Mexico’s c.2,500 municipalities, which were empowered by decentralizing reforms in the late 1980s and especially in the mid 1990s (see Wellenstein, Núñez and Andrés 2006). Municipal mayors now preside over 20% of total government spending, and are responsible for providing basic public services, infrastructural investments and (in the majority of municipalities) local police forces. At the same time, mayors—who will only become eligible for re-election for the first time in 2018—face relatively few checks on their actions in office. Since mayors could not be re-elected, a notable feature of this study is the focus on party-level accountability. Mexican political parties remain the focal point of Mexican politics, and while voters are barely aware of individual candidates they generally know what the main parties represent and which is in power locally and nationally (Chong et al. 2 2015; Larreguy, Marshall and Snyder 2016). To support local social and economic development, voters must avoid electing mayors that are corrupt, unwilling or unable to implement effective policies, or reliant upon providing expensive clientelistic benefits to retain office. However, many voters remain poorly informed about politics (e.g. Chong et al. 2015), media outlets vary in their engagement with and independence from politics (Lawson 2002), and local political competition is often limited (Larreguy, Marshall and Snyder 2016). The task of selecting competent politicians is thus far from trivial. Considering municipal incumbent party performance in terms of addressing three salient valence issues—corruption, poverty reduction and public security—this dissertation specifically examines how information consumption affects local electoral accountability in Mexico’s low-information political environment. I argue that voters are both able and willing to sanction local governments upon receiving incumbent performance indicators. More specifically, I cast voters as Bayesians updating their beliefs based on how politically-relevant information relates to their beliefs before new information was received. However, the extent to which voters ultimately hold municipal governments to account reflects the incentives for voters to acquire, and media outlets to report, such politically-relevant news. As a consequence, electoral accountability follows what I term political information cycles: information in the news just before elections, when voter acquisition and media reporting incentives align, powerfully influencing electoral accountability. Over the course of the four papers in this dissertation, I develop these theoretical arguments and map them to appropriate empirical tests that cement the interpretation of the findings and clarify the mechanisms. Blending a range of empirical techniques, from field experiments and quasi-experimental designs to insights from the field, I aim to illuminate the incentives underpinning the supply and consumption of political news concerning these core issues, how voters update from this information, and ultimately how it impacts vot- 3 ing behavior. Although my analysis exploits subnational variation within Mexico, I do not believe my findings are specific to Mexico; throughout, I highlight parallels with both developed and developing contexts across the world.

1.1 Dissertation overview

In “Priors rule,” which is coauthored with Eric Arias, Horacio Larreguy and Pablo Querubín, my dissertation first seeks to establish how voters update from and act on information documenting mayoral malfeasance in office. While these questions have rightfully received considerable attention, existing research has struggled to impose a clear theoretical framework on the mixed empirical findings. For example, it remains unclear why media revealing mayoral malfeasance reduces the likelihood an incumbent party is re-elected in some instances (e.g. Ferraz and Finan 2008), but has no impact in others (e.g. Banerjee et al. 2014; de Figueiredo, Hidalgo and Kasahara 2013). Similarly, while malfeasance revelations appear to induce systemic disengagement in some contexts (Chong et al. 2015), they instead enhance participation in another (Banerjee et al. 2011). We seek to rationalize these mixed findings in a simple Bayesian model emphasizing voter prior beliefs and incorporating a cost of turning out. Theoretically, we show that electoral punishment may be rare among voters already pessimistic about the incumbent party’s competence, even when relatively substantial malfeasance is revealed. The implications for turnout reflect a more subtle non-linearity when voter partisan attachments are bimodally distributed. While relatively unsurprising information can reduce turnout by inducing a large mass of voters to abstain because their preference between the parties no longer exceeds the costs of turning out, sufficiently surprising revelations—whether positive or negative about the incumbent—can increase turnout by inducing supporters around one mode to switch parties.

4 These predictions are tested empirically using a large-scale field experiment in the central states of Guanjuato, México, Querétaro and San Luis Potosí around the 2015 municipal elections. Specifically, we examined how voters responded to learning the outcome of independent audits assessing the extent to which municipal governments correctly spent federal transfers from a major government program designated solely for social infrastructure projects benefiting the poor. In the average municipality between 2007 and 2015, 8% of funds stipulated for social infrastructure projects benefiting the poor were not spent on projects that actually benefited the poor, while a further 6% were spent on unauthorized projects. Reflecting voters’ bleak prior beliefs, relatively high malfeasance revelations increased the incumbent party’s vote share on average. Consistent with our voter learning model, this increase in incumbent party support decreases with positive prior beliefs, the strength of priors, the severity of malfeasance revelations, and the extent of negative updating. Po- tentially squaring the diverging extant findings regarding information’s impact on turnout, we find clear evidence that surprising information mobilizes turnout while relatively unsurprising information cultivates reduces turnout. In contrast, we find no evidence to suggest that voters become disillusioned with democratic politics. The preceding experimental evidence illustrates the potential importance of information for understanding voter beliefs and behavior. However, to understand whether voters respond similarly when information is not directly provided to them, a key question regards whether voters will consume such information without external intervention. This question is important in the face of Downs’s (1957) powerful and pessimistic free-riding logic that individual voters have strong incentives to leave information acquisition to others and thus remain “rationally ignorant.” Since Mexico, like many developing democracies, in practice retains relatively weak institutional protections against misuse of office, an informed electorate is central in sustaining accountable government. 5 In my second paper, “Signaling sophistication,” I address a key dimension explaining when and why voters choose to become politically informed: social approval. To endogenize political information, I propose a social signaling model of information acquisition where voters strategically acquire costly political information to cultivate a reputation among their peers as politically sophisticated. I thus resolve Downs’s (1957) collective action problem by arguing that, rather than acquiring information to maximize the probability that the best candidate is elected, individuals consume political information for selfish reasons emanating from their desire for approval among their peers. The model highlights two motivations that drive voters with varying levels of political sophistication within a group to acquire political information: (1) the desire to meet a minimum standard that separates less sophisticated voters from the least sophisticated, and (2) a social differentiation mo- tive whereby increases in information acquisition among less sophisticated voters forces causes more sophisticated voters to acquire more information to continue to differentiate themselves. I test this theory using both a small-scale field experiment around the 2015 elections and upcoming local elections to generate exogenous variation in social stimuli. These tests vary the likelihood that an individual’s peers are able observe an individual’s political knowledge. The results show that such social incentives significantly increase political knowledge among voters nested in groups that collectively value political knowledge, and likely assign large reputational benefits to politically sophisticated members. This increase is particularly prevalent among relatively unsophisticated voters within a group seeking to reach a minimum standard, but is also evident among more sophisticated voters acquiring news to differentiate themselves from less-informed peers. Although this social signaling logic likely coexists alongside a mass of voters that likely enjoy consuming political news, these findings provide the first evidence of which I am aware that political information acquisition is caused by social pressure. 6 From the perspective of supporting electoral accountability, these findings demonstrate the importance of social networks in supporting informed political participation. Con- versely, they also highlight how large groups of voters can get stuck in low-information equilibria, or “information traps,” where they have little incentive to consume news because their peers attach little value to political discussion. A growing body of global research, including from Africa (e.g. Casey 2015), Europe (e.g. Adams and Ezrow 2009) and the United States (e.g. Bartels 2008; Snyder and Strömberg 2010), indicates that such information differentials can have important policy and distributional consequences as politicians distribute resources toward informed voters. The analysis thus far ignores the reporting of news by broadcast and print media, the primary sources of news for Mexican voters. Given that many voters face weak incentives to actively search out political information for themselves, the potential importance of media outlets making relevant and (at least somewhat) impartial information accessible to voters is substantial. However, media outlets did not necessarily maximize the public’s capacity to hold elected politicians to account. Rather, newspapers, radio and television stations compete for audience shares and advertising revenues (e.g. Gentzkow and Shapiro 2006; Mullainathan and Shleifer 2005). In the third paper, “Revealing malfeasance,” which is coauthored with Horacio Lar- reguy and James Snyder, I investigate the impact of broadcast media on the electoral sanctioning of the municipal malfeasance examined in the first paper. We argue that media outlets only supply their audience with incumbent performance information when they face economic incentives to do so, namely a large interested audience. We exploit two sources of plausibly exogenous variation to isolate the effect of broadcast media outlets where audit reports are published: the timing of the release of municipal audit reports with respect to around elections, and within-neighboring precinct geographic variation in media coverage. Our findings emphasize the importance of media market structure in supporting elec- 7 toral accountability. In particular, we identify that voters electorally sanction the party of malfeasant mayors, but only in precincts covered by local media stations emitting from within the precinct’s municipality. Each additional local radio or television station reduces the vote share of an incumbent party revealed to be either corrupt or neglectful of the poor by around one percentage point. Demonstrating the importance of audience-based economic incentives, the effect of a local media station increases with the share of an outlet’s market located inside the municipality. Moreover, we find no evidence that non-local media stations contribute to the electoral sanctioning of malfeasant mayors. Rather, non-local media crowds out electoral sanctioning by attracting audiences to outlets less likely to report local mayoral malfeasance. The final paper, “Political information cycles,” brings together the insights of the preceding chapters to demonstrate how the forces underpinning news consumption patterns and information processing influence the sanctioning of local homicides. In the context of unprecedented violent crime, alongside the economy, public security has represented the most important policy issue among Mexican voters of the last decade. Consequently, the incumbent party’s effectiveness at addressing homicides rates is a key performance metric. Building on the propensity of voters to acquire, and media outlets to report, politically-relevant news before elections, I argue that short-term performance indicators in the news prior to elections may powerfully shape the voting behavior of citizens. My simple Bayesian learning model argues that even signals only weakly linked to longer-term performance can heavily influence the beliefs of poorly informed voters whose weak prior beliefs reflect high levels of political disengagement outside of election campaigns. Drawing on fine-grained survey and electoral data and multiple identification strategies, I document empirical support for the key elements of this political information cycles the- sis. First, I show that voters indeed consume more news before local elections, and that homicides just before such elections increase the salience of public security concerns and 8 reduce voter confidence in the mayor. Second, electoral returns confirm that pre-election homicide shocks reduce the incumbent party’s re-election probability by around ten percentage points. In contrast, incumbent electoral performance is barely impacted by longer- term homicide rates. Consistent with voter learning, sanctioning is limited to mayor controlling police forces, is greater where opposition parties did not experience such a shock prior to the previous election, and is not evident among state and federal incumbents. Third, the punishment of homicide shocks requires, and increases with, access to local broadcast media stations. These effects are pronounced only among less-informed voters that principally engage with politics around elections. Together, the results illustrate the importance of when voters consume news, and may thus explain the electoral volatility and mixed electoral accountability often observed outside consolidated democracies and in federal systems.

1.2 Implications and future research

This dissertation thus examines, piece by piece, how voters obtain and use information to hold their local governments to account in a major non-consolidated democracy. Combining these pieces unearths the finding that electoral accountability reflects political information cycles. With respect to corruption, misallocated spending on projects not actually benefiting, and especially local homicides, I show that voters are highly responsive to information in the news just before elections. A key conclusion is therefore that, given the political engagement and media environment in Mexico and other similar developing democracies, incumbent performance indicators are only likely to meaningfully impact voting behavior—and ultimately electoral outcomes—when their coverage in the media coincides with voters actually consuming news. My research suggests that this equilibrium simultaneously reflects the incentives for

9 voters to acquire politically-relevant information and for media outlets to report such information. On the voter side, I find that social approval drives information acquisition. The incentives to meet a minimum standard or differentiate oneself within a social group are especially pronounced around elections when relatively politically unsophisticated voters consume news for the first time. At the same time, whether the media provides politically- relevant information reflects economic incentives. In particular, local outlets whose audience principally resides within the municipality are most likely to provide voters with access to incumbent performance information. While the extant literature has highlighted the importance of access to news (e.g. Banerjee et al. 2011; Casey 2015; Ferraz and Finan 2008), this dissertation illuminates the mechanisms underpinning the production of news and, most importantly, shows that only when access is paired with voter consumption will political news impact vote choice. This dynamic provides a novel explanation for the mixed evidence of electoral sanctioning observed both in Mexico (Larreguy, Marshall and Snyder 2015; Vivanco et al. 2015) and among poorly-informed electorates more broadly (Achen and Bartels 2004b; Brollo 2009; Chang, Golden and Hill 2010; Roberts and Wibbels 1999). Another central conclusion is that Mexican voters generally behave as Bayesians upon receiving political information. With respect to both malfeasance in office and local homicides, voters update their beliefs from new information. Perhaps unsurprisingly, given the importance of these political issues, such updating ultimately influences actual vote choices: I consistently find that what voters regard as poor (good) performance induces electoral punishment (rewards). Beyond influence the direction of voter updating, prior beliefs also play a crucial role in conditioning the extent to which voters respond. More surprising is evidence of subtle updating, such as taking into account performance and the partisanship of the incumbent at the previous election, or recognizing that local politicians not controlling local police forces may not be responsible for local violence. There are still limits though: I find that voters do not utilize comparisons with neighboring municipali- 10 ties and may over-estimate the precision of recent news reports. Nevertheless, my analysis highlights how voters in a major developing context are not lacking in their capacity to hold politicians to account. The dissertation similarly cast light on Mexican municipal politics. A simple but important conclusion is that local politics matters to voters: even in a complex federal system, voters perceive local politicians as important actors and sanction them accordingly. My specific findings reinforce the salience of local government spending and violence—two key issues across Latin America. While the importance of rising crime may come as no surprise to scholars of Mexico, my emphasis on homicide spikes before elections through the prism of political information cycles make sense of the lack of correlation between general homicide trends and electoral outcomes. The findings with regard to incumbent malfeasance in office and local homicides similarly add nuance by showing that accountability requires local media, whose information appears to primarily influence voter beliefs rather than coordinate collective responses by generating common knowledge. These findings suggest mixed normative implications for democracy and electoral accountability. On one hand, voters demonstrate consistent willingness to vote on the basis of incumbent performance metrics, and exhibit impressive capacity to process the information they encounter. There is thus little doubt that voters armed with clear information have the capacity to hold local governments to account on key political issues. On the other hand, only a limited portion of the electorate obtains politically-relevant information. Unfortunately, this likely generates three equilibrium pathologies. First, limited political engagement may sustain the low voter expectations that mean that levels of malfeasance that are far from trivial may even be rewarded. Without holding politicians to higher standards, career-oriented politicians face weak incentives to perform in office while parties face little pressure to root out politicians that could represent electoral liabilities. Second, because low levels of political information foster weak prior beliefs, voters can be strongly 11 influenced by new revelations. Particularly in the case of local homicides, voters may thus “over-weight” relatively uninformative signals at the expense of informative signals they did not consume. This possibility chimes with recent theoretical research arguing that an informed electorate necessarily enhances electoral accountability (Ashworth and Bueno de Mesquita 2014a). Third, certain types of voter are systematically poorly informed, and thus politicians are likely to strategically allocate resources toward informed voters. This potential skew may undermine democratic representation. These normative concerns challenge policy makers seeking to strengthen democracy. One clear implication is that NGOs and government seeking to enhance electoral accountability must recognize political information cycles to enhance their information dissemination campaigns. Here, it is important to provide longer-term information beyond proximate security or economic shocks that accurately reflect incumbent performance. However, addressing the deeper issue of low expectations is more difficult. To achieve this, efforts to alter voter prior beliefs through civic education campaigns—whether in school, through the media or elsewhere—may be helpful. From an alternative perspective, politicians must believe that competence is an election-winning campaign strategy. The introduction of re-election in 2018 could aid this process by individualizing campaigns, while the independent candidates first permitted to run in 2015 similarly present an opportunity to alter the political agenda. Furthermore, there is clear evidence for this in Italy (Kendall, Nannicini and Trebbi 2015), although engaging Mexican politicians is likely to require compelling local evidence. My dissertation also raises interesting questions that will drive my future research. In particular, more work is required to understand equilibrium dynamics along several dimensions. First, while social groups play a fundamental role in explaining voter political information levels, a key question remains: what induces voters to enter high information social networks? This dissertation has focused on illustrating how social groups create in- 12 formation acquisition incentives, but deeper questions concern the drivers of social group formation and how politics becomes a focal discussion topic in some networks rather than others. Second, there is also much to be discovered about the production of politically-relevant news. This dissertation consistently shows that media coverage is a necessary condition for electoral accountability, but a systematic account of how much and what types of political news are reported is still missing. My ongoing research aims to address these questions by accumulating an unprecedented archive of local print and broadcast news coverage. Experi- mental tests will identify the relative importance of prohibitive search costs, news coverage cascades within media markets and upcoming local elections in explaining the extent of political news production. Similarly, I intend to explore how the timing of malfeasance revelations differentially influences electoral outcomes. Nevertheless, profit or audience maximization (e.g. Gentzkow and Shapiro 2006; Mullainathan and Shleifer 2005) may only tell part of the story. A key and often neglected component is the interaction between politicians and journalists. It remains essential to understand the empirical relevance of media capture (e.g. Besley and Prat 2006) in preventing a shift toward a high-accountability equilibrium. Third, an outstanding question is what about broadcast media produces such large effects on voter behavior. Media is distinctive from leaflets in important respects, including providing context, a tone of presentation, informational credibility, and the expectation that others will also have been similarly informed. Separating the sources of media’s effects is an important challenge for our understanding and for policy makers seeking to maximize the reach of the information provided. Finally, little is know about the extent to which one-shot and sustained information provision influence political equilibria. How do politicians respond to and learn from the sanctioning of high rates of crime or corruption revelations? Will voters remember infor- 13 mation from previous elections and update it accordingly before the next election? Can learning of reports documenting malfeasance encourage media outlets to continue reporting on such issues? These questions have profound consequences for social welfare, but remain notable omissions from this dissertation. Moving forward, I intend to examine how policy decisions, voter consumption choices and media outlet reporting choices are shaped by one-shot and sustained provision of politically-relevant information. Given information’s significant influence on electoral behavior, illustrated throughout this dissertation, the optimist in me hopes that it can also help to drive improving government provision and increase voter expectations of their politicians.

14 2| Priors rule: When do malfeasance revelations help and hurt incumbent parties?1

Co-authored with Eric Arias, Horacio A. Larreguy and Pablo Querubín

2.1 Introduction

Elected politicians across the world put in place policies to alleviate poverty and support economic development. However, the implementation of these policies is often beset by political rent seeking, including bribery (e.g. Hsieh and Moretti 2006), preferential con- tracting (e.g. Tran 2009) and misallocated spending (e.g. Larreguy, Marshall and Snyder 2015). While policy-makers and NGOs have increasingly sought to design institutions to mitigate such agency losses, political accountability ultimately rests upon citizens choosing to elect competent politicians and sanction incompetent politicians (e.g. Barro 1973; Fearon 1999; Ferejohn 1986; Rogoff 1990). Given that malfeasance in ofﬁce still represents a major challenge in many developing contexts (e.g. Kaufmann, Kraay and Mastruzzi 2009; Mauro 1995), a key question is then: when do voters hold their governments to ac-

1We are extremely grateful to Alejandra Rogel, Adriana Paz, Anais Anderson and the Data OPM and Qué Funciona para el Desarrollo team for their implementation of this project. We thank Chappell Lawson, Ken Shepsle, participants at the WESSI workshop at NYU Florence and the organizers and other team members of the EGAP Metaketa project for illuminating discussions and useful comments. This research was ﬁnanced by the Evidence in Governance and Politics (EGAP) Metaketa initiative, and was approved by the Harvard Committee on the Use of Human Subjects (IRB15-1068) and the New York University Committee on Activ- ities Involving Human Subjects (IRB15-10587). Our pre-analysis plan was pre-registered with EGAP, and is publicly available here.

15 count by punishing incumbent parties for malfeasant behavior in office? A growing political economy literature has emphasized the importance of providing voters with incumbent performance information. Negative information, such as reports revealing corruption, is typically expected to cause voters to electorally sanction those responsible. However, in practice, the evidence is mixed. On one hand, Ferraz and Finan (2008), Larreguy, Marshall and Snyder(2015) and Chang, Golden and Hill(2010) find that local media revealing mayoral malfeasance reduces their likelihood of re-election in Brazil, Mexico and Italy respectively. On the other, Banerjee et al.(2011, 2014), de Figueiredo, Hidalgo and Kasahara(2013) and Humphreys and Weinstein(2012) find that disseminating negative incumbent performance information in India, Brazil and Uganda often does not affect incumbent electoral prospects. The effects on turnout are similarly mixed: while Chong et al.(2015) suggest that negative information may even induce systemic disengagement in Mexico, Banerjee et al.(2011) instead observe increased turnout in India. From both a theoretical and applied policy perspective, it thus remains difficult to anticipate when or how providing information about incumbent performance will impact voters (see Lieberman, Posner and Tsai 2014). We argue that previously overlooked voter prior beliefs can rationalize these mixed findings, and ultimately explain when and how incumbent performance information impacts turnout and vote choice. We highlight the importance of the direction and magnitude of belief updating from new information in a simple two-party model where voters form beliefs about the competence of the incumbent party, receive expressive benefits from voting and are subject to fixed partisan attachments (see also Kendall, Nannicini and Trebbi 2015). Specifically, if voters already believe their incumbent party is malfeasant, even relatively severe corruption revelations can increase incumbent support because voters positively update from information that is better than expected. The implications for turnout reflect a more subtle non-linearity. Under empirically-appropriate bimodal distributions of partisan 16 attachments, relatively unsurprising information reduces turnout by inducing a large mass of voters located around one mode to abstain because their relative preference between the parties no longer exceeds the costs of turning out. However, sufficiently surprising revelations—in either direction—increase turnout by inducing a large mass of supporters around one mode to switch parties. We test these theoretical predictions—which are codified in our pre-analysis plan— using a large-scale field experiment conducted in Mexico around the 2015 elections. Be- yond its large population and shift toward multi-party democracy, Mexico’s relatively high but substantially varying levels of corruption and distrust in elected politicians mark it as an important location to test the implications of our argument. Extending two recent empirical studies with markedly different findings (Chong et al. 2015; Larreguy, Marshall and Snyder 2015), we examine how voters respond to learning the outcome of independent audits assessing the extent to which municipal governments correctly spent federal transfers from a major government program designated solely for social infrastructure projects benefiting the poor. Although parties are strong, Mexico represents an interesting case because mayors could not seek re-election.2 Across 678 electoral precincts in 26 municipalities from 4 central Mexican states, we randomized the dissemination of leaflets indicating the results of the municipal audit reports to up to 200 households in each treated precinct. We provided voters with one of two measures of incumbent performance: the share of funds spent on projects not benefiting the poor, or the share of funds spent on unauthorized projects. Such malfeasance ranged from 0% to 58% in our sample, with a mean of 21%. To improve our understanding of when information impacts voter beliefs and behavior, we also experimentally varied (1) whether the leaflet included a comparison with the performance of mayors from other parties within

2Re-election will become possible for incumbent mayors from 2018 in most states.

17 the state, and (2) whether leaflet distribution was accompanied by a loud speaker announc- ing the widespread delivery of the leaflets in the precinct. The former variant seeks to isolate the role of relative performance information, while the public mode of transmission seeks to generate tacit or explicit voter coordination through common understanding that others also received the information (Adena et al. 2015; Yanagizawa-Drott 2014). Consistent with the theory, we find that voter responses depend on how the information relates to their prior beliefs. On average, audit report information increased the incumbent party’s vote share by almost 3 percentage points. This positive effect reflects the common perception that the incumbent party is highly corrupt and relatively unwilling to support the poor. Somewhat depressingly, this suggests that voter expectations are so low that poor performance is often rewarded. However, these average effects mask substantial heterogeneity in the response of a Mexican electorate already highly skeptical that local politicians allocate funds as legally required. Illustrating the importance of understanding voter prior beliefs, we demonstrate that the increase in incumbent support induced by our treatment is concentrated among voters in municipalities with lower levels of malfeasance, voters with relatively negative prior perceptions of the incumbent party, and voters that positively updated about their perceptions of the incumbent party’s malfeasance upon receiving the information. Conversely, for egregious cases of malfeasance, and where voters update most negatively about their incumbent’s performance, voters are more likely to punish the incumbent party at the polls. Moreover, we find support for the non-linear prediction of malfeasance information’s impact on electoral turnout. In particular, relatively unsurprising information—20-30% of funds spent on projects not benefiting the poor or unauthorized projects—depresses turnout by around 1 percentage point. Conversely, extreme cases of malfeasance—both 0% and above 50%—mobilize turnout by 1-2 percentage points. This non-linearity, which is in line with a bimodal distribution of voters updating from new information, further underscores 18 the key role of voter priors in explaining voting behavior. In contrast, we find little evidence to suggest that revealing more severe cases of malfeasance to voters reduces turnout or confidence in the capacity of elections to select competent politicians. Turning to our variants of the information treatment, there is little evidence that either comparative performance information or low-intensity public information dissemination moderate voter responses. First, if anything, providing voters with a benchmark from other parties produced weaker effects. This is consistent with previous studies in Mexico failing to observe benchmarking across municipalities (Marshall 2016a), and likely reflects voter comprehension constraints and the fact that our information reaffirmed voters’ prior perceptions that challengers are less malfeasant. Second, the loud speaker—intended to induce common knowledge among voters about the information, and thereby facilitate coordinated action—neither increased social responses nor accentuated voter sanctioning. This null finding contrasts with studies isolating large magnification effects of local broadcast media (Ferraz and Finan 2008; Larreguy, Marshall and Snyder 2015; Marshall 2016a), and suggests that a more powerful public mechanism is required. However, we find evidence suggesting that voter responses are mediated by party reactions to revealing incumbent party performance. In addition to our distribution team, voters in treated precincts recalled that both incumbent and challenger local party organizations were more likely to combat or incorporate malfeasance reports in their campaigns, especially where the leaflets informed voters of high levels of malfeasance. We explore how this impacted election strategies by exploiting a discontinuity in the number of polling stations in a precinct, which previous research has found to enable parties to more effectively extract voter mobilization effort from local brokers by increasing the party’s ability to infer broker performance after receiving the signal of an additional electoral return (Larreguy, Marshall and Querubín 2016). The results indicate that the incumbent party is particularly able to counteract bad information about its performance in the precincts that possess an 19 additional polling station. This suggests that the relatively weak punishment of even severe malfeasance revelations in part reflects the incumbent’s capacity to combat negative belief updating with brokered voter mobilization. In sum, the results provide strong support for our simple prior-oriented theory of voter responses to incumbent performance information. In rationalizing the mixed evidence that malfeasance information impacts voting behavior, we make three main contributions. First, we demonstrate that the nature of voter sanctions (rewards) depends on the extent to which voter priors were relatively positive (negative). While relatively negative priors entailed rewards for many incumbents in practice in our study, lower expectations of incumbent malfeasance may explain the apparent greater willingness of voters to sanction in more developed contexts (e.g. Chang, Golden and Hill 2010; Eggers 2014). These results extend Banerjee et al.(2011), who start from a similar theoretical framework, by explicitly focusing on voters prior and posterior beliefs. Nevertheless, the strongest evidence of sanctioning outside consolidated democracies comes from studies examining the role of broadcast media (Ferraz and Finan 2008; Larreguy, Marshall and Snyder 2015; Marshall 2016a). Although our loud speaker did not accentuate the impact of our treatment, radio and television coverage may entail a considerably more powerful signal that drowns out voter priors and primes voters to vote on the issue. More generally, our emphasis on voter priors complements Kendall, Nannicini and Trebbi(2015), who go to great lengths to elicit prior beliefs and demonstrate their importance for understanding how Italian voters respond to campaign information. Our findings similarly reinforce studies showing that corruption information only affects voter attitudes when delivered by a credible source (e.g. Botero et al. 2015). Second, we reinterpret extant findings suggesting that negative campaigning and malfeasance revelations can engender disengagement (Ansolabehere and Iyengar 1995; Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013). Our non-linear explanation for the 20 relationship between performance information and turnout instead relies on the distribution of voters and their beliefs. In our context, we demonstrate that information both increases and decreases turnout. This may thus explain the mixed existing findings (e.g. Banerjee et al. 2011; Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013; Humphreys and Weinstein 2012). Nevertheless, our theory cannot explain the simultaneous decline in incumbent and challenger support that Chong et al.(2015) observe. This remains an important question for further study, but could simply reflect an idiosyncratic feature of their data, given that—consistent with our argument—they only observe a decline in incumbent partisanship without a commensurate decline in challenger partisan attachment. Third, that voter behavior may be conditioned by challenger and especially incumbent strategies highlights the importance of integrating political actors into models examining the effects of information provision. Our findings pertaining to incumbent responses complement results from the Philippines and Sierra Leone, where respectively Cruz, Keefer and Labonne(2015) observe that distributing information about government projects also increases vote buying and Casey(2015) shows that providing voters with information about their candidates lead politicians to readjust their distributive strategies in favor of a more equitable allocation of resources. Using less nefarious means, Cole, Healy and Werker (2012) also find that Indian voters are less likely to punish incumbents for adverse weather shocks when the incumbent responds effectively to the crisis. The remainder of the paper is structured as follows. Section 2.2 first describes the Mexican municipal context in which we build and test our argument. Section 2.3 presents a simple model highlighting the conditions under which we expect new information to increase and decrease a voter’s propensity to turn out and cast a ballot for the incumbent or challenger party. Section 2.4 explains the experimental design, and uses individual-level surveys to validate our interpretation of the treatments. Section 2.5 presents the main findings, while section 2.6 examines the channels through which information impacts voting 21 behavior. Section 2.7 concludes.

2.2 Malfeasance, audits and elections in Mexican munici-

palities

Mexico’s federal system is divided into 31 states (and the Federal District of Mexico City) containing around 2,500 municipalities and 67,000 electoral precincts. Following major decentralization reforms in the 1990s (see Wellenstein, Núñez and Andrés 2006), municipal governments—the focus of this paper—play an important role in delivering basic public services and managing local infrastructure, and now account for 20% of total government spending. Municipalities are governed by mayors typically elected to three- year non-renewable terms.3

2.2.1 Federal audits of municipal spending

A key component of a mayor’s budget is the Municipal Fund for Social Infrastructure (FISM), which represents 24% of their total budget. According to the 1997 Fiscal Coordi- nation Law (LCF), FISM funds are direct federal transfers legally designated for infrastructure projects that benefit the population living in extreme poverty. Eligible projects include investments in the water supply, drainage, electrification, health infrastructure, education infrastructure, housing and roads. However, voters remain poorly informed about both the resources available to mayors and their responsibility to provide basic public services (Chong et al. 2015). The use of FISM transfers is subject to independent audits. Responding to high levels of perceived mismanagement of public resources, the Federal Auditor’s Office (ASF) was es-

3Re-election will become possible for incumbents from 2018.

22 tablished to audit the use of federal funds in 1999. Although the ASF reports to Congress, its management autonomy is constitutionally enshrined and it has the power to impose fines, recommend economic sanctions and file or recommend criminal cases against public officials (see Larreguy, Marshall and Snyder 2015). The ASF selects around 150 municipalities for audit each year, based primarily on the relative contribution of FISM transfers within the municipal budget, historical performance and factors that raise the likelihood of mismanagement, and whether the municipality has recently been audited (including concurrent federal audits of other programs) (Auditoría Superior de la Federación 2014). The municipalities to be audited in a given year are announced after spending has occurred. Audits address the spending, accounting and management of FISM funds from the previous fiscal year. Although the ASF’s reports categorize the use of FISM funds in various ways, we focus on two key dimensions of mayoral malfeasance documented in the audit reports (that are not necessarily mutually exclusive): the share of funds spent on projects not directly benefiting the poor and the share of funds spent on unauthorized projects. Spending not benefiting the poor entails the allocation of FISM funds to social infrastructure projects that do not principally benefit voters in extreme poverty. Unauthorized spending primarily includes the diversion of resources for non-social infrastructure projects (e.g. personal expenses and election campaigns) and funds that are accounted for. Such spending is akin to the corruption identified by Brazilian audit reports (Ferraz and Finan 2008). The results for each audited municipality are reported to Congress in February the year after the audit was conducted, and are made publicly available on the ASF’s website. Mayoral malfeasance is relatively high. Between 2007 and 2015, 8% of audited funds were spent on projects not benefiting the poor, while a further 6% were spent on unauthorized projects. In one case, the mayor of Oaxaca de Juárez created a fake union to collect payments, presided over public works contracts without offering open tender, diverted ad-

23 vertising and consulting fee payments, and failed to document quantities of spending.4 In another instance, 9 municipal governments in the state of Tabasco—Centro, Balancán, Cár- denas, Centla, Jalapa, Jonuta, Macuspana, Tacotalpa and Tenosique—diverted resources to fund the 2012 electoral campaigns of their parties’ candidates.5 It is thus unsurprising to ﬁnd that 50% of voters do not believe that municipal governments honestly use public resources (Chong et al. 2015).

2.2.2 Municipal elections

Traditionally, local political competition has been between either the populist Institu- tional Revolutionary Party (PRI) and the right-wing National Action Party (PAN), or the PRI and its left-wing offshoot Party of the Democratic Revolution (PRD). Due to regional bases of political support and highly localized influence within municipalities, local politics is typically dominated by one or two main parties. In the municipal elections we study, the effective number of political parties by vote share at the precinct and municipal levels remain consistently around 2.5. To win office, the three large parties often subsume small parties into municipal-level coalitions.6 In 2015, the National Regeneration Movement (MORENA) stood for the first time, and made headway against this hegemony with 9% of the federal legislative vote. Although economic and criminal punishments for misallocating funds are relatively rare, there are good reasons to believe that voters will hold the incumbent party responsible, even without the re-election of individual mayors. First, voters are considerably better informed about political parties than individual politicians (e.g. Chong et al. 2015; Lar- reguy, Marshall and Snyder 2015). Crucially for political accountability, 80% of voters in

4BBM Noticias, “ASF: desvió Ugartchechea 370.9 mdp,” October 21st 2013, here. 5Tabasco Hoy, “Pagaron pobres campañas 2012,” March 6th 2014, here. 6These smaller parties typically benefit by receiving sufficient votes to maintain their registration. 24 our survey can correctly identify the party of their municipal incumbent. Second, Mex- ico’s main parties retain entrenched and differentiated candidate selection mechanisms (Langston 2003). For example, 74% of voters in our survey believe that if the current mayor is malfeasant another candidate from within the same party is at least somewhat likely to also be malfeasant. Third, Marshall(2016 a) also finds that, at least when local media is available, Mexican voters punish their municipal incumbent parties for elevated pre-election homicide rates. However, existing evidence of electoral sanctions against the incumbent party is mixed. As noted above, Larreguy, Marshall and Snyder(2015) observe large electoral penalties among voters with access to broadcast media outlets incentivized to report local news. Exploiting exogenous variation in the release of audit reports prior to elections and access to radio and television stations, they find that each additional local media station—but especially an additional local television station—decreases the vote share of an incumbent party revealed to be malfeasant just before the election by around 1 percentage point. This evidence supports the standard electoral accountability model (e.g. Barro 1973; Fearon 1999; Ferejohn 1986; Rogoff 1990). Chong et al.(2015) find different results in a field experiment conducted in 12 municipalities across 3 states. Disseminating to voters leaflets containing audit report outcomes, they instead find that while incumbent support declines when the incumbent is revealed as highly malfeasant, challenger support declines at least as much. They speculate that such broad-based disengagement, also observed through reduced partisan attachment to the incumbent, reflects an equilibrium where voters disengage because they believe that all politicians are corrupt.7 The disjuncture between these accountability and disengagement findings, which cover

7In the context of our model below, we could think of this as increasing the cost of turning out or inducing the difference in expected competence between candidate parties to zero.

25 the same information over the same period, illustrate the need for a theoretical logic able to identify when different types of information impact voters differently. With a bimodal distribution of voters, our learning model offers an explanation for voter disengagement among incumbent supports that we test using a ﬁeld experiment.

2.3 Prior beliefs and voting behavior

We now explore how information about incumbent malfeasance may impact electoral accountability. A key insight of our simple learning model is that the impact of information on voter beliefs and vote choice rests on how the information relates to voters’ prior beliefs. While high levels of malfeasance are clearly bad news, it is not obvious whether voters will reward or punish incumbent parties for low but non-zero levels of malfeasance. Moreover, with a cost to voting and a bimodal distribution of voters, information relatively close to voter priors can reduce turnout, while major departures can cause wholesale shifts in support from incumbent to challenger (or vice versa). The model derives the key comparative static predictions that guide our experimental design.

2.3.1 Theoretical model

We consider a simple decision-theoretic model where a unit mass of voters update about party competence for ofﬁce from informative signals, and choose between voting for the incumbent I,8 voting for the challenger C and abstaining.9 As noted above, two party competition holds in most parts of Mexico. Voters receive expressive utility from voting for their favored party, rather than believe

8The theory can be easily adjusted to allow for imperfect within-party candidate correlations. 9In the model, we do not consider the possibility that parties react to the provision of information to counteract its effect. Empirically, however, we provide some evidence of this below.

26 that their choice will impact the electoral outcome (e.g. Brennan and Hamlin 1998). This includes a competence and a ﬁxed partisan component, such that the utility to voter i of voting for party p ∈ {I,C} is given by:

 h i E −exp − (θ + δ ) if p = I p  I i Ui = (2.1)  E[−exp(−θC)] if p = C

where θp is the competence of party p and δi ∈ Γ ⊆ R is a positive or negative partisan bias toward the incumbent. For analytical simplicity in incorporating risk aversion, we employ a standard exponential utility function satisfying constant absolute risk aversion utility, where competence and partisanship are perfect substitutes. The partisan shock δi is independently and identically distributed across the electorate according to cumulative distribution function F. This shock, which is uncorrelated with perceptions of candidate competence, could reﬂect durable partisan attachments, clientelistic transfers or shocks occurring before the election and after voters receive information. Let c > 0 be the cost of turning out to vote. Voters thus turn out if the utility from voting for their preferred party p is sufﬁciently large relative to their utility from voting for the other party −p, i.e. p −p Ui −Ui > c.

However, voters are uncertain about the competence θp of both the incumbent and challenger parties, and learn from common information pertaining to party competence such as malfeasance revelations in a Bayesian fashion. In particular, we assume that all voters share the same normally distributed prior beliefs over the competence of each party

2 p, distributed according to N(µp,σp ). Focusing on the case where voters only receive an audit report documenting malfeasance that pertains to the incumbent, voters observe a

2 signal sI drawn from a normal distribution of signals N(θI,τI ) centered on the incumbent’s true but unknown competence level θI. For simplicity, we consider the case where the

27 competence of each party p is independently distributed; voters thus assume that this signal is uninformative about the challenger (c.f. Kendall, Nannicini and Trebbi 2015).10 The

2 known variance of this signal, τI , reﬂects the fact that the audit report may only capture one dimension of an incumbent’s competence or malfeasance. We discuss the extension to

also receiving a signal sC of challenger quality below.

After receiving a signal of incumbent competence sI, voters form their posterior belief over the incumbent’s competence using Bayes’ rule:

2 N µI + κI(sI − µI),κIσI (2.2)

2 τI where κI ≡ 2 2 captures the strength of the signal relative to the prior. Higher values σI +τI of κI indicate that the signal is relatively more precise than voter prior beliefs. Voters thus update positively about expected incumbent competence if s > µ, and particularly when the signal is precise relative to the prior. Because the beliefs about the competence of both

parties are independent, voters do not update about θC. New information also reduces the 2 2 variance of voter posterior beliefs, given that κIσI < τI .

A positive incumbent performance signal (i.e. sI > µI), increases the utility of voting for I by both increasing the incumbent’s expected competence and reducing i’s uncertainty over policy outcomes. This is reﬂected in the following expected utility received by voter i:

 2  κIσI −exp − µI + κI(sI − µI) − 2 + δi if p = I p  Ui = (2.3) 2 −exp − µ − τC if p = C  C 2

10At the cost of mathematical complexity this could be relaxed, and would yield similar results for a sufﬁciently small correlation between sI and θC.

28 where the negative term inside the parentheses reﬂects voter risk-aversion. Integrating over the distribution of voter partisan biases, it is easy to derive the share of voters Vp turning out for each party and the share of abstainers A:

2 2 τ − κIσ V = 1 − F c − ∆ − κ ∆ − C I , (2.4) I µ I I 2 2 2 τ − κIσ V = F −c − ∆ − κ ∆ − C I , (2.5) C µ I I 2 2 2 2 2 τ − κIσ τ − κIσ A = F c − ∆ − κ ∆ − C I − F −c − ∆ − κ ∆ − C I ,(2.6) µ I I 2 µ I I 2

where ∆µ ≡ µI − µC and ∆I ≡ sI − µI are respectively the difference in (the expectation of) voter priors between incumbent and challenger and the extent of updating about the 2 2 τC−κIσI incumbent’s competence from the signal, while 2 represents the difference in risk associated with C relative to I. These vote shares are perhaps most intuitively illustrated graphically in Figure 2.1, which plots the distribution of voters by their relative preference 2 2 I C τC−κIσI ∆U ≡ Ui −Ui = ∆µ + κI∆I + 2 +δi for the incumbent. We can analyze how voting behavior is impacted by the key parameters in our model by simply shifting the distribution along the ∆U axis. As illustrated by the dotted distribution, a relatively weak positive signal of incumbent competence relative to voters’ priors beliefs entails a small increase in ∆I and a reduction in the relative risk of voting for I, and thus a commensurate shift in the distribution of relative voter preferences to the right. This un- equivocally increases the number of voters supporting I and decreases the number of voters supporting C. Similarly, an increase in the difference between incumbent and challenger priors (i.e. greater ∆µ ), or the precision of a relatively positive signal (i.e. greater κI, or 2 lower σI ), also shift the distribution to the right and increase the incumbent party’s vote share. While the incumbent vote share results hold for any distribution F of partisan attach-

29 Vote C Abstain Vote I Density of voters

-c 0 c

Preference toward I (∆U)

Prior Small positive update Large positive update

Figure 2.1: Vote choice and distributions of voters

ments, the implications of providing information for turnout depend upon the shape and position of F and the extent to which information induces updating. Without loss of gener-

ality, consider the case of receiving sI > µI.A positive signal about the incumbent has two effects, again by shifting voter expectations and reducing aversion. First, it induces some voters that would not have otherwise voted to turn out for I; and if the signal is sufﬁciently powerful, some C may also shift to I. Second,the positive signal also induces some voters (that would have otherwise voted for C) not to turn out. The relative masses of these conﬂicting effects on turnout determine whether turnout increases or decreases. To produce sharp empirical predictions, we focus on the case where voter partisan attachments are bimodally distributed. In addition, we assume that the voters at each mode turn out for different parties. In many electoral contexts, including Mexico, this is a reasonable approximation. As noted above, the geographic dispersion of party strength that

30 ensures most races are effectively two-party races. Furthermore, our survey indicates that, of the two-thirds of voters with a partisan attachment, 73% rate the strength of their attachment at 5 or higher on a 7-point scale. The non-linear effect of information in turnout is best illustrated graphically using the example in Figure 2.1. The dotted distribution shifted slightly to right demonstrates that a small update in favor of the incumbent causes more initial C voters to abstain than initial abstainers to vote I. This is easy to see by comparing the mass under each distribution over the interval [−c,c]. However, a sufﬁciently large positive update about the incumbent— which shifts the distribution further to the right—induces initial C supporters to vote I rather than abstain. More generally, it is easy to see that such non-linear predictions hold for any bimodal distribution where the voters at each mode initially turn out for different parties.11

2.3.2 Empirical implications

The model generates various comparative static predictions, some more particular to our model than others. We focus on the impact of providing voters with a signal of incumbent performance, sI, via a treatment containing incumbent performance information pertaining to mayoral malfeasance. We now enumerate the key hypotheses that our experiment is designed to test empirically; all hypotheses were registered in our pre-analysis plan before the experiment was conducted. We first consider how incumbent malfeasance information affects posterior beliefs regarding the incumbent party’s competence and vote choice. As equation (2.2) shows, the direction of updating from a given signal sI depends on the prior expectation µI. The effect is thus context-dependent, reﬂecting both the nature of the information provided and voter

11To see this, consider the derivative of the density function at the turn out cutoffs.

31 prior beliefs regarding incumbent competence. Given the relatively negative incumbent performance information we will typically provide voters with, and despite the reduction in incumbent uncertainty that this provides, we anticipated that on average:

H1. Incumbent malfeasance information increases voter posterior beliefs that the incumbent party is malfeasant and decreases the incumbent party’s vote share.

However, the most important implications of the model capture how the effect of incumbent malfeasance information varies with voter prior beliefs. These are theoretically unambiguous in our model. First, if voters already believe that the incumbent party is malfeasant (i.e. low µI), a signal of high malfeasance has a weaker impact on negative posterior beliefs and the incumbent party vote share:

H2. The effect of incumbent malfeasance information on voter posterior beliefs that the incumbent party is malfeasant is decreasing in voter prior beliefs that incumbent is malfeasant, while the effect of incumbent malfeasance information on incumbent party vote share is increasing in such voter prior beliefs.

Second, voters already possessing strong prior beliefs about the incumbent’s competence

(i.e. low κI) are less responsive to new information:

H3. The effect of incumbent malfeasance information on voter posterior beliefs that the incumbent party is malfeasant and incumbent vote share is weaker among voters with strong prior beliefs.

Third, voters update more positively (negatively) about the incumbent party’s competence upon hearing that the incumbent is relatively clean (malfeasant):

H4. The effect of incumbent malfeasance information on voter posterior beliefs that the incumbent party is malfeasant is increasing in the level of reported malfeasance, 32 while the effect of incumbent malfeasance information on incumbent party vote share is decreasing in the level of reported malfeasance.

Finally, given that the extent of voter belief updating reﬂects the difference between the signal and voter prior beliefs,

H5. The effect of incumbent malfeasance information on voter posterior beliefs that the incumbent party is malfeasant is increasing in the extent to which the information is worse than voter priors, while the effect of incumbent malfeasance information on incumbent party vote share is decreasing in the extent to which the information is worse than voter priors.

In absolute terms, and given voter prior beliefs, the direction of updating produces different responses:

H6. Good (bad) news about the incumbent party decreases (increases) voter posterior beliefs that the incumbent party is malfeasant and increases (decreases) the incumbent party’s vote share. Neutral news leaves both outcomes unchanged.

As demonstrated above, the effect of new information on turnout is non-linear when voters are bimodally distributed with voters at each mode initially turning out for different parties, as the evidence suggests is the case in Mexico. In particular, shockingly positive or negative revelations induce voters to switch parties, while relatively unsurprising but nevertheless informative positive (negative) information induces challenger (incumbent) partisans to become relatively indifferent between the parties and instead abstain when faced with a cost of turning out. This logic does not yield clear predictions for the average effect of new information or its linear interaction with the level of malfeasance reported. However, a clear prediction of the theory is that:

33 H7. High and low levels of reported incumbent malfeasance information increase electoral turnout, while relatively unsurprising middling levels of reported malfeasance decrease electoral turnout.

2.3.3 Variation in information’s impact

Our model can be extended to incorporate the impact of simultaneously providing challenger information, and thus permitting relative performance comparisons. We anticipate that such information will impact voter posterior beliefs about the competence of the challenger akin to equation (2.2). However, the effect on vote choice depends on the relationship between incumbent and challenger signals. Empirically, we will restrict attention to the case where challenger party information is provided alongside either notably worse or notably better incumbent party malfeasance information. Comparative performance information thus always provides a stronger signal, by compounding a single signal with a clear benchmark for comparing sI to sC. For average malfeasance revelations, in our experimental sample where the majority of voters receive information indicating stronger challenger performance, we expected that:

H8. Comparative malfeasance information on average decreases the incumbent party’s vote share more than just incumbent malfeasance information.

Furthermore, this anti-incumbent party effect was expected to be particularly large when voters learn of especially poor incumbent performance—in both absolute and relative terms:

H9. The effect of comparative malfeasance information on incumbent party vote share decreases more quickly in the level of reported incumbent malfeasance, and the difference relative to the reported malfeasance of the challenger, than just incumbent malfeasance information.

34 Another potentially important factor in explaining when information supports political accountability is the method of information transmission. Public modes of transmission— through which voters become aware that other voters have also received a given piece of information—could produce powerful effects by inducing explicit or tacit voter coordination based on their common knowledge (e.g. Adena et al. 2015; Arias 2015; Yanagizawa- Drott 2014). Explicit discussion of the information may result in voters engaging heavily with the information received, and in turn consolidating their beliefs around such information. This could even result in agreement to harmonize vote choices. Alternatively, tacit coordination only relies on the delivery of information engendering the (higher-order) belief that others also received the information, and will likely act on such information. Such coordination may in part explain the large effects of local media found in Mexico (Larreguy, Marshall and Snyder 2015; Marshall 2016a). Both such cases of coordinated behavior induce shifts that could not be achieved by providing the same information using private modes of information transmission. In the context of our model, where we do not explicitly model voter discussion and coordination, public delivery mechanisms are perhaps best interpreted as a clearer signal

2 of performance (lower σI ). We thus expected the social treatment to magnify any impact of incumbent malfeasance information:

H10. The average and heterogeneous effects of incumbent malfeasance information on the incumbent party’s vote share (H1-H7) are greater when the information is delivered through a public mechanism.

2.4 Experimental design

We designed a field experiment to test our theoretical argument. Specifically, we focus on Mexico’s municipal elections held on Sunday 7th June 2015, and seek to identify 35 the effect of providing voters in 678 electoral precincts with the results of audit reports documenting the municipal use of federal transfers designated for infrastructure projects benefiting the poor. Municipal elections were held concurrently with state and federal legislative elections. We first explain our sample selection, before outlining our information interventions and how voters interpreted them.

2.4.1 Sample selection

Our study focuses on 26 municipalities in the central states of Guanajuato (7 municipalities), México (14 municipalities), San Luis Potosí (4 municipalities) and Querétaro (1 municipality). These are shown in Figure 2.2. Beyond holding elections in 2015,12 these states were chosen for logistical reasons, because they contain internal variation in the municipal incumbent party and because they broadly represent Mexico as a whole. The 26 municipalities were selected from those where an audit was released in 2016 according to three criteria. First, the safety of voters and our distribution and survey teams. After immediately receiving threats upon entering Aquismón and Villa Victoria, these municipalities were replaced by Atlacomulco, Temoaya and additional precincts from Tlalnepantla de Baz in the state of México. Importantly, our blocking strategy—explained in detail below—ensures that all blocks are contained within the same municipality, and thus excluding these problematic municipalities does not affect the internal validity of our study. Second, we only selected municipalities over-performing, and especially under-performing, relative to the state average of opposition parties by at least 2 percentage points. Finally, municipalities were chosen to match the distribution of incumbent parties across audited municipal governments across the 4 states. Of our 26 municipalities, 17 were governed by the PRI (including 16 in coalition with the Teacher’s (PANAL) and Green (PVEM) parties), 5 by

12Municipal elections reﬂect state electoral cycles, which are staggered across years. On 7th June 2015, 15 states and the federal district held simultaneous local elections. 36 Figure 2.2: The 26 municipalities in our sample

the PAN (including 2 in coalition with PANAL), 2 by the PRD, and 1 by the Citizen’s Movement (MC). Within each municipality, we selected up to one third of the electoral precincts, over- sampling precincts from municipalities with particularly high and low levels of incumbent malfeasance and strong contrasts with opposition party performance within the state. We first prioritized accessible rural precincts, where possible, in order to minimize cross- precinct spillovers and maximize the probability that voters would not receive the information through other means. Moreover, to maximize the share of households we could reach with a fixed number of leaflets, attention was restricted to precincts with fewer registered voters (subject to our blocking design detailed below). In urban areas, where we had more precincts to chose from, we restricted our sample to precincts with at most 1,750 registered voters, and designed an algorithm to minimize the number of neighboring urban precincts

37 in our sample.13

2.4.2 Treatments

In partnership with the non-partisan Mexico NGO Borde Político,14 we sought to evaluate the impact of providing voters with incumbent performance information. Our primary treatment distributed leaflets to voters documenting the use of FISM funds in their municipality. For each municipality, the leaflet focused on either spending that does not benefit the poor or unauthorized spending. All treatments were delivered at the electoral precinct level, Mexico’s lowest level of electoral aggregation. Our leaflet was designed to be non-partisan, accessible and sufficiently intriguing that voters would not discard it.15 The front page of the leaflet explains that Borde Político is a non-partisan organization and that the information contained in the leaflet is based on the ASF’s official audit reports available online. Figure 2.3 provides an example of the main page from a leaflet focusing on a severe case of unauthorized spending in the municipality of Ecatepec de Morelos in the state of México. The leaflet first states that FISM funds should only be spent on social infrastructure projects, and provides graphical examples of such projects on the right. The leaflet then informs recipients of the total amount of money their municipality received (146.3 million pesos, in this case), and the percentage of this money spent in an unauthorized way by their government (45%). To avoid suspicions of

13The algorithm started with the set of neighboring precincts surrounding each precinct and identified all neighboring precincts that were eligible for our sample; we then iteratively removed the precinct with most “in-sample” neighbors until we reached the required number of precincts for that municipality. In most municipalities, the algorithm ensured that our sample contained no neighboring precincts. 14Borde Político is a leading NGO seeking to increase voter knowledge about the actions of their politicians in office, with significant experience in the development of web-based platforms providing politically-relevant information to individuals (see borde.mx). 15The particular design was produced by a graphic artist based on feedback from multiple focus groups. We also sought legal advice to ensure that our leaflets did not constitute political advertisements, and thus were not subject to distribution restrictions.

38 EL DINERO DEL FISM, FONDO DE INFRAESTRUCTURA AGUA DRENAJE CAMINOS LUZ SOCIAL MUNICIPAL, DEBE POTABLE

GASTARSE EN OBRAS DE ESCUELAS CLÍNICAS VIVIENDA INFRAESTRUCTURA

LOS GASTOS QUE NO SEAN EN OBRAS DE INFRAESTRUCTURA DEBEN SER 0%

EN 2013, EL PARTIDO QUE

GOBIERNA ECATEPEC RECIBIÓ AGUA DRENAJE CAMINOS LUZ POTABLE 146.3 MILLONES DE PESOS DEL FISM Y GASTÓ 45% EN COSAS ESCUELAS CLÍNICAS VIVIENDA QUE NO DEBE

GASTÓ COMO NO DEBE

45 PARTIDO QUE GOBIERNA ECATEPEC

DE ¡PIÉNSALO! EL EL VOTO ¡COMPÁRTELO! 7 JUNIO DEPENDE DE TI

Figure 2.3: Example of local-information only leaﬂet in the municipality of Ecatepec de Morelos in the state of México

political motivation, neither the mayor nor their party is referred to directly, although as noted above the vast majority of voters can correctly identify the party of their incumbent mayor. Figure 2.4 provides an example from the municipality of Salamanca in Guanaju- ato, where all 54.1 million pesos were correctly allocated to social infrastructure projects benefiting the poor. To investigate how the provision of information impacts voters, we varied leaflet dissemination along two theoretically important dimensions. First, to identify the impact of providing voters with a benchmark against which to compare their municipality’s performance, we produced a comparative leaflet. In contrast with the local leaflet shown in

39 EL DINERO DEL FISM, FONDO DE INFRAESTRUCTURA SOCIAL MUNICIPAL, DEBE GASTARSE EN OBRAS QUE BENEFICIEN A LOS QUE MENOS TIENEN.

¡LOS GASTOS EN OBRAS QUE NO BENEFICIAN A LOS QUE MENOS TIENEN DEBEN SER 0%

EN 2013, EL PARTIDO QUE GOBIERNA SALAMANCA RECIBIÓ 54.1 MILLONES DE PESOS DEL FISM Y GASTÓ 0% EN OBRAS QUE NO BENEFICIAN A LOS QUE MENOS TIENEN.

¡COMPAREMOS CON LOS GASTOS DE OTROS PARTIDOS!

MUNICIPIOS DE TU ESTADO GOBERNADOS POR OTROS PARTIDOS GASTARON EN PROMEDIO 16% EN OBRAS QUE NO BENEFICIAN A LOS QUE MENOS TIENEN.

GASTOS QUE NO BENEFICIAN A LOS QUE MENOS TIENEN

0 16 PARTIDO QUE OTROS GOBIERNA PARTIDOS SALAMANCA EN TU ESTADO

DE ¡PIÉNSALO! EL EL VOTO ¡COMPÁRTELO! 7 JUNIO DEPENDE DE TI

Figure 2.4: Example of a comparative information leaﬂet in the municipality of Salamanca in the state of Guanajuato

Figure 2.3, the comparative leaflet in Figure 2.4 involved providing information about the mean outcome among all audited municipalities within the same state but governed by a different political party in the second and third panels. In this example, voters in Salamanca were informed that while their government spent all FISM funds on projects not benefiting the poor, audited municipal governments run by other parties in Guanajuato spent 16% of funds on projects not benefiting the poor. Second, to vary the extent to which the distribution of the leaflets is common knowledge among voters within the precinct, we also varied whether the leaflets was delivered in a private or public manner. For the public mode of delivery, the door-to-door delivery de-

40 hi w uiiaiyad9 nmncplte ihntersae oendb te par- other by governed states their within within spending municipalities malfeasant in 21% 9% and of municipality informed own was their precinct average The state. municipalities the audited within governing parties other among average the and municipality voter’s and share to provided. them information encouraged the and information discuss mayor, receive municipal also their would of neighbors performance their the concerning that voters on informed message message second The 30 a loop. playing while leaflets distributing members team other alongside single a messages, campaign blaring of back the on member. carried team speaker a loud portable powerful a by accompanied was above scribed oiia apin iial ekn obodattermessage. their broadcast to seeking similarly campaigns political iue25 uiiaiisadpeicsb hr fmlesn pnigi u sample our in spending malfeasant of share by precincts and Municipalities 2.5: Figure 16 sostedsrbto fmlesn pnigi u ape ohwti a within both sample, our in spending malfeasant of distribution the shows 2.5 Figure fpeicsi u apefo htmunicipality. that from sample our in precincts of Notes eprhsdteemdfidrcsc odsekr rmavno nMxc iyta loserves also that City Mexico in vendor a from speakers loud rucksack modified these purchased We ahpiti n for2 uiiaiis h ieo onscrepnst h number the to corresponds points of size The municipalities. 26 our of one is point Each :

Incumbent malfeasant spending (other parties in state) 16 0 .05 .1 .15 .2 knt h eilscmol rvn rudbfr eia elections Mexican before around driving commonly vehicles the to Akin 0 Incumbent malfeasantspending perifonista .2 41 akdtruhtesreso ahprecinct each of streets the through walked .4 .6 ties, while these variables respectively ranged from 0% to 58% and 0% to 18%; 74% of precincts received relatively bad news about their incumbent party.

2.4.3 Block randomization, compliance and data

Our sample of 678 precincts was allocated according to the factorial design with a pure control shown in Table 2.1. To maximize our ability to differentiate the effects of comparative information and public dissemination, the 400 treated precincts were equally divided between the four variants of the information treatment. The control group comprising 278 electoral precincts reflects our sampling and block randomization. For the randomization, precincts were first stratified into rural or urban blocks of 6 or 7 similar precincts within a given municipality.17 Precinct similarity was defined by the Mahalanobis distance between 23 social, economic, demographic and political variables provided by Mexico’s National Statistical Agency (INEGI) and the National Electoral In- stitute (INE).18 Within each block, we then randomly assigned precincts to each of the treatment conditions and, depending on the availability of an additional precinct, either 2 or 3 pure control precincts. Block randomization ensures that different municipalities do not receive different treatment dosages and maximizes the power of the experiment by minimizing differences between treated and control precincts. Leaflets were distributed by our implementing partners Data OPM and Qué Funciona para Desarollo using precinct maps provided by state electoral institutes. Starting from a random location, our distribution team delivered one leaflet to a maximum of 200 households within each treated precinct. Where possible, leaflets were delivered in person with

17Subject to there existing sufficient precincts, and the total treated precincts not exceeding one third of all precincts, we used blocks of 7 precincts. 18We used the R package blockTools to assign precincts to blocks. The algorithm is greedy in that it creates the most similar group first. Where a surplus of potential precincts were available, we used the most similar blocks to maximize statistical efficiency.

42 Table 2.1: Factorial design with a pure control Control Private Public Control 278 precincts Local 100 precincts 100 precincts Comparative 100 precincts 100 precincts

a short message explaining the leaflet’s provenance. When no adult was available, leaflets were left in mail boxes or taped to the recipient’s front door in a water-proof bag. This process typically took several hours per precinct, and was implemented over two weeks up until the no-campaigning period a week before the election. Our team recorded where leaflets were distributed for our follow-up survey. While compliance with the delivery of our treatments was very good in general, we nevertheless encountered some issues in the field.19 In a couple of cases, some leaflets were delivered to voters outside the precinct or adverse weather conditions and poor road conditions prevented us from reaching a precinct. To preserve the randomization, we focus on estimating intent to treat (ITT) effects. Such estimates are also most policy-relevant. Finally, we collected two additional sources of data to measure our principal outcomes. First, we downloaded and requested official precinct-level electoral returns from each state’s electoral institute. We drop 3 precincts in our sample that were merged with another precinct, due to having fewer than 100 registered voters, leaving a final sample of 675 electoral precincts.20 Second, we conducted a post-election survey interviewing 10 voters from each of the treated precincts and 100 randomly selected control precincts.21 This survey included questions regarding voting behavior, perceptions of incumbent and

19In addition, a new block in Chimalhuacán replaced a block where treatments were misassigned. We move the misassigned block from our sample, but the results are unaffected by its inclusion. 20In two of these cases, the small precinct was merged with another precinct that remains in our sample; where the treatment condition conﬂicts, we retain the larger precinct’s treatment status. 21Budgetary constraints prevented a baseline survey.

43 challenger parties, social coordination, the responses of political parties to our information, and receipt of the leaﬂet.22

2.4.4 Estimation and balance

To estimate the average ITT effect of providing any type information—local or comparative, or public or private—we estimate OLS regressions of the form:

Ypbm = βTreatmentpbm + ρbm + εpbm, (2.7)

where Ypbm is an outcome for electoral precinct p within randomization block b in municipality m. For individual-level survey outcomes, Yipbm also includes an i subscript.

Block fixed effects, ρbm, are included to enhance the efficiency of our estimates by en- suring that we only exploit variation in treatment assignment within blocks of similar precincts. Throughout, standard errors are clustered at the municipality-treatment level, while precinct-level observations are weighted by the inverse of the share of voters to whom we delivered a leaflet (in control precincts, we use the average number of leaflets delivered to treated precincts). This weighting scheme—the only departure from our pre-registered regression specifications—permits more precise estimates by de-weighting large precincts where only a small fraction of voters could receive the leaflet. We show similar results weighting each precinct equally as a robustness check in Table 2.6 below. We also use this baseline specification to validate the randomization. Unsurprisingly, as Table A.1 in the Appendix demonstrates, the treatment is well balanced across 46 precinct

22Despite its success in the administration of the Mexico Panel Surveys and Comparative Study of Electoral Systems modules, our attempts to gauge vote choice by simulating the electoral process with an urn during the survey failed. Many voters felt uneasy in the aftermath of surprising electoral results, and refused to participate believing that our survey team were working on behalf of a party to identify individual vote choices or to trick voters into casting a different ballot. Unfortunately, we had to discard this question from our analysis.

44 and survey respondent characteristics. As usual, there are some significant differences, most notably with respect to incumbent vote share. However, in Table 2.6 we demonstrate that the results are robust, and if anything more precisely estimated, when controlling for pre-treatment variables. Our main estimates pool municipalities receiving information about spending on projects not benefiting the poor and unauthorized spending. If voters evaluate these dimensions of malfeasance similarly, as the findings of Larreguy, Marshall and Snyder(2015) suggest, this maximizes the power of the experimental design. Since voters could plausibly respond to negligent and corrupt spending differently, we also examine the types of information separately in Table 2.5 and observe similar responses. To analyze the mechanisms underpinning the information treatment, we estimate analogous specifications separating out the dimensions of our information treatment. To examine how the effects of providing audit result information vary with how the information is perceived by voters, we estimate interactive specifications of the form:

Ypbm = βTreatmentbm + γ Treatmentpbm × Xbm + ρbm + εpbm, (2.8)

where Xbm is a block-level measure of voter prior beliefs, leaflet content, or how leaflet content causes voters to update their beliefs; the lower-order term Xbm is absorbed by the block fixed effects. These interaction variables are described below.

2.4.5 How do voters interpret the information treatment?

Before turning to our main voting outcome results, we ﬁrst examine how the information treatment affected voter actions and perceptions using our post-election survey. Since, within treated precincts, the survey was designed to interview only the voters that received

45 leaflets, survey estimates remain unweighted.23 We first conduct several “manipulation checks” to ensure that treated voters indeed experienced the treatment as intended. The 4 self-reported outcomes in Table 2.2 provide clear evidence that voters received and engaged with the information that we provided. Column (1) demonstrates that treated voters are 24 percentage points more likely to report remembering receiving our leaflet, relative to a control mean of 9% of voters. Moreover, column (2) confirms that voters in treated precincts were 17 percentage points more likely to report having read the leaflet, while column (3) shows that treated voters were 14 percentage points more likely to correctly recall the spending the leaflet pertained to.24 Given that voters may only hazily recall receiving a pamphlet around the election, these differences are likely lower bounds. Moreover, the majority of people believed the leaflet came from an NGO—far more than from any government source or political party. Finally, column (4) indicates that 7% of treated voters reported that the leaflet influenced their vote choice, 5 percentage points more than voters located in control precincts. While a considerable number of voters report that the information contained in the leaflet influenced their vote, we expect differences in information content to impact voters differently. For example, it is not obvious whether misallocated spending totaling 10% represents good or bad news without comparing information content with voters expectations. Accordingly, we asked voters to rate, on a five-point scale from very low to very high, the level of corruption or the level of interest in supporting the poor from each major party (depending on the type of provided in the leaflets).25 Without a baseline survey, we

23Predicting the likelihood that respondents reported receiving the leaflet, we find no significant interaction between the treatment and the share of voters in the precinct that could be treated. 24In addition to the both types of spending, respondents were also given the opportunity to say that the leaflet contained unemployment or public security information. 25Because we did not ask explicitly about the MC, which was the sole incumbent in Apaseo el Alto, the 24 precincts from this municipality are dropped from analyses using prior beliefs.

46 Table 2.2: Effect of information treatment on self-reported engagement with leaﬂet

Remember Remember Correctly Leaflet leaflet reading remember influenced leaflet content vote (1) (2) (3) (4) Information treatment 0.247*** 0.171*** 0.138*** 0.051*** (0.022) (0.018) (0.019) (0.010)

Outcome range {0,1} {0,1} {0,1} {0,1} Control outcome mean 0.09 0.05 0.06 0.02 Control outcome std. dev. 0.28 0.22 0.25 0.14 Treatment mean 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 R2 0.11 0.09 0.10 0.06 Observations 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, and are estimated using OLS. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. proxy for the prior beliefs of treated voters using the average perception among the voters surveyed in the control precinct within the same block.26 Akin to the respondents in Chong et al.(2015), the distribution of such beliefs about the (pre-election) municipal incumbent party’s competence in the control group, shown in Figure 2.6, demonstrates that voters are pessimistic about the incumbent party. Respondents are essentially equally likely to report that the incumbent party engages in high or very high levels of corruption or misallocated spending as low or very low levels of corruption or misallocated spending. This indicates that voters expect their municipal mayors to engage in non-trivial levels of malfeasance in office. To capture the extent to which the treatment caused voters to negatively update their perceptions of the incumbent party, we showed voters in control precincts our leaflet at

26We pre-specified that priors would be defined by control voters at the municipal level. However, focus on the block level controls to produce more precise measures, but nevertheless show in Tables A.12 and A.13 in the Appendix that this choice does not affect our results. 47 eaie oee,oraayi rmrl oue ntesoe ..cmaigteipc ftetreatment the generally of was impact the updating comparing voter negatively. challenger, i.e. less slope, the or more the than relatively on updated well focuses voters less primarily where far analysis our doing However, incumbent negative. the depicted average, on Our party. incumbent the about positively update to voters 45% ex- caused of were spending learning priors unauthorized even voter and where incumbent, case PRI-PVEM-PANAL in odd their México, an about of negative is state ceedingly Morelos the de in Ecatepec Morelos, corner. de right Ecatepec bottom of the case circled the of leaflet. exception the the in with malfeasance presented of information level audit the with the correlated by positively revealed is updating shows negative 2.7 of Figure extent Unsurprisingly, the leaflet. that the control reading block’s after the and among before given perceptions respondents a malfeasance survey in voters in treated difference all average of the updating negative with for block proxy we high), (very -2 2 from to ranging low) scale (very five-point a Opera- as party. perceptions incumbent malfeasance the party perceived they incumbent how tionalizing asking again before survey the of end the iue26 ecie nubn orpinadntsedn ntepo mn oesin voters among poor the on spending not and corruption incumbent Perceived 2.6: Figure 27 l epnet rmtecnrlgopwr hw h oprtv nomto eflt ie htthis, that Given leaflet. information comparative the shown were group control the from respondents All

Frequency 0 100 200 300 Very low Perceived incumbentmalfeasance oto precincts control Low 48 Medium High 27 hsrltosi sstrong, is relationship This Very high 3 2 1

Ecatepec

Negative updating de Morelos 0 -1

0 .2 .4 .6 Share malfeasant spending

Figure 2.7: Scatter plot of audit report outcomes and control group negative updating

robustness check in Table 2.6 shows that removing Ecatepec de Morelos further strengthens our ﬁndings. To understand whether voters update negatively or positively after receiving our information treatments before the election, we examine how the treatment impacted the posterior belief of treated voters regarding the incumbent party’s level of malfeasance. The negative but negligible coefﬁcient in column (1) of Table 2.3 indicates that, if anything, the average treated voter slightly improved their perceptions of their incumbent party. This contrasts with our expectation entering the study that the malfeasance information provided would cause voters to update negatively about the incumbent on average (H1). Nevertheless, it is consistent with the pessimistic priors of Mexican voters, and voters in other countries (e.g. Chong et al. 2015; Lieberman, Posner and Tsai 2014). However, this slightly positive average perception masks substantial heterogeneity in responses across voter priors and municipal context. Consistent with H2, the treatment’s interaction with average voter prior beliefs in column (2) demonstrates that treated voters 49 Table 2.3: Effect of information treatment on voter beliefs about incumbent party malfeasance

Perceived incumbent party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) Information treatment -0.001 -0.031 0.874** 0.016 -0.165*** -0.119* (0.040) (0.047) (0.327) (0.067) (0.059) (0.064) × Incumbent malfeasance prior -0.275*** (0.040) × Strength incumbent prior -0.270** (0.103) × Incumbent malfeasant spending -0.083 (0.214) × Negative incumbent updating 0.178*** (0.036) × Good news -0.241** 50 (0.091) × Bad news 0.234** (0.090)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.14 -0.14 -0.14 -0.14 -0.14 -0.14 Control outcome std. dev. 1.48 1.48 1.48 1.48 1.48 1.48 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.24 0.21 0.89 0.04 / 0.52 Interaction std. dev. 0.89 0.37 0.17 1.07 0.20 / 0.50 R2 0.29 0.30 0.29 0.29 0.30 0.30 Observations 4,624 4,624 4,624 4,624 4,624 4,624

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Respondents from Apaseo el Alto are removed from all specifications. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. in blocks with negative perceptions of the incumbent positively update about the incumbent, while voters in blocks with positive perceptions of the incumbent become substantially more likely to report perceiving their incumbent as highly corrupt or neglectful of the poor. Moreover, column (3) shows that voters in precincts where prior beliefs are relatively weak update negatively about their incumbent. However, also supporting H3, the negative interaction term indicates this effect declines as the strength of voter priors increases. The insignificant interaction in column (4) initially provides little evidence that the share of funds revealed not to have been correctly allocated differentially impacts the beliefs of treated voters (H4). However, this changes once we account for how the information provided relates to prior beliefs. Interacting the treatment indicator with our measure of average voter updating in the control precinct of each block, the large and statistically significant positive coefficient on this interaction in column (5) demonstrates that treated voters in blocks where voters negatively (positively) update about the incumbent display substantially more negative (positive) opinions of the incumbent party. We thus find strong support for H5. The final column documents similar results when information is classi- fied as bad, good or neutral. Specifically, we define bad (good) news by a 0.5 increase (decrease) in the average change in the control group perception of the incumbent upon reading our leaflet. While neutral news slightly reduces voter perceptions of malfeasance, the interactive coefficients show that good news further decreases perceptions of incumbent party malfeasance by around a sixth of a standard deviation while bad news substantially increases such perceptions by one seventh of a standard deviation. Our information treatment could also impact beliefs about challengers (see Kendall, Nannicini and Trebbi 2015), particularly where we also provided information about challenger malfeasance. Tables A.3-A.8 in the Appendix similarly show that treated voters with negative priors about the challenger are more likely to positively update about their posterior beliefs about the challengers’ competence. Given that such effects are similar across 51 the local and comparative treatments, this suggests that voters are primarily updating about challengers from the information they receive about the incumbent, believing them to be positively correlated types.28 However, voter beliefs about challengers are highly correlated with voter beliefs about incumbents, and thus it remains difficult to separate these explanations. Moreover, we show in Tables A.9-A.11, that voting behavior is ultimately driven by how the treatment relates to voter beliefs about the incumbent rather than challengers. Together, these results confirm that voters meaningfully updated their beliefs in response to our treatment content. However, the direction of updating varies substantially across voters, depending on how the information received relates to prior beliefs. Turning to our primary precinct-level electoral outcomes, we now examine whether voter updating translates into actual vote choices.

2.5 Aggregate election results

We now present our two key precinct-level ﬁndings. First, in line with voters’ relatively pessimistic perceptions of incumbent competence, we ﬁnd that our information treatment increased the incumbent’s vote share on average. Consistent with our theoretical model, this effect is greatest where voters updated positively about the incumbent from the information received. Second, we identify a non-linear effect of information on electoral turnout such that medium levels of malfeasance reduce turnout but extreme levels—high or low—increase turnout.

28On 5-point scales of perceived similarity of candidates within parties and across parties, ranging from not all probable to extremely probable that they will behave similarly in ofﬁce, there is a 0.57 correlation in the control group. Similarly, conditional on believing that the incumbent is at least somewhat incompetent, 69% of control group voters believe that challengers are also somewhat incompetent.

52 2.5.1 Heterogeneous effects of information on incumbent vote share

We first examine the average ITT effect of disseminating malfeasance information across our sample. Consistent with the fact that for many voters our information represents good news about the incumbent’s performance, or at least reduced the uncertainty associated with re-electing the incumbent party, the information treatment increases the incumbent party’s vote share on average.29 Column (1) of panel A in Table 2.4 demonstrates that our information increased the incumbent party’s vote share, as a proportion of those that turned out, by 2.9 percentage points on average. Column (1) of panel B similarly shows that this translates into a 1.6 percentage point increase in the incumbent party’s vote share, as a proportion of all registered voters in the precinct. The latter finding indicates that the information caused the incumbent to gain more voters, rather than simply demobilize challenger supporters. Relative to the mean vote share in the control group, the treatment increased incumbent party vote share by 8%. Although voters somewhat surprisingly interpreted our treatment as positive information about the incumbent party on average, voter responses vary with the content of the information received exactly as theorized and in line with our survey data. First, supporting H2, our information treatment’s largest effects are in precincts where voters were initially most negative about the incumbent’s performance in office. Across both panels, column (2) shows that the increase in incumbent party vote share caused by receiving malfeasance information is significantly greater (smaller) in blocks where the control group exhibits more negative (positive) prior beliefs regarding the incumbent’s level of malfeasance. These results indicate that moving from a block with the most positive incumbent prior (-1.6) to

29It is important to reiterate that the model in our pre-analysis explicitly acknowledged this possibility, noting that the expected negative effect is premised on the empirical expectation “that on average the information reported in the leaﬂets will exceed voters’ priors (though this is something that we will be able to test).”

53 Table 2.4: Effect of information treatment on incumbent party vote share

Incumbent party vote share (1) (2) (3) (4) (5) (6) Panel A: Incumbent party vote share (share of turnout) Information treatment 0.029*** 0.027*** 0.076 0.046*** 0.034*** 0.020*** (0.006) (0.006) (0.072) (0.008) (0.007) (0.007) × Incumbent malfeasance prior 0.010* (0.006) × Strength incumbent prior -0.015 (0.022) × Incumbent malfeasant spending -0.080*** (0.027) × Negative incumbent updating -0.009* (0.005) × Good news 0.084** (0.034) × Bad news 0.005 (0.014)

Outcome range [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] Control outcome mean 0.38 0.39 0.39 0.38 0.39 0.39 Control outcome std. dev. 0.12 0.12 0.12 0.12 0.12 0.12 R2 0.63 0.62 0.62 0.63 0.62 0.62 Panel B: Incumbent party vote share (share of registered voters) Information treatment 0.016*** 0.015*** 0.006 0.027*** 0.020*** 0.014*** (0.004) (0.004) (0.038) (0.006) (0.004) (0.004) × Incumbent malfeasance prior 0.008** (0.003) × Strength incumbent prior 0.003 (0.012) × Incumbent malfeasant spending -0.050*** (0.016) × Negative incumbent updating -0.006*** (0.002) × Good news 0.038** (0.015) × Bad news -0.002 (0.009)

Outcome range [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] Control outcome mean 0.19 0.20 0.20 0.19 0.20 0.20 Control outcome std. dev. 0.07 0.07 0.07 0.07 0.07 0.07 R2 0.66 0.66 0.65 0.66 0.66 0.66 Treatment mean 0.59 0.59 0.59 0.59 0.59 0.59 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 0.49 Interaction mean -0.08 3.24 0.21 0.89 0.04 / 0.52 Interaction std. dev. 0.89 0.37 0.17 1.08 0.20 / 0.50 Observations 675 651 651 675 651 651

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in columns (2), (3), (5) and (6) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

54 the block with the most negative incumbent prior (1.8) increases the effect of providing information on the incumbent party’s vote share from 1.3 to 4.7 percentage points, and the effect on the incumbent’s share of registered votes from 0.3 to 3.0 percentage points. Weakly supporting hypothesis H3, column (3) in panel A identifies a smaller effect of the information in precincts where the average control respondents has stronger prior beliefs, although this is not statistically significant. Second, the extent to which the incumbent party is rewarded declines with the level of malfeasance revealed to voters. In line with H4, the significant negative interaction in column (4) between the treatment and the share of malfeasant spending shows that voters substantially reward incumbents for better performance, but may sanction incumbent parties in the most severe cases. As illustrated in Figure 2.8, revealing any level of malfeasant spending below 40% actually significantly increases the incumbent’s vote share. We show below that this performance slope becomes starker if the municipality of Ecatepec de More- los is dropped. Third, and combining the preceding heterogeneous effects, the increase in incumbent party vote share caused by providing incumbent performance information decreases with the extent to which the information induces negative updating. Consistent with hypotheses H5 and H6, this relationship is evident from columns (5) and (6) of both panels. Column (5) demonstrates a significant negative interaction between the treatment and our measure of negative updating in the block’s control group. Comparing the effect of good and bad news to information consistent with voter priors, column (6) similarly documents a substantially larger increase in incumbent party vote share when the information represents good news. However, given that it is difficult to interpret the level of updating in the control group (see above) and our model predicts that even voters which do not update may increase their support for the incumbent because they are risk-averse, it is hard to interpret why receiving neutral information increases incumbent vote share. Accordingly, we focus on the 55 i h pedxsossmlrrslsfor results similar shows Appendix the in A.14 Table specifications. alternative several to malfeasance. incumbent of responses Larreguy, measures similar across of qualitatively findings imply together the estimates with these Consistent (2015), Snyder voters. and registered Marshall of share a as incumbent share the vote for party results similar show (4)-(6) column Unsurprisingly, direction subsamples. same of across the in extent point the consistently and coefficients the spending differ, malfeasant somewhat updating of negative share with malfeasant of interactions types the both Although across treatment 2.5 spending. our Table of of effect B average and positive A clear panels a show of again (1)-(3) Columns results. similar broadly find we spending, (5). emphasizing column by in negative slope more the relatively was information the where places of comparison iue28 agnlefc fifrainteteto nubn oesae ysaeof share by share, vote incumbent on treatment information of effect Marginal 2.8: Figure dmntae htteicmetpryvt hr eut r robust are results share vote party incumbent the that demonstrates 2.6 Table Finally, unauthorized and poor the benefiting not spending concerning information Comparing

Marginal effect of information treatment -.02 0 .02 .04 .06 0 afaatsedn 9%cndneintervals) conﬁdence (95% spending malfeasant Incumbent voteshare(turnout) Share malfeasancespending .2 .4 .6 56

Marginal effect of information treatment -.01 0 .01 .02 .03 .04 0 Incumbent voteshare(registeredvoters) Share malfeasancespending .2 .4 .6 Table 2.5: Effect of information treatment on incumbent party vote share, by type of malfeasance

Incumbent party vote share (turnout) Incumbent party vote share (registered) (1) (2) (3) (4) (5) (6) Panel A: Share of spending not spent on the poor Information treatment 0.026*** 0.048*** 0.029 0.015** 0.030*** 0.028* (0.009) (0.011) (0.019) (0.006) (0.008) (0.015) × Incumbent malfeasant -0.110*** -0.073*** spending (0.028) (0.017) × Negative incumbent -0.005 -0.010 updating (0.009) (0.007)

Outcome range [0.09,0.85] [0.09,0.85] [0.09,0.85] [0.03,0.47] [0.03,0.47] [0.03,0.47] Control outcome mean 0.40 0.40 0.41 0.20 0.20 0.20 Control outcome std. dev. 0.12 0.12 0.11 0.07 0.07 0.06 Treatment mean 0.60 0.60 0.59 0.60 0.60 0.59 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 0.49 Interaction mean 0.22 1.58 0.22 1.58 Interaction std. dev. 0.18 0.84 0.18 0.84 R2 0.56 0.56 0.52 0.60 0.61 0.59 Observations 407 407 383 407 407 383 Panel B: Share unauthorized spending Information treatment 0.034*** 0.037*** 0.030*** 0.017*** 0.017** 0.015*** (0.007) (0.013) (0.006) (0.003) (0.007) (0.003) × Incumbent malfeasant -0.015 0.001 spending (0.031) (0.018) × Negative incumbent -0.042*** -0.020** updating (0.014) (0.009)

Outcome range [0.07,0.71] [0.07,0.71] [0.07,0.71] [0.03,0.44] [0.03,0.44] [0.03,0.44] Control outcome mean 0.35 0.35 0.35 0.19 0.19 0.19 Control outcome std. dev. 0.12 0.12 0.12 0.07 0.07 0.07 Treatment mean 0.58 0.58 0.58 0.58 0.58 0.58 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 0.49 Interaction mean 0.21 -0.09 0.21 -0.09 Interaction std. dev. 0.15 0.41 0.15 0.41 R2 0.72 0.72 0.72 0.75 0.75 0.75 Observations 268 268 268 268 268 268

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in columns (3) and (6) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

57 incumbent party vote share using registered voters in the denominator. First, we drop the municipality of Ecatepec de Morelos, where voters positively updated about the incumbent party when presented with 45% unauthorized spending. Suggesting that Ecatepec de Morelos is an anomaly, panel A further strengthens estimates of the conditional effects underpinning our theory. Moreover, Table A.2 similarly shows that our survey estimates in Table 2.3 are strengthened by dropping Ecatepec de Morelos. Second, to address the potential concern that our results are driven by imbalances remaining after random assignment, panel B simultaneously controls for all 46 covariates we assess balance over. The results are substantively indistinguishable and more precisely estimated. Third, despite up-weighting large precincts where proportionately few voters actually received the treatment, panel C shows that the results are robust to equally weighting all precincts. In this speciﬁcation, the interaction with the strength of voter prior beliefs becomes statistically signiﬁcant, and thus bolsters support for H3.

2.5.2 Non-linear effects of information on turnout

A distinctive feature of our theory is the non-linear relationship between the extent of malfeasance and turnout. In particular, we predicted that extremely low or high malfeasance levels would induce voters to switch parties. The results in Table 2.7 strongly support this non-linear prediction. Providing clear evidence consistent with H7, column (3) shows that for very low levels of malfeasance the lower-order treatment term indicates that turnout signiﬁcantly increases by 1.2 percentage points, while the negative linear and positive quadratic interactions with the share of malfeasant spending demonstrate that turnout decreases for interim levels of malfeasant but increases for high levels of malfeasance. Figure 2.9 depicts this non-linearity graphically. Consistent with such heterogeneity in response, column (1) shows that, on average, providing information does not impact turnout.

58 Table 2.6: Robustness of information treatment on incumbent party vote share (share of turnout)

Incumbent party vote share (share of turnout) (1) (2) (3) (4) (5) (6) Panel A: Removing Ecatepec de Morelos’ 64 precincts Information treatment 0.030*** 0.029*** 0.072 0.048*** 0.037*** 0.019** (0.007) (0.007) (0.075) (0.008) (0.009) (0.009) × Incumbent malfeasance prior 0.011 (0.007) × Strength incumbent prior -0.014 (0.023) × Incumbent malfeasant spending -0.097*** (0.029) × Negative incumbent updating -0.010* (0.005) × Good news 0.125*** (0.022) × Bad news 0.007 (0.015) Panel B: Controlling for covariates Information treatment 0.022*** 0.021*** 0.076 0.039*** 0.031*** 0.019** (0.006) (0.006) (0.072) (0.008) (0.008) (0.008) × Incumbent malfeasance prior 0.011** (0.005) × Strength incumbent prior -0.015 (0.022) × Incumbent malfeasant spending -0.077*** (0.023) × Negative incumbent updating -0.012*** (0.004) × Good news 0.098* (0.054) × Bad news -0.004 (0.010) Panel C: Unweighted precinct estimates Information treatment 0.020*** 0.019*** 0.113** 0.031*** 0.025*** 0.018*** (0.004) (0.004) (0.043) (0.006) (0.005) (0.005) × Incumbent malfeasance prior 0.007 (0.004) × Strength incumbent prior -0.029** (0.013) × Incumbent malfeasant spending -0.052** (0.023) × Negative incumbent updating -0.007** (0.004) × Good news 0.072*** (0.023) × Bad news -0.006 (0.008)

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated (except those in panel C), and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in columns (2), (3), (5) and (6) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

59 Table 2.7: Effect of information treatment on turnout and conﬁdence in the electoral process

Panel A: Turnout Turnout rate (1) (2) (3) Information treatment -0.001 -0.000 0.012** (0.004) (0.006) (0.006) × Incumbent malfeasance spending -0.005 -0.181*** (0.019) (0.058) × Incumbent malfeasance 0.345*** spending squared (0.109)

Outcome range [0.21,0.79] [0.21,0.79] [0.21,0.79] Control outcome mean 0.50 0.50 0.50 Control outcome std. dev. 0.10 0.10 0.10 Treatment mean 0.59 0.59 0.59 Treatment std. dev. 0.49 0.49 0.49 Interaction mean 0.21 0.21 Interaction std. dev. 0.17 0.17 R2 0.71 0.71 0.71 Observations 675 675 675 Panel B: Conﬁdence in system Elections help to select competent candidates (did not help at all - helped a lot) (1) (2) (3) (4) (5) (6) Information treatment 0.008 0.002 0.155 0.052 -0.066 -0.093 (0.042) (0.042) (0.493) (0.078) (0.059) (0.070) × Incumbent malfeasance prior -0.054 (0.043) × Strength incumbent prior -0.045 (0.153) × Incumbent malfeasant spending -0.209 (0.255) × Negative incumbent updating 0.081** (0.035) × Good news 0.309*** (0.102) × Bad news 0.164* (0.083)

Outcome range {1,2,3,4,5} {1,2,3,4,5} {1,2,3,4,5} {1,2,3,4,5} {1,2,3,4,5} {1,2,3,4,5} Control outcome mean 2.86 2.86 2.84 2.86 2.86 2.86 Control outcome std. dev. 1.40 1.40 1.41 1.40 1.40 1.40 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.24 0.21 0.89 0.04 / 0.53 Interaction std. dev. 0.89 0.37 0.17 1.07 0.20 / 0.50 R2 0.06 0.06 0.06 0.06 0.06 0.06 Observations 4,615 4,615 4,615 4,615 4,615 4,615

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in columns (2), (3), (5) and (6) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

60 rae ofiec nterl feetos hsmyrflc u nigta oesuethis use voters that candidates. finding between our adjudicate reflect to may information This elections. of developed role malfeasance the high in discovering confidence voters greater anything if that candidates. show good (6) select and can (5) elections Columns that of faith levels voter high alter that significantly finds not again do (4) column malfeasance B, panel in malfeasant se- politicians to competent of help and elections share honest that lect the trust voter in of decreasing scale five-point significantly a examining not (2) Furthermore, Column is spending. turnout ). 2014 Tsai that and shows Posner A Lieberman, panel ; 2015 of al. disengagement et Chong of form (e.g. general system a the with induces information malfeasance that suggest to dence iue29 agnlefc fifrainteteto unu,b hr fmalfeasant of share by turnout, on treatment information of effect Marginal 2.9: Figure vtrpirepcain,w n oevi- no find we expectations, prior voter emphasizing model our supporting Beyond

Marginal effect of information treatment -.02 0 .02 .04 .06 0 pnig(5 ofiec intervals) confidence (95% spending Share malfeasancespending .2 61 .4 .6 2.6 How information influences voting behavior

We now examine the extent to which belief updating’s effects on voting behavior may vary. While we ﬁnd that neither comparative performance information nor public dissemination meaningfully alter the effect of information provision, we ﬁnd evidence highlighting a more complex equilibrium where both opposition and especially incumbent parties respond to relatively severe malfeasance revelations.

2.6.1 Weaker effects of comparative performance information

Theoretically, we expected that providing voters with a benchmark—especially one that contrasts with the incumbent party’s performance, like we provide (see Figure 2.5)—might elicit stronger responses to our information treatment. In particular, if comparative performance plays a key role in helping voters to differentiate party competence, we expecteda particularly strong response to treatments revealing high levels of incumbent malfeasance. Before testing this hypothesis, we first verify that voters absorbed such information. Based on our post-treatment survey, column (1) of Table 2.8 shows that voters receiving both the local and comparative treatments were more likely to recall receiving information about other parties in their state. The effect in precincts receiving just information about their own municipal government indicate fuzzy recall. Nevertheless, voters receiving the comparative information treatment were 2 percentage points more likely to recall receiving comparative information. Although relative small, this differential is statistically significant, as demonstrated at the foot of column (1). Rather than accentuate the effect of providing local incumbent performance information, Table 2.9 reports weaker effects for comparative performance information. The positive coefficients in column (1) are statistically indistinguishable, and thus—in contrast with

62 Table 2.8: Effect of variants of information treatment on self-reported treatment engagement

Remember Remember Share of opposition loud community benchmark speaker received (1) (2) (3) Local information treatment 0.046*** (0.010) Comparative information treatment 0.066*** (0.010) Private information treatment 0.006 0.483*** (0.007) (0.062) Public information treatment 0.057*** 0.566*** (0.008) (0.064)

Outcome range {0,1} {0,1} {1,2,3,4,5} Control outcome mean 0.03 0.03 1.45 Control outcome std. dev. 0.18 0.16 1.01 Local/private treatment mean 0.39 0.39 0.39 Local/private treatment std. dev. 0.49 0.49 0.49 Comparative/public treatment mean 0.38 0.38 0.38 Comparative/public treatment std. dev. 0.49 0.49 0.49 Test: same treatment effect (p value) 0.03 0.00 0.06 R2 0.05 0.05 0.09 Observations 4,958 4,958 4,929

63 hypothesis H8—suggest that both types of information equally impact voter behavior on average. The fact that comparative information if anything has a larger effect is surprising given that the vast majority of treatments informed voters that the challenger was performing better. Furthermore, offering little support for H9, the interactive estimates suggest that voters are less likely to respond to comparative information in a Bayesian fashion: column (2) shows that voters punish incumbents less for high rates of malfeasance when presented with comparative performance information, column (3) shows similar results when we instead examine the difference in malfeasance spending between the incumbent and the average challenger elsewhere in the state, and columns (4) and (5) show little difference in response to the comparative treatment by voter updating. Unreported estimates for turnout similarly suggest that local and comparative information impacted turnout similarly. The lack of additional effect associated with providing comparative information could reﬂect several possibilities. First, voters may not believe that the performance of parties in other municipalities represent a good proxy for how such parties would perform in their municipality. Second, this null ﬁnding is also consistent with voters struggling to com- prehend comparative information, as some of our enumerators reported when they were conducting the survey. Consistent with both such explanations, unreported survey estimates show no differential updating about the incumbent party’s competence across local and comparative treatments. Third, as a comparison of Figures 2.6 and 2.10 illustrates, voters in the control group already believed the main local challenger—the party that placed second in the last election—to be less malfeasant. Consequently, the information we provided broadly coincided with voter prior beliefs, and may thus have induced limited voter updating. Unfortunately, we cannot distinguish between these potential explanations.

64 Table 2.9: Effect of local and comparative information treatments on incumbent party vote share

Incumbent party vote share (1) (2) (3) (4) (5) Panel A: Incumbent party vote share (share of turnout) Local information treatment 0.026*** 0.048*** 0.049*** 0.037*** 0.020** (0.008) (0.012) (0.012) (0.008) (0.009) Local × Incumbent malfeasant spending -0.104** -0.107** (0.048) (0.047) Comparative × Incumbent malfeasant spending -0.053 (0.037) Comparative × Difference in malfeasance spending -0.057 (0.037) Local × Negative incumbent updating -0.014* (0.007) Comparative × Negative incumbent updating -0.004 (0.006) Local × Good news 0.087* (0.047) Local × Bad news 0.003 (0.020) Comparative × Good news 0.079** (0.037) Comparative × Bad news 0.008 (0.015) Panel B: Incumbent party vote share (share of registered voters) Local information treatment 0.015*** 0.028*** 0.028*** 0.023*** 0.014** (0.005) (0.009) (0.009) (0.005) (0.006) Local × Incumbent malfeasant spending -0.059** -0.062** (0.028) (0.029) Comparative × Incumbent malfeasant spending -0.040* (0.021) Comparative × Difference in malfeasance spending -0.043** (0.020) Local × Negative incumbent updating -0.009*** (0.003) Comparative × Negative incumbent updating -0.004 (0.003) Local × Good news 0.033 (0.022) Local × Bad news -0.000 (0.014) Comparative × Good news 0.045** (0.021) Comparative × Bad news -0.003 (0.009)

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. For summary statistics, see Table 2.4. The smaller sample in columns (4) and (5) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality- treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

65 so mle nraei nubn oesaeascae ihpublic with associated share vote incumbent in increase smaller and a A show panels 2.10 of Table (1) in Column B responses. behavioral weaker somewhat but similar broadly fraction large a that leaflets. believe the (3) to received likely column community more their of of also foot were the precincts at such in tests voters the that Moreover, indicate speaker. loud a were leaflets by the accompanied that recall delivered correctly to likely more public points to percentage subject 6 precincts were in dissemination voters However, speaker. than loud likely a more recall no to were respondents control treatment information private the 2.8 receiving (2) Table voters Column that responses. expected confirms effective. the more elicited treatment is dissemination transmission public the public that that shows possible is it information, mance dissemination information public of impact additional Limited 2.6.2 iue21:Prevdcruto n o pnigo h oro eodpae at in party second-placed of poor the on spending not and corruption Perceived 2.10: Figure ept nraigcmo wrns forlaes ulcdseiainproduced dissemination public leaflets, our of awareness common increasing Despite perfor- comparative by magnified not are effects treatment’s information our While

Frequency 0 100 200 300 400 h lcinaogvtr ncnrlprecincts control in voters among election the Very low Perceived incumbentmalfeasance Low 66 Medium High Very high dissemination. While this could reflect a less sanguine response to generally high levels of malfeasance, columns (4) and (5) also document similar or smaller slopes with respect to the level of malfeasance reported and belief updating. In sum, we find little evidence supporting hypothesis H10. The limited impact of adding a loud speaker contrasts with the large effects of media found in similar contexts (Ferraz and Finan 2008; Larreguy, Marshall and Snyder 2015; Marshall 2016a). One potential explanation for this limited voter response is the greater capacity of broadcast media to foster either explicit or tacit coordination through common knowledge (Adena et al. 2015; Yanagizawa-Drott 2014). Consistent with the relative weakness of loud speakers as an explicit social coordination device, columns (1)-(3) in Ta- ble 2.11 find no significant difference in discussion of the leaflet, vote coordination on the basis of the leaflet or changes in voting behavior on the basis of discussions of the leaflet between the private and public forms of dissemination. However, a second dimension of difference from existing studies examining the role of the media is the direction of updating: Larreguy, Marshall and Snyder(2015) find that local media access induces punishment of considerably lower levels of malfeasance than we document here. Given that television coverage of audit reports is not uncommon, it is possible that voter priors have already been shaped by news reports, although followers of the news were no less likely to respond to our treatment and possess similar posterior beliefs. More plausibly, television reports simply represent more effective signals, due to their relatively high credibility, broad reach in the population and capacity to hold voters’ attention.

2.6.3 Party responses to malfeasance revelations

While our analysis has focused on how information impacts voter beliefs, political parties may also respond to the provision of malfeasance information. By altering campaign

67 Table 2.10: Effect of private and public information treatments on incumbent party vote share

Incumbent party vote share (1) (2) (3) (4) (5) (6) Panel A: Incumbent party vote share (share of turnout) Private information treatment 0.040*** 0.037*** -0.019 0.067*** 0.044*** 0.029*** (0.009) (0.009) (0.069) (0.014) (0.010) (0.009) Public information treatment 0.018** 0.017* 0.175** 0.024* 0.024** 0.012 (0.008) (0.008) (0.085) (0.014) (0.011) (0.008) Private × Incumbent malfeasance prior 0.011 (0.008) Public × Incumbent malfeasance prior 0.009 (0.008) Private × Strength incumbent prior 0.017 (0.023) Public × Strength incumbent prior -0.049* (0.026) Private × Incumbent malfeasant spending -0.129*** (0.037) Public × Incumbent malfeasant spending -0.030 (0.045) Private × Negative incumbent updating -0.009 (0.006) Public × Negative incumbent updating -0.009 (0.007) Private × Good news 0.056* (0.032) Private × Bad news 0.011 (0.021) Public × Good news 0.119** (0.050) Public × Bad news -0.001 (0.014) Panel B: Incumbent party vote share (share of registered voters) Private information treatment 0.024*** 0.022*** -0.035 0.039*** 0.028*** 0.018*** (0.007) (0.007) (0.039) (0.010) (0.007) (0.005) Public information treatment 0.008* 0.008 0.048 0.013* 0.012** 0.009** (0.005) (0.005) (0.044) (0.008) (0.006) (0.004) Private × Incumbent malfeasance prior 0.010** (0.005) Public × Incumbent malfeasance prior 0.006 (0.004) Private × Strength incumbent prior 0.017 (0.013) Public × Strength incumbent prior -0.013 (0.013) Private × Incumbent malfeasant spending -0.075*** (0.025) Public × Incumbent malfeasant spending -0.025 (0.022) Private × Negative incumbent updating -0.007** (0.003) Public × Negative incumbent updating -0.005 (0.003) Private × Good news 0.032* (0.016) Private × Bad news 0.003 (0.015) Public × Good news 0.045** (0.021) Public × Bad news -0.007 (0.008) Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. For summary statistics, see Table 2.4. The smaller sample in columns (2), (3), (5) and (6) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

68 Table 2.11: Effect of variants of information treatment on social transmission

Social Discussion Discussion discussion created vote of leaﬂet of leaﬂet coordination changed vote (1) (2) (3) Private information treatment 0.111*** 0.022*** 0.028*** (0.015) (0.008) (0.007) Public information treatment 0.125*** 0.030*** 0.030*** (0.014) (0.008) (0.008)

Outcome range {0,1} {0,1} {0,1} Control outcome mean 0.05 0.02 0.02 Control outcome std. dev. 0.23 0.13 0.12 Test: same treatment effect (p value) 0.22 0.26 0.82 R2 0.08 0.07 0.06 Observations 4,958 4,958 4,958

Immediate and campaign reactions

Qualitatively, our intervention incited push back from incumbent parties and challenger attempts to exploit the information in various instances. While these reactions are also incorporated in the preceding point estimates, the overall effect of providing information represents the primary policy parameter of interest. In many cases we encountered a strong response from the incumbent authorities, as Borde Político went on to explain publicly.30

30See their press release here.

69 In fact, the response was sufficiently strong that the Permanent Commission of the Mex- ican Congress, a group of Deputies and Senators taking care of urgent matters during the electoral recess, passed a resolution exhorting federal authorities to investigate municipal abuses of power. Most notably, in 4 municipalities across the states of Guanajuato and Méx- ico members of our team were arrested after local incumbent party operatives complained that they disrupted public order and (falsely) accused our team of illegally distributing political advertisements. All team members were quickly released. Moreover, in both Ixta- paluca and Juventino Rosas, the PRI and PAN respectively arrived shortly after the leaflets were delivered to confiscate them. However, incumbent responses were sometimes more subtle. In Apaseo el Alto in Guanajuato, people from the ruling MC attempted to justify their behavior to voters. Most creatively, the PRI incumbent in Cuatiltlán Izcalli in the state of México forged our leaflet and proceeded to disseminate a leaflet blaming local PAN politicians for government debt and accusing them of supporting local drug cartels. The opposition in some cases also responded by magnifying the impact of the information. For example, in Tamasopo in San Luis Potosí, an opposition party called a meeting to discuss the leaflet. To examine this more systematically, we asked voters whether incumbents and challengers referenced the information we provided in leaflets, campaign activities, visits from local political actors, adverts or through the media. Around 17% of voters reported an incumbent response and 16% reported a challenger response. According to our respondents, incumbents most frequently claimed that all parties are equally bad, while opposition parties were somewhat more likely to try to emphasize the content of the leaflets. Examining the total number of such activities reported by each individual, column (1) of panels A and B in Table 2.12 identifies no insignificant change in incumbent responses and a slight but significant increase in challenger responses on average across treatment and control

70 precincts.31 However, these modest increases in political activity mask the sharp increase where malfeasance revelations were most severe. The large and significant positive interactions in column (2) demonstrate that, for both incumbents and challengers, party activity increased substantially in municipalities where high malfeasance was revealed. In a treated precinct within a municipality with 50% malfeasant spending, activity almost doubled relative to a municipality with 0%. A comparison of columns (2) and (3) demonstrates that political responses are driven by the level of malfeasance reported, rather than the extent to which voters negatively update from this information. This suggests that incumbent parties may over-estimate the extent to which voters expect their representatives to engage in minimal malfeasant spending in office. Decomposing the activities index, columns (4)-(8) show that increased party activity in high-malfeasance treated precincts often occurs through the media. The incumbent, in particular, is also significantly more likely to engage in on-the- ground campaign and visiting tactics. Unreported estimates indicate that parties did not differentially respond where the treatment was delivered publicly.

Increased vote and turnout buying

In addition to increased broadcast and campaign activity, it is also possible that an incumbent facing relatively severe malfeasance revelations may resort to increased vote or turnout buying. To address this question, we examine the effect of a precinct containing an additional polling station.32 Since polling station-level electoral returns are publicly available, and voters are effectively randomly distributed across polling stations by surname

31The non-zero number of activities in the control group likely reﬂects cross-precinct spillovers or recall failures. 32We also conducted a list experiment, but failed to detect any evidence of vote buying. Given that we actually observed vote buying on the ground on election day, and vote buying is generally common in our sample, the list experiment most likely failed to perform as expected.

71 Table 2.12: Effect of information treatment on political party responses

Total party activities Leaﬂets Campaigning Party visits Adverts Media (1) (2) (3) (4) (5) (6) (7) (8) Panel A: Incumbent reactions Information treatment 0.032 -0.131* -0.012 -0.026 -0.020 -0.026 -0.013 -0.015 (0.043) (0.077) (0.065) (0.018) (0.013) (0.017) (0.017) (0.015) × Incumbent malfeasant spending 0.766*** 0.085 0.096** 0.156** 0.076 0.100** (0.258) (0.059) (0.044) (0.068) (0.053) (0.049) × Negative incumbent updating 0.048 (0.035)

Outcome range {0,1,2,3,4,5} {0,1,2,3,4,5} {0,1,2,3,4,5} {0,1} {0,1} {0,1} {0,1} {0,1} Control outcome mean 0.43 0.43 0.43 0.06 0.07 0.06 0.06 0.07 Control outcome std. dev. 1.18 1.18 1.18 0.24 0.25 0.23 0.24 0.26 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean 0.21 0.89 0.21 0.21 0.21 0.21 0.21 Interaction std. dev. 0.17 1.05 0.17 0.17 0.17 0.17 0.17 2

72 R 0.12 0.12 0.12 0.15 0.15 0.14 0.17 0.16 Observations 4,958 4,958 4,958 4,958 4,958 4,958 4,958 4,958 Panel B: Challenger reactions Information treatment 0.102** -0.024 0.085 -0.007 0.005 0.003 -0.011 -0.010 (0.039) (0.060) (0.055) (0.011) (0.012) (0.012) (0.011) (0.013) × Incumbent malfeasant spending 0.591*** 0.060* 0.025 0.079 0.078* 0.086** (0.204) (0.035) (0.051) (0.052) (0.039) (0.042) × Negative incumbent updating 0.019 (0.032)

Outcome range {0,1,2,3,4,5} {0,1,2,3,4,5} {0,1,2,3,4,5} {0,1} {0,1} {0,1} {0,1} {0,1} Control outcome mean 0.40 0.40 0.40 0.07 0.06 0.04 0.06 0.06 Control outcome std. dev. 1.17 1.17 1.17 0.26 0.25 0.20 0.23 0.24 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean 0.21 0.89 0.21 0.21 0.21 0.21 0.21 Interaction std. dev. 0.17 1.05 0.17 0.17 0.17 0.17 0.17 R2 0.12 0.12 0.12 0.15 0.15 0.14 0.17 0.16 Observations 4,958 4,958 4,958 4,958 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. (Cantú 2014), each additional polling station provides a clearer signal of broker effort, and thus facilitates turnout buying by increasing the party’s ability to extract turnout buying effort through its enhanced monitoring capacity (Larreguy, Marshall and Querubín 2016). To examine whether our treatment increased brokerage in treated precincts, we follow Larreguy, Marshall and Querubín(2016) by exploiting discontinuity created by Mexico’s electoral rule providing an additional polling station for every multiple of 750 voters registered in a given precinct.33 The difference-in-difference regression discontinuity estimates in Table 2.13 reinforce the graphical evidence. Column (1) first reports a local linear specification restricting the sample to precincts within 100 voters of the discontinuity and including linear trends in the running variable—the number of voters from receiving an additional polling station—either side of the discontinuity. The significant positive coefficient on the triple interaction between the information treatment, an additional polling station and the share of malfeasant spending indicative of a large turnout buying effect. Specifically, a standard deviation increase in malfeasant spending in a treated precinct with an additional polling station increases incumbent party vote share by around 11 percentage points. Conversely, in a treated precinct without an additional polling station, a standard deviation increase in the share of malfeasant spending reduces the incumbent party’s vote share by 8 percentage points. This stark difference suggests that incumbent parties strongly target the precincts where they have greatest capacity to undo potential detrimental effects of voters learning of high malfeasance. As columns (2) and (3) show, even in the small samples closest to the discontinuity, we find robust evidence that an additional polling station increases incumbent party vote share where voters learn of relatively high levels of incumbent malfeasance.

33Larreguy, Marshall and Querubín(2016) provide strong evidence to suggest that there is no sorting around this discontinuity in Mexico, and also rule out various alternative potential explanations for changes in voting behavior arising from an additional polling station within a precinct.

73 Table 2.13: Heterogeneous effect of an additional polling station on the effect of negative information on incumbent party vote share and turnout

Incumbent party vote share Turnout rate (share of registered voters) (1) (2) (3) (4) (5) (6) Information treatment 0.092* 0.101** 0.118* -0.043 0.038 0.063 (0.054) (0.041) (0.063) (0.054) (0.059) (0.158) Additional polling station 0.047 0.086 0.081 -0.023 0.039 -0.042 (0.065) (0.061) (0.082) (0.083) (0.084) (0.205) Information treatment × Additional polling station -0.145 -0.152** -0.195*** -0.009 -0.065 -0.071 (0.087) (0.070) (0.063) (0.107) (0.097) (0.158) Information treatment × Incumbent malfeasant spending -0.425** -0.851*** -0.675 0.513** 0.665 0.272 (0.201) (0.183) (0.426) (0.225) (0.485) (1.069) Additional polling station × Incumbent malfeasant spending -0.266 -0.663*** -0.622* 0.354 0.427 0.673 (0.233) (0.194) (0.315) (0.302) (0.276) (0.790) Information treatment × Additional polling station 0.635** 1.074*** 0.980** -0.375 -0.638 -0.251 74 × Incumbent malfeasant spending (0.285) (0.237) (0.426) (0.329) (0.515) (1.069)

Bandwidth (number of voter from additional polling station) 100 50 25 100 50 25 Linear trends XX Outcome range [0.04,0.36] [0.06,0.36] [0.06,0.36] [0.24,0.70] [0.38,0.68] [0.42,0.68] Control outcome mean 0.19 0.20 0.22 0.50 0.52 0.54 Control outcome std. dev. 0.07 0.08 0.08 0.10 0.08 0.09 Treatment mean 0.58 0.63 0.62 0.58 0.63 0.62 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 0.49 Additional polling station mean 0.58 0.56 0.55 0.58 0.56 0.55 Additional polling station std. dev. 0.50 0.50 0.50 0.50 0.50 0.50 Interaction mean (second) 0.19 0.19 0.18 0.19 0.19 0.18 Interaction std. dev. (second) 0.18 0.18 0.17 0.18 0.18 0.17 R2 0.80 0.92 0.94 0.83 0.88 0.90 Observations 173 80 47 173 80 47

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. In columns (1) and (4), interactions between all variables and the running variable are omitted. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. In contrast, columns (4)-(6) report no differential impact on overall turnout of an additional polling station in precincts informed of high malfeasance. This suggests that effective mobilization of support is concentrated among incumbents. Given the increase in the incumbent party vote, the negative triple interaction—which is almost statistically significant—in fact suggests a decline in the total non-incumbent vote. Although it is hard to verify, this could reflect a substitution of opposition mobilization efforts toward precincts not subject to more effective turnout buying technologies, or the incumbent’s mobilization drive inducing opposition supporters to switch toward the incumbent. The preceding evidence of increased party activity, especially the increased incumbent vote where an additional polling station exists, indicates that party responses play a key role in understanding how information dissemination impacts incumbent party support. The differential turnout buying response by the incumbent may thus in part explain why the extent of sanctioning in especially malfeasant municipalities is limited. Moreover, given that the relatively low-scale of our intervention allows incumbent parties to specifically target the precincts receiving negative information, these incumbent responses offer another potential explanation as to why broadcast media—which reach sufficiently many voters that incumbent parties cannot easily target all voters—may more effectively support the electoral sanctioning of mayoral malfeasance than leaflet dissemination.

2.7 Conclusion

Our results demonstrate the importance of understanding voter priors for understanding how incumbent performance information will affect voting behavior. We ﬁnd that Mexican voters—who, like voters in many developing contexts, are pessimistic that their incumbents correctly allocate resources—on average actually reward municipal incumbent parties revealed to have engaged in non-trivial levels of malfeasance in ofﬁce. However, consistent

75 with voters possessing negative priors, we find considerable support for our simple learning model. In particular, rewards are concentrated among voters with especially negative prior beliefs, that learn of relatively low incumbent malfeasance, and that update most positively about the incumbent. Moreover, turnout responds to information non-linearly, with surprising information increasing turnout by shifting voters between parties, and relatively unsurprising information shifting voters toward indifference. By emphasizing voter prior beliefs, these findings explain the mixed evidence that information induces electoral sanctioning or impacts political participation in developing democracies (e.g. Banerjee et al. 2011, 2014; Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013; Ferraz and Finan 2008; Humphreys and Weinstein 2012; Larreguy, Marshall and Snyder 2015). However, the implications for electoral accountability are nevertheless troubling. The fact that voters are sufficiently pessimistic that 40% of funds being misallocated represents good news is worrying for good governance proponents. This suggests the need for greater voter expectations of their elected representatives, which could induce politicians to perform better in office (Barro 1973; Ferejohn 1986), the need for better politicians to stand for office, or more effective audits and legal sanctions. With respect to altering voter performance expectations, civic education or a critical media may be required to enable voters to understand what good performance entails (e.g. Botero et al. 2015). Encour- aging high-quality candidates to stand for office is more challenging, and some evidence suggests that increased wages can help (Ferraz and Finan 2011). Although it is difficult to discern selection from monitoring effects, the extensive use of audits may reduce incumbent malfeasance (Bobonis, Fuertes and Schwabe 2014; Olken 2007; Zamboni and Litschig 2014).

76 3| Signaling sophistication: How social expectations can increase political information ac- quisition1

3.1 Introduction

Considerable evidence now shows that access to reliable political information plays a key role in enabling voters to hold governments to account for their performance in ofﬁce (see Ashworth 2012; Pande 2011), helping voters to vote in line with their distributive policy interests (Bartels 2008; Iversen and Soskice 2015), and coordinating collective action (Kuran 1991; Tarrow 1994). However, voters are often poorly informed about politics (e.g. Delli Carpini and Keeter 1996; Pande 2011), which can create a deﬁcit of informed political participation. Particularly in non-consolidated democracies, where democratic institutions and expectations are weakly ingrained, low levels of voter information may harm political representation where it is already weakest. Despite the foundational role of voter information in theories of political behavior, surprisingly little is known about what motivates voters to inform themselves about politics.

1I thank Jim Alt, Eric Arias, Donal Cahill, Francisco Cantú, Diego Domínguez, Jorge Domínguez, Mauri- cio Fernández Duque, Raissa Fabregas, Hernán Flom, Jeff Frieden, Saad Gulzar, Sebastian Garrido de Sierra, Torben Iversen, Joy Langston, Horacio Larreguy, Soeren Henn, Chappell Lawson, Rakeen Mabud, Pablo Querubín, James Robinson, Jennifer Sheehy-Skeffington, Jim Snyder, Edoardo Teso, Julie Weaver, and workshop or conference participants at presentations at Harvard, MIT, NEWEPS and MPSA for illuminating discussions and useful comments. This research received financial support from the Eric M. Mindich Research Fund for the Foundations of Human Behavior and the Institute for Quantitative Social Science, and approved by the Harvard Committee on the Use of Human Subjects (IRB15-1068). 77 Since Downs(1957) seminally argued that “rationally ignorant” voters face strong incentives to leave costly information acquisition to others, given the low probability that their informed choice will affect political outcomes, researchers have struggled to explain why individual voters would choose to become politically informed.2 Experimental evidence examining the effect of providing voters with newspapers indicates that low levels of political knowledge are not simply a supply constraint (Gerber, Karlan and Bergan 2009). Prominent demand-side explanations suggest that voters consume political information because they find such news entertaining (Hamilton 2004), have a strong sense of civic duty (Blais 2000; Feddersen and Sandroni 2006a), require such information for their job (Larci- nese 2005), or are subject to spillovers within their social or vocational networks (Iversen and Soskice 2013; Prior 2007).3 However, beyond the lack of compelling causal evidence, such accounts struggle to explain variation in demand for news across time or context because they reflect exogenous preferences or incidental externalities. In this paper, I develop and test the theory that voters strategically acquire political information to cultivate a desirable reputation among their peers as politically sophisticated.4 In the social signaling model that I propose, voters differ in their underlying political sophistication. Such sophistication is unobserved by their peers, but greater sophistication allows voters to acquire information relatively cheaply. Based on the quantity of information that an individual acquires, social groups collectively form beliefs about an individual’s

2Although there is evidence that reducing the costs of acquiring information can increase information acquisition (e.g. Larcinese 2007; Prior 2007; Snyder and Strömberg 2010), these studies abstract from the beneﬁts of acquiring information—the focus of this paper. 3The evidence that turnout incentives increase political knowledge is mixed (Loewen, Milner and Hicks 2008; Shineman 2013). 4Similar signaling arguments have been proposed for charitable giving (e.g. Bénabou and Tirole 2006; Glazer and Konrad 1996). Abrams, Iversen and Soskice(2011) and Iversen and Soskice(2015) similarly propose that social approval motivates electoral turnout and information acquisition. However, they neither test this argument nor consider heterogeneity within or across groups. Aldashev(2010) develops a decision- theoretic model that posits that the beneﬁts of social network size and exchange increase with information acquisition, but does not endogenize these production functions.

78 political sophistication. The resulting semi-separating equilibria highlight two forces facing voters possessing different levels of underlying political sophistication. An increase in the probability that members of a voter’s social group observe their information acquisition accentuates: (1) the desire to meet a minimum standard that separates less sophisticated voters from the least sophisticated that continue to face prohibitively high costs of acquiring information, and (2) a differentiation motivation where increased information acquisition among less sophisticated voters causes more sophisticated voters to also acquire more information in order to differentiate themselves from the newly-informed less sophisticated voters. Both effects are especially pronounced among voters in social groups where other members highly value political knowledge. I test the model’s predictions in Mexico, using both experimental and observational empirical designs to identify the effects of voters’ political information acquisition being revealed to their peers. As in many non-consolidated democracies, Mexican voters are relatively uninformed about politics (e.g. Castañeda Sabido 2011; Chong et al. 2015; Lawson 2004; Marshall 2016a). However, recent studies have shown that politically- relevant information can play a key role in supporting electoral accountability (Larreguy, Marshall and Snyder 2015; Larreguy, Marshall and Trucco 2015; Marshall 2016a) and reducing support for locally dominant parties (Larreguy, Marshall and Snyder 2016). Con- sequently, especially as Mexican NGOs seek to increase transparency in the wake of political scandals—including the disappearance of 43 students in the municipality of Iguala in 2014—understanding the demand side of voter information acquisition is particularly pertinent. Using a field experiment conducted at an elite Mexican university, I first show that the prospect of friends observing an individual’s political sophistication can increase political information acquisition. Students were offered the opportunity to participate in a panel study consisting of a baseline survey three weeks before the 2015 national legislative elec- 79 tions and a post-election survey that students were informed in advance would include a quiz about the election campaign and results. To vary social incentives to acquire information, treated students were informed at the end of the first survey that their quiz results would be emailed to three of their friends.5 On average, treated students did not exhibit significantly higher post-election quiz scores. However, this masks important heterogeneous effects. Supporting the social signaling model’s prediction that less-sophisticated students acquire information to reach a minimal standard when they expect their social group to receive a clear signal of their political sophistication, I find that students with low initial levels of political knowledge substantially increased their information acquisition in preparation for the quiz. Furthermore, this change was concentrated among students whose friends were more interested in politics. In line with the model, the lack of differentiation effects among relatively politically sophisticated students may reflect the high marginal cost of acquiring additional information to differentiate oneself from some of Mexico’s most informed voters. Using a nationally representative survey, the observational design tests the general- izability of the model and explores the model’s implication that voters face differential incentives to acquire information before elections. Specifically, I exploit staggered state election cycles using a difference-in-differences design to identify the effects of upcoming local elections on political information acquisition. The likelihood that peers observe a voter’s political knowledge increases as elections approach, because politics becomes a more salient discussion topic (e.g. Eifert, Miguel and Posner 2010; Huckfeldt and Sprague 1995; Walsh 2004). Consistent with this greater probability of type revelation, I find that individuals nested in politically-engaged social networks acquire significantly more topical political information in the run-up to elections. Placebo tests demonstrate that political

5The email addresses of these friends were elicited earlier in the survey. Participants were informed that the receipt of prizes would be conditional on these friends verifying their friendship.

80 knowledge relating to information which is unlikely to be covered in the news is unaffected, while elections do not differentially increase information acquisition in civically-oriented social networks. Further robustness checks indicate that the results are not driven by voters facing more competitive elections, with a greater sense of civic duty, or living in municipalities with greater access to media sorting into more politically-oriented groups. Combined, these experimental and observational empirical results indicate that social group dynamics play a key role in explaining information acquisition across the Mexican electorate. This occurs both among unsophisticated voters aiming to reach a minimum standard and sophisticated voters seeking to differentiate themselves. However, these effects operate principally among voters nested in politically-oriented networks, and thus suggest that a lack of politically-engaged friends can generate political information traps among voters in networks with limited interest in politics. Such political information traps are not only likely to be detrimental to political representation in general(Casey 2015), but especially so among disadvantaged groups which typically possess lower levels of political engagement. In developed contexts, Iversen and Soskice(2013, 2015) provide evidence suggesting that social incentives induce high and low information equilibria that impact electoral outcomes. They find that the decline of union membership correlates with lower political engagement and more centrist voting among low-income voters across 20 advanced democracies. The study makes several theoretical and empirical contributions. Theoretically, I propose a signaling model that endogenizes the acquisition of political information and highlights how responses to social incentives to acquire political information vary with voter sophistication. By focusing on social approval motives, my strategic argument stands in contrast with the vast majority of the information acquisition literature emphasizing incidental spillovers or exogenously-given consumption preferences as explanations for why voters become politically informed. The signaling dimension differentiates the model from 81 Abrams, Iversen and Soskice(2011) who highlight the importance of political interest within groups but not types of individual, and do not ultimately examine the effects of social approval on information acquisition. Similarly, my strategic account differs from decision- theoretic models positing that voters acquire information to increase political discussion within their social group (Aldashev 2010; Iversen and Soskice 2013, 2015) or expand their network (Aldashev 2010). Perhaps most importantly, the study provides the first evidence of which I am aware that political information acquisition is caused by social interactions. Previous studies, which have focused heavily on consolidated democracies, are vulnerable to the concern that informed voters select into certain types of network. In this respect, the findings complement and extend the recent literature using field experiments to demonstrate that social pressure increases the likelihood that an individual turns out to vote (DellaVigna et al. 2014; Gerber et al. 2011; Gerber, Green and Larimer 2008; Nickerson 2008) or participates in a protest (McClendon 2014). In contrast with these studies, I demonstrate that differences in social context play a key role in our understanding of when different types of voters will acquire political information in a major non-consolidated democracy, and thus when voters are likely to effectively hold politicians to account (e.g. Manin, Przeworski and Stokes 1999). The results also provide quantitative evidence that reinforces the insights of studies in the United States observing the importance of political discussion within social groups (e.g. Huckfeldt and Sprague 1995; Walsh 2004). Furthermore, as noted above, understanding when voters acquire political information may be particularly important in consolidating democracies with limited institutional protections. The results thus chime with Brollo(2009) and Marshall(2016 a), who respectively show in Brazil and Mexico that voters primarily punish governments for corruption and high homicide rates publicized in the news just before elections. These articles rely on voters consuming more news before elections, but are unable to identify whether such an 82 increase in consumption reflects increased demand for, or increased supply of, politically- relevant information. The remainder of this paper is structured as follows. Section 3.2 presents the model, demonstrating that an increase in the likelihood that the group observes a voter’s information acquisition induces voters—especially those embedded in more politically-engaged groups—to themselves acquire more political information. To test the model, section 3.3 presents the experimental findings before section 3.4 reinforces these results exploiting variation in upcoming elections. Section 3.5 considers the broader implications of the findings.

3.2 Voter information acquisition within social groups

In this section I model voter acquisition of political information. I propose that— independent of changes in the supply of information—voters strategically acquire political information in order to signal their status as politically sophisticated within their social group. Bénabou and Tirole(2006) and Glazer and Konrad(1996) propose similar arguments for charitable giving. The model predicts that an increase in the likelihood an individual’s information acquisition is observed by members of their social group pushes the least sophisticated voters to acquire news for the ﬁrst time to meet a minimum standard. This in turn induces more sophisticated voters to acquire more news to distinguish themselves from less sophisticated voters consuming for the ﬁrst time.

3.2.1 Model

Voters are nested within social groups. For simplicity, I focus on a generic group with a political interest level of w > 0. The group contains a continuum of voter types θ ∈ [θ,θ] ⊆ (0,∞) deﬁning different levels of latent political sophistication, and distributed 83 according to cumulative distribution function F(θ). Although homophily is likely to drive similar individuals to form social groups, group members still differ with respect to political characteristics (e.g. Sinclair 2012).

To acquire political information, voters decide to consume n ∈ [0,∞) news programs.

However, consuming n programs costs c(n,θ), where c(0,θ) = 0, c(n,θ) is convex in n,

cθ < 0, and cnθ < 0 (where subscripts denote partial derivatives). This cost could reflect opportunity costs, the price of accessing news content, or the cognitive cost of comprehension. The final assumption reflects the “single-crossing” condition stating that information

acquisition is cheaper for voters with higher levels of political sophistication θ. A strategy for a voter of type θ is thus n : [θ,θ] 7→ [0,∞). With probability p ∈ (0,1], all members of the group observe the level of information n that each voter acquires through social interaction, e.g. discussion of politics. This probability can also be interpreted broadly as the precision or frequency of the signal.

However, an individual’s type θ is only known to themselves. Individual voters can thus signal their political sophistication to their peers through the amount of political news that they acquire. In this model, voters signal their type because acquiring costly political information allows them to distinguish themselves within their social group by developing a coveted reputation as politically knowledgeable. For example, groups may collectively prize political knowledge because members intrinsically value political discussion (Huckfeldt and Sprague 1995) or because groups serve to aggregate information for their members (e.g. Huckfeldt, Johnson and Sprague 2004; McKelvey and Ordeshook 1985). Alternatively, individuals may enjoy competing to demonstrate their political sophistication, or conversely suffer from exhibiting their ignorance (e.g. Baumeister and Leary 1995).6 For such a repu-

6It is possible that demonstrating sophistication could induce rejection from their peer group in some setting (Austen-Smith and Fryer 2005), although this phenomena is likely to be rare in the population, and

84 tation, the group collectively bestows rewards wβ, where β : (0,∞) 7→ [θ,θ] is the group’s belief about an individual’s type (based on observing n, but without observing θ). When with probability (1 − p) information acquisition is not revealed, reputational beneﬁts of

E[θ] are assigned to all types. The reputational beneﬁts accruing to type θ are thus given by:

Z θ pwβ (n(θ)) + (1 − p)w θdF(θ), (3.1) | {z } θ Utility when information | {z } acquisition is revealed Utility when information acquisition is not revealed

where, as noted above, w captures the collective importance attached to political sophistication by the group. The game’s timing can be summarized as:

1. Voters learn their type θ, which is private information.

2. Voters choose to acquire information from n news programs.

3. With probability p, the social group observes n. If this occurs, the group then col-

lectively forms posterior beliefs β over voter types, and assigns reputational beneﬁts accordingly.

3.2.2 Equilibrium and comparative statics

I search for perfect Bayesian equilibria. I ﬁrst identify the unique fully separating equilibrium, where the different levels of information acquired by different types reveal all types in equilibrium.7 Given the single crossing condition, Mailath(1987) shows that such an equilibrium exists and that optimal news consumption, n∗(θ), is a strictly monotonic

especially among the elite set of students I study experimentally. 7Pooling equilibria may also exist, although for analytical interest I focus only on separating equilibria. 85 (and thus continuously differentiable) function of voter type.8 In other words, each level of voter sophistication is associated with a unique level of information acquisition. Incentive

compatibility then requires that each type θ ∈ [θ,θ] chooses to consume n ∈ n∗(θ) news programs to maximize their utility given the group’s correctly anticipated posterior belief

β = (n∗)−1(n) about their type when the equilibrium strategy n∗(θ) is revealed:

( Z θ ) n∗(θ) = argmax pw(n∗)−1(n) + (1 − p)w θdF(θ) − c(n,θ) . (3.2) n∈n∗(θ)≥0 θ

This incentive compatibility condition requires that, for any level of political sophistication, a voter has no incentive to deviate from the equilibrium strategy n∗(θ) to mimic another type. This allows their social group to correctly adduce each voter’s underlying sophistication when the level of news they consume is revealed. In the fully separating equilibrium, the ﬁrst-order condition associated with equation (3.2) binds for all types and is given by:9

dn∗(θ) pw = ∗ . (3.3) dθ cn(n (θ),θ)

To derive equilibrium information acquisition by a generic type θ, I exploit the initial value for type θ, n∗(θ; p,w), and integrate over θ to yield:

Z θ ∗ ∗ pw n (θ; p,w) = n (θ; p,w) + ∗ dθ. (3.4) θ cn(n (θ; p,w),θ)

8The initial value condition is discussed below, and is satisﬁed in a fully separating equilibrium. 9 d f −1 x 1 n n∗ n∗ −1 n To see this, note that dx ( ) = f 0( f −1(x)) , and = (θ) implies θ = ( ) ( ).

86 The lowest type, θ, maximizes equation (3.2) by choosing n∗(θ; p,w) = 0.10 In the fully separating equilibrium, all other types acquire n∗(θ; p,w) > 0 to differentiate themselves from the least politically sophisticated voter. However, if pw is not large enough to overcome the costs of acquiring information,

a semi-separating equilibrium occurs where all types θ < θ˜(p,w) pool at n = 0. There may thus exist some low types that never choose to acquire political information. Types

θ ≥ θ˜(p,w) continue to separate as the reputational beneﬁts of signaling their political understanding are sufﬁciently high, where type θ˜(p,w) chooses n∗(θ˜(p,w), p,w) > 0 once they weakly prefer this outcome to pooling with the lowest types θ < θ˜(p,w). In other words, only types above a certain sophistication cutoff have an incentive to distinguish themselves. The preceding analysis is summarized in the following proposition:

Proposition 1. (Equilibrium characterization) A unique separating or semi-separating

perfect Bayesian equilibrium hn∗(θ; p,w),β (n∗(θ; p,w))i exists, and is characterized by the following strategies and beliefs:

 n∗ ˜ p w p w R θ pw d if > ˜ p w >  (θ( , ), , ) + θ˜(p,w) c (n∗(θ;p,w),θ) θ θ θ( , ) θ, n∗(θ; p,w) = n (3.5)  0 if θ ≤ θ ≤ θ˜(p,w).   ∗ θ if n (θ; p,w) > n(θ˜(p,w), p,w), β (n∗(θ; p,w)) = (3.6) ˜(p w)  1 R θ , θdF(θ) if n∗(θ; p,w) ≤ n(θ˜(p,w), p,w).  F(θ˜(p,w)) θ

All proofs are provided in the Appendix. The following proposition identiﬁes several key comparative static predictions, in terms

10This ensures that the individual rationality constraint is satisﬁed. If personal incentives to acquire information were included in the model, it is possible that n∗(θ; p,w) > 0. These likely exist, and explain why people do not literally possess zero political knowledge, but are excluded to emphasize the model’s core social insights. 87 of both information acquisition and voting.

Proposition 2. (Comparative statics) In the perfect Bayesian equilibrium described in Proposition1, the following comparative statics hold:

1. The proportion of voters that acquire any information, 1−F(θ˜(p,w)), is increasing in the probability that information acquisition is observed by members of the social groups (p), the political interest of the group (w), and (if F00 is not too large) their interaction.

2. Information acquisition among those that acquire information is also increasing in the probability that information acquisition is observed by members of the social groups (p), the political interest of the group (w), and (if F00 is not too large) their interaction.

The ﬁrst part of Proposition5 establishes that individual voters start to acquire news when other members of their social group are more likely to observe their signal of sophistication, when the social group collectively rewards political sophistication, and especially when both factors are salient.11 This reﬂects an increase in the value of developing a reputation as politically knowledgeable, which induces even relatively unsophisticated types to differentiate themselves from the lowest types by acquiring some news. While increases in p and w induce low types to start acquiring information, the second part of proposition5 shows that more sophisticated voters seek to separate themselves. By increasing the returns to developing a desirable reputation, higher values of p and w increase incentives to mimic higher types, and thus force more sophisticated voters to further differentiate themselves until the point where no lower type is willing to pay the cost of

11The condition that F00 is not too large is not unreasonable, given the political sophistication distributed is likely to be positively skewed (like the education and income distribution in almost every country in the world). 88 mimicking them. In groups where many voters already acquire considerable amounts of news, the differentiation required to separate may be small because the convexity of the

∗ costs ensures that the marginal cost of acquiring additional news—cn(n (θ; p,w),θ) in the denominator of the second term in equation (D.1)—is high. Consequently, the model predicts greater differentiation effects among voters nested in groups attaching lower reputational beneﬁts to political sophistication.

3.2.3 Testable implications

Proposition5 suggests various testable implications. First, a key prediction is that political news consumption will increase with the likelihood that an individual’s information acquisition is revealed to members of their social group. In contrast, non-strategic explanations for information acquisition based on consumption preferences and incidental spillovers do not suggest that voters should respond to such stimuli. Averaging across all types of voter, the model predicts that:

H1. (Knowledge revelation effect) A greater probability that an individual’s political knowledge is revealed to their social group induces voters to acquire more news about politics.

Empirically, I examine two such situations where political knowledge is revealed: most directly, when the results of a political knowledge quiz are sent to three close university friends; and during an electoral campaign when politics becomes a more salient discussion topic. The social signaling model also suggests two motivations for news acquisition that differentially affect different types of voter. First, to obtain reputational beneﬁts by meeting a minimum standard, relatively politically unsophisticated voters (i.e. θ < θ˜(p,w)) within a social group acquire news for the ﬁrst when faced with an increased likelihood that their 89 political knowledge is revealed to members of their social group:

H2. (Meeting a minimum standard effect) A greater probability that an individual’s political knowledge is revealed to their social group induces relatively unsophisticated voters within a given social group to acquire political news for the ﬁrst time.

It is important to emphasize that voters in many social settings may not literally acquire news for the first time. This represents an analytical simplification of the model. In practice, voters increase their acquisition from relatively low levels within their group if they also consume news for other reasons such as consumption preferences or through social or vocational externalities. Nevertheless, I expect to observe a significant jump in acquisition. Empirically, I will operationalize voter sophistication in terms of prior levels of political knowledge.

Second, relatively sophisticated voters (i.e. θ ≥ θ˜(p,w)) consume just enough additional news to continue to differentiate themselves from less sophisticated voters:

H3. (Differentation effect) A greater probability that an individual’s political knowledge is revealed to their social group induces relatively sophisticated voters within the group to acquire relatively more political news than before.

This “ratchet” effect, whereby all types (except those that never acquire information) slightly increase their news consumption, is an important implication of the signaling model. In- tuitively, this reflects the fact that increasing the benefits of reputation induces unsophisticated voters to acquire information for the first time (because the reputational gains now exceed the costs), which in turn requires more sophisticated voters to acquire more information in order to continue differentiating themselves from lower types. Consequently, I expect sophisticated voters to disproportionately increase acquisition at high levels, while unsophisticated voters may jump from acquiring essentially no news to consuming a small

90 quantity. For a given reputational benefit, lower marginal costs of acquiring information imply larger differentiation effects in social groups with lower levels of information acquisition. Finally, a defining feature of the social signaling model is the importance of differences across social groups. In particular, the effects of revealing political knowledge are expected to be greater among voters embedded in social groups characterized by higher average levels of interest in politics where political knowledge is rewarded with greater reputational benefits:

H4. (Differential network effect) The effect of an individual’s political knowledge being revealed to their social group is greater among among individuals nested in more politically-oriented social networks.

Conversely, I expect that a reputation for political knowledge will be less valuable in civic organizations such as neighborhood, voluntary and sporting groups, where political knowledge is less pertinent.

3.3 Experimental evidence

I ﬁrst test the implications of the social signaling model using a ﬁeld experiment conducted around the 2015 Mexican elections. The experiment is designed to identify the effect on political information acquisition of voters anticipating that their level of political knowledge, as measured by a post-election political quiz, will be revealed to their friends.

3.3.1 Design

I recruited students from an elite university in Mexico City just before the 2015 elections held on June 7th. The student population, of around 5,000 undergraduate students, 91 includes many of Mexico’s most politically engaged students. All undergraduate students, all of whom are of legal voting age, were offered the opportunity to participate in the study via a mass email sent by university administrators.12 The experiment consisted of a baseline and endline survey. Crucially, before beginning the baseline survey, which could be reached by clicking through from the recruitment email, students were informed of the study’s structure. A clear understanding of expectations is an essential element of the experimental design.13 The survey preamble explained that participants would undertake one survey immediately and a second survey containing questions about the 2015 elections just after the election. Furthermore, participants were informed that they would be required to list the email address of three friends also attending the university, and that the results of their quiz may be sent to their friends (but nobody else, including the university).14 The number of friends was chosen to ensure a non-negligible social cost without encumbering students to find many friends’ email addresses. Although the prospect of facing a political knowledge quiz may reduce participation, such a quiz is not subject to the social desirability biases commonly associated with subjective measures of information acquisition. To incentivize participation, all students that completed the baseline survey would enter a prize draw to win one of three Best Buy gift cards with sufficient value (MXN$6,600) to purchase an iPad Air 16GB, while students that also completed the second survey would enter a second (and independent) draw to win one of five such gift cards. Ultimately, 754 students

12Students were informed that study was a collaboration between their home institution and a major U.S. university. 13In contrast to some experiments, there was little value in either deceiving participants or obfuscating the purpose of the experiment. 14When listing friends, participants were told that the receipt of any prize would be contingent upon ver- iﬁcation that the owner of the email address was indeed a fellow student at the university and knew the participant in question.

92 completed the baseline survey.15 The baseline survey was conducted around three weeks before the election. In addition to the email addresses of three friends, the baseline survey elicited each participant’s demographic details and subject of study, interest in politics, friends’ interest in politics, behavior in political discussions, organizational participation, political news consumption, knowledge of recent events and Mexican political institutions, and political partisanship. The average participant respectively consumed 5.6, 2.5, 2.5 and 2.4 hours of political news a week through the internet, newspapers, radio and television. By consuming almost two hours of news a day, the students in the sample are among the most politically informed in the country. Moreover, 69% of participants answered all three topical political knowledge questions correctly.16 While politics is clearly a salient issue among such students, these initially high levels also imply high marginal costs of consuming additional political news. At the end of the survey, students were randomly assigned into one of two conditions. Control participants received the following (translated) message explaining that their performance on the post-election quiz—the quartile of the distribution in which they fell— would not be sent to their three friends:17

Once you have completed the quiz in the second survey, your performance on the quiz will NOT be sent to the three friends that you listed at the beginning of this survey.

Conversely, treated students received the following almost-identical message explaining that their friends would be informed of their performance on the quiz:

15Several students were excluded because they entered invalid email addresses that prevented me from recontacting them. 16These multiple choice questions asked: (1) which party was ﬁned MXN$180 million for violating the electoral law in April 2015 (the Green party); (2) how long federal deputies serve in ofﬁce (three years); and (3) from which party was José Luis Abarca Velázquez, the mayor of Iguala (PRD). 17In Spanish, performance was expressed by desempeño. 93 Once you have completed the quiz in the second survey, your performance on the quiz will be sent to the three friends that you listed at the beginning of this survey.

The treatment condition was thus designed to generate social incentives to acquire information by creating the expectation that friends would learn about treated respondents’ level of political sophistication. To mitigate potential spillovers, participants were asked not to discuss the study with other students. Prior to the election, participants received two emails reminding them about the upcoming election quiz and reiterating their treatment status. The post-treatment survey indicates that 64% of students correctly identified the number of people that would be informed of their quiz results. Linking participants across surveys by their email addresses, the second survey was sent to students on June 9th—two days after the election. The key component of the survey was the election quiz, which contained ten multiple-choice questions each offering four possible answers. The questions varied in difficulty, covering national and Mexico City- specific political events that had occurred during the campaign since the closing date of the baseline survey up until the election day results themselves.18 To ensure that students did not look up the answers online, they were given 20 seconds to answer each question. For the election to remain fresh in the mind of respondents, my final sample includes only students that completed the second survey within a week of the election.19 I use the political quiz score as my main measure of political information acquisition. The average student answered 5.6 questions correctly.20 In addition to the quiz, participants were again asked

18The Appendix details the exact questions. 19In total, 659 students completed the second survey. However, respondents taking the quiz more than a week after the election scored substantially less well on the political quiz and were substantially less likely to correctly identify their treatment condition. The inclusion of late respondents does not substantively alter the results, but reduces estimate precision and induces significant differences in attrition across treatment and control groups. 20Consistent with the variation in difficulty, only 28% answered the most difficult question correctly, while 94 about their interest in politics, friends’ interest in politics, and political news consumption, but also new questions probing the mechanisms underpinning their behavior. To validate the experimental design, I examine attrition rates and balance across treatment conditions. Of the 754 initial participants (of which 374 were treated and 380 were not), 260 treated and 280 control students completed the post-treatment survey within a week of the election. There was no statistically significant difference in attrition across treatment conditions (p = 0.21), or its interaction with potential proxies—baseline political knowledge, hours of news consumption or political interest—for how well a student may expect to perform on the quiz.21 Furthermore, Table B.1 in the Appendix confirms that differences across 49 pre-treatment characteristics are consistent with chance.22 The lack of differential attrition or imbalance indicate that the integrity of the randomization was preserved across the panel surveys. To identify the effects of social incentives to acquire political news, I estimate the following regression using OLS:

Yi = βSocial treatmenti + Xiγ + εi, (3.7)

where Yi is student i’s score out of ten on the post-election quiz in the main analysis. The vector of covariates, Xi, contains an indicator for male students, year of birth ﬁxed effects, baseline interest in politics, and an indicator for membership of a party organization.23

87% answered the easiest question correctly. The other questions were relatively uniformly distributed between these limits. Respondents ranged from answering all incorrectly to all correctly. 21Half of the baseline respondents were encouraged to complete the second survey by being offered entry to a separate draw for an additional gift card. This measure, intended to guard differential attrition by providing an additional source of exogenous variation, significantly increased participation by 5.6 percentage points, and may have helped to prevent significant differences in attrition rates. 22Table B.2 shows initial balance across respondents that completed the baseline survey. 23Year of birth fixed effects absorb the constant term.

95 These controls were included to address slight imbalances on variables that could affect information acquisition, and entailed dropping one respondent with a missing value for interest in politics. My ﬁnal sample thus contains 539 respondents.

3.3.2 Results

Table 3.1 presents estimates of the average treatment effect of the social incentive on the election quiz score, and thus tests hypothesis H1. Although the estimate in column (1) is positive, as predicted by the model, it is small in magnitude and not statistically significant. In particular, the social treatment only increases the likelihood that an individual answers a given question correctly by 0.3 percentage points. This effect slightly increases to 0.5 percentage points in column (2) when the social treatment is used to instrument for an indicator for respondents that believed that they had been treated. Nevertheless, among politically engaged students at an elite university, the results clearly indicate that social incentives have limited effects on information acquisition on average. However, I now demonstrate that these null ﬁndings mask important heterogeneity consistent with the model’s predictions. Table 3.2 tests hypotheses H2 and H3—that politically unsophisticated students acquire information to meet a minimum standard, while sophisticated students ratchet up their information acquisition to differentiate themselves—by interacting the treatment with measures of politically sophistication. Principally, I deﬁne political sophistication using an indicator for the 69% of participants that correctly answered all three political knowledge questions on the baseline survey correctly. Politically sophisticated students consume more than six hours of internet news—their primary source of news—a week, 2.5 hours more than non-sophisticated students. The results in column (1) present a very different picture. Consistent with H2, the

96 Table 3.1: Effect of the social treatment on election quiz scores

Political quiz score (1) (2) Social treatment 0.030 (0.169) Believe received treatment 0.052 (0.287)

Observations 539 539 Outcome mean 5.65 5.65 Outcome range 0 to 10 0 to 10 Outcome standard deviation 2.07 2.07 Treatment mean 0.48 0.36 First stage F statistic 281.7

Notes: All specifications control for an indicator for being male, year of birth fixed effects, an individual’s level of political interest, and an indicator for membership of a political party organization. The specification in column (1) is estimated using OLS. The specification in column (2) uses the social treatment to instrument for an indicator for respondents that believed they had been treated, and is estimated using 2SLS. Robust standard errors are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

97 Table 3.2: Effect of the social treatment on election quiz scores, by student political sophistication

Political Political Political quiz score quiz score quiz score (1) (2) (3) Social treatment 0.639** 0.981** 0.424 (0.289) (0.494) (0.264) Sophisticated 1.326*** (0.248) Social treatment × Sophisticated -0.934*** (0.351) Follow national news 1.600*** (0.322) Social treatment × Follow national news -1.080** (0.525) Hours of internet news a week (baseline) 0.080*** (0.027) Social treatment × Hours of internet news -0.072** a week (baseline) (0.036)

Observations 539 539 539 Outcome mean 5.65 5.65 5.65 Outcome range 0 to 10 0 to 10 0 to 10 Outcome standard deviation 2.07 2.07 2.07 Social treatment mean 0.48 0.48 0.48 Interaction mean 0.69 0.88 5.59 Test: Social treatment + Interaction = 0 (p value) 0.14 0.57

Notes: All speciﬁcations control for an indicator for being male, year of birth ﬁxed effects, an individual’s level of political interest, and an indicator for membership of a political party organization, and are estimated using OLS. Robust standard errors are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

98 significant positive coefficient on the social treatment demonstrates that the treatment increased the number of questions correctly answered by unsophisticated students by 0.64 questions, or almost one third of a standard deviation. In terms of the model, this result suggests that unsophisticated students substantially increase their political knowledge in order to signal to their peers that they are not the least sophisticated in their social group. Conversely, the significant negative coefficient on the interaction between the social treatment and politically sophisticated students, together with the test at the foot of the table, shows that the treatment did not significantly impact the information acquisition of sophisticated students. This finding provides no evidence that, as suggested by H3, politically sophisticated students acquire more knowledge to differentiate themselves from the less well-informed students that invested in increasing their knowledge. The inability to detect such differentiation could reflect the fact that the elite students in this sample, who are highly politically engaged, have already credibly demonstrated high levels of political sophistication, or as noted above face greater marginal costs of acquiring additional information from already high levels. The observational design using a nationally representative sample tests this relationship in the general population. Columns (2) and (3) present similar results using alternative measures of political sophistication based on pre-treatment news consumption. Column (2) uses an indicator for respondents that follow the national news—the primary element of the election quiz—and even more starkly than with the first sophistication indicator highlights that the treatment only produced a large significant effect on quiz scores among students not following the national news. Finally, showing that the treatment had greatest impact among students with relatively low levels of prior news consumption, column (3) identifies a significant negative interaction between the treatment and the number of hours of internet news consumed

99 a week.24

3.3.3 Mechanisms and alternative interpretation

To further support the signaling interpretation, I test various additional implications of the theory. First, if social signaling drives information acquisition, we should expect to find that unsophisticated students also learned more about politics in preparation for the quiz. Examining an indicator for students that claimed to have learned more about politics before the election, column (1) of Table 5.6 demonstrates that the social treatment increased the likelihood that politically unsophisticated students learned more by 13 percentage points. Again, politically sophisticated students were unaffected. Second, although the estimates are somewhat noisy, column (2) shows that the social treatment had a substantially larger effect among the 7% of students that openly stated in the baseline survey that they acquire political information in order to demonstrate this knowledge to their friends. Third, for students that incorrectly claim to be well informed relative to their friends, and may thus feel stronger pressure to perform, the social treatment may be especially powerful. I test this possibility by interacting the social treatment with both and individual’s political sophistication and a five-point variable measuring the extent to which they believe that they know more about politics than their friends. The significant positive interaction between the social treatment and this belief in column (3) shows that the effect of the social treatment on information acquisition is increasing in the (likely incorrect) belief that politically unsophisticated students know more about politics than their friends. A fourth, and particularly important, test examines how the effect of the treatment varies with the political interest of the three friends respondents listed in the baseline survey. Hypothesis H4 predicts that the social treatment should have its greatest impact among

24A negative but insigniﬁcant interaction holds for television, which is a less important source of news among participating students. 100 Table 3.3: Mechanisms underpinning the effect of the social treatment on election quiz scores

Learned Political Political Political more quiz score quiz score quiz score (1) (2) (3) (4) Social treatment 0.125* -0.030 -0.968 -0.050 (0.074) (0.172) (0.803) (0.552) Sophisticated 0.168*** 0.359 -0.587 (0.064) (0.793) (0.847) Social treatment × Sophisticated -0.134 0.259 1.385 (0.088) (1.089) (1.011) Demonstrate knowledge -0.200 (0.611) Social treatment × Demonstrate knowledge 0.826 (0.758) Know more than friends -0.217 (0.160) Social treatment × Know more than friends 0.467** (0.214) Sophisticated × Know more than friends 0.260 (0.207) Social treatment × Sophisticated -0.356 × Know more than friends (0.284) High interest friends -0.656 (0.456) Social treatment × High interest friends 0.943 (0.650) Sophisticated × High interest friends 2.127** (0.884) Social treatment × Sophisticated -2.705** × High interest friends (1.086)

Observations 533 539 528 529 Outcome mean 0.61 5.65 5.65 5.65 Outcome range {0,1} 0 to 10 0 to 10 0 to 10 Outcome standard deviation 0.49 2.07 2.06 2.07 Social treatment mean 0.48 0.48 0.48 0.48 Sophisticated mean 0.69 0.69 0.69 Other interaction mean 0.07 3.72 0.88 Test: Social treatment + Social treatment 0.01 × High interest friends = 0 (p value)

101 individual whose social groups collectively prize political knowledge. I test this using a (pre-treatment) indicator for respondents that list their friends’ interest in politics at 5 or greater on a scale from 0 to 10.25 Column (4) affirms the model’s expectation: while the social treatment has no significant effect on unsophisticated students with low-interest friends, the test at the foot of the tables demonstrates that combining the social treatment with high interest friends produces a significant positive effect. Combined with the negative triple interaction, this indicates that only sophisticated students in politically-oriented social groups were impacted by social incentives to acquire political knowledge. An alternative mechanism is that unsophisticated treated students did not in fact acquire more political news themselves, but rather were more likely to cheat on the quiz by asking their friends about the questions. Given that students are unlikely to report having “cheated,” I exploit a list experiment to differentiate these possibilities. After completing the quiz, all respondents were asked to list the total number of the following activities that they had engaged in during recent weeks: attended a campaign activity; watched the news on television; written an article about politics on the internet; and, in the case of a random subset of students, talked about the questions on the quiz with a friend. Among students that did not receive the additional option, the average student engaged in 1.5 of these activities. Comparing the number of items listed by students that did and did not receive the additional option (see Blair and Imai 2012), column (1) of Table 3.4 indicates that 36% of students consulted their friends about the quiz questions. However, interacting the additional option with the social treatment and the indicator for political sophistication, the results in column (2) provide no evidence that unsophisticated treated students were more likely to discuss the questions. While the substantial proportion of students discussing the questions is likely to downwardly bias the effects of the social treatment (by helping ev-

25Contrary to the possibility that the treatment affected this perception, column (11) of Table 3.4 also indicates that the social treatment did not affect respondents’ appraisals of their friends’ interest in politics.

102 Table 3.4: Alternative interpretations

Items Items Interest Acquire to Acquire to Acquire Estimated Political listed listed in politics choose best due to due to friend interest candidate interest duty score of friends (1) (2) (3) (4) (5) (6) (7) (8) List experiment treatment 0.360*** 0.444** (0.065) (0.180) Social treatment 0.035 0.074 -0.056 -0.071 -0.040 -0.332 0.016 (0.141) (0.174) (0.043) (0.069) (0.068) (0.316) (0.277) Sophisticated -0.094 0.370** -0.073** 0.094 -0.028 0.237 0.404* (0.126) (0.154) (0.036) (0.062) (0.057) (0.261) (0.227) Social treatment × Sophisticated -0.001 -0.329 0.071 0.069 0.037 0.417 0.063 (0.163) (0.206) (0.054) (0.083) (0.081) (0.373) (0.334)

103 Social treatment × List -0.162 experiment treatment (0.227) Sophisticated × List -0.073 experiment treatment (0.210) Social treatment × Sophisticated 0.117 × List experiment treatment (0.274) Observations 529 529 538 539 539 539 488 522 Outcome mean 1.67 1.67 7.64 0.89 0.55 0.74 5.72 6.78 Outcome range 0 to 4 0 to 4 0 to 10 {0,1}{0,1}{0,1} 0 to 10 0 to 10 Outcome standard deviation 0.80 0.80 2.04 0.31 0.50 0.44 1.93 1.92 List experiment treatment mean 0.48 0.48 Social treatment mean 0.48 0.48 0.48 0.48 0.48 0.47 0.48 Sophisticated mean 0.69 0.69 0.69 0.69 0.69 0.69 0.69

Notes: All specifications control for an indicator for being male, year of birth fixed effects, an individual’s level of political interest, and an indicator for membership of a political party organization, and are estimated using OLS. Sample size differences reflect minor missingness on non-common variables. Robust standard errors are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. eryone reach the correct answer), this demonstrates that the effect among unsophisticated students is not driven by such cheating. While the discussion of questions with friends is not the mechanism driving my findings, the relatively high levels of discussion indicate the perceived social importance of performing well on the quiz. Another potential interpretation of the results is that the treatment affected the broader role that political news plays in the lives of students. However, there is little evidence for such a change. First, column (3) indicates that treated respondents—whether politically sophisticated or unsophisticated—do not register greater interest in politics on a scale from 0 to 10. Second, columns (4)-(6) show that the treatment did not induce unsophisticated students to acquire information about politics at the beginning of the second survey (but before the quiz) for non-social reasons. In particular, the treatment did not increase the likelihood that respondents said they acquired political information to cast an informed ballot, due to interest in politics or to fulfill a civic duty. Third, the social treatment did not affect student appraisals of their friends’ political knowledge. Columns (7) and (8) respectively report no differences in the number of correct answers students expected that their friends would provide to the same quiz questions or the level of political interest of their friends. Together, these results further support the social signaling model. Rather than by causing students to cheat on the quiz or alter their view of acquiring political information, the evidence suggests that politically unsophisticated students were induced to perform better on the quiz through social expectations or individual characteristics that make them more susceptible to social pressures. To examine the generality of these findings, and to better examine the differentiation effect where the marginal cost of acquiring additional information is lower, I now test the social signaling argument in a more nationally representative sample.

104 3.4 Observational evidence

The experimental ﬁndings point to the importance of social incentives to acquire political information among less politically sophisticated students in social networks containing members with high levels of political interest. However, the results pertain to an unrepresentative subset of the Mexican population. To examine the implications of social signaling in the broader population, I now identify how upcoming local elections—that Marshall(2016 a) shows increase political discussion in Mexico, and are likely to increase the beneﬁts of cultivating a reputation for political sophistication—differentially affect the information acquisition of voters embedded in different types of social networks.

3.4.1 Data

To test the implications of the social signaling model, I use four waves of the government- run National Survey of Political Culture and Civil Practices (ENCUP) conducted in 2001 (November), 2003 (February), 2005 (December), and 2012 (August).26 Each wave draws stratiﬁed random samples of around 4,000 eligible Mexican voters for face-to-face interviews from within urban and rural strata deﬁned by the electoral register, and was designed by the government to be broadly nationally representative (see Marshall 2016a). The pooled sample includes 17,213 respondents from an unbalanced panel of 539 municipalities across all 31 states and the federal district of Mexico City. My analysis utilizes four key types of variable. First, I measure political news consumption by the frequency with which voters watch or listen to the news, programs about politics, or programs about public affairs.27 Relating to the model’s predictions, I focus

26The 2008 survey could not be used because it lacks both geographic identiﬁers and comparable questions. 27Like Marshall(2016 a), I focus on radio and television. Unlike the experimental study, these are by far the most prevalent sources of political information in Mexico at large. According to the 2010 Census, only

105 on two measures of consumption: an indicator for the 87% of respondents that report ever consuming political news, which captures meeting a minimum standard; and, to capture the differentiation effect, I computed a five-point consumption intensity scale ranging through never, at some point, at least monthly, at least weekly, and daily.28 News consumption was not elicited in the 2001 survey. Second, like the experimental study, political knowledge is the principal outcome variable. Political knowledge is defined by the first (standardized) factor from a set of indicators coding correct responses to simple factual questions. To focus on the acquisition of recent information, I use only topical political knowledge questions asking about contemporary political debates and movements or the party identity of the state governor (see Appendix for details).29 The former type of question may be important for identifying which party represents a voter’s interests, while the latter is a necessary condition for electoral accountability. The average respondent answered around half these basic questions correctly. As a placebo test, I examine knowledge of Mexican political institutions that are unlikely to be covered directly in the news. Third, the treatment variable is an upcoming local election. Based on the municipality in which the respondent lives, I code an indicator for respondents facing a municipal, and typically a simultaneous state legislative election, within the five months following the survey (see also Marshall 2016a). This period approximates the length of a typical election campaign.30 Mexican states have traditionally followed distinct electoral cycles, both in around a quarter of Mexican households have internet access at home. 28Table B.5 in the Appendix shows similar results when examining these intensities separately. 29In 2001, 2003, 2005 and 2012 respectively, there were 3, 2, 2 and 2 topical questions, and 4, 1, 1 and 2 institutional questions asked of respondents. 30Table 4.7 shows that the results are robust to using a continuous measure—months until the election— and Table B.6 in the Appendix shows that the results are similar when defining an upcoming local election as any number of months between 1 and 10 before the election.

106 terms of the month and year in which elections are held, although the months when election are held within a given year were homogenized following a major electoral reform in 2007 (Serra 2013). Unlike federal elections, a key advantage of utilizing state-level elections is the ability to isolate plausibly exogenous variation in the likelihood that peers learn about an individual’s political sophistication through the heightened salience of politics. Finally, although no suitable pre-treatment measure of voter political sophistication was available,31 I use a respondent’s participation in political groups to proxy for the level of political engagement in their social group. This measure is designed to test the hypothesis that the prospect of an individual’s political information acquisition being revealed to their peers has a bigger effect on pre-election news consumption in politically-oriented social networks (H4). Speciﬁcally, I created a summative rating scale containing three (standardized) variables: the number of politically-oriented organizations an individual is a member of;32 the number of such organizations at which the respondent attended a meeting in the last year;33 and a three-point scale capturing the extent to which the community discusses local problems. Since the ﬁnal component of the scale and cooperative organizations were not asked about in the 2012 survey, I multiply imputed responses over ten datasets using a variety of pre-treatment covariates (e.g. King et al. 2001). Supporting its conceptual coherence, the scale has a Cronbach’s alpha of 0.57.

3.4.2 Design

I exploit a difference-in-differences design to identify how the impact of upcoming local elections on political information acquisition differs across individuals from different

31Self-reported political interest is itself a function of upcoming local elections, and thus susceptible to post-treatment bias. 32In this category I include political, party, and cooperative organizations. 33I use the sum of the following two indicators: attended a meeting at a political or party organization, and attended a community or cooperative meeting. 107 social groups. In particular, I leverage state-speciﬁc election cycles to compare changes in information acquisition within states facing an upcoming local election to changes within states that are not, across voters in more or less politically engaged social networks. The ENCUP surveys do not track federal elections like the Mexico Panel Studies (e.g. Lawson et al. 2013), and each wave was conducted in a different month of the year. Given that at least one survey from virtually all states was conducted in the same year as an election, there is little reason to believe that the surveys were strategically timed. Consistent with this claim, Table B.3 in the Appendix shows that the occurrence of upcoming local elections is well-balanced across various individual, municipal and state level characteristics, while Table B.4 shows that neither upcoming local elections nor recent violence predict the inclusion of a municipality in a given survey wave. To assess the social signaling model’s implications, I examine how the effect of an upcoming local election varies with the political interest of a respondent’s social network by estimating the following interactive difference-in-differences speciﬁcation for respondents i in state s during survey year t:

Yist = β1Upcoming local electionst + β2Political network scaleist (3.8) +β3 Upcoming local electionst × Political network scaleist + ηs + µt + εist,

where Yist is a measure of political news consumption or political knowledge. Respectively,

ηs and µt are state and survey ﬁxed effects that absorb all time-invariant state factor and common period effects such as the availability of news or trends in political behavior. The effect of an upcoming local election, and how it varies with the political orientation of an individual’s network, is identiﬁed under the parallel trends assumption that without an upcoming local election trends in Yist would not have differed between states with and without upcoming local elections. Standard errors are clustered by state.

108 I focus on the Upcoming local electionst × Political network scaleist interaction term because the baseline difference-in-differences estimate of the effect of an upcoming local election is hard to interpret. In particular, non-social factors such as increased political news coverage or greater interest around elections in politics could also increase information acquisition before elections (Marshall 2016a). However, a positive interaction term would suggest that—consistent with H4—upcoming elections differentially induce voters to acquire political information in social contexts where the increase in the salience of politics before elections is particularly important for establishing reputations among peers. Such individual-level heterogeneity in response to upcoming local elections within a given state is unlikely to reﬂect changes in access to political information that impact all voters in the state. Nevertheless, I providing several robustness tests below suggesting that alternative mechanisms through which upcoming elections could affect news consumption and political knowledge are not driving the main results.

3.4.3 Results

Columns (1) and (2) of Table 3.5 estimate the differential effects of upcoming local elections on political news consumption, at different levels of intensity. The positive but insignificant (lower-order) coefficients for upcoming local elections indicate that a voter with the mean political network scale score generally consumes more news just before elections. However, the key test of hypothesis H4 relates to the interaction. Supporting H4, the significant positive interactions with my index of political-oriented social networks in both columns (1) and (2) indicate that the effect of increasing the salience of politics on active news consumption is particularly pronounced among respondents in such networks. For both voters consuming political news for the first time and voters ratcheting up their consumption before local elections, a standard deviation increase in politically-oriented

109 Table 3.5: Difference-in-differences estimates of the effect of upcoming local elections on information acquisition, by political network engagement

Watch and Watch and Topical Institutional listen to news listen to news political political ever scale knowledge knowledge (1) (2) (3) (4) Upcoming local election 0.009 0.174 0.189 -0.131 (0.025) (0.112) (0.133) (0.106) Political network scale 0.022** 0.122** 0.101* 0.080* (0.008) (0.039) (0.048) (0.036) Upcoming local election 0.033** 0.192*** 0.173** 0.034 × Political network scale (0.015) (0.055) (0.062) (0.041)

Observations 13,030 13,030 17,213 17,213 Outcome mean 0.87 2.58 0.00 0.00 Outcome std. dev. 0.34 1.47 1.00 1.00 Upcoming local election mean 0.19 0.19 0.16 0.16 Political network scale mean 0.00 0.00 0.00 0.00 Political network scale std. dev. 1.00 1.00 1.00 1.00 Survey year outcome was not asked 2001 2001

Notes: All specifications include survey fixed effects, and are estimated across ten multiply imputed datasets using OLS. Several underlying elements of the political network scale were imputed in 2012 (see Appendix for details). Standard errors clustered by state are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. social networks from its mean of zero more than doubles the average effect of upcoming local elections. Given the quantity of television news programming—the primary source of news in the population—has not increased before elections on television (Marshall 2016a), the increase in the share of voters ever consuming news suggests that the results do not simply reflect an increase in the supply of political news before elections. These large estimates again provide evidence that relatively unsophisticated voters in more politically-engaged social networks are consuming information about politics for the first time. The results also indicate that more sophisticated voters, with higher levels of initial political news consumption, are considerably more likely to consume political news around local elections when they are nested in politically-engaged networks. In addition to

110 reinforcing the experimental evidence that social incentives induce less sophisticated voters to acquire news (H2), these results for the consumption scale—which hold even for daily news consumption (see Table B.5 in the Appendix)—also support the differentiation effect proposed in H3. Columns (3) and (4) confirm that news consumption translates into topical political knowledge. Supporting H4, column (3) first shows that a standard deviation increase in the political network scale again doubles the mean effect of an upcoming local election on a voter’s topical political quiz score. Given the increases in individual news consumption in columns (1) and (2), an important component of increased political knowledge likely reflects an voter’s own consumption rather than spillovers from discussions within their social group. If voters are watching the news to learn about current affairs for conversations with their peers, they are far more likely to consume topical news as opposed to institutional knowledge of the political system that is less likely to be covered. Consistent with this argument, column (4) demonstrates that there is no significant interaction between local elections and politically-oriented social networks in the case of institutional knowledge questions. This placebo test also suggests that the results are not driven by politically engaged voters—who are also likely to score highly in terms of institutional knowledge— both consuming more news before elections and selecting into generally knowledgeable social groups.

3.4.4 Robustness checks

A key concern is that voter sorting, or the characteristics correlated with membership of politically engaged networks, can account for these ﬁndings. In particular, sorting could explain the results if voters with a taste for news about elections (Hamilton 2004), or a greater sense of civic duty to cast an informed ballot when elections arise (e.g. Blais

111 2000; Feddersen and Sandroni 2006a), are also more likely to enter certain social networks. Such concerns are most convincingly disproved using the type of experimental variation exploited above. Nevertheless, in addition to the institutional knowledge placebo test discussed above, I now present evidence contrary to the sorting concern. First, I conduct a second placebo exercise using civically-oriented, as opposed to politically- oriented, social networks. Consistent with a group’s political interest driving increased information acquisition around elections, panel A of Table 4.7 shows that the effects of local elections on political information acquisition do not vary with a civic network scale combining the number of civic organizations a respondent has been a member of and the number of meetings they attended in the last year.34 This null relationship indicates that it is not simply participation in any group that drives information acquisition around elections, or that the results reﬂect differential civic engagement between individuals in more or less politically-oriented networks. However, this test cannot ensure that voters interested in following elections do not sort into more political groups. Second, to address the concern that certain types of voters populate politically-oriented social groups, I simultaneously control for the interaction of indicators of political interest and civic duty with upcoming local elections. Given that political interest is itself a function of upcoming local elections, and would thus risk post-treatment bias, I proxy for interest in local elections using two pre-treatment measures of political competition: the municipal incumbent’s victory margin and the effective number of parties (ENPV) at the previous municipal election. Panels B and C of Table 4.7 show that the interaction for neither indicator of political interest differentially increase political news consumption or political knowledge just before municipal elections. Conversely, the signiﬁcant interactive effect of politically-oriented networks is highly robust to controlling for either potential confound.

34Civic organizations include pensioner, professional, labor, social, voluntary, religious, neighbor, cultural, sporting, parents and citizen organizations.

112 Table 3.6: Robustness of the differential difference-in-differences estimates

Watch and listen Watch and listen Topical political to news ever to news scale knowledge (1) (2) (3) Panel A: Civic network scale placebo test Upcoming local election 0.009 0.177* 0.211*** (0.018) (0.098) (0.077) Upcoming local election × Civic 0.005 0.024 -0.009 network scale (0.007) (0.033) (0.040) Observations 13,030 13,030 17,213

Panel B: Control for local election-incumbent victory margin interaction Upcoming local election × Political 0.033** 0.197*** 0.172** network scale (0.014) (0.054) (0.062) Upcoming local election × Municipal 0.087 0.694 -0.037 incumbent win margin (last election) (0.116) (0.525) (0.520) Observations 13,030 13,030 17,213

Panel C: Control for local election-ENPV interaction Upcoming local election × Political 0.033** 0.191*** 0.172** network scale (0.014) (0.055) (0.061) Upcoming local election × Municipal 0.007 -0.034 0.192 ENPV (last election) (0.023) (0.112) (0.127) Observations 13,030 13,030 17,213

Panel D: Control for local election-prior turnout interaction Upcoming local election × Political 0.031** 0.185*** 0.149** network scale (0.013) (0.054) (0.056) Upcoming local election × Voted 0.003 -0.000 0.066 for mayor since 2000 (0.025) (0.087) (0.084) Observations 13,030 13,030 13,030

Panel E: Control for local election-media interaction Upcoming local election × Political 0.033** 0.201*** 0.183** network scale (0.015) (0.058) (0.066) Upcoming local election × Media -0.000 0.004 0.001 stations within municipality (0.001) (0.004) (0.007) Observations 13,030 13,030 17,213

Panel F: Continuous measure of upcoming local election Months until local election -0.001* -0.012** -0.004 (0.001) (0.004) (0.005) Months until local election × Political -0.001* -0.004** -0.004* network scale (0.000) (0.002) (0.002) Observations 13,030 13,030 17,213

Notes: All specifications include state and survey fixed effects, and are estimated using OLS in the case of panel A, and estimated across ten multiply imputed datasets using OLS in the cases of panels B-G Lower-order terms and interactive controls are omitted to save space or when the interpretation is changed by the inclusion of the controls. Standard errors clustered by state in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. 113 To test the duty-based explanation, panel D similarly demonstrates that citizens that have voted in at least one mayoral election since 2000 are no more likely to acquire information in the run up to an election. Again, the results for politically-oriented networks are unaffected. Together, these robustness checks provide no evidence to suggest that the larger effect of upcoming local elections among individuals in more politically-oriented networks reflects the individual-level characteristics that could most plausibly explain sorting into such groups. Another possibility is that individuals in politically-oriented social groups are more likely to passively consume news when it is in greater supply around elections. To assess this potential concern, I examine whether the effect of upcoming local elections is especially large where voters have greater access to media outlets broadcasting news. In particular, panel E controls for the interaction of an upcoming local election with the total number of AM radio stations, FM radio stations and television stations located in the municipality. The lack of a significant interaction with such access to media, combined with the continuing significant interaction with politically-engaged networks, suggests that the results are not driven by differences in access to news. Finally, I demonstrate that the results are robust to using alternative definitions of upcoming local elections. Panel F similarly shows that voters in politically-engaged networks continue to consume differentially most news as elections approach when I examine months until the upcoming local election, instead of the upcoming local election indicator employed above. Table B.6 in the Appendix shows broadly similar results when defining an upcoming local election as an indicator for voters facing an election within any number of months between one and ten.

114 3.5 Conclusion

This paper endogenizes the acquisition of political information, and shows that social incentives play a key role in inducing voters to strategically acquire news about politics. The social signaling model highlights how, especially in groups that collectively value knowledge about politics, the prospect of an individual’s knowledge about politics being revealed causes politically unsophisticated voters to start acquiring information to meet a minimum standard of knowledge, while politically sophisticated voters differentiate themselves by acquiring more information. Exploiting both experimental and observational research designs, I provide evidence that such incentives significantly impact voter information acquisition in Mexico, a major non-consolidated democracy. Across both analyses, I find clear evidence that relatively uninformed voters can be induced to acquire political news, particularly when nested in high-interest social networks. The results provide further evidence that social interactions play an important causal role in political behavior. In addition to previous experimental studies showing that social factors increase turnout, my findings suggest that elections serve as a focal point for political discussion within groups. Such socially-induced informed participation chimes with recent work similarly demonstrating that political concerns are transmitted through social networks, and can substantially alter political preferences (Alt et al. 2016; Jennings, Stoker and Bowers 2009; Newman 2014; Sinclair 2012). However, my findings also raise questions for future research. In particular, a deeper understanding of group dynamics would benefit from studies identifying how incentives to acquire political information vary with group heterogeneity and the extent to which political discussion is factually or ideologically oriented. The findings thus have implications for policy-makers seeking to increase informed political participation. First, the results highlight the importance of timing: information 115 dissemination is only likely to be effective when, e.g. just before elections, voters have strong social incentives to consume the news available to them. This conclusion reinforces the findings of Marshall(2016 a), who shows that voters are only likely to punish incumbents for failing to maintain public security when homicides are reported in the news before elections. In tandem with the conclusions of Gerber, Karlan and Bergan(2009), the results suggest that variation in political news consumption over time may predominantly reflect voter demand, rather than media supply, constraints. Second, the theory also points to an “information trap,” whereby information acquisition is perpetuated by politically-engaged social groups that assign reputational benefits to being informed. Such groups are likely to disproportionately contain well-educated and already well-informed elites. This produces a trap in that poorly informed voters may consequently fail to support parties representing their interests. As Bartels’s (2008) findings suggest in the United States, the least informed voters may fail to support beneficial tax policies. Iversen and Soskice(2015) similarly document a centrist shift among uninformed low-income voters in advanced democracies with low levels of political engagement. Con- sequently, to increase political knowledge across the entire electorate, the results suggest the need for government and the media to engage all types of citizens with politics, which may require visible interactions, public outreach and civic education programs. However, while beyond the scope of this study, a deeper understanding of differences in political engagement across groups remains a central question for future research. Panel studies examining migration between groups may be especially informative. Third, the results may also explain why political parties differentially target policies toward better-informed groups. A key implication of information traps is that parties have limited incentive to represent voters that lack the information required to identify their patrons, or even recognize that policies were targeted to help them. This is consistent with considerable evidence from both developed (e.g. Adams and Ezrow 2009; Gilens 2005) and 116 developing contexts (e.g. Bates 1981; Casey 2015) that policy is more responsive to the interests of elite subgroups of voters, who are typically also better informed about politics. Although further research is required to validate this link, such potential biases in political representation may compound accountability losses due to the lack of political information in politically disengaged networks.

117 4| Publicizing malfeasance: When media facilitates electoral accountability in Mexico1

Co-authored with Horacio A. Larreguy and James M. Snyder Jr.

4.1 Introduction

A large body of scholarship in political economy asserts that in large democracies: (i) elections are one of the key institutions for producing political accountability; (ii) in order for elections to function well, voters must be adequately informed; and (iii) the mass media play an essential role in informing voters. One important application of this trio is the electoral sanctioning of malfeasant behavior such as corruption and diverting funds away from the projects for which they are earmarked. This is particularly relevant in developing democracies, where corruption is prevalent (e.g. Mauro 1995) and voters are relatively uninformed (Greene 2011; Lawson and McCann 2005; McCann and Lawson 2003). This paper seeks to identify when broadcast media effectively hold local government to account by sanctioning malfeasant political parties. The media often cover corruption scandals and other cases of malfeasance (e.g. Puglisi

1We thank Daron Acemoglu, Rakeen Mabud, Ben Olken, Jesse Shapiro, and David Stromberg, as well as participants at the Harvard Comparative Politics Workshop and MIT Political Economy Lunch for useful comments. Thanks to Andrea Ortiz and Daniel Silberwasser for excellent research assistance, and to ASF ofﬁcials for providing information about the auditing process. Horacio Larreguy acknowledges ﬁnancial support from the IQSS Undergraduate Research Scholars Program. All errors are our own.

118 and Snyder 2011). However, there is little solid evidence establishing (a) whether such information causes voters to punish politicians at the polling booth, and (b) what kinds of media station induce such punishment. Estimating these effects is challenging because identification requires exogenous variation in both malfeasance revelations and voter access to media coverage. Ferraz and Finan(2008) find that incumbent mayors in Brazil who are randomly revealed to be corrupt just before an election suffer more at the polls in municipalities with more AM radio stations. However, as the authors acknowledge, without exogenous variation in media access the study cannot separate the effects of AM radio from its many correlates, such as education and demand for political news.2 Several studies similarly show that randomizing access to information about incumbent representatives can induce voters to sanction poor performance in office (Banerjee et al. 2011; Humphreys and Weinstein 2012).3 However, because they lack exogenous variation in malfeasance revelations, such studies cannot separate the effects of the information provided from their mode of transmission. We combine these literatures and overcome such concerns by exploiting plausibly exogenous variation in both the release of audit reports and access to media. Similarly, there is limited evidence that the structure of the media market influences the electoral sanctioning of malfeasant politicians in developing contexts (Larreguy and Mon- teiro 2014). First, not all voters receive politically-relevant news from the media stations that cover them, given that media stations vary in their incentives to supply politically- relevant news to segments of their audience. While Snyder and Strömberg(2010) examine

2Table C.2 in the Appendix shows that media access is significantly greater in more urban, literate, and economically developed areas in Mexico. This is likely to induce bias, given that such variables themselves have important political consequences. For example, Klašnja(2011) finds evidence that voters with greater “political awareness” are more likely to punish incumbents in corruption scandals in the U.S., while Weitz- Shapiro and Winters(forthcoming) find similar results for voters with higher literacy in Brazil. 3Observational studies find that corrupt politicians are more likely to be punished electorally when their corruption is covered in the news, or when political corruption is more salient (Chang, Golden and Hill 2010; Costas, Solé-Ollé and Sorribas-Navarro 2011; Eggers 2014). However, such studies are vulnerable to the concern that the presence of media coverage is correlated with the severity of malfeasant behavior.

119 the implications of media market and electoral boundary overlap for behavior in ofﬁce in the U.S., how such media supply incentives affect the electoral punishment of malfeasant behavior is an outstanding question. Second, surprisingly little empirical attention has been devoted to the role of market concentration. Since it is easier to capture media in concentrated markets (Besley and Prat 2006), voters may require multiple signals to substantially update their beliefs (Gentzkow, Shapiro and Sinkinson 2014) and new media stations may engage new segments of the market (Prat and Strömberg 2005), an additional media stations providing similar news content could considerably increase voter sanctioning. Despite the potential importance of market concentration, existing research has largely focused on identifying the electoral implications of access to a new media station or market offering distinct news content (DellaVigna and Kaplan 2007; Enikolopov, Petrova and Zhuravskaya 2011; Snyder and Strömberg 2010). In this paper, we identify large electoral effects of local broadcast media stations— which emit within an electoral precinct’s municipality—publicizing revelations of mayoral malfeasance in Mexico, particularly when these media stations primarily serve local audiences. To do so, we utilize detailed geographic data and exploit two sources of plausibly exogenous variation. First, as in Ferraz and Finan(2008), we leverage variation in the timing of the release of municipal audit reports around elections. In particular, we use a difference-in-differences design to compare mayors revealed to have engaged in malfeasant behavior—either corruption or diverting funds to other projects that do not beneﬁt the intended poor recipients—before an election to comparable mayors whose audit reports are not published until after the election. Second, and moving beyond existing studies, we also leverage within-neighbor variation in commercial quality radio and television signals. These signals differ across urban electoral precincts from within the same municipality due to plausibly exogenous factor such as antenna power and geographic features lying between

120 the antenna and particular precincts.4 Mexican voters rely largely on such local media, particularly television, to learn about malfeasance in the use of public funds (Castañeda Sabido 2011). In Mexico’s federal system, a significant proportion of government spending is administered by municipal mayors. Amidst widespread concerns about corruption, the Mexican Congress passed a law institutionalizing independent audits of the use of federal funds in 1999. We focus on audit reports pertaining to the Municipal Fund for Social Infrastructure (FISM), a major social program—representing about 25% of mayors’ annual budgets—that provides mayors with funds for infrastructure projects required to benefit impoverished citizens. Audits are announced the year after the funds have been allocated, and reports reveal the share of FISM money spent “in an unauthorized manner,” as well as the share spent on projects “not benefiting the poor.” The first figure clearly represents malfeasance, and usually actual corruption. The second figure indicates malfeasance of a different sort— diverting funds from their intended targets. By law, 100% of FISM projects must benefit the poor, so any money not spent on the poor represents illegal misallocation. Provided that voters care about malfeasance, we are thus able to also address the important question of which types of malfeasance matter most. Our results first demonstrate that each additional local media station substantially increases voter punishment of the party of mayors revealed to be either corrupt or neglectful of the poor. Although mayors cannot seek re-election (due to term limits), voters primarily choose between parties since individual candidates are not well known (Larreguy, Marshall and Snyder 2016) and internal selection procedures ensure that candidates types within parties are highly correlated (Langston 2003). Our point estimates imply that each

4Speciﬁcally, we provide estimates of the “intention-to-treat” voters with access to media, because commercial quality coverage boundaries reduce the likelihood that voters receive a media signal but cannot pre- clude coverage entirely. Such issue are discussed below. Since we lack the data required to compute a ﬁrst stage and the exclusion restriction might not hold, we focus on providing reduced form estimates.

121 additional local radio or television station reduces the vote share of the incumbent political party whose mayor was revealed to be corrupt by more than 0.7 percentage points. The effects of failing to spend funds on the poor are even larger: if the incumbent party’s mayor was revealed to have misallocated funds away from the poor, each additional local radio or television station covering a given precinct reduces the party’s vote share by up to 1.2 percentage points, depending on the severity of the misallocation. Like Ferraz and Finan (2008), we also find that voters reward good performance in office. When the incumbent party’s mayor did not engage in corruption or correctly spent the money on the poor, an audit report released before an election increases the party’s vote share by almost 0.7 percentage point for each additional local media station. Differentiating media formats, we find that exposure to an additional local television station—the most prevalent source of political information in Mexico—has substantially larger effects on electoral sanctioning than an additional FM radio station, which has similar signal range.5 While there are large effects for local media stations, we find no evidence that non-local media stations contribute to electoral sanctioning. Although this could simply be because non-local media stations fail to capture audiences outside their municipality, we find that the presence of an additional non-local media station in fact crowds out the sanctioning effect of local media. Furthermore, consistent with the extent of news coverage of mayoral performance increasing with audience demand for such information, we find—akin to Sny- der and Strömberg(2010)—that the effect of a local media station is larger when the media station’s audience principally resides inside the municipality. These findings suggest that electoral accountability requires that the media market is structured to ensure that stations have incentives to supply politically-relevant information to their audiences.

5We are unable to estimate the effect of an additional AM radio station since we lack a sufﬁciently large sample of neighboring precincts within the same municipality that differ in AM radio coverage. This reﬂects the high power of AM radio antennae, which often cover all urban areas in their municipality.

122 These findings contribute to the literature in a variety of ways. First, we demonstrate the particular importance of local media for municipal electoral accountability, rather than media in general.6 This is a salient consideration because local radio and television are often the only way in which isolated voters can learn about the performance of their incumbent politicians. In Mexico, 45%, 50% and 56% of electoral precincts are respectively not covered by a single AM radio, FM radio or television station emitting from within their municipality. Moreover, understanding the role of local media in supporting electoral accountability is relevant as local media markets in many countries shrink.7 This trend is particularly worrying in light of the fact that our findings provide a clear rationale for malfeasant politicians to exploit the weakening economic position by seeking to control it (Besley and Prat 2006), purchasing radio stations (Boas and Hidalgo 2011) or introducing regulations to outlaw “defamation” (Stanig 2015). Second, we demonstrate that media market structure plays a key role in explaining the extent to which local media matter for electoral accountability. Our results, in a developing context, thus buttress the finding that the electoral composition of media markets has important implications for the types of political news that media stations report (Ansolabehere, Snowberg and Snyder 2006; Snyder and Strömberg 2010; Strömberg 2004), which in turn affects electoral accountability (Snyder and Strömberg 2010). In addition to comparing the relatively efficacy of more and less locally-oriented broadcast media stations, we also show that the presence of non-local media stations detracts from electoral accountability by crowding out politically-relevant news. Third, we show that voters respond to different types of malfeasance. With the excep-

6In contrast, Larreguy and Monteiro(2014) show the importance of regional and national media networks for national electoral accountability. 7Over the last 15 years, Mexico has experienced a 40% decline in the share of individuals claiming to read political news in newspapers. 60% of Latinobarometer respondents claimed reading political news in newspapers 1996, compared to 36% in 2009. In the U.S., daily newspaper circulation dropped from just over 1.0 newspapers per household in 1950 to about 0.3 per household in 2010. 123 tions of Banerjee et al.(2011) and Humphreys and Weinstein(2012), who respectively analyze the discretionary allocation of funds to slums rather than urban areas and the relative parliamentary performance of legislators, the bulk of the literature on political accountability in developing democracies has focused on corruption. Our results are most similar to Banerjee et al.(2011), who also find that voters punish politicians for diverting funds away from the poor. However, the relatively large effects that we document—which are slightly greater for revelations that politicians did not spend on the poor than for corruption revelations—may reflect the fact that in Mexico such diversion is a direct violation of FISM program rules.8 Finally, given that audits are announced after FISM funds have been allocated, our results suggest that the positive likelihood of being audited is insufficient to prevent municipal mayors from engaging in malfeasance. Our study thus complements previous research suggesting that audits can be effective at reducing corruption only if politicians know prior to spending that the reports could result in criminal prosecution (Olken 2007) or will be released before an election (Bobonis, Fuertes and Schwabe 2014). Rather, the corruption levels we observe in Mexico are broadly similar to those found in Brazil (Ferraz and Finan 2008), where the municipal audit scheme was only announced after spending had occurred. This is consistent with the dynamic optimizing behavior observed in India (Niehaus and Sukhtankar 2013), but partially contrasts with recent findings from Brazil suggesting that increasing the probability of audit reduces corruption but does not affect spending patterns (Zamboni and Litschig 2014). Our findings ultimately suggest that media-induced electoral accountability can mitigate adverse selection problems by reducing the probability that the parties of malfeasant mayors are re-elected. This paper proceeds as follows. Section 4.2 first outlines why an additional media

8Conversely, legislators in India are free to allocate their discretionary project funds anywhere in their districts.

124 station may support electoral accountability. Section 4.3 provides a brief overview of local governments in Mexico, the FISM funds that we study, the audit of such funds, and local media in Mexico. Sections 4.4 and 4.5 detail our data and identiﬁcation strategy. Section 4.6 presents our main results and robustness checks. Section 4.7 concludes.

4.2 Why the intensity of media coverage matters

An influential literature has emphasized the political importance of access to a new media station or media market offering distinct content, both in terms of the focus of their news (e.g. Ansolabehere, Snowberg and Snyder 2006; Snyder and Strömberg 2010) and their ideological stance (e.g. DellaVigna and Kaplan 2007; Enikolopov, Petrova and Zhu- ravskaya 2011). However, despite significant theoretical interest in the role of media market concentration (e.g. Besley and Prat 2006; Gentzkow and Shapiro 2006), little is known about the marginal effect of an additional media station providing relatively similar content on electoral accountability. Combining existing arguments, we suggest that access to more media stations providing similar content may also play an essential role in informing voters about the behavior of their politicians in office. This is of particularly relevance in light of the mixed evidence regarding whether voters punish malfeasant behavior in office (e.g. Banerjee et al. 2011; Chong et al. 2015; Humphreys and Weinstein 2012; Ferraz and Finan 2008). In theory, the existence of a single news-reporting media outlet could be sufficient to support electoral accountability. Consider a situation where the incumbent is corrupt, but voters may only learn this through the news. Abstracting from demand-side biases, for one news-reporting media station to be sufficient to generate electoral sanctioning, many voters would have to receive credible information from such an outlet, sufficiently update their beliefs about their incumbent (party), and ultimately vote according to such beliefs.

125 In practice, many voters may not consume information, or the information consumed may not cause voters to substantially update their beliefs. We argue that, once these conditions are violated, an additional news-reporting media station may substantially affect voter behavior. In their model of media capture, Besley and Prat(2006) show that political capture of media stations decreases with the number of stations in the market, and in turn increases turnover of low-quality politicians by ameliorating adverse selection concerns. Intuitively, this is because high market concentration makes it costlier for politicians to buy off many media stations that must be paid monopoly prices to resist the commercial profits associated with revealing information. Consistent with this model, Djankov et al.(2003) provide suggestive evidence that state media ownership reduces government performance in developing countries, while Boas and Hidalgo(2011) find that Brazil politicians actively seek to control local media stations once in office. Without political capture, media firms and journalists may have incentives not to reveal politically-relevant information (see e.g. Anderson and McLaren 2012; Baron 2006; Gentzkow and Shapiro 2006; Mullainathan and Shleifer 2005). For example, Gentzkow and Shapiro(2006) suggest that the media may bias its news coverage toward the priors of their consumers in order to increase their reputation as a high-quality news outlet. However, greater competition and diversity in the media market increases the probability that com- petitors will expose the factual inaccuracies of a report pandering to audience priors, and thus ensures that voters receive truthful signals about political performance.9 Although newspaper biases are correlated with the biases of their readers (Gentzkow and Shapiro 2010), there is some evidence that media market competition indeed reduces biased reporting (Gentzkow and Shapiro 2006; Puglisi and Snyder 2011), although it may come

9Anderson and McLaren(2012) make a similar argument in a model where media owners are politically biased.

126 at the cost of journalistic quality (Cagé 2014). Moreover, receiving the same information from multiple sources should increase voter certainty regarding incumbent performance, especially when the information is consistent across outlets with different political biases (Gentzkow, Shapiro and Sinkinson 2014). The entry of a new media station may also affect media consumption, and consequently increase consumption of politically-relevant news. Even in relatively large markets, new media can expose new listeners. Supporting this claim, Prat and Strömberg(2005) show that the introduction of commercial television disproportionately attracts relatively uninformed voters, causing them to become more politically knowledgeable and increase their political participation. In particular, media stations with specific qualities may be particularly effective at gaining new audiences, given that voters are willing to change their media consumption patterns in response to changes in the types of available media (Durante and Knight 2012). Similarly, the introduction of media stations with different partisan biases can attract sufficiently large audiences to alter electoral behavior (DellaVigna and Kaplan 2007; Enikolopov, Petrova and Zhuravskaya 2011). Below, we offer suggestive evidence that the availability of additional local media station increases news consumption in our context of Mexico. Finally, even if voters consume unbiased news coverage, such coverage must still be politically relevant to affect electoral accountability. Profit-maximizing media firms face strong incentives to tailor their programming to consumer demands, and may thus focus on catering to particular tastes (Strömberg 2004). In diverse media markets, small consumer groups may simply not be served. For example, television stations predominantly cover in-state politicians even when the media market includes voters outside the state (An- solabehere, Snowberg and Snyder 2006; Snyder and Strömberg 2010). This phenomenon may detract from electoral accountability, not only by failing to provide information about incumbent performance, but also by crowding-out media stations that do provide such con- 127 tent. There are thus good reasons to believe that reducing media market concentration may enhance electoral accountability, by increasing voter exposure to credible and informative signals. This paper aims to causally identify the impact of an additional broadcast media station providing news about incumbent party malfeasance. Furthermore, to capture the role of market structure, we differentiate local and non-local media stations and examine how the effects of media coverage change with media market composition.

4.3 Political accountability in Mexico

Following 70 years of Institutional Revolutionary Party (PRI) hegemony, national and local politics in Mexico have become relatively competitive. Elections to the Chamber of Deputies, the lower house of Mexico’s national legislature, are held every three years, while the President and Senate are concurrently elected to six-year tenures.10 State and municipal elections are instead staggered across the electoral cycle and typically held every three years. Over the period of our sample, three main political parties competed for political control: the left-wing Party of the Democratic Revolution (PRD), the populist PRI, and the right-wing National Action Party (PAN).11 Competition in most parts of the country was generally between only two of these parties, with the PRI performing best in rural areas and the PAN and PRD performing best in urban areas (Larreguy, Marshall and Querubín 2016).

10In the Chamber of Deputies, 300 members are elected via plurality rule from single-member districts and 200 members are elected via proportional representation. The Senate comprises 128 Senators, where three are allocated from each state (including the Federal District) such that the largest party receives two Senators and the second largest receives one Senator, and a further 32 are allocated according to the national vote share. 11The PRD’s candidate in the 2006 and 2012 presidential elections—Andrés Manuel López Obrador—left the PRD to form a new party, the National Regeneration Movement (MORENA). MORENA stood for the ﬁrst time in 2015.

128 4.3.1 Municipality audits

In 2010, Mexico’s 31 states contained 2,435 municipalities, varying in population from 102 to 1.8 million with a mean of 46,134. States and municipalities exercise significant control over local policy. Following major fiscal decentralization reforms in the 1990s, the average municipality’s annual budget has been around nine million U.S. dollars, which constitutes 20% of total government spending.12 Municipal governments are led by mayors, who are responsible for delivering basic public services and managing local infrastructure. Mayors are normally elected every three years, although they serve four-year terms in some states, but will only become eligible for re-election for the first time in 2018. An important component of a mayor’s budget is the Municipal Fund for Social In- frastructure (FISM). On average, this represents 24% of a municipality’s total resources. FISM funds, which are allocated to municipalities according to the Fiscal Coordination Law (LCF) passed in 1997, are direct federal transfers provided exclusively for the funding of public works, basic social actions, and investments that directly benefit the socially disadvantaged population living in extreme poverty. Spending may be allocated in any of the following categories: potable water, sewage, drainage and latrines, municipal urbanization, electrification or rural and poor suburban areas, basic health infrastructure, basic education infrastructure, improvement of housing, rural roads, and rural productive infrastructure. The use of FISM funds is subject to independent audits by Mexico’s Federal Audi- tor’s Office (ASF). The ASF, which was established in 1999 in response to widespread concerns regarding the mismanagement of public resources, is an independent body with constitutionally-enshrined powers to audit the use of federal funds by the federal, state and municipal governments. In each year since 2000, the ASF has audited FISM spending in

12Education and health were decentralized between 1992 and 1996 and the decentralization of infrastructure projects followed in 1997 (Wellenstein, Núñez and Andrés 2006).

129 multiple municipalities per state. Audits focus on the spending and management of FISM resources in the prior fiscal year, and the list of municipalities to be audited in a given year is announced the year after the spending occurred. Between 2007 and 2012, 14% of Mexi- can municipalities were audited at least once, with around 120 municipalities being audited each year. Although municipalities are not randomly selected for audit, the timing of an audit is essentially random. The ASF’s summary report makes clear that only the following criteria affect the probability of audit in a given year: the financial importance of FISM funds to the municipality, relative to the municipal budget; historical performance indicators and institutional weaknesses that raise the likelihood of misallocation; whether FISM spending has recently been audited; whether other federal audits are occurring simultaneously; and where a specific mandate exists to examine a particular municipality.13 Direct communication with the ASF confirmed these criteria, and clarified that an audit should not have occurred within the last two years, and that—for logistical reasons—they often simultaneously audit neighboring municipalities.14 Crucially, given that our identification strategy exploits the timing of audits, our correspondence with the ASF also confirmed that the selection of municipalities for audit does not depend upon the electoral cycle or the government’s political identity. Our balance tests below confirm that the timing and content of audit reports are uncorrelated with a wide variety of political, demographic, and socioeconomic indicators. Independent ASF auditors check that officials abide by the rules established for the

13This information is formally stated on page 240 of their 2014 summary report, “Informe del Resultado de la Fiscalización Superior de la Cuenta Pública 2012,” available here. Federal auditors may also audit Funds for the Strengthening of Municipalities and Federal Demarcations of the Federal District (FORTAMUNDF) funds allocated in proportion to the number of citizens and intended to strengthen municipal social spending, or Subsidy for Security in Municipalities (SUBSEMUN) funds allocated to support public security. 14Based on a personal interview with the Licentiate Jaime Alvarez Hernández, General Director of Re- search and Evaluation of the Special Audit of Federal Spending, in July 2012, and formal email correspondence with the ASF. 130 management of FISM resources (e.g., procurement rules and accounting procedures), that the status of the funded projects is in accordance with the books, and that funds are used as intended. Reports categorize the use of FISM funds across multiple dimensions; an example report is provided in Figure C.1 in the Appendix. Most importantly, reports state the percentage of FISM funds spent on projects not benefiting the poor and the percentage of funds spend on unauthorized projects. Spending that does not benefit the poor ranges from the diversion of resources to support agricultural production in areas without poverty to paving the streets of relatively rich urban areas. We interpret unauthorized spending, which includes the diversion of resources for personal expenses of the mayor or electoral campaigning and funds that are unaccounted for, as corruption.15 Audit reports are publicly released two years after the spending actually occurred, when they are presented in Congress by the last working day of February each year and made publicly available online at the ASF’s website.16 Relative to previous social programs, FISM funding has been comparatively successful at targeting resources at the poor (Wellenstein, Núñez and Andrés 2006). However, funds are often misallocated. Unlike previous studies focusing on corruption in more general programs (e.g. Bobonis, Fuertes and Schwabe 2014; Ferraz and Finan 2008), the specific targeting of FISM funds allows us to examine the electoral response by voters to both corruption and the misuse of funds intended to serve a disadvantaged population. The ASF can impose a variety of punishments on malfeasant public officials. In particular, the ASF can inflict fines on the municipality to recover FISM funds, recommend that the Ministry of Public Function removes, suspends or imposes economic sanctions on

15This deﬁnition resembles Ferraz and Finan(2008) in that we focus on violations that include procurement fraud, diversion and over-invoicing, but differs in that we quantify the relative importance of such corruption. Rather than the percentage of unauthorized spending, Ferraz and Finan(2008) count the number of corruption violations. 16Guía para el ciudadano. ¿Qué es y qué hace la Auditoría Superior de la Federación?

131 officials, or file (or recommend) a criminal case against culpable individuals. Between De- cember 2006 and July 2012, the Ministry of Public Function only recovered two million U.S. dollars, sanctioned 9,000 public employees for serious misdemeanors, and incarcer- ated one hundred officials.17 However, the largest punishment may be electoral. Since Mexican mayors cannot stand for re-election, any electoral penalty hits the party of a malfeasant mayor. This feature differentiates this paper from many preceding studies (e.g. Banerjee et al. 2011; Ferraz and Finan 2008; Humphreys and Weinstein 2012). However, there are good reasons to believe that a mayor’s political party may be punished by voters at the next election. First, although some voters are aware of particular candidates, Mexico’s strong party system ensures that party labels play a key role in voting decisions. Previous studies show that voters are poorly informed about local politics, and that voters are willing to punish the party’s of local mayors for their actions in office (e.g. Chong et al. 2015). Second, the top-down internal structure of Mexican parties at the state level ensures that within-party candidate choice is highly correlated at the local level (Langston 2003).

4.3.2 Local broadcast media

As in many developing countries, radio and television stations are the principal source of news in Mexico. Within Mexico’s large media networks, the majority of entertainment content is common. Such content sharing among major radio networks such as Radiorama, ACIR, Radiocima, Multimedios Radio and MVS Radio, and Mexico’s two main television networks, Televisa and TV Azteca, allows Mexico’s many small media stations to support themselves through local advertising revenues including government advertising expendi- tures. However, news programming is more locally differentiated. Conditional on provid-

17El Universal, “A la cárcel, solamente 100 ex servidores,” 29th May 2014, link.

132 ing news, radio and television stations provide around 2 hours of news coverage a day.18 Based on our content analysis of television schedules, around half of news programming is local. Voters are generally unaware of mayoral responsibilities and the use of public funds (Chong et al. 2015), as well as the institutions that are responsible for auditing the use of public resources (Castañeda Sabido 2011). Most public spending is invisible and in- accessible to most voters. A study by Castañeda Sabido(2011) indicates that only 34% of surveyed individuals believe that municipal governments are transparent about the use of public resources.19 Moreover, only 25% of surveyed individuals can mention a public institution in charge of auditing the use of public funds, and only 1.4% of those individuals mention the ASF as the main institution responsible for that task. Voters learn about public spending primarily through media coverage. Figures from the 2009 Latinobarometer indicate that 83% of respondents gather political information from television, 41% gather political information from radio, 30% gather political information from newspapers, and 41% gather political information from family, friends and colleagues (many of whom, of course, gather their information from television, radio and newspapers).20 Internet is not widespread: according to the 2010 Census, only 24% of households in the average electoral precinct have internet access. Additionally, according to Castañeda Sabido(2011), 83% of individuals report that they receive information about malfeasance in the management of public resources through media, and 61% regard such

18These ﬁgures are based on IFE monitoring of a non-random sample of 200 radio and television stations providing news coverage during the 2012 Mexican Federal election. 19In principle, local governments are required to inform the public about the arrival of FISM funds. How- ever, only about 50% comply with this requirement. Moreover, among those that do comply, the main communication channels used are newspapers and the internet—i.e., two types of mass media. Furthermore, media relies extensively on the ASF audit reports since governments are extremely reluctant to release information about their expenses to the public (Lavielle, Pirker and Serdán 2006). 20These are the only four responses to an open-ended question that received a non-negligible number of mentions.

133 information as reliable. We thus expect that television may be the most the important media source of incumbent performance information. Furthermore, the likelihood that a voter follows the news increases with the availability of local media. Using data on media consumption from the 2009 CIDE-CSES Survey, Ta- ble C.1 in the Appendix demonstrates a strong positive correlation between access to local media stations (defined in detail below) and consumption in a predominantly urban sample. Specifically, an additional local radio and television station respectively increases the probability that an individual regularly watches a news program by 0.4 and 1.4 percentage points, as well as significantly increasing the total number of news programs watched regularly. In total, 80% of respondents listed at least one television program, and 30% listed one radio program. Conversely, we find no evidence that the number of non-local media stations, which are less likely to provide local news coverage, are associated with news consumption. The release of municipal audit report results each February is a major annual media event, especially for television stations. As in many other countries (see Pande 2011), revelations about political malfeasance are frequently reported, and can remain salient news for many months. News reports generally cover mayors within the local vicinity, and typically focus on cases of corruption and mayors not spending FISM funds on projects targeting the poor.21 Most reports accurately cite the proportions of unauthorized spending and spending on projects not targeted at the poor, and some dig deeper to describe the nature of the malfeasance. There are many such examples available online.22 For example, in 2013 BBM radio station reported that Oaxaca de Juarez’s mayor had created a fake union to

21Little mention was made of other features of the reports such as the the degree of participation of the community in the allocation of funds or the share of FISM funds that were spent. 22For example, see: BBM Noticias, “ASF: desvió Ugartchechea 370.9 mdp,” 21st October 2013, here; El Informador, “Hallan irregularidades en gasto tapatío contra pobreza,” 28th February 2013, here; Revolución Tres Punto Cero, “En 2012, se desviaron a campañas 29 millones de pesos para combate a la pobreza en Tabasco,” 6th March 2014, here. 134 collect payments, presided over many public works contracts without offering open tender, diverted payments for advertising and consulting fees, and failed to provide details of considerable quantities of spending.23 While this particular case represents one of the most corrupt mayors, such behavior was not uncommon: many media reports pointed to mayors diverting payments, using FISM funds for personal and family expenses and manipulating tender processes. Failures to spend FISM funds on the poor were also common in media reports. In other cases, public works projects were undertaken in urban and afﬂuent parts of the city, never materialized despite being paid for, or were diverted toward alternative uses such as supporting local candidates from the incumbent’s party. To more systematically demonstrate the substantial coverage of the audit report releases, Figures 4.1a and 4.1b show spikes in Google searches for the ASF and FISM around February and March of each year. We now turn to our data, and to the empirical strategy we use to identify the effects of the release of audit reports in electoral precincts covered by an additional media station.

4.4 Data

This section describes our main sources of data: incumbent electoral performance at the electoral precinct level; municipal audit reports released just before and just after an election; and precinct-level radio and television coverage.

4.4.1 Mayoral election outcomes

Mexico’s municipalities are divided into approximately 67,000 electoral precincts. Us- ing data from the Federal Electoral Institute (IFE) and State Electoral Institutes, we col-

23BBM Noticias, “ASF: desvió Ugartchechea 370.9 mdp,” October 21st 2013, here.

135 h aacvrGol erhsi eiofrtepro sdi u sample. our in used period the for Mexico in searches Google cover data The Notes iue41 ogesace eae oadtrprsb ot,2006-2012 month, by reports audit to related searches Google 4.1: Figure n1t uy2014. July 15th on (http://correlate.googlelabs.com) Correlate Google using Extracted :

Standardized search activity (sigma) Standardized search activity (sigma) -1 0 1 2 3 -2 -1 0 1 2 3 Feb 06 Feb 06 Feb 07 Feb 07 Feb 08 Feb 08 b erhsfrFISM for Searches (b) a erhsfrASF for Searches (a) Feb 09 Feb 09 136 Feb 10 Feb 10 Feb 11 Feb 11 Feb 12 Feb 12 lected electoral returns for every available precinct in each municipal election between 2004 and 2012. We thus accumulated up to four election results per electoral precinct, which enabled us to identify the municipal incumbent party and their past vote share in all the elections in our period of analysis, 2007-2012.24 We focus on two main outcomes: the change in incumbent party’s vote share at the precinct level, and whether the incumbent party was re-elected at the municipal level. The former measure quantifies the extent of precinct-level voter sanctioning, while the latter captures the municipality-level electoral outcome. For most of this paper, we exploit fine- grained variation in media coverage across precincts, and thus solely focus on changes in the incumbent party vote share. We define the vote share by the proportion of voters that turned out.25 The incumbent party received 48% of votes in the average electoral precinct in our sample, which represents a 4.7 percentage point decline in their vote share. Since Mexican mayors could not stand for re-election, we focus on the party of the incumbent mayor. However, municipal politics often entail the formation of local coalitions between political parties, and this can change across elections. For example, in 2009 the incumbent mayor of the municipality of Colima represented a two-party coalition containing the PRI and the Green Party (PVEM). However, the 2009 election saw six groups stand for election: the PT, PVEM and PC all stood separately against three coalitions, PAN-ADC, PRI-PANAL and PRD-PSD. To classify such cases where the incumbent coalition split at the next election, we determined the party affiliation of the mayor by researching their identity and party ties.

24By ending our sample in 2012, our sample does not feature any mayors that can seek re-election. 25We obtain similar results when measuring vote share as a proportion of registered voters.

137 4.4.2 Audit reports

Since audit reports are released with a two year lag, reports released in February of a municipal election year generally refer to the first year the incumbent mayor was in office.26 Since municipal elections take place later in the calendar year, we define a pre-election audit report release by whether an audit was released in February of an election year.27 Typically, the report is released four months before the election. Our control group will be mayors in municipalities where the audit was released in February the year following the election. In such cases, the audit report generally pertains to their second year in office. The results of audit reports, which quantify the use of FISM funds, are publicly available on the ASF’s website. We extracted the proportion of funds spent in an unauthorized manner and the proportion of funds not spent on projects benefiting the poor from every available report between 2007 and 2012. This yielded 742 municipal audits relatively evenly spread across years and covering 351 unique municipalities. Of these, 470 reports from 321 different municipalities were released in an election year or the year after. We henceforth restrict attention to this subsample of audits, which are shaded by their level of malfeasance in Figure 4.2. We operationalize malfeasance using indicators to capture severity. For corruption, we define indicators for precincts with mayors in the third and fourth quartiles of the distribution of unauthorized FISM spending. Unauthorized spending in the third quartile ranges from 0.6% to 11.2% of available FISM funds with a mean of 5.1%, while spending in the fourth quartile exceeds 11.2% with a mean of 30.7%. For neglectful spending, we similarly define indicators for mayors in the third and fourth quartiles with respect to FISM funds not

26In Coahuila, where mayors have recently been elected to four-year terms, the report instead refers to the second year of their term. 27Although states differ in the month in which they hold elections, only the state of Baja California Sur holds elections before mid-February. We adjust for Baja California Sur accordingly.

138 Figure 4.2: Distribution of audit report outcomes by municipality Notes: Only the 268 municipalities in our ﬁnal sample are included. Where more than one audit occurs, we take the average audit outcome.

139 allocated to spending on the poor. Not spending on the poor in the third quartile contains any positive value up to 12.9% of available FISM funds with a mean of 5.7%, while the fourth quartile exceeds 12.9% with a mean of 38.0%. Since around 50% of precincts did not experience corruption or not spending on the poor, the 25th percentile of each distribution is 0%; hence we do not use an indicator for the second quartile. These indicators are more ﬂexible than linear measures, and may better reﬂect standard theoretical models suggesting that voter sanctioning involves cut-off rules (e.g. Barro 1973; Ferejohn 1986). Furthermore, our examination of media reports indicates that only relatively serious cases are widely reported. Nevertheless, our robustness checks show similar results when using a linear measure of performance or different cutoff levels of malfeasance.

4.4.3 Radio and television coverage

In addition to ﬁne-grained electoral data, a key feature of this study is our detailed media coverage data. Following a major media reform in 2007 (see Serra 2012), the IFE required that every AM and FM radio station and every television station in the country provide signal coverage data.28 For each media station we are thus able to identify the municipality from where it broadcasts and deﬁne the commercial quality coverage range of its signal.29 Inside a station’s coverage area the signal is of high quality, so precincts inside the area have good access to the station’s broadcasts. Precincts outside the coverage area experience sharply declining coverage as the distance from the boundary increases. See Larreguy, Marshall and Snyder(2016) for further details of the coverage data.

28For only a small number of FM and television stations did the same station broadcast from multiple municipalities. No electoral precincts received the same signal from multiple antennae. 29The IFE deﬁnes the boundary of the coverage area using a 60 dBµ threshold for signal strength. This is the threshold commonly used to determine a radio station’s audience and sell advertising space commercially in the U.S., where it “is recognized as the area in which a reliable signal can be received using an ordinary radio receiver and antenna” (NTIA link).

140 Figure 4.3: AM radio signal coverage areas (source: IFE)

Figures 4.3-4.5 map the location and coverage of each of the 852 AM, 1,097 FM and 1,255 television stations. Although media coverage is extensive, with most precincts receiving at least one media signal and most municipalities containing at least one media station, there is considerable variation in the number of media stations covering each precinct that emit from within the precinct’s own municipality.30 The ﬁgures also clearly indicate that the commercial quality coverage range of AM radio is substantially greater than for FM and television. However, as discussed in detail below, we restrict attention to more urban precincts that primarily differ in terms of FM and television signals. Our principal measure of local media coverage is the total number of AM, FM or television stations covering a given electoral precinct that broadcast from within the precinct’s municipality. To avoid counting signals that barely cover a given precinct, we use (ur-

30Since the number of radio and television stations has remained constant between 2003 and 2010, we cannot exploit temporal variation in media coverage.

141 Figure 4.4: FM radio signal coverage (source: IFE)

Figure 4.5: TV signal coverage (source: IFE)

142 ban) block-level and (rural) locality-level population data from the 2010 Census to deﬁne a precinct as covered by a given media station only if at least 20% of its population lies in side the commercial quality coverage boundary. The average precinct is covered by 4.4, 5.8 and 2.5 local AM, FM and television stations respectively, while the total number of local media stations covering a precinct ranges from 0 to 44. We also examine the effects of each type of media station separately. To compare the effects of local media to non-local media, we also consider the number of media stations that cover a precinct but transmit from outside their municipality. The average precinct receives roughly as many FM and television signals from inside their municipality as outside, although the greater signal range of AM stations means that precincts are typically covered by three times as many AM stations emitting from outside their municipality.

4.5 Empirical strategy

To identify the effect of local media coverage of municipal audit reports on the incumbent party’s electoral performance, we exploit exogenous variation in both the timing of audit report releases around elections and access to local media. In particular, we combine the difference-in-differences design of Ferraz and Finan(2008) with plausibly exogenous variation in the number of local media stations covering neighboring electoral precincts.

4.5.1 Identifying the effects of audit reports

The difference-in-differences (DD) component of our design rests upon exogenous variation in the timing of audit report releases. To first identify the effects of audits we compare municipalities where an audit report was released just before a municipal election to a control group of municipalities where the audit was released after a municipal election. We 143 then move beyond this first difference by also comparing municipalities where the mayor is corrupt or neglectful. The DD sample contains 47,938 precinct-election observations. Conditional on our sample of municipalities that have been audited at least once, the identifying assumption required to estimate the effects of audits released just before an election is that the timing of the releases is effectively random. Supporting the ASF’s claim that selection is independent of electoral considerations, Table 4.1 confirms that differences between electoral precincts in municipalities where an audit was released in the year before an election and precincts where an audit was released the following year across 49 political, demographic, media, and economic characteristics are consistent with chance. Specifically, we find only one statistically significant difference (which is only statistically significant at the 10% level). The final 26 variables are from the precinct-level Census data from 2010, and are described in detail in the Appendix. A second potential concern is that FISM spending decisions—the content of audit reports—might differ across reports released before and after elections. This could reflect differences in the behavior of either auditors or mayors. First, even though the timing of audits is effectively random, auditors could still be more lenient or more meticulous in the knowledge that a report will be released in an election year. Second, mayors anticipating the release of an audit report in an election year may spend more appropriately (Bobonis, Fuertes and Schwabe 2014). Alternatively, the relative inexperience of first-year mayors— about whom audit reports are released before an election—may induce them to allocate their funds differently from second-year mayors (Chattopadhyay and Duflo 2004). However, we find no empirical support for such concerns. Table 4.1 shows that there is no systematic correlation between pre-election audits release and either the proportion of unauthorized spending or the proportion not spent on the poor. This holds regardless of whether we measure such spending on average or in terms of being in the third or fourth quartile of either distribution. Furthermore, Figure 4.6 compares the report outcome distri- 144 Table 4.1: Summary statistics by audit status (DD sample)

Control (no audit) mean Audit difference Unauthorized spending 0.106 -0.028 (0.030) Corrupt Q3 0.186 0.062 (0.065) Corrupt Q4 0.322 -0.110 (0.080) Spending not on the poor 0.097 0.028 (0.033) Not poor Q3 0.229 0.026 (0.075) Not poor Q4 0.263 -0.012 (0.079) PAN incumbent 0.354 0.022 (0.085) PRI incumbent 0.464 0.073 (0.086) PRD incumbent 0.171 -0.083 (0.058) Coalition partners 1.751 0.134 (0.228) Municipal incumbent victory margin (lag) 0.151 0.010 (0.019) Municipal effective number of political parties (lag) 2.615 -0.038 (0.077) Incumbent party vote share (lag) 0.484 -0.003 (0.012) Effective number of political parties (lag) 2.485 -0.031 (0.063) Registered voters (log) 7.450 -0.026 (0.060) Turnout (lag) 0.472 0.030 (0.018) Distance to municipal head from precinct border (log) 8.160 -0.015 (0.067) Distance to municipal head from precinct centroid (log) 8.460 -0.045 (0.054) Area (log) 1.039 -0.077 (0.115) Population (log) 7.688 0.026 (0.055) Population density (log) 7.932 0.159 (0.210) Local media 10.446 3.743* (2.182) Non-local media 27.626 -3.349 (4.566) Average children per woman 2.298 -0.012 (0.034) Share households with male head 0.749 -0.002 (0.005) Share indigenous speakers 0.041 -0.001 (0.008) Average years of schooling 8.947 0.209 (0.154) Share illiterate 0.052 -0.001 (0.005) Share no schooling 0.058 -0.001 (0.004) Share incomplete primary schooling 0.942 0.001 (0.004) Share complete primary schooling 0.828 0.006 (0.010) Share incomplete secondary schooling 0.674 0.011 (0.013) Share complete secondary schooling 0.623 0.011 (0.013) Share higher education 0.385 0.018 (0.015) Share economically active 0.408 0.007 (0.006) Share without health care 0.321 -0.015 (0.014) Share state workers health care 0.052 0.007 (0.005) Average occupants per dwelling 3.925 0.041 (0.038) Average occupants per room 1.068 -0.008 (0.024) Share non-dirt ﬂoor 0.944 0.003 (0.005) Share toilet at home 0.961 0.003 (0.005) Share running water 0.898 0.014 (0.016) Share drainage 0.923 0.011 (0.011) Share electricity 0.981 0.001 (0.002) Share washing machine 0.711 0.020 (0.018) Share fridge 0.857 0.015 (0.013) Share cell phone 0.686 0.015 (0.016) Share car 0.468 0.014 (0.023) Share computer 0.325 0.018 (0.019)

Notes: The audit difference results are from regressions of the outcome variables on the left-hand-side of the table on an indicator for an audit being released the year before an election, where standard errors clustered by municipality election are in parentheses. There are 47,938 observations for each variable. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

145 i,peicsaewihe ytenme frgsee oes euea pncnkvkernel Epanechnikov an use analy- We main 0.025. voters. the of registered with bandwidth of As a number election. with the an by after weighted just are and precincts before sis, just audit an received that nicipalities Notes iue46 itiuino ui eotrsls yrlaeaon election around release by results, report audit of Distribution 4.6: Figure h ui ucmskre est itiuin r ae neetrlpeicsi mu- in precincts electoral on based are distributions density kernel outcomes audit The :

Density Density 0 5 10 15 20 0 5 10 15 20 25 0 0 .2 .2 b pnignto h poor the on not Spending (b) a nuhrzdspending Unauthorized (a) Proportion ofspendingnotonthepoor Proportion ofspendingunauthorized After election After election .4 .4 146 .6 .6 Before election Before election .8 .8 1 1 butions across audit reports released just before and just after an election.31 The distributions of unauthorized spending and spending not on the poor are very similar, and in neither case does a Kolmogorov-Smirnov test reject equality of distribution.32 Combined with our randomization check, this indicates that the audits reports released in election years are typical of “normal” FISM spending. Given the exogeneity of the timing of audit report releases, and their content, we estimate the following simple DD speciﬁcation to identify the effect of revealing a mayor to be corrupt or neglectful before an election:

Ypmt = β1auditmt + β2outcome Q3mt + β3outcome Q4mt +β4 auditmt × outcome Q3mt + β5 auditmt × outcome Q4mt + εpmt,(4.1)

where Ypmt is the incumbent party’s vote share in precinct p in municipality m in year t (or whether the incumbent party won the election in municipality m), auditmt is an indicator for an audit being released in the year before the election, and outcome Q3mt and outcome Q4mt are indicators for municipalities in the third and fourth quartiles of the distributions of corrupt or neglectful mayors (regardless of whether the audit was released before or after the election). Since precincts differ in electorate size, we weight each observation by the number of registered voters.33 Throughout, we cluster by municipal election to account for spatial correlation between precincts receiving the same audit report.

Our main coefﬁcients of interest are β4 and β5, which identify the effect of an audit conditional upon it revealing corruption, or that the mayor did not—as legally required—

31The graphs are very similar if we compare audits from election years to all non-election years (i.e. not just including audits released in the year after a municipal election). 32Collapsing to the municipality-audit level (with 470 observations) to avoid duplication across precincts, the p values of the corruption and not spending on poor distributions are respectively 0.39 and 0.99. 33Given that electoral precincts were all designed to contain up to 1,500 voters, precinct numbers remain relatively similar, and the unweighted results yield very similar estimates. 147 spend FISM money on the poor. We are also interested in β1, which identiﬁes the effect of an audit conditional upon it revealing no malfeasance.

4.5.2 Identifying the effects of local media stations revealing audit re-

ports

To examine how media affects voter punishment of malfeasant behavior, Ferraz and

Finan(2008) further interact auditmt × outcomemt with the number of AM stations located in a municipality. If the number of AM stations were effectively randomly assigned, then this would estimate the average effect of an audit report being released for each additional local media station.34 However, in general, media stations are not randomly assigned across electoral precincts. As we show in the Appendix, the number of local media stations received by a precinct in our sample is signiﬁcantly correlated with almost every Census characteristic in Table 4.1. Local media is signiﬁcantly more prevalent in more highly developed, urban and politically uncompetitive precincts. These correlations may upwardly bias our estimates of local media’s effects if, for example, the better educated and informed citizens in such precincts are more willing or able to sanction incumbent mayors revealed to be malfeasant (e.g. Alt, Lassen and Marshall 2016; Weitz-Shapiro and Winters forthcoming).35 To plausibly identify the effects of local media coverage, we compare neighboring electoral precincts from within a given municipality that differ in the total number of local AM radio, FM radio and television stations that they are covered by (for similar designs, see Ansolabehere, Snowberg and Snyder 2006; Enikolopov, Petrova and Zhuravskaya 2011;

34Furthermore, unlike our precinct-level data, this strategy rests upon between-municipality differences in media coverage. 35In theory, these correlations could also downwardly bias our estimates if such precincts contain voters with stronger prior beliefs about their incumbent’s quality (Zaller 1992).

148 Fergusson 2014; Snyder and Strömberg 2010). A media station is local to a precinct if it emits from within the same municipality. As explained above, we define a precinct as covered by a given media station if at least 20% of its population lie inside the commercial quality boundary. Since broadcast signals decay continuously and whether any given household receives a signal may depend upon the quality of their receiver, discrete differences in commercial quality signal coverage by our definition do not imply that neighboring precincts differ strictly between receiving or not receiving a station’s signal. Rather, our design estimates an “intent to treat” effect of differences in the proportion of an electoral precinct that can receive commercial quality local media.36 Our estimates will thus understate the effect of an terminal drop in coverage. Similarly, any spillovers (e.g. arising from driving to work across radio coverage boundaries) would further attenuate our estimates toward zero. Our identifying assumption is that neighboring precincts differ only in their local media coverage. Restricting attention to within-neighbor variation removes a wide variety of potential confounds. Ideally, otherwise similar precincts near the commercial quality coverage boundary only vary due to exogenous and fixed signal impediments or facilitators such as large physical objects, terrain, and salt water that affect ground conductivity (in the case of AM long waves) and line of sight (in the case of FM and television waves). As

36Ideally, we could also identify the electoral effect of receiving or consuming an additional media station using instrumental variable techniques. To estimate the relevant first stage, we would need to measure either the proportion of voters in each precinct that can access all media stations or the proportion of voters that actually listen to each radio stations or watch each television station. Unfortunately, such detailed individual- level data is not available. Survey datasets typically cover only 1-2% of all electoral precincts and never ask specifically about which radio or television stations voters have access to or actually consume. Although the Comparative Study of Electoral Systems and Mexican Panel surveys did ask whether respondents listen to the radio or television, the surveys are predominantly urban and cover only 1-2% of electoral precincts. The Latinobarometer, which also asks basic questions about media consumption, does not provide precinct-level identifiers for its respondents. Even if such surveys had greater coverage, none of the surveys could identify the number or identity of the media available to voters—such measures would be necessary to compute the relevant first stage. Furthermore, since voters are likely to discuss the news that they receive with their friends and family, the exclusion restriction requiring that a commercial quality coverage signal only affects electoral outcomes through either access or especially consumption is hard to sustain.

149 noted above, the coverage maps provided by the IFE account for such obstacles and facilitators, as well as the frequency and power of transmitters signals. While the conductivity of AM signals is sensitive to variable weather conditions and the night-time ionosphere, FM radio and television coverage is relatively constant because such waves travel by line of sight. Furthermore, focusing on within-municipality neighbors removes municipal political differences and ensures that neighboring precincts vote on the same incumbent. Nevertheless, strategic sorting represents an important concern. Our estimates could be biased if certain types of voters choose to live in areas with better local media coverage or media stations strategically choose the strength or location of their emitter to exclude certain types of voters. However, such sorting is unlikely. First, if voters were migrating to guarantee high-quality signal coverage, they would likely choose a location close to the antennae rather than near the commercial quality coverage boundary where coverage remains imperfect. Second, media stations lack the technology to precisely differentiate neighboring precincts: beyond the fact that excluding voters is challenging when signals are not discontinuous, the antennae strengths that media stations purchase are highly discrete.37 To maximize the plausibility of our identiﬁcation strategy, we focus on relatively urban precincts. Speciﬁcally, our sample only contain precincts with at most an area of 10km2. Between such neighbors, strategic sorting is particularly unlikely because media stations cannot plausibly choose technologies to separate markets. Given the substantial reach of AM stations, a second key advantage of this design is that more urban areas almost exclusively differ with respect to FM radio and television coverage. Removing the most volatile media signal maximizes the accuracy of our coverage measures. Moreover, this area restriction prevents rural-urban comparisons, which could be problematic because distance from urban areas may be correlated with other politically-relevant characteristics and me-

37The power output in watts for the AM, FM and television stations in our sample are almost exclusively round thousands and divisible by 5.

150 dia signals decay at different rates in rural areas. Supporting the validity of this approach, our balance tests show that only two of our 36 precinct-level balancing variables are significantly correlated with local media (see below).38 Figure 5.7 illustrates our empirical strategy graphically. Electoral precincts 1571 and 1583 in the municipality of Villa de Tututepec de Melchor Ocampo are neighbors, but differ because only precinct 1583 is covered by a television station emitting from within the municipality. Operationally, we define a “treated” precinct as one which differs from at least one neighboring precinct in terms of the number of local media stations that it receives. For each such precinct, we then collect all possible neighboring “control” precincts that receive a different number of local media stations. Since some neighbors differ by more than one local media station,39 we include neighbor set fixed effects to ensure that our estimates are not confounded by differences in the types of places where neighbors differ by one as opposed to two (or more) local media stations. This yielded a sample containing 17,312 observations, where neighbors typically differ by one or two local media stations. To combine variation in the timing of audit and the number of media stations covering a given electoral precinct, we estimate the following triple-difference (DDD) specification using our neighboring precincts sample:

Ypkmt = β1auditmt + β2outcomemt + β3mediapm +β4 auditmt × outcome Q3mt + β5 auditmt × outcome Q4mt +β6 auditmt × mediapm + β7 auditmt × outcome Q3mt × mediapm +β8 auditmt × outcome Q4mt × mediapm + ξk + εpkmt, (4.2)

38Relative to Table 4.1, we removed municipality level variables and, naturally, local media. 3954% of neighbor pairs differ by more than one local media station. Therefore, we cannot simply compare treated and control units because the difference in the number of media stations between the two groups (or the treatment intensity) is not constant.

151 Figure 4.7: Identiﬁcation strategy example Note: Both precincts are from the municipality of Villa de Tututepec de Melchor Ocampo in the state of Oaxaca. While precinct 1583 is covered by the television emitting from within the municipality, 1571 is not.

where mediapm is the total number of local media stations. Including a neighbor set ﬁxed effect ξk ensures that our DDD estimates identify only out of within-neighbor variation in media coverage (given that we restricted attention to neighbors within a given municipality). In addition to weighting by the number of registered voters in the precinct, we further divide this weight by the number of control units per comparison to give each neighbor comparison equal weight. Using a similar design, we also examine the effect of the total number of non-local media stations, i.e. those broadcasting from outside the municipality.

The coefﬁcients β6, β7 and β8 respectively identify the electoral effect of an additional media station for a precinct in a municipality where the audit report reveals the incumbent party not to be malfeasant, to be in the third quartile of malfeasance, and to be in the fourth quartile of malfeasance. The coefﬁcients β4 and β5 estimate the effect of an audit report

152 in the absence of media. These coefficients thus capture how media coverage supports electoral accountability—both in terms of punishing malfeasant incumbent parties and rewarding incumbents that correctly allocate federal transfers. By exploiting within-neighbor variation, we can plausibly identify the average effect of an additional media station. Un- fortunately, however, the demanding structure of our identification strategy means that we lack the power to non-parametrically estimate media’s effect.40 To assess the plausibility of the design, we test whether local media predicts pre- treatment covariates using specifications akin to equation (4.2). Despite our restriction to relatively urban areas and our arguments above, a key concern is that precincts receiving more local media are likely to be closer to the municipal center, where antennas are typically located, and thus differ in politically salient ways. To examine this possibility, Table 4.2 tests for balance over distance to the centroid of the municipal head (from both a precinct’s border and its centroid), area, the total population and number of registered voters, and population density (all as natural logarithms). The results are consistent with local media’s random assignment across neighboring precincts: for none of these variables we register a statistically significant difference, while the variation in coefficients directions is consistent with chance rather than access to local media correlating with distance from relatively large urban developments. Crucially for distinguishing the effects of local and non-local media, we also find no relationship between the number of local and non-local media stations. Furthermore, we find no evidence that the number of local media stations

40The ideal non-parametric approach would be to allow media’s effect to vary for number of media stations. However, by requiring 4 coefﬁcients for each of the 40 levels in our data, we quickly lose power and rely on cells with very little support. A second best approach could involve using multiple categories (e.g. at least ten, twenty, thirty media stations etc.) or quantiles. However, since 90% of neighbors only differ by one or two local media stations, using broad categories eliminates most of this variation because most neighbor comparison will not cross the threshold deﬁning media intensity categories. Rather than taking the average marginal effect, this approach would estimate the marginal effect around a given threshold, e.g. among neighbors with 9 and 10 and 9 to 11 media stations. By ignoring within-neighbor variation away from the threshold, this simply yields an under-powered average effect for an additional media station for a somewhat arbitrary threshold.

153 covering a precinct predicts proxies for socioeconomic development such as education, basic housing needs, or owning more luxurious items like computers, washing machines or cell phones. Similarly, we find no correlation between the presence of local media and prior political behavior: an additional media station is insignificantly and negligibly associated with incumbent electoral performance, local political competition (proxied by the effective number of political parties) and electoral turnout. Together, our wide range of balance tests thus offer substantial support for our identification strategy.

4.6 Results

We first briefly examine whether audits revealing the incumbent party’s mayor to be corrupt or neglectful before an election affects the party’s vote share and probability of retaining office. However, our main contribution is to then identify the heterogeneous effect of local media and the pre-election release of incumbent performance information in holding the incumbent party to account.

4.6.1 Audits and electoral accountability

Table 4.3 presents our DD estimates of the average effect of an audit revealing a mayor to be corrupt or neglectful of the poor on the mayor’s electoral prospects. The outcome in columns (1) and (2) is the change in the incumbent party’s vote share at the precinct level, while the outcome in columns (3) and (4) is whether the incumbent party was re-elected in the municipality. The results suggest that an audit report released may have important electoral implications. Although the point estimates are relatively large, they are imprecise and never statistically signiﬁcant. Column (1) suggests that revealing the incumbent party’s mayor to be in the most corrupt quartile before the election, on average, reduces the party’s vote 154 Table 4.2: Linear balance over local media (neighbor sample)

Incumbent Effective Registered Turnout Distance to Distance to Area Population Population party vote number of voters (log) (lag) municipal municipal (log) (log) density share (lag) political head from head from (log) parties (lag) border (log) centroid (log) Local media 0.0004 0.0005 -0.0040 0.0003 0.0022 -0.0024 -0.0077 -0.0027 0.0068 (0.0004) (0.0018) (0.0056) (0.0004) (0.0069) (0.0040) (0.0068) (0.0048) (0.0071) Non-local Average Share Share Average Share Share no Share Share media children households indigenous years of illiterate schooling incomplete complete per woman with male speakers schooling primary primary head schooling schooling Local media -0.1172 -0.0021 -0.0006*** -0.0001 0.0344 -0.0002* -0.0002 0.0002 0.0005 (0.2114) (0.0023) (0.0002) (0.0002) (0.0279) (0.0001) (0.0001) (0.0001) (0.0005) 155 Share Share Share Share Share Share Average Average Share incomplete complete higher economically without state occupants occupants non-dirt secondary secondary education active health workers per per ﬂoor schooling schooling care health care dwelling room Local media 0.0008 0.0009 0.0012 0.0000 -0.0003 0.0003 -0.0028 -0.0019 0.0002 (0.0009) (0.001) (0.0014) (0.0002) (0.0005) (0.0002) (0.0032) (0.0015) (0.0003) Share Share Share Share Share Share Share Share Share toilet running drainage electricity washing fridge cell car computer at home water machine phone Local media 0.0000 0.0007 0.0005 0.0001 0.0008 0.0002 0.0006 0.0006 0.0016 (0.0003) (0.0006) (0.0003) (0.0001) (0.0006) (0.0004) (0.0007) (0.0011) (0.0013)

Notes: Each coefﬁcient is from a separate OLS regression including neighbor ﬁxed effects, are estimated using OLS, and weight by electorate size divided by the number of matches per neighbor set. Each regression contains 17,312 observations. Standard errors are clustered by municipal election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table 4.3: The effects of audits revealing malfeasance before an election (DD sample)

Change in incumbent Incumbent party party vote share re-elected (1) (2) (3) (4) Constant -0.057*** -0.066*** 0.483*** 0.621*** (0.015) (0.015) (0.098) (0.090) Audit 0.005 0.023 0.031 -0.025 (0.019) (0.021) (0.122) (0.113) Corrupt Q3 -0.008 0.068 (0.027) (0.159) Audit × Corrupt Q3 0.022 0.136 (0.037) (0.191) Corrupt Q4 0.014 0.152 (0.036) (0.153) Audit × Corrupt Q4 -0.045 -0.291 (0.045) (0.208) Not poor Q3 0.014 -0.200 (0.022) (0.156) Audit × Not poor Q3 -0.026 0.242 (0.035) (0.196) Not poor Q4 0.034 -0.117 (0.041) (0.162) Audit × Not poor Q4 -0.068 -0.171 (0.047) (0.203)

Observations 47,938 47,938 47,938 47,938 Outcome mean -0.05 -0.05 0.55 0.55 Outcome std. dev. 0.15 0.15 0.50 0.50

Notes: All speciﬁcations weight by the number of registered voters, and are estimated using OLS. Similar estimates for the neighbor sample are provided in Table C.3 in the Appendix. The omitted category for corruption and not spending on the poor is Q1 and Q2. Standard errors are clustered by municipal election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

156 share by 4.5 percentage points. Although the effect is not statistically significant, the magnitude represents 9% of the average incumbent’s initial vote share. Column (2) suggests that revealing the incumbent party’s mayor to have been neglectful similarly reduces their vote share: the vote share of mayors in the third quartile declines by 2.6 percentage points, while mayors in the fourth quartile lose a further 4.2 percentage points. In both cases, the audit coefficient in the first row—which captures the baseline category of essentially zero or negligible malfeasance—suggests that the parties of mayors whose reputations are not negatively affected by the audits may increase slightly their vote share, especially when they actually spent the FISM funds on its intended poor recipients. Looking at the probability of re-election similarly suggests that voters may severely punish mayoral malfeasance. Measured at the municipal level, the change in incumbent vote share maps to large but very imprecise reductions in the probability of being re-elected. Column (3) finds that revealing a mayor as one of the most corrupt reduces their re-election probability by 29 percentage points, although this is not quite statistically significant.41 Column (4) shows that the publication of an audit report showing that a mayor did not spend FISM federal transfers on the poor is 17 percentage points less likely to be re-elected. For both audit outcomes, the effect is much larger for mayors in the fourth quartile relative to the third, where the estimate is surprisingly positive. The electoral sanctioning suggested by these results is broadly similar in magnitude to that found by Ferraz and Finan(2008) in Brazil, although we measure corruption in terms of stolen funds rather than the number of corrupt spending violations. In addition, our results suggest that incorrectly spending money earmarked for the poor could also evoke sanctioning of similar magnitude to corruption.42 However, our DD estimates are

41The particular relative lack of precision for the incumbent win probability reflects the fact that we have 481 audited municipalities, of which only 51 had mayors that were revealed to be corrupt before the election. 42Table C.3 in the Appendix reports quantitatively similar, but far noisier, estimates for the neighbors sample, which includes only a selection of precincts from 181 municipality elections and only identifies out 157 very noisy. A plausible explanation for the lack of precision is substantial heterogeneity reflecting the possibility that audit reports only affect voter behavior when the information is effectively conveyed by local media stations.

4.6.2 Broadcast media audit report coverage

We now address the central question of this paper: is the party of a malfeasant mayor more likely to be sanctioned by voters who live in areas covered by media stations that publicize audit reports revealing the mayor’s behavior in ofﬁce? Combining our DD and within-neighbor sources of variation, we ﬁrst estimate the effects of local media station— those emitting from the municipality that the electoral precinct belongs to—before turning to non-local media stations. Since mayoral corruption and neglect are primarily important local issues, we expect that local media is more effective at facilitating electoral accountability.

Importance of local media

Table 4.4 provides estimates of the sanctioning effect of an additional media station emitting from the precinct’s own municipality. Since we now focus on precinct-level variation in media coverage, our analysis focuses only on the change in the incumbent’s vote share at the precinct level. Columns (1) and (2) provide our preferred estimates, while columns (3) and (4) show that the results are robust to controlling for the presence of non- local media stations. The results indicate that local media play a key role in supporting electoral accountability. Voters only punish the incumbent party when an audit report released before an election reveals the incumbent mayor to be corrupt in the presence of sufﬁcient local me- of within-neighbor variation.

158 Table 4.4: Effects of local media and audits reports revealing malfeasance before an election (neighbor sample)

Change in incumbent party vote share (1) (2) (3) (4) Audit -0.029 0.043 -0.041 0.044 (0.054) (0.050) (0.069) (0.080) Audit × Local media 0.007*** 0.006* 0.008*** 0.005 (0.002) (0.003) (0.002) (0.003) Audit × Corrupt Q3 0.146* 0.072 (0.074) (0.085) Audit × Corrupt Q3 × Local media -0.007* -0.006** (0.004) (0.003) Audit × Corrupt Q4 0.104 0.199** (0.074) (0.095) Audit × Corrupt Q4 × Local media -0.007* -0.008** (0.004) (0.004) Audit × Not poor Q3 0.123 0.031 (0.098) (0.139) Audit × Not poor Q3 × Local media -0.010* -0.008 (0.006) (0.006) Audit × Not poor Q4 -0.042 0.020 (0.076) (0.101) Audit × Not poor Q4 × Local media -0.011*** -0.012*** (0.004) (0.004) Audit × Non-local media 0.000 0.001 (0.003) (0.001) Audit × Corrupt Q3 × Non-local media 0.002 (0.003) Audit × Corrupt Q4 × Non-local media -0.003 (0.004) Audit × Not poor Q3 × Non-local media 0.002 (0.004) Audit × Not poor Q4 × Non-local media -0.002 (0.002)

Observations 17,312 17,312 17,312 17,312 Outcome mean -0.04 -0.04 -0.04 -0.04 Outcome std. dev. 0.13 0.13 0.13 0.13 Local mean 20.60 20.60 20.60 20.60 Local std. dev. 10.57 10.57 10.57 10.57

Notes: All speciﬁcations include neighbor ﬁxed effects, are estimated using OLS, and weight by electorate size divided by the number of matches per neighbor set. The omitted category for corruption and not spending on the poor is Q1 and Q2. Media, corruption and non poor spending lower order terms are omitted. Standard errors are clustered by municipal election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

159 dia stations. Column (1) shows that, on average, mayors in the third and fourth quartiles of the corruption distribution experience a significant loss in their vote share—almost 1 percentage point for each additional local media station.43 A standard deviation increase in the number of media stations, which entails 10.6 more media stations, thus reduces the vote share of an incumbent revealed to be corrupt by around 7.5 percentage points. This represents half of the average municipal victory margin (in the sample) of 15 percentage points. However, the insignificant positive interactions between our audit dummy and a mayor’s corruption quartile indicate that revealing a mayor to be corrupt is not punished electorally in precincts covered by no (or few) local media stations. On the other hand, clean incumbent parties are only meaningfully rewarded for their good performance in the presence of local media stations. The statistically significant coefficient on the interaction between being audited before an election and the number of local media stations shows that an additional local media station publicizing that the incumbent mayor did not engage in unauthorized spending increases the incumbent party’s vote share by 0.7 percentage points. In precincts covered by more local media, we observe slightly greater sanctioning of revelations that funds were not spent on projects benefiting the poor. Moreover, such punishment is increasing in the severity of a mayor’s neglect. Column (2) shows that an additional local media station reduces a neglectful mayor’s vote share by 0.8 percentage points for mayors in the third quartile and 1.2 percentage points for mayors in the most neglectful quartile. A standard deviation increase in the number of local media stations thus entails an 13 percentage point decrease in the vote share of the most neglectful mayors if their behavior is revealed before an election. This represents a 25% reduction in their precinct-level vote share. Again, the insignificant interaction between the pre-election audit release and

43By contrast, there is no signiﬁcant effect associated with the (unreported) interactions between local media and audit report outcomes released after the election, for either corruption or not spending on the poor. These coefﬁcients are omitted to save space.

160 not spending on the poor shows that in locations with zero local media stations the party of the mayor is not meaningfully electorally sanctioned. Furthermore, the significant positive interaction between revealing an audit and the total number of local media stations shows that parties that were revealed to have correctly spent all their FISM funds on projects benefiting the poor are boosted at polls, although not as much as malfeasant mayors are punished. It is possible that voters are more likely to believe that their municipal representatives are corrupt than neglectful, and thus positive information that contradicts this prior is rewarded more. It is important to remember that the average proportion of funds not spent on the poor in the third and fourth quartiles are slightly higher than for corruption. Nevertheless, the point estimates suggest that the electoral impact of an additional local media station reporting a fixed proportion of misallocated funds is, on average, around 25% higher for not spending on the poor than corruption. The findings thus suggest that local-media revelations about misallocated spending are at least as as important to voters as revelations about unauthorized spending. Which types of media platform drive such electoral accountability? Table 4.5 reports the main results for FM radio and television when we only identify out of neighbors that respectively differ in the number of such media stations that they receive.44 Consistent with television’s status as the main source of news consumption, the effect of local media appears to be driven primarily by additional local television stations. For both corruption and neglect of the poor, television has very large effects on incumbent electoral performance: while each additional local television reduces a malfeasant incumbent party’s vote share by 3-5 percentage points, a high-performing incumbent party benefits by a similar margin. Although television stations are far less prevalent than radio stations, a standard devia-

44Given the extensive reach of AM radio stations and our restriction to urban areas, the AM sample cannot provide informative estimates and is thus omitted. The available sample size drops to 1,038.

161 Table 4.5: Effects of local FM radio and television stations and audits reports revealing malfeasance before an election (neighbor sample)

Change in incumbent party vote share (1) (2) (3) (4) Audit × Local FM media 0.005 -0.003 (0.004) (0.004) Audit × Corrupt Q3 × Local FM media 0.003 (0.006) Audit × Corrupt Q4 × Local FM media -0.009 (0.007) Audit × Not poor Q3 × Local FM media 0.013** (0.007) Audit × Not poor Q4 × Local FM media -0.001 (0.005) Audit × Local TV media 0.037*** 0.037*** (0.005) (0.008) Audit × Corrupt Q3 × Local TV media -0.031*** (0.008) Audit × Corrupt Q4 × Local TV media -0.047*** (0.011) Audit × Not poor Q3 × Local TV media -0.037** (0.015) Audit × Not poor Q4 × Local TV media -0.042** (0.018)

Observations 9,194 9,194 9,686 9,686 Outcome mean -0.03 -0.03 -0.05 -0.05 Outcome std. dev. 0.12 0.12 0.14 0.14 FM/television Media mean 11.39 11.39 4.09 4.09 FM/television Media std. dev. 5.22 5.22 2.37 2.37

Notes: Both the FM radio and television neighbor samples were computed in the same way as for local media (see text for details). All speciﬁcations include neighbor ﬁxed effects, are estimated using OLS, and weight by electorate size divided by the number of matches per neighbor set. The omitted category for corruption and not spending on the poor is Q1 and Q2. Media, corruption and non poor spending lower order terms are omitted. Standard errors are clustered by municipal election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

162 tion increase in local television stations nevertheless translates into reducing the incumbent party’s vote share by almost 25% in the most egregious cases. With a similar sample size, there is no clear sanctioning effect associated with FM radio.

No role for non-local media

We now consider the role of non-local media. Columns (3) and (4) of Table 4.4 add interactions with non-local media to our main estimates for the interaction between audit reports released before an election and local media. First, and unsurprisingly given that local media is well balanced across the number of non-local stations, we find that the local media point estimates are unaffected. If anything, the estimates slightly increase and are more precisely estimated. Second, the non-local media interactions show that non-local media does not significantly influence incumbent party vote share. This suggests that local media is the key driver of electoral accountability, and thus any media coverage is not sufficient to punish the parties of malfeasant mayors. However, because our design does not isolate exogenous variation in non-local media, this finding does not prove that non- local media serves to accountability function. To better test whether non-local media also facilitates electoral sanctioning, we again implement the same identification strategy to isolate plausibly exogenous variation in non- local media coverage. The results are presented in Table 4.6, and due to the larger sample size offer a more powerful test than for non-local media. However, the triple interaction between revealing an audit before the election, the audit’s outcome and non-local media provides no evidence that the number of media stations broadcasting from outside the municipality affect the incumbent party’s vote share. Comparing these estimates to those for local media in Table 4.4, the effect of such non-local media is indistinguishable from zero

163 Table 4.6: Effects of non-local media and audits reports revealing malfeasance before an election (non-local media neighbor sample)

Change in incumbent party vote share (1) (2) Audit -0.111 -0.011 (0.070) (0.051) Audit × Non-local media 0.004* 0.000 (0.002) (0.001) Audit × Corrupt Q3 0.083 (0.086) Audit × Corrupt Q3 × Non-local media -0.003 (0.003) Audit × Corrupt Q4 -0.010 (0.103) Audit × Corrupt Q4 × Non-local media -0.003 (0.003) Audit × Not poor Q3 0.021 (0.084) Audit × Not poor Q3 × Non-local media 0.003 (0.002) Audit × Not poor Q4 -0.153 (0.111) Audit × Not poor Q4 × Non-local media 0.003 (0.002)

Observations 40,007 40,007 Outcome mean -0.06 -0.06 Outcome std. dev. 0.14 0.14 Non-local media mean 32.97 32.97 Non-local media std. dev. 22.73 22.73

Notes: This neighbor sample was computed in the same way as for local media (see text for details). All speciﬁcations include neighbor ﬁxed effects, are estimated using OLS, and weight by electorate size divided by the number of matches per neighbor set. The omitted category for corruption and not spending on the poor is Q1 and Q2. Media, corruption and non poor spending lower order terms are omitted. Standard errors are clustered by municipal election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

164 and always smaller than for local media.45 This evidence reinforces our claim that the presence of local media is essential for electoral accountability.

Robustness checks

We now demonstrate that our main findings for local media are robust to a variety of robustness tests. Table 4.7 presents the results of these specification and variable definition checks, focusing on the triple interactions identifying the effects of local media revealing mayoral malfeasance. First, although the number of local media is well balanced across pre-treatment variables, we nevertheless ensure that our results are not driven by the most plausible potential confounds. Columns (3) and (4) of Table 4.4 showed that our results are robust to including an interaction with the number of non-local media. As a more general test, columns (1) and (2) of panel A in Table 4.7 simultaneously include the interaction of audit results with second-order polynomials for four important background indicators: non-local media, distance to the municipal head (from the precinct centroid), average years of education, and the proportion of households with a car. For both corruption and neglectful incumbents, if anything we find larger and precisely estimated effects. A second potential concern is that our results are driven by restricting attention to relatively small neighboring precincts. Accordingly, columns (3)-(6) in panel A of Table 4.7 show similar, if not larger, estimates when our 10 km2 restriction is relaxed to 25km2 and 50km2.46 Third, to address the possibility that media and election dynamics in municipalities receiving audit reports after the election are different, we drop the post-election comparison. We thus focus only on precincts in municipalities where audit reports were released be-

45We also ﬁnd the same results when using the neighbors sample used to estimate the effect of local media. 46Similar results are obtained when using all possible neighbors. 165 Table 4.7: Robustness checks (neighbor sample)

Panel A: Speciﬁcation tests Change in incumbent party vote share (1) (2) (3) (4) (5) (6) (7) (8) Audit × Corrupt Q3 × Local media -0.006* -0.009** -0.009** -0.004*** (0.003) (0.004) (0.004) (0.001) Audit × Corrupt Q4 × Local media -0.017*** -0.007 -0.008** -0.008*** (0.003) (0.004) (0.004) (0.001) Audit × Not poor Q3 × Local media -0.011* -0.010** -0.010** -0.008** (0.006) (0.005) (0.004) (0.003) Audit × Not poor Q4 × Local media -0.013*** -0.015*** -0.017*** -0.008** (0.004) (0.005) (0.005) (0.003)

Area restriction <10km2 <10km2 <25km2 <25km2 <50km2 <50km2 <10km2 <10km2 Quadratic interactive controls XX Pre-election audits only XX Observations 17,312 17,312 22,489 22,489 28,147 28,147 11,285 11,285 Panel B: Malfeasance deﬁnitions Change in incumbent party vote share (1) (2) (3) (4) (5) (6) (7) (8)

166 Audit × Local media × Corruption -0.022 (0.023) Audit × Local media × Not poor -0.039** (0.017) Audit × Local media × Corrupt ≥ 5% 0.001 (0.006) Audit × Local media × Not poor ≥ 5% -0.009** (0.004) Audit × Local media × Corrupt ≥ 10% -0.006* (0.003) Audit × Local media × Not poor ≥ 10% -0.007** (0.003) Audit × Local media × Corrupt ≥ 20% -0.028** (0.013) Audit × Local media × Not poor ≥ 20% -0.016*** (0.004)

Observations 17,312 17,312 17,312 17,312 17,312 17,312 17,312 17,312

Notes: All specifications include neighbor fixed effects, are estimated using OLS, and weight by electorate size divided by the number of matches per neighbor set. All other interactions are omitted. In Panel A, the controls are the number of non-local media stations, (log) distance to the municipal head from the precinct’s centroid, average years of schooling, and the share that own a car. Standard errors are clustered by municipal election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. fore an election, and identify entirely off variation in local media before the election. The results in columns (7) and (8) of panel A show that we obtain similar results. Although we sacrifice the DD design, this approach increases the precision of our estimates despite dropping 40% of the sample. Fourth, we consider alternative operationalizations of malfeasance. Columns (1) and (2) of panel B in Table 4.7 first report linear measures of corruption and not spending on the poor. In both cases the effects are negative and considerable in magnitude: for each additional 10 percentage points of the FISM budget not spent correctly, a standard deviation increase in local media stations reduces the incumbent vote share by 2.2 percentage points in the case of corruption and 3.9 percentages points in the case of not spending on the poor. However, the estimate for corruption is not statistically significant. Together with our previous findings, this suggests that corruption’s effects are non-linear. To further examine when the impact of malfeasance kicks in, columns (3)-(8) sequentially operationalize malfeasant behavior using an indicator for corruption or funds not spent on the poor exceeding 5, 10 and 20% of FISM funds. At each level, not spending on the poor is punished, although the magnitude of punishment is greatest for the 20% cutoff. For corruption, the evidence suggests that only the highest levels of corruption—greater than 20%—are severely punished. This is consistent with voters possessing more negative priors about incumbent corruption. Fifth, at the cost of losing randomization in local media, we estimate equation (4.1) in the larger DD sample to check the external validity of the neighbor sample estimates. The results, provided in Table C.4 in the Appendix, are broadly similar to our neighbor sample estimates. For both corruption and spending not on the poor, a mayor revealed to be in the most malfeasant quartile experiences a significant decrease in their vote share for each additional local media station. However, malfeasance in the third quartile is not punished. Although the samples differ in terms of both composition and quality of identification, the similarity of the results is encouraging. 167 4.6.3 Media market structure

Our main ﬁnding is that only local media stations facilitate electoral accountability. To better understand when media support this important social function, we further investigate the media market structure. In particular, we focus on two important features: crowd-out by non-local media and the geographic composition of a station’s potential audience.

Non-local media crowd out local media

Since non-local media are less likely to cover the relevant political actors, and the preceding analysis demonstrated that non-local media stations do not affect incumbent electoral performance, an additional non-local media station could crowd-out the effects of local media. This possibility rests upon the likelihood that a new media station attracts listeners or viewers away from local media stations, plausibly because voters consume media for reasons other than acquiring politically-relevant information and are unable to increase consumption of all stations commensurately. To examine this possibility empirically, we interact the number of local media stations with the number of non-local stations in Table 4.8. The results are consistent with crowd- out: for both and low performance incumbent mayors, coverage by an additional non-local media station weakens the impact of a local media station. Particularly for not spending on the poor, the signiﬁcant quadruple interactions show that each additional non-local media station reduces the effect of an additional local media station by 0.1-0.16 percentage points. In the average precinct, covered by nearly 15 non-local media stations, the impact of an additional local media station thus falls by around 2 percentage points. These positive interactions are not statistically signiﬁcant in the case of corruption. This evidence thus suggests that the presence of uninformative non-local media stations drowns out the electoral accountability facilitated by local media, especially when it comes to investments

168 Table 4.8: Effects of local media and audits reports revealing malfeasance before an election, conditional on the number of non-local media stations (neighbor sample)

Change in incumbent party vote share (1) (2) Audit × Local media 0.014*** 0.020*** (0.005) (0.005) Audit × Non-local media × Local media -0.0003 -0.0005*** (0.0002) (0.0001) Audit × Corrupt Q3 × Local media -0.011 (0.008) Audit × Corrupt Q3 × Non-local media × Local media 0.0002 (0.0002) Audit × Corrupt Q4 × Local media -0.015* (0.009) Audit × Corrupt Q4 × Non-local media × Local media 0.0004 (0.0003) Audit × Not poor Q3 × Local media -0.031*** (0.008) Audit × Not poor Q3 × Non-local media × Local media 0.0016*** (0.0004) Audit × Not poor Q4 × Local media -0.039*** (0.008) Audit × Not poor Q4 × Non-local media × Local media 0.0010*** (0.0002)

Observations 17,312 17,312 Non-local media mean 14.85 14.85 Non-local media std. dev. 16.33 16.33

169 not beneﬁting the poor. Implicitly, the fact that an additional non-local media station reduces the effect of local media suggests that voters indeed use local media to follow the news—otherwise, an additional media station should not affect voter behavior.47

Media markets determine news content

To further explore the mechanism driving local media’s effect, we consider the composition of the media market. As Snyder and Strömberg(2010) show, the congruence of political boundaries and media markets plays a key role in the content that broadcasters choose to provide. In our context, local media stations that predominantly cover consumers within their municipality have the strongest incentives to cover audit reports relating to their municipality. However, despite emitting from a given municipality, some media stations may primarily serve audiences in other municipalities, and adjust their news coverage accordingly. To test this argument, we employ a similar approach to the congruence measure in- troduced by Snyder and Strömberg(2010). Speciﬁcally, for each local media station, we calculate the proportion of the population that receives a commercial quality signal that reside inside the municipality of the emitter. We then computed the average such share across all AM radio, FM radio and television stations covering a given electoral precinct. In our sample, 63% of the average local media station’s audience is from within the municipality. Higher values imply that a given media station has a strong incentive to cover audit report outcomes in depth. As with our crowd-out analysis, we interact this measure of the local market with local media to examine how the accountability-enhancing effects of local media depend upon the media market.

47This could arise because an additional non-local media station causes pre-existing ﬁrms to alter their content, or because voters shift toward the new non-local station. Without detailed media consumption data, we cannot differentiate these explanations.

170 Table 4.9: Effects of local media and audits reports revealing malfeasance before an election, conditional on local market media coverage share (neighbor sample)

Change in incumbent party vote share (1) (2) Audit × Local media 0.002 -0.006** (0.005) (0.003) Audit × Corrupt Q3 × Local media -0.005 (0.006) Audit × Local media × Local market 0.017** 0.027*** (0.008) (0.009) Audit × Corrupt Q3 × Local media × Local market -0.014 (0.015) Audit × Corrupt Q4 × Local media -0.000 (0.006) Audit × Corrupt Q4 × Local media × Local market -0.010 (0.013) Audit × Not poor Q3 × Local media 0.007 (0.007) Audit × Not poor Q3 × Local media × Local market -0.029** (0.012) Audit × Not poor Q4 × Local media 0.001 (0.004) Audit × Not poor Q4 × Local media × Local market -0.037*** (0.013)

Observations 17,312 17,312 Local market mean 0.63 0.63 Local market std. dev. 0.28 0.28

171 The results in Table 4.9 indicate that a media’s station’s market has important effects on electoral accountability. Consistent with news coverage depending upon the extent to which a station’s market is local, we find that a larger local market share increases the reward and punishment of incumbent parties. This is particularly true in the case of not spending on the poor, where the large and statistically significant negative quadruple interaction implies that each additional media station with an entirely local market share would reduce the incumbent party’s vote share by 2.9 percentage points for mayors in the third quartile and 3.7 percentage points for mayors in the fourth quartile. Similarly, mayors that correctly spend FISM funds on projects benefiting the poor gain 2.7 percentage points for an additional media station exclusively serving their market. Although the analogous coefficients for corruption are not statistically significant, they paint a similar picture where a mayor from the third quartile is punished by 1.9 percentage points and a mayor from the fourth quartile is punished by 1 percentage point. The party of a non-corrupt mayor receives a significant boost of 1.7 percentage points for each additional local media station serving only the local municipal market.

4.7 Conclusion

Many scholars call media “the fourth estate,” due to its potential to inform voters about the behavior of politicians in office. Both national and local media are needed: while national media outlets cover national level actors, local media are necessary to inform voters about the performance of local politicians. While the influence of national media has received considerable attention, this paper demonstrates the importance of local media for local electoral accountability in a federal setting where local governments play an increasingly important role in service delivery. Furthermore, we show that effective electoral accountability requires more than just a single local media outlet: we find that additional

172 local media stations have large effects, even in relatively congested media markets. Using detailed local data, and an identification strategy that exploits both the timing of the release of audit reports with respect to elections and differences in signal coverage across neighboring electoral precincts, we identify the impact of the media environment on electoral accountability. We show that voters punish the party of malfeasant mayors, but only in electoral precincts covered by local radio or television stations. In particular, we find that each additional local media station reduces the vote share of an incumbent political party revealed to be corrupt by nearly 0.7 percentage points. Local media similarly reduces the vote share of an incumbent political party revealed to have diverted funds away from the poor by around 1.2 percentage points. However, we find no effect of non-local media stations that are based in other municipalities. Delving further into the structure of media markets, we show that non-local media crowd out the effects of local media—plausibly by attracting away consumers of local news—and that even local media may not as effectively support electoral accountability when their audience contains many consumers from outside the municipality. Our findings thus illustrate the importance of media, especially local media, in supporting local electoral accountability by facilitating the sanctioning of malfeasant behavior in office.

173 5| Political information cycles: When do voters sanction incumbent parties for high homicide rates?1

5.1 Introduction

Electoral accountability rests upon informed voters electing competent representatives. Given the difﬁculty of directly observing candidate competence, performance on salient issues like the economy, public security or corruption represent key signals of an incumbent party’s continuing suitability for ofﬁce (Fearon 1999; Ferejohn 1986). However, particularly in developing contexts, voters are often poorly informed about politics and current affairs (e.g. Keefer 2007; Pande 2011), or do not receive relevant information pertaining to their incumbent party (Snyder and Strömberg 2010).2 This lack of politically-relevant information may explain why voters often do not hold incumbents to account on available

1I thank Jim Alt, Eric Arias, Abhijit Banerjee, Allyson Benton, David Broockman, Bruce Bueno de Mesquita, Melani Cammett, Francisco Cantú, Ana De La O, Jorge Domínguez, Oeindrila Dube, Raissa Fabregas, Nilesh Fernando, Hernán 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, Otis Reed, James Robinson, Rob Schub, Gilles Serra, Jim Snyder, Chiara Superti, Edoardo Teso, Julie Weaver, and workshop or conference participants at presentations at APSA, Caltech, Columbia, Harvard, LSE, MIT, MPSA, NYU, NEWEPS, Stanford GSB and Yale for illuminating discussions and useful comments. I thank Horacio Larreguy for kindly sharing municipal election data. 2Ill-informed electorates pose similar dilemmas in established democracies (Bartels 2008; Delli Carpini and Keeter 1996), although the risks of electing incompetent politicians are greater with weak democratic institutions.

174 performance metrics, or struggle to assign responsibility across layers of government.3 While many scholars have argued that access to media facilitates electoral accountability (see Ashworth 2012; Pande 2011),4 the availability of media coverage does not imply that voters actually consume politically-relevant news. I instead argue that how voters hold incumbent parties accountable for their performance in office reflects the timing of their news consumption. In particular, if voters primarily consume news before elections—when they are more likely to actively seek out politically-relevant information (Hamilton 2004; Marshall 2016b), and media outlets are more likely to supply such information (Prior 2007)—electoral sanctioning will reflect salient indicators of incumbent performance reported in the media at this time. Conse- quently, incumbent performance indicators only impact election outcomes if they are covered in the media when voters consume news before elections. I expect the effect of such political information cycles on vote choice to be most pronounced among poorly-informed voters, whose weak prior beliefs over the incumbent’s continuing suitability for office are most responsive to available performance indicators (Zaller 1992). This paper provides evidence for the central elements of this theory in the context of examining how Mexican voters hold different levels of government to account for local violent crime. Following significant democratizing and decentralizing reforms in the 1990s,

3Evidence that voters, especially in consolidating democracies, punish or reward politicians for economic performance (Remmer 1991; Roberts and Wibbels 1999; Singer and Carlin 2013), levels of public security (Kronick 2014; Vivanco et al. 2015), or malfeasance in ofﬁce (Arias et al. 2016; Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013; Ferraz and Finan 2008; Larreguy, Marshall and Snyder 2015) is mixed. The developed country literature similarly highlights mixed or conditional effects (see Anderson 2007; Golden 2010; Healy and Malhotra 2013). 4Well-identiﬁed studies show that news or advertising content affect voting behavior (e.g. Enikolopov, Petrova and Zhuravskaya 2011; DellaVigna and Kaplan 2007; Gerber et al. 2011; Larreguy, Marshall and Snyder 2016; Snyder and Strömberg 2010), while others emphasize that such information must have a credible source (Alt, Lassen and Marshall 2016; Chiang and Knight 2011; Lenz and Ladd 2009) and may require a minimum level of sophistication to process complex information (Alt, Lassen and Marshall 2016; Gomez and Wilson 2006). Experimental studies primarily focus on providing information just before elections, and rarely mimic typical newscasts.

175 Mexican elections have become relatively competitive at the national and sub-national levels and responsibility for administering key public services such as policing is shared across municipal, state and federal governments. Public security is a major concern among Mexican voters, as in many other Latin American nations, and thus local homicides represent a salient incumbent performance metric. However, since voters are largely poorly informed about politics and a significant portion of the electorate lacks strong partisan ties (Greene 2011; Lawson and McCann 2005; McCann and Lawson 2003), there is scope for pre-election crime reports to influence voting behavior. Although incumbents cannot seek re-election, voters hold Mexico’s strong parties responsible for incumbent performance in office (Chong et al. 2015; Larreguy, Marshall and Snyder 2015; Larreguy, Marshall and Trucco 2015). Leveraging plausibly exogenous variation in temporal proximity to local elections, municipal-level homicide shocks just before elections, and access to local broadcast media stations, I test the political information cycles argument at the individual and electoral precinct levels. First, to identify the effect of upcoming local elections on news consumption and voter beliefs, I exploit the irregular timing of survey waves and staggered electoral cycles across Mexican states to isolate variation in the proximity of municipal and state legislative elections (see Eifert, Miguel and Posner 2010). Rather than separate supply and demand explanations for information consumption, I focus on consumption cycles in equilibrium.5 I first show that voters indeed consume more political news through television and radio just before local elections, and also demonstrate greater political knowledge. While better informed voters increase the intensity of their news consumption, less informed voters effectively consume political news for the first time over the electoral cycle. Suggesting that voters internalize the news in the media at the time, I find that homicides

5Marshall(2016 b) seeks to tease apart explanations for increased consumption.

176 that occur just before elections both increase concern about public security and reduce confidence in the municipal incumbent by around 10 percentage points. Such responses to short-term performance indicators are concentrated among less educated voters, who consume less news outside election campaigns. This implies that only voters with weaker prior beliefs about the incumbent party’s performance on violent crime update from pre- election performance indicators. In contrast with the “recency bias” literature (see Healy and Lenz 2014), homicide shocks that do not occur before elections do not affect voter beliefs. Second, using precinct-level electoral returns between 1999 and 2013, I show that pre- election homicide shocks substantially harm the municipal incumbent party’s electoral performance. To identify the effects of homicide spikes before elections—when voters consume most news—I exploit idiosyncratic month-to-month volatility in municipal homicide counts. Specifically, I compare “shocked” municipal elections that experienced more homicides in the two months preceding the election to “control” elections from within the same municipality that experienced more homicides in the two months after the election.6 Such homicide shocks are consistent with random sample variability and well balanced over 104 covariates, while I find no evidence that drug trafficking organizations (DTOs) or politicians manipulate homicide rates around elections. Despite being uncorrelated with broader homicide levels and trends, and thus representing weak indicators of long-term performance, pre-election homicide shocks reduce the incumbent party’s vote share by 2.2 percentage points and its probability of winning by 11 percentage points. Consistent with poorly-informed voters relying on the pre-election signals available when they actually consume news, voters do not hold incumbents accountable for longer- run homicide rates. Using a difference-in-differences design, I find no evidence that homi-

6I include only municipalities experiencing at least one homicide over the two months before and after an election.

177 cide rates over the prior year or electoral cycle—arguably more desirable indicators of incumbent performance (Healy and Lenz 2014), but which may not be observed by voters predominantly consuming information just before elections—affect incumbent electoral performance. Furthermore, my analysis of concurrent state and federal elections, and comparisons between municipalities with and without their own police force, indicates that the parties of incumbent mayors, rather than higher levels of government, are held to account for local homicide shocks (see also Vivanco et al. 2015). Further illustrating voters’ capacity to learn, sanctioning of homicide shocks is greatest where a different incumbent party at the previous election did not oversee a pre-election homicide shock or the same incumbent party experiences consecutive shocks. Finally, I link the individual and aggregate-level findings by showing that local broadcast media play a crucial role in the transmission of these effects. Local broadcasters report politically-relevant news that voters may not be able to access otherwise. Using detailed media coverage maps for every radio and television station in the country, I exploit within- neighboring precinct variation in coverage to identify the effects of broadcast media. I find that, on average, each additional local media station based within a precinct’s own municipality reduces the incumbent’s vote share by 0.23 percentage points, and is driven by local television. However, the 21% of precincts covered by least local media stations do not significantly punish homicide shocks, while an additional non-local station does not affect voter sanctioning. Moreover, consistent with the survey results showing that news around elections disproportionately affects voters with weak priors, the effects of local media are largest in precincts containing voters with lower levels of higher education. My finding that political information cycles induce voters to heavily weight recent performance indicators in ways that can substantially alter electoral outcomes makes several theoretical and empirical contributions. First, it shows that, in addition to voters requiring

178 access to performance indicators,7 such indicators are only likely to meaningfully alter voting behavior when their coverage in the media coincides with voters actually consuming news. Political information cycles may thus explain the mixed evidence of electoral accountability across different performance metrics and release dates (e.g. Achen and Bartels 2004b; Brollo 2009; Chang, Golden and Hill 2010; Roberts and Wibbels 1999). Second, in contrast with theories resting on myopic voters (e.g. Downs 1957; Nord- haus 1975) or short memories (e.g. Zaller 1992), my argument implies that the common ﬁnding that voters respond positively to economic performance just before elections could instead reﬂect a poorly-informed electorate updating from the only information they possess at the time. In this respect, my results micro-found and extend Healy and Lenz’s (2014) conclusions from U.S. survey experiments across many elections in a major developing democracy. Furthermore, the sharp political information cycles observed in Mexico depict a more extreme “recency bias” than observed in the U.S. (Achen and Bartels 2004b; Healy and Lenz 2014), and may thus explain the relatively high levels of electoral volatility documented across Latin America (Roberts and Wibbels 1999) in comparison with developed democracies (Dalton and Wattenberg 2000). However, the sophisticated updating and differential punishment across levels of government exhibited by Mexican voters suggests that such behavior cannot be accounted for by voters blindly responding to events beyond the incumbent’s control (see Achen and Bartels 2004a; Healy, Malhotra and Mo 2010). Third, I contribute evidence that voters attempt to hold governments accountable for public security. Alongside the economy, public security represents the other key valence issue—an issue on which all voters broadly agree on what good performance constitutes— in many developing contexts, and particularly in Latin America.8 By illustrating how the

7See e.g. Banerjee et al.(2011), Casey(2015), Chang, Golden and Hill(2010), Ferraz and Finan(2008), Healy and Lenz(2014), Larreguy, Marshall and Snyder(2015), and Snyder and Strömberg(2010). 8While existing studies have suggested that high levels of violence can affect incumbent electoral performance (e.g. Berrebi and Klor 2006; Cummins 2009; Kronick 2014; Ley 2014), incumbent evaluations 179 timing of news consumption affects sanctioning, my results help explain why Mexican scholars have struggled to verify the widely-held belief that voters respond to local crime (e.g. Vivanco et al. 2015). Fourth, in a complex federal system where shared control of policing reduces “clarity of responsibility” (Anderson 2006; Powell and Whitten 1993), I show that Mexican voters are able to assign responsibility across layers of government, at least in their own minds, for local homicides. In contrast with previous studies suggesting that voters hold national- level politicians accountable for major policy outcomes (e.g. Rodden and Wibbels 2011), I ﬁnd that voters perceive municipal rather than state or federal incumbents to be responsible for local violent crime. The paper proceeds as follows. Section 5.2 analyzes how political information cycles affect electoral accountability. Section 5.3 describes the Mexican context used to test my argument. The empirical analysis then identiﬁes the three central components of my theory: sections 5.4-5.6 respectively examine information consumption and voter beliefs, incumbent electoral peformance, and the moderating role of media coverage. Section 5.7 discusses the normative and policy implications.

5.2 Theoretical argument

Electoral accountability is underpinned by retrospective (Ferejohn 1986) or prospective voting (Fearon 1999) models premised on the assumption that voters re-elect incumbent parties based on relevant publicly available information. In light of the mixed evidence that governments are held accountable for their performance in ofﬁce (see e.g. Anderson 2007;

(García-Ponce, Wantchekon and Zeitzoff 2014), or increase informed participation among victims (e.g. Bate- son 2012; Bellows and Miguel 2009), this study explains why responses to violence primarily relate to recent shocks, rather than standard longer-term measures. Furthermore, by isolating plausibly exogenous variation in homicide shocks, the timing of news consumption, and access to local media, the results advance a literature that has predominantly correlated indicators of violence with electoral outcomes. 180 Ashworth 2012; Healy and Malhotra 2013), scholars have pointed to the importance of voter access to politically-relevant information in the media (e.g. Chang, Golden and Hill 2010; Larreguy, Marshall and Snyder 2015; Snyder and Strömberg 2010). However, such studies assume that voters actually consume the available information, and by focusing on performance metrics released before elections fail to capture how the timing of politically- relevant news may differentially impact voters. I instead propose that the impact of news on voting behavior depends upon the timing of voter information consumption. In particular, I argue that news consumption follows a cycle where poorly-informed voters only seriously engage with politically-relevant news just before elections, and thus heavily weight the incumbent performance indicators in the news at this time.

5.2.1 Political information cycles

There are various reasons to believe that consumption of politically-relevant news increases before elections. First, to the extent that politics is a pure consumption good (e.g. Hamilton 2004), electoral campaigns are likely to particularly interest voters. While coverage of political events out of election season is often more focused on specific policy issues, campaigns are specifically designed to appeal to voters. Second, Marshall(2016 b) shows that voters in social networks that collectively value knowledge about politics may strategically acquire more information to cultivate a reputation as politically sophisticated, especially around elections when political discussion is more common. Third, voters may feel a civic duty to become informed around elections (Blais 2000; Feddersen and Sandroni 2006b). Fourth, even voters without intrinsic motivations may consume more information around elections simply because more air time is devoted to political news, advertising, and specific election programming. Even for uninterested voters, such information becomes

181 difficult to avoid, especially where there are relatively few channels available (Prior 2007) and performance indicators are increasingly placed in the context of appraising incumbent politicians and parties (Semetko and Valkenburg 2000). Regardless of whether increased consumption is driven principally by voter demand for information or political and media supply of information, consumption of politically- relevant news is likely to reflect a political information cycle with a clear spike prior to elections. Such increased consumption is likely to cause many voters—especially in developing contexts where baseline political knowledge is relatively low (Pande 2011)—to engage with the news for the first time in an electoral cycle. Politically engaged voters are likely to to ratchet up their consumption. Since the timing of news consumption is a key building block of my argument, it is essential to first establish the existence of such political information cycles:

H1. Voters start to consume, or consume relatively more, news just before elections.

Akin to voters learning about candidate positions over an electoral campaign (e.g. Al- varez 2001; Druckman 2005; Hirano et al. 2015; Lenz 2009, 2013), increased information consumption is likely to translate into greater voter awareness of current affairs including salient valence issues.

5.2.2 Implications for electoral accountability

Political information cycles may impact voters’ political beliefs by emphasizing politically- relevant information in the news when consumption is greatest. Events in the news may prime the salience of valence issues, or induce voters to learn about the performance of incumbent politicians (Lenz 2013). For example, if there are many homicides in the news just before an election, voters may become more concerned about public security and negatively update their beliefs about the incumbent party’s ability to address this issue. This requires 182 that voters believe the signal to be informative about the party’s competence. These issue priming and learning mechanisms imply:

H2a. Voters care more about issues that are in the news just before elections.

H2b. Negative information about incumbent performance on salient issues just before elections reduces conﬁdence in the incumbent party’s competence for ofﬁce.

Especially at local elections, voters are poorly informed about incumbent performance in office.9 There is thus good reason to believe that voters may substantially update their beliefs in line with the news that they consume before elections. With weak prior beliefs, voters may still update substantially when recent news is a weak signal of performance over the full electoral cycle. For such beliefs to translate into voting behavior, the issues in the news before elections must be sufficiently important to voters (e.g. Krosnick and Kinder 1990).10 The model in the Appendix, which derives the hypotheses enumerated in this section, shows that the electoral impact of increasing the salience of an issue can be ambiguous, depending upon posterior beliefs about the relative competence of incumbent and challenger parties on that issue.11 However, if the news substantially updates voters’ posterior beliefs about the suitability of a candidate for office—as the literature examining voter learning about candidate positions suggests (Hirano et al. 2015; Lenz 2009)—the learning effect dominates and implies that:

9Even in the U.S., Healy and Lenz(2014) provide evidence that voters use recent economic performance over the last year to proxy for cumulative performance over a Presidential term. 10Rather than policy issues, voters may rely heavily on partisanship, ethnicity or social ties (e.g. Brader and Tucker 2012; Casey 2015; Chandra 2007). 11The model shows that bad news about the incumbent on a given issue induces voters to negatively update about the party’s suitability for ofﬁce, and that this negative learning effect is compounded by increasing the issue’s salience if the incumbent already scored badly on the issue. If the incumbent scored well, the negative learning effect dominates the opposing priming effect when the increase in salience is small or news departs substantially from a voter’s prior. 183 H3. Negative information about incumbent performance on salient issues just before elections reduces the incumbent party’s vote share.

The voting implications of news consumption also depend upon how voters attribute responsibility for performance across different levels of government. Particularly in federal systems where powers are shared or assigned according to non-transparent rules, disentan- gling such responsibility can be challenging (Anderson 2006; Powell and Whitten 1993). Although the federal government is often the primary target of sanctions (e.g. Rodden and Wibbels 2011), voters may also rely on visible local indicators of responsibility such as the frequency with which they engage with local as opposed state or federal police forces. Changes in voting behavior induced by recent news are then most pronounced when voters can clearly assign responsibility:

H4. Negative information about incumbent performance on salient issues just before elections only reduces the vote share of incumbent parties deemed responsible for such performance.

Exactly how voters perceive responsibility across layers of government on different issues is ultimately an empirical question. However, the electoral impact of news reports before elections depends upon the availability and content of the news. Voters rely upon media sources such as television and radio to supply credible and politically-relevant news—news that pertains to a voter’s incumbent party. Locally-based media are more likely to report information about relevant local incumbents (e.g. Snyder and Strömberg 2010). Furthermore, given that media sources may vary in the extent to which they distort or under-report certain types of news (e.g. Besley and Prat 2006; Gentzkow and Shapiro 2006; Mullainathan and Shleifer 2005), and market segmentation may mean that not all stations reach all types of voters (Barabas and Jerit 2009; Prat and Strömberg 2005), the ultimate impact of information about incumbent per- 184 formance on voting behavior should increase with the number of local media stations as the likelihood that voters consume relevant news increases (Larreguy, Marshall and Snyder 2015). I thus hypothesize that:

H5. Negative information about incumbent performance on salient issues just before elections only reduces the incumbent party’s vote share if voters can access relevant information via the media, and is increasing in the level of such exposure.

Moreover, events in the news before elections differentially impact different types of voters. Voters with weak priors, who may be consuming politically-relevant information for the ﬁrst time in months or years, are most susceptible to news because their beliefs are most malleable (e.g. Hirano et al. 2015; Lawson 2004; Zaller 1992). Conversely, voters consuming news throughout the electoral cycle possess stronger prior beliefs over the issues they regard as important and how well different parties address such issues. They are also well-placed to distinguish informative from uninformative signals. These arguments are demonstrated formally in the Appendix, and receive empirical support from Da Silveira and De Mello(2011) and Larreguy, Marshall and Snyder(2016), who ﬁnd that relatively uneducated voters in Brazil and Mexico respectively respond most to political advertising. For incumbent performance indicators, I therefore hypothesize that:12

H6. The consumption of negative information about incumbent performance on salient issues just before elections impacts most the beliefs and voting behavior of voters with weak priors.

The empirical analysis will use education as a proxy for the strength of voter priors. Nevertheless, even if voters primarily consume political information before elections, this does not entail that they entirely lack prior beliefs. Incumbent performance infor-

12The results follow most straight-forwardly from common priors. A sufﬁciently strong correlation between the position of a voter’s prior distribution and education could reverse this result. 185 mation available around the previous election may represent an important component of priors beliefs that voters can benchmark current indicators of incumbent party performance against. For example, if voters received a positive signal of incumbent performance on a given issue relating to a previous incumbent party’s competence before the last election, a negative information in the new before the current election may cause voters to negatively update about the current incumbent party. Similarly, consecutive negative or positive signals about the same incumbent party may strengthen posterior beliefs. Conversely, voters should update less when two different parties experience the same signal, or the same party experiences different signals across time; these cases point to common shocks or high signal variance, and thus imply that voters should downweight the signal.

H7. The impact of negative information about incumbent performance on the incumbent party’s vote share is greater where voters previously received signals that opposition parties performed well or where such information compounds previous negative information about the incumbent party.

To the extent that voters observe events in neighboring municipalities and believe that their governments would behave similarly in their own municipality, a similar logic may hold for spatial comparisons. I now test these hypotheses in the context of local violent crime reported in the media— a prevalent performance indicator on a highly salient issue—before Mexican municipal elections.

186 5.3 Violent crime and political accountability in Mexico

Mexico was an archetypal competitive authoritarian regime until the late 1990s, but has since democratized signiﬁcantly.13 Three main political parties—the right-wing National Action Party (PAN), left-wing Party of the Democratic Revolution (PRD), and previously- hegemonic Institutional Revolutionary Party (PRI)—have competed for power in recent decades.14 Despite growing partisan alignment, often induced by clientelistic exchanges, many Mexicans have not developed 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 government: approximately 2,500 municipalities, 31 states (excluding the Federal District of Mexico City), and the federal government. Constitutional reforms in the mid-1990s substantially increased mayoral autonomy over the provision of local public services (see Diaz- Cayeros, González and Rojas 2006), inducing municipal spending to rise to around 20% of total government spending.15 The median municipality contains 13,000 people, although large cities with the exception of the Federal District—which is divided into delegations— represent singular municipal bodies. Municipal mayors are typically elected every three years to non-renewable terms,16

13The PRI had retained a stranglehold on power since 1929 by implementing populist policies, establishing strong clientelistic ties, effectively mobilizing voters, creating barriers to political entry, and manipulating electoral outcomes (e.g. Cornelius 1996; Greene 2007; Magaloni 2006). 14After attaining legislative pluralities in the 1990s, the long-time conservative opposition PAN fully broke PRI control when Vicente Fox won the Presidency in 2000. In 2006, Felipe Calderón retained the Presidency for the PAN, narrowly defeating the PRD’s Andrés Manuel López Obrador. After regaining legislative ma- jorities, but never relinquishing regional control (Langston 2003), the PRI candidate Enrique Peña Nieto won the Presidency in 2012. 15This expansion of spending included policing, although most municipalities already had their own police forces. Presidents Fox and Calderón also increased federal transfers to municipalities for policing in the 2000s (Sabet 2010). 16From 2018, re-election will become possible for legislators, and mayors in most states.

187 and enter ofﬁce between three and seven months after election day. Municipal elections are almost always held in tandem with state legislative elections, although gubernatorial elections are often held separately. I refer to simultaneous municipal and state legislative elections as “local elections.” Congressional elections to the House and Senate are held every three years, while the President is elected every six years. The majority of local elections do not coincide with these federal elections.

5.3.1 Public security forces

Responsibility for public security is shared across levels of government, and like other federal systems like the United States, mayors play an important role in fighting crime. Both state and federal laws can be used to prosecute criminals, and uniformed (preventive) and investigative police forces exist at both the federal level and in each of Mexico’s 31 states and the Federal District of Mexico City. State and federal police, and increasingly the army, are responsible for investigating major crimes in different jurisdictions. State police investigate state crimes such as homicides, while federal officers focus on organized crime (Reames 2003). However, around three-quarters of municipalities also possess their own police force. Although such forces support higher-level operations and supply information, their principal role is preventive: they primarily patrol the streets and maintain public order, address administrative issues, and respond first to criminal incidents (Reames 2003; Sabet 2010). Nevertheless, municipal police are by far most numerous, accounting for more than half of Mexican enforcement personnel (Sabet 2010). A large central police force controlled by the mayor of Mexico City covers all delegations within the Federal District. Mayors choose the local police chief and set local policies. PAN mayors have played a key role in supporting Calderón’s crackdown on Mexico’s DTOs (Dell 2015), while political alignment across neighboring municipalities has reduced rates of violent crime (Du-

188 rante and Gutierrez 2015). The widely-publicized case of Iguala represents a particularly egregious example of political control of the local police, where the mayor and his wife exploited their links with law enforcement to cover up and possibly instigate the murder of 43 student protesters in 2014. Given the number of municipal police offers and their “on-the-streets” presence, it is not surprising that municipal police are the foremost police force in the minds of voters. When asked which law enforcement authorities they can identify, the 2010 National Survey About Insecurity (ENSI) ﬁnds that while 75% of voters could identify municipal police, only 38% and 48% respectively recognized state and federal forces. The focus groups that I conducted suggested that voters primarily blame local government for crime in their community.

5.3.2 Trends in violence

According to the United Nations Office on Drugs and Crime (UNODC), Mexico suffers one of the world’s highest homicide rates. In 2012, 21.5 people per 100,000 were intentionally murdered (UNODC 2013). This represents the 20th highest homicide rate in the world—slightly less than South Africa, Colombia and Brazil, and slightly more than Nigeria, Botswana and Panama. Using data from the National Institute of Statistics, Geography, and Information (IN- EGI), Mexico’s autonomous statistical agency, Figure 5.1 plots the number of homicides per month afflicting residents of the average municipality between 1999 and 2013. IN- EGI defines an intentional homicide as an unnatural death, as determined by a coroner’s report.17 Month of death is also based on the coroner’s report. The monthly homicide

17Although there is no ofﬁcial measure of drug-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, Rodríguez Ferreira and Shirk 2014). Furthermore, voter fears about public security are not only linked to organized crime, but could reﬂect equally-prevalent non-drug related homicides (of 189 Calderon's War on Drugs 1.2 1 .8 .6 .4 Average number of homicides .2 1999 2002 2005 2008 2011 2014

Month (January of year) Figure 5.1: Trends in monthly homicides in the average Mexican municipality, 1999-2013 rate has been substantial throughout this period, but increased dramatically after President Calderón entered office in December 2006 and began Mexico’s “War on Drugs” (see Dell 2015). Although these figures may understate the true homicide rate if unnatural causes cannot be detected (México Evalúa 2012), it is unlikely the coroners reports are falsified because aggregated data is publicly released with a lag of several years.18 The homicide clearance rate is only 20% (México Evalúa 2012), and drug-related homicides—which are regionally concentrated—represent 50% of homicides over this period (Heinle, Ro- dríguez Ferreira and Shirk 2014). However, many municipalities only rarely experience a homicide; in fact, only one homicide occurs in the median municipality each year. Unsurprisingly, voters are concerned about Mexico’s high rates of violent crime. Like many other Latin American countries (Blanco 2013), Figure 5.2 shows that the number of which only a tiny fraction occur within families). Nevertheless, Table D.13 shows broadly similar results for the sample utilizing only drug-related homicides. 18Heinle, Rodríguez Ferreira and Shirk(2014) discuss Mexican homicide metrics in depth.

190 detsn eeust upr hmevs n hsfc togicnie otio their tailor to incentives on strong rely face outlets thus media and Private themselves, support below). to revenues 5.6 advertising Figure (see cover areas generally geographic which stations, small television relatively 1,255 and stations FM home. 1,097 at stations, radio internet radio a the AM own to respectively access have households 20% of only 92% while and television, 79% or Census, 2010 the on Based icans. coverage Media 5.3.3 last the of economy. the most of ahead For voters for issue rates. salient most homicide the as with registered security line public decade, in broadly increased country the facing problem important most the as security public citing respondents Latinobarómetro Mexican 19 iems eeoigcutis racs ei stemi oreo esfrMex- for news of source main the is media broadcast countries, developing most Like rm/ulcscrt,du rfcig ilne rtroimpltclviolence. terrorism/political or violence, trafﬁcking, drug security, crime/public Notes n21 n 01 0 fLtnbrmtorsodnsrpre htte ee s h internet. the use never they that reported respondents Latinobarómetro of 60% 2011, and 2010 In iue52 oeslsigciea h otipratpolmta eiofaces, Mexico that problem important most the as crime listing Voters 5.2: Figure aafo h 0121 aioaóer uvy.Cieicue ocrsabout concerns includes Crime surveys. Latinobarómetro 2001-2011 the from Data :

Proportion answering that crime is most important 0 .1 .2 .3 .4 2001 2002 2003 2004 2001-2011 2005 191 2006 2007 2008 2009 2010 19 eiocnan 852 contains Mexico 2011 programming to local audiences (Larreguy, Marshall and Snyder 2015). Most stations form part of broader regional and national radio or television networks, most prominently Televisa and TV Azteca, and are often fully-owned subsidiaries. Within networks, entertainment content is bought from or relayed by network providers (and is thus identical across stations). News is an important exception: while national news is typically centrally provided, many affiliates and regional subdivisions 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 which schedules were available in 2015, the average channel broadcasts 3.6 hours of news coverage each weekday (both before and after the June elections). Slightly less than half of this news is devoted to state or city-specific programming. While the pro-PRI coverage biases that characterized elections before the late 1990s have somewhat dissipated (e.g. Hallin 2000; Lawson 2002), television stations and networks continue to vary in their level of bias (Hughes and Lawson 2004). Homicides are not always reported in great depth, but are regularly covered in print (Ley 2014; Osorio 2015) and “omnipresent” in the local broadcast media (Trelles and Carreras 2012).20 Based on the 2010 ENSI survey, 87% of respondents report learning about public security in the country and in their state from television news programs, while 34%, 29% and 9% respectively report learning from radio news programs, periodicals or newspapers, and the internet.21 Furthermore, 82% of voters believe that television news has the most important influence on public opinion.

20For example, see this local news report of an gangland-style murder in Monterrey. The same news station also reports on domestic murders, e.g. this case of a woman strangling her partner. News programs may continue to report on arrests and cases that go to trial, especially when the defendants are found guilty. For example, on 17th February 2015, W Radio reported here that two men were convicted of intentional homicide and sentenced to 27 years. 21Although 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%). 192 Despite claiming reasonable levels of attention to political news, knowledge of public affairs is limited. I show below that only half of Mexicans can answer basic questions about politics, while Castañeda Sabido(2011) and Chong et al.(2015) ﬁnd that voters are particularly unaware of local politics and mayoral performance in ofﬁce. Only 19% of CIDE-CSES respondents could name any of their candidates for federal deputy in 2009. This lack of knowledge suggests that many voters possess weak priors, and may thus internalize information when they actually follow the news.

5.4 Local elections, news consumption, and political per-

ceptions

The starting block of my theoretical argument is that voters consume more information just prior to elections. While some voters start consuming politically-relevant information for the ﬁrst time over an electoral cycle, others will consume more than before. Given the salience of crime in Mexico, greater attention to news in the run-up to an election inevitably entails learning about local crime. However, exposure to information about local violence is only likely to affect voting behavior if it affects how voters perceive the salience of public security and/or the ability of different parties to effectively address this key valence issue. This section provides survey-level empirical support for each element of this argument.

5.4.1 Data

To test these hypotheses, I use four waves of the National Survey of Political Culture and Civil Practices (ENCUP) conducted over several weeks in November 2001, Febru- ary 2003, December 2005, and August 2012.22 Each round draws stratiﬁed random sam-

22The 2008 wave does not provide a respondent’s municipality and asks different questions about media consumption. The ENCUP surveys are preferred to the Latinobarómetro, which covers more years, but does 193 ples of around 4,500 Mexican voters for face-to-face interviews from pre-selected electoral precincts within urban and rural strata defined by the electoral register.23 The survey was commissioned by the Interior Ministry and designed to be broadly nationally representative, and focuses on the country’s political culture rather than more contentious questions about elections.24 Each survey wave was conducted in a different month of the year, and the irregularly-spaced waves do not correspond with the federal electoral cycle. The pooled sample thus includes up to 15,976 respondents across 523 municipalities.25 I focus on three types of outcome. First, I measure political news consumption by the frequency with which voters watch or listen to the news, programs about politics, or programs about public affairs.26 To understand the margins at which consumption is changing, I examine various consumption intensities: never, at some point, at least monthly, at least weekly, and daily. I also compute a 5-point scale (from 0 to 4). Because these consumption measures cannot distinguish changes in demand for information among voters from changes in the supply of political news provided by media outlets, this paper emphasizes political information cycles in equilibrium. Second, I assess whether politically-relevant news consumption translates into greater not always ask about media consumption and especially political information, and does not contain municipal identifiers (only states). Nevertheless, the Latinobarómetro returns similar results. My identification strategy is not compatible with the rich Mexican Panel Surveys that have been conducted around the 2000, 2006 and 2012 presidential elections. 23In 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 were surveyed from each precinct according to specific directions (see the 2012 methodological manual here). Although such detailed sampling information is not available for the earlier surveys, the overall design is similar. 24The study was implemented by INEGI in 2001 and 2003, and private firms in 2005 and 2012. The specific objectives of the study, which does not address elections at all, are enumerated here. 25The Federal District, where policing is not administered at the delegation-level, is excluded. Several usos y custombres in Oaxaca are included in the survey analysis, although the results are robust to their exclusion. 26I focus on radio and television, which are by far the most prevalent sources of political information in Mexico.

194 political knowledge. Political knowledge is the first (standardized) factor from a set of indicators coding correct responses to simple factual questions regarding topical political news and basic (national and local) knowledge of political institutions and incumbents parties.27 The average respondent answered around half the questions correctly. A key advantage of this measure is that voters cannot falsely inflate their knowledge. To separate topical news from institutional knowledge that voters are less likely to learn about, I also divide the quiz into topical and institutional sub-indices. Third, I measure the salience of violence and changes in beliefs about the competence of the incumbent party. To test whether homicides around elections translate into concern about public security, I define an indicator for respondents citing crime and insecurity, drug trafficking, violence, or vandalism as the most important problem for their community to solve.28 I also create an indicator of low confidence in the municipal mayor, defined by respondents ranking their confidence at 5 or below on a scale from 0 to 10 (for 2012) or expressing no, or almost no, confidence in the given institution (in 2001).29

5.4.2 Identiﬁcation strategy

Using a similar design to Eifert, Miguel and Posner(2010), I exploit the timing of survey administration with respect to state-speciﬁc election cycles to identify the effects of an upcoming election. Operationally, I code an indicator for respondents facing an upcoming municipal election, and typically a simultaneous state legislative election, within ﬁve months of the survey. Although there is some state-level discretion over campaigning

27In 2001, 2003, 2005 and 2012 respectively, respondents were asked 6, 3, 3 and 4 questions (see Ap- pendix). 28Along with all other political concerns, I include as zeroes respondents that do not cite a problem. 29The 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. Unfortunately, vote intentions were not elicited in any survey wave.

195 in such local elections, campaigns last around ﬁve months.30 As noted above, the ENCUP surveys are administered irregularly—both in terms of the month in which the survey was conducted and the number of years between survey rounds. In conjunction with the fact that states historically follow different electoral cycles, varying both in the month and year of their elections (including over time within states, especially due to a recent constitutional reform),31 whether the surveys were administered just before or after elections or whether an election was held recently at all effectively occurs by chance. There is no evidence to suggest that these surveys, which explicitly do not address elections, were strategically timed with respect to state elections. This is reinforced by the fact that 27 of Mexico’s 32 states (including the Federal District) register an election in one of the survey years, of which 14 hold an election in a survey year and after the survey occurs. Supporting the plausibility of these arguments, Table D.2 shows that individuals surveyed prior to local elections are well-balanced over individual and municipal level characteristics: only 2 of 20 tests reported a signiﬁcant difference at the 10% level. Table D.3 shows that neither changes in upcoming local elections nor violence predict changes in whether a municipality is included in the survey sample. To identify the effect of an upcoming local election, I estimate the following regression for respondents i in municipality m at survey year t:

Yimt = βUpcoming local electionmt + µt + εimt, (5.1)

30Political advertising slots are specifically allocated for these purposes five months prior to federal elections (Larreguy, Marshall and Snyder 2016). The results are robust to defining upcoming elections by any number of months between 1 and 12. 31See Table D.1 in the Appendix for a full list of municipal elections by month. In Chiapas, Coahuila, Guerrero, Michoacán, Quintana Roo, Veracruz, and Yucatán, the typical 3-year cycle was adjusted over the sample period by switching 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 cycle coincides (Serra 2014). 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. 196 where Yimt is a measure of politically-relevant news consumption, political knowledge, concern about public security, or confidence in the police or municipal mayor. The survey

ﬁxed effects, µt, capture common period effects that might arise from concurrent federal elections (in 2003 and 2012), presidential elections (in 2012), or national trends in political behavior. To explore how the effects of upcoming local elections vary with respondent education and proximate events, I also interact the election indicator with educational attainment and various measures of municipal-level homicides. Throughout, standard errors are clustered by municipality.

5.4.3 Upcoming elections increase news consumption and knowledge

The results in panel A of Table 5.1 demonstrate that local elections increases both the likelihood that a voter consumes any political information at all as well as the level of information consumed.32 Column (1) shows that an upcoming local election significantly increases the probability that a voter listens or watches news at all by five percentage points. Although this shift could still reflect an increase in the supply of political news on television and the radio, consuming information for the first time is likely to represent a conscious acquisition choice by voters that had previously avoided the relatively extensive news coverage that is available even outside election campaigns. Columns (2)-(4) also report large increases in news and political program consumption at higher viewing and listening intensities. The effect on the scale in column (5) is also strongly positive. These findings suggest that virtually all types of voters, ranging from the least to most engaged, consume more political information prior to statewide elections. Panel B shows that the cycle of increased information consumption prior to elections mirrors changes in political knowledge. Contrary to potential concerns about social desir-

32Comparable questions from 2001 were not available.

197 Table 5.1: The effect of upcoming local elections on political news consumption and knowledge

Panel A: Political news Watch and listen to news and political programs... consumption ...ever ...monthly ...weekly ...daily ...scale (1) (2) (3) (4) (5) Upcoming local election 0.051*** 0.083*** 0.081*** 0.064** 0.280*** (0.014) (0.022) (0.023) (0.026) (0.077)

Observations 11,983 11,983 11,983 11,983 11,983 Outcome mean 0.86 0.68 0.62 0.38 2.55 Outcome range {0,1} {0,1} {0,1} {0,1} {0,1,2,3,4} Upcoming local election mean 0.19 0.19 0.19 0.19 0.19 Survey year without data 2001 2001 2001 2001 2001 Panel B: Political Political Topical Institutional knowledge knowledge knowledge knowledge quiz questions questions (1) (2) (3) Upcoming local election 0.283*** 0.307*** 0.096*** (0.051) (0.056) (0.031)

Observations 15,976 15,976 15,976 Outcome mean 0.00 0.00 0.00 Outcome standard deviation 1.00 1.00 1.00 Outcome range [-1.8,1.7] [-2.1, 1.6] [-1.5,2.2] Upcoming local election mean 0.16 0.16 0.16

Notes: All specifications include survey-year fixed effects, and are estimated using OLS. Each coefficient identifies the average change in an outcome associated with an upcoming local election within five months of the survey. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. ability bias, column (1) demonstrates that voters facing an upcoming local election are more than a quarter of a standard deviation, or nine percentage points, more likely to correctly answer a question testing their political knowledge.33 This increase indicates that voters actively engage by internalizing political information—a key prerequisite for political in-

33Like Hirano et al.(2015), this result shows that voters learn more about politics during electoral campaigns, even around lower-level elections. Interacting upcoming local elections with with baseline covariates for both consumption and knowledge outcomes indicates that the same voters consuming more news are also become more knowledgeable about politics.

198 formation to affect electoral accountability. Columns (2) and (3) show that although voters also learn about Mexican political institutions, increases in topical political knowledge— contemporary policy issues and knowledge of incumbent state governors—are more than three times greater in magnitude. The Appendix demonstrates the robustness of these findings. First, Table D.9 shows that the results are unaffected by controlling flexibly for education—the only variable showing imbalance over upcoming local elections. Second, at considerable efficiency cost, Table D.10 reports similar results when including state fixed effects. Third, to show that the results do not depend upon on the definition of the upcoming election indicator, Table D.11 confirms that voters exhibit greater news consumption and political knowledge as the number of months until the next local election decreases. Further analysis points to a non-linear effect where increased consumption is concentrated during the electoral campaign, and accelerates within the five months before the election.

5.4.4 Homicides before elections increase public security concerns and

reduces conﬁdence in municipal institutions

I now examine whether the type of information in the news during local election campaigns affects voter perceptions of political performance. Since homicides are widely reported in local news, and I just demonstrated that voters consume more politically-relevant information prior to local elections, local violence before elections may cause voters to alter the importance they attach issues of public security and update their beliefs about the performance of their municipal institutions. To test these prerequisites for electoral accountability, I extend equation (5.1) by interacting the presence of a local election with municipal homicide data. I compare short- and longer-run homicide measures to differentiate recent news from information over a longer

199 time span. I employ two short-term measures: an indicator for one or more homicides occurring within a voter’s municipality during the month before the survey;34 and a homicide shock indicator for municipalities where, conditional on experiencing at least one homicide over the two months either side of the survey, the number of homicides in the two months prior to the survey exceeds the number of homicides in the two months after the survey. The latter measure is described in greater detail below and is shown to be exogenous to a wide variety of covariates in the aggregate analysis. Table D.2 indicates that neither measure is signiﬁcantly correlated with upcoming local elections. In both cases, voters with weak priors are most likely to respond to short-term measures because they possess little other information. The longer-term measures—the average number of homicides per month in the year prior to the survey, and the three years priors to the survey—are less likely to be covered directly in the news, although well-informed voters may have already internalized such information.35 The results in panel A of Table 5.2 demonstrate that homicides just before an election substantially increase fears regarding public security. The insigniﬁcant effect of the lower- order local election term, combined with the large positive interactions in columns (1) and (2), show that upcoming local elections only increase concerns about security among voters in municipalities experiencing at least one homicide during the month before the survey or a short-run homicide shock.36 Relative to the sample mean, short-term homicides increase security concerns by almost 50% before elections. Such priming effects of media

34Since 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 thus use an indicator for a homicide occurring in either the month of the survey or the prior month. 35The balance checks in Table D.2 show that upcoming elections are not significantly correlated with each measure. 36The positive interactions are robust to simultaneously controlling for the interaction of an upcoming election with pre-treatment municipal characteristics. The positive effect of local elections in columns (3) and (4) reflect the fact that the positive interactions in columns (1) and (2) are being averaged across in these specifications.

200 Table 5.2: Heterogeneous effects of upcoming local elections on concern for public security and institutional conﬁdence, by short-run and long-run municipal homicide measures Homicide measure: Homicide Homicide Homicides Homicides within shock per month per month last month last year last 3 years Panel A: Public Public insecurity the major problem in the community security concerns (1) (2) (3) (4) Upcoming local election 0.002 0.024 0.034** 0.034** (0.016) (0.021) (0.016) (0.016) Homicide measure 0.061*** -0.011 0.009*** 0.009*** (0.009) (0.012) (0.002) (0.002) Upcoming local election 0.073*** 0.092*** 0.002 0.002 × Homicide measure (0.023) (0.028) (0.003) (0.003)

Observations 12,541 9,764 12,541 12,541 Outcome mean 0.11 0.13 0.11 0.11 Outcome range {0,1} {0,1} {0,1} {0,1} Upcoming local election mean 0.20 0.22 0.20 0.20 Homicide measure mean 0.66 0.46 2.99 3.06 Survey year without data 2012 2012 2012 2012 Panel B: Conﬁdence Low conﬁdence in the municipal mayor in municipal mayor (1) (2) (3) (4) Upcoming local election -0.035 -0.019 0.000 0.001 (0.036) (0.042) (0.036) (0.038) Homicide measure 0.003 -0.048** 0.001 0.001*** (0.024) (0.026) (0.001) (0.000) Upcoming local election 0.077* 0.114** -0.001 -0.002 × Homicide measure (0.045) (0.048) (0.009) (0.015)

Observations 7,236 5,925 7,236 7,236 Outcome mean 0.34 0.34 0.34 0.34 Outcome range {0,1} {0,1} {0,1} {0,1} Upcoming local election mean 0.03 0.02 0.03 0.03 Homicide measure mean 0.72 0.42 6.33 6.98 Survey years without data 2003, 2005 2003, 2005 2003, 2005 2003, 2005

Notes: All speciﬁcations include survey year ﬁxed effects, and are estimated using OLS. The number of observations drops in column (2) because voters in municipalities failing to experience any homicide over the two 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.

201 are strongly supported in the political communication literature (e.g. Gilliam and Iyengar 2000; Iyengar 1994; Iyengar, Peters and Kinder 1982). However, the insignificant lower- order homicide shock term suggests that voters do not update about public security from short-run homicide spikes during periods of lower news consumption. It remains possible that voters that only become informed around the election are actually responding to longer-run homicide rates. The specifications in columns (3) and (4) interact an upcoming local election with the longer-term homicide measures to assess this possibility. The lower-order terms show that public security is a greater concern in more violent municipalities. However, the interaction of upcoming elections with the average number of homicides over the last year, or last 3 years, does not significantly affect the salience of public security for voters. Turning to beliefs about institutional competence, panel B shows that homicides coinciding with upcoming local elections also cause voters to reduce their confidence in the municipal incumbent. Columns (1)-(4) in panel B demonstrate that local elections only increase the likelihood that a respondent expresses low confidence in the incumbent mayor in violent municipalities. The reductions in mayoral confidence are respectively 8 and 11 percentage points—or 23% and 34% of the sample mean—greater following a homicide within the last month or a recent spike in homicides. The insignificant negative coefficients for local elections in columns (1) and (2) tentatively suggest that voters may positively update about their incumbents when the number of homicides just before an election is low. Again, long-run homicide trends do not differentially affect voters around elections. Table D.12 in the Appendix reports similar results for confidence in the police. Finally, to test whether voters with weaker priors respond more to performance indicators in the news around elections, I examine how the effects of short-run homicide shocks vary by education—a popular proxy for information. My education indicator for the 25% of respondents that completed high school at age 18 is positively correlated with news 202 a agnlefc fucmn oa lcin npbi neuiybigtemjrpolmi the in problem major the being insecurity public on elections local upcoming of effect Marginal (a) iue53 feto poiglcleetoso ocr bu ulciscrt n low and insecurity public about concern on elections local upcoming of Effect 5.3: Figure ofiec nicmetmyr,b oiiesokadcmlto fuprsecondary upper of completion and shock homicide by mayors, incumbent in confidence h eaieipeiinrflcsteuaalblt feuainmaue n2001. in measures measures. education shock of unavailability homicide the short-run reflects and imprecision relative education, The secondary lower completing for indicator Notes b agnlefc fucmn oa lcin nlwcndnei h uiia mayor municipal the in confidence low on elections local upcoming of effect Marginal (b) neatn oa lcin,an elections, local interacting (D.1) equation of versions modified from are estimates All :

Marginal effect of upcoming local election Marginal effect of upcoming local election

-.1 0 .1 .2 -.05 0 .05 .1 .15 Incomplete Incomplete Negative homicideshock Negative homicideshock dcto 9%cndneintervals). confidence (95% education Complete Complete community 203 Incomplete Incomplete Positive homicideshock Positive homicideshock Complete Complete consumption and political knowledge. The marginal effects of an upcoming local election reported in Figure 5.3 suggest that increased concern about public insecurity and reduced confidence in municipal incumbents are most pronounced among less educated voters.37 Both panels show that a homicide shock only significantly increases the concerns of voters that did not complete upper secondary schooling.

5.5 Local violence and electoral accountability

Given that voters are better informed just before an election, and given that homicides just prior to an election increase concerns about public security and reduce conﬁdence in the municipal mayor, voters may punish incumbent parties for poor performance at the ballot box—even if such poor performance is not indicative of performance over the entire election cycle. However, electoral accountability for performance on this highly salient issue is further complicated by the fact that responsibility for policing, and other social policies, is shared across layers of Mexican government. I now seek to disentangle voter perceptions of political accountability using detailed precinct-level electoral returns to both identify the electoral effect of a pre-election homicide shock and understand how voters assign responsibility across the municipal, state, and federal incumbent parties.

5.5.1 Data

To examine the electoral effects of violence in the run-up to elections I utilize two main sources of data. First, electoral returns for municipal, state and federal elections covering Mexico’s c.67,000 electoral precincts were assembled from the Federal Electoral Institute (IFE),38 state electoral institutes, and freedom of information requests. I focus on municipal

37The homicide within the last month measure produced similar but less stark results. 38IFE recently became the National Electoral Institute (INE). 204 elections between 1999 and 2013, for which data is widely available across all states; state and federal electoral returns are used to parse out accountability channels. Since municipal elections generally occur every three years, the full dataset contains around five elections per municipality. Second, I combine the electoral data with INEGI’s monthly municipal homicide data (described in section 5.3). As noted above, incumbent mayors could not seek re-election. However, differences in candidate selection mechanisms across parties at the state level ensure that candidate choices differ across party but are highly correlated within parties (Langston 2003), while voters are substantially better informed about parties than individual politicians (Chong et al. 2015; Larreguy, Marshall and Snyder 2016). Previous research has shown that voters frequently punish parties for an incumbent’s performance in office (Chong et al. 2015; Larreguy, Marshall and Snyder 2015). I therefore focus on the electoral performance of the incumbent party.39 Reflecting their persisting power at the state level (see Langston 2003), 53% of incumbent mayors in the sample are from the PRI; 29% are from the PAN and 12% are from the PRD. The empirical analysis focuses on two measures of incumbent electoral performance: change in the incumbent’s vote share at the precinct level, and an indicator for whether the incumbent won the municipal election.40 Unlike the U.S. (e.g. Lee, Moretti and Butler 2004), but consistent with the incumbency disadvantage found in Brazil (Klašnja and Titiu-

39In 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 vote share at the following election. In general, coalitions are dominated by large parties; the three main parties represent more than 90% of incumbents. I show below that the results are robust to restricting attention to incumbents containing the three largest parties and single-party incumbents. 40I show similar results for the incumbent vote share level as a robustness check, but prefer the precision associated with changes in vote. For comparability, precincts where the election winner is available but the lagged incumbent vote share is not are dropped. Given that the analysis weights by the number of registered voters in each precinct and standard errors are clustered by municipality, the precinct-level estimates are almost identical to municipal-level estimates. Because some precincts are missing within municipalities, a precinct-level analysis has slightly more power. Subsequent analyses also exploit precinct-level variation.

205 nik forthcoming), the average incumbent party experiences a 5.7 percentage point decline in their vote share, but still wins 54% of races. Turnout rates typically hover around 60%. As in the survey analysis, the Federal District is excluded.

5.5.2 Identiﬁcation strategy

While the national homicide rate has changed relatively smoothly over time, this masks considerable inter-month volatility within municipalities. At the height of the War on Drugs in 2010, the monthly homicide count in Ciudad Juárez, Chihuahua, oscillated dramatically, e.g. from 394 in September to 477 in October before falling to 242 in November and again rising to 309 in December. Conversely, the median municipality experiences no homicides in any given month, and one homicide over a year. More generally, month- to-month ﬂuctuations appear fairly random: since 1999, the average municipal homicide count registering a 0.002 month-on-month increase is dwarfed by a 1.7 homicide standard deviation. Moreover, the count is almost exactly as likely to increase as decrease—both in the run-up to elections and in the middle of a government’s tenure.41 To identify the electoral effects of homicides occurring at the times when voters consume most politically-relevant news, I exploit short-term deviations from trend in the number of homicides around local elections.42 In particular, I compare otherwise similar elections experiencing a short-term spike relative to the municipal homicide rate before the election to elections experiencing a short-term relative decrease. This strategy is preferred

41Across municipalities since 1999, the monthly homicide change was positive in 11.86% of cases and negative in 11.79% of cases. In election months, these shares are 11.90% and 11.87%. In the month preceding an election, the shares 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. 42A beneﬁt of the unavailability of the number of homicides reported by broadcast media stations is that the empirical strategy is not susceptible to differential media reporting biases. Table D.4 shows that delayed homicide registration, a potential indicator of strategic registration and missing homicides, is balanced over homicide shocks.

206 to a design examining changes in pre-election homicides rates across elections because such changes are highly correlated with the longer-run homicide trends that I seek to differentiate pre-election shocks from.43 I deﬁne a municipality as “treated” by a pre-election homicide shock if, conditional on experiencing at least one homicide over the four months spanning the penultimate month before an election and the second month after an election, a municipality m experiences

more homicides in the two months before an election in month τ than the two months after the election:44

 1 if Homicidesm − + Homicidesm − > Homicidesm  τ 2 τ 1 τ   τ+1  +Homicidesmτ+1 and ∑ Homicidesmτ0 > 0,  0  τ =τ−2   0 if Homicidesmτ−2 + Homicidesmτ−1 < Homicidesmτ Homicide shockmτ ≡ τ+1 (5.2)   +Homicidesmτ+1 and ∑ Homicidesmτ0 > 0,  0  τ =τ−2   . if Homicidesmτ−2 + Homicidesmτ−1 = Homicidesmτ   τ+1   +Homicidesmτ+1 or ∑ Homicidesmτ0 = 0. τ0=τ−2

Akin to Ferraz and Finan’s (2008) comparison of Brazilian municipalities where federal audit reports were randomly released just before and just after municipal elections, conditioning on the occurrence of a homicide around elections produces the relevant counter-

43Nevertheless, Table D.14 shows similar results using identification strategies that instead capture the intensity of pre-election shock by exploiting differences in pre-election deviations from homicide trends or using the proportional change in the number of homicides around elections. 44Elections typically occur on the first Sunday of the month, so I define the treatment using the two months prior to the election month (i.e. τ − 2 and τ − 1). For the 9% of elections held on the 16th or later, I use months τ − 1 and τ. I show below that the results are robust to dropping elections not held during the first half of the month. The results are also robust to including cases of no change in the number of homicides as either positive or negative shocks.

207 factual by extracting homicide trends.45 I use a two-month window to capture the most intensive moments 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 are plausibly random. This homicide shock equates to an increase of around three homicides per month in shocked municipalities relative to control municipalities. I show below that the results are robust to using one or three month bandwidths instead. After dropping the 30% of municipalities where no homicide is registered in either the two months before or after the election, Figure D.1b in the Appendix highlights the final sample. The key identifying assumption is that, within municipalities, the timing of homicides around elections is effectively random. Leveraging within-municipality variation controls for all time-invariant municipal-level features, and also reduces noise generated by differences in homicide histories among municipalities receiving the same treatment.46 However, although month-to-month changes in homicide rates are highly idiosyncratic, this does not necessarily imply that short-term changes in homicide shocks across elections in a given municipality occurred by chance. I first show that homicide shocks are uncorrelated with a wide variety of pre-treatment covariates and homicide pre-trends. First, Table D.4 in the Appendix examines 104 pre- treatment covariates capturing demographic, socioeconomic, media coverage, electoral threat, and political features of each electoral precinct. Only 6 differences are statistically significant at the 10% level.47 In particular, homicide shocks are equally distributed across

45These municipalities differ signiﬁcantly from even municipalities experiencing only a single homicide over the 4-month window. Municipalities without a homicide are less developed, less politically competitive, and more violent. 46The robustness checks below show that a simple difference-in-means yields similar results. 47Table D.5 shows similar balance once ﬁxed effects are excluded for 2010 Census or other time-variant variables.

208 violence. the preceding months Furthermore, 3 shocks. homicide and with vary 6 significantly 12, not do the treatment in the defining homicides period of number average the that confirm significant. define statistically to never used and period shocks the preceding homicide months 10 untreated the and over treated constant relatively between both homicides is monthly precincts in difference the violence. that of shows levels figure underlying The different or rates homicide in pre-trends differential exhibit shocks negative attainment. and educational positive receiving and municipalities radio, whether or examines 5.4 television Figure a Second, own that num- share the the to stations, respect media with of also ber but identity, party of incumbent’s the competitiveness and the election, share, previous the vote previous incumbent’s the including variables political iue54 ifrnei h ubro oiie ewe uiiaiiswt positive a with municipalities between homicides of number the in Difference 5.4: Figure 48 n eaiehmcd hc,b ot ni h lcin(5 ofiec intervals) confidence (95% election the until month by shock, homicide negative and ee oiie ntetdmncplte ol onadyba siae fvtr acingreater sanction voters if estimates bias downwardly would municipalities treated in homicides Fewer esti- main the to akin the regression defining a from period is the estimate to (5.3). Each equation prior mating months election. the ten around the shock in homicide municipalities shock) (negative control and Notes ahbrdntstedfeec ntenme fhmcdsi rae pstv shock) (positive treated in homicides of number the in difference the denotes bar Each :

Difference in monthly homicides -10 -5 0 5 12 11 10 Months untiltheelection 9 209 8 7 48 also D.4 Table in tests balance The 6 5 4 3 I show below that a placebo shock defined 6 months prior to the election does not affect electoral outcomes, and that the results are robust to controlling for homicide levels. Third, consistent with sampling variability, Figure D.2 shows that the distributions of homicides prior to the period defining a shock, two months before elections and two months after elections are very similar. I also find no evidence for the more specific concern that DTOs alter the number of homicides around elections to oust unpopular incumbent parties. Table D.4 shows that homicide shocks are not significantly correlated with the number of drug-related homicides in the prior year over the 2006(Dec.)-2011 period when monthly data was made publicly available by the Mexican National Security System. Homicide shocks are also balanced across municipalities registering more than one drug-related homicide in any pre-election year over this period, and no more likely to occur in the 5% of precincts designated by the IFE as high-risk (typically locations with high DTO activity). In addition, I show below that the results are robust to removing municipalities with high levels of drug-related homicides and states with high DTO presence. Nevertheless, the types of homicide could change without affecting overall levels. Gang- land killings are typically concentrated among young and uneducated men, and are often committed using firearms or more gruesome methods—particularly if intended as signals. Using the International Classification of Diseases codes in INEGI’s coroner reports, Table D.4 also examines the causes of death and victim characteristics of homicides occurring in the two months before an election. In particular, homicide shocks are well balanced with respect to the share of homicides caused by a firearm, cutting implements, hanging, chemical substances, or drowning. Furthermore, there is no evidence that such homicides disproportionately afflict young, male, unmarried, or uneducated individuals that are most likely to be involved with organized crime. A different potential concern is that homicide rates change around elections because 210 effective governments can crack down on crime prior to win votes (e.g. Levitt 1997), or because election outcomes themselves impact post-election homicides rates. In practice, none of the municipal politics experts I interviewed believed that municipal governments crack down on local homicides. However, I also examine this concern more systematically. First, proxies for state capacity—including municipality size and budget, police per voter, alignment with governors or the president, and neighbor homicide shocks—are uncorrelated with homicide shocks, while if anything mayors controlling a municipal police force are more likely to experience a shock. This indicates no differential ability to engage in such crackdowns. Second, using Osorio’s (2015) newspaper-based measures of actual DTOs crackdowns, Table D.4 demonstrates that homicide shocks are uncorrelated with violent enforcement, drug-related arrests, asset seizures, drug seizures and gun seizures in the two months before the election. Moreover, Table D.6 finds no change in the number of security force employees per voter in election years.49 Third, although the new mayor does not enter office until more than two months after the election, Table D.7 shows that neither pre-election political variables nor election outcome variables predict violence levels in the two months after an election.50 The definition of homicide shocks including post-election homicides is thus unlikely to induce bias. Ultimately, I estimate the effects of homicide shocks just prior to the election in precinct p of municipality m in election year t using regressions of the form:

Ypmt = βHomicide shockmt + ηm + µt + εpmt, (5.3)

49Although such data are only available from the National Census of Municipal Governments (ENGM) on an annual basis over five waves, I was able to impute 9,655 municipal-years. 50Dell(2015) finds that homicide rates increased where PAN mayors were elected during the Calderón administration, but such mayors do not enter office within two months of the election. More generally, post-election homicides are also uncorrelated with interactions between election outcomes and background covariates.

211 where Ypmt is either the change in the incumbent party’s vote share relative to the previous election or an indicator for the incumbent party winning the election. Municipality and year ﬁxed effects are respectively denoted by ηm and µt. All precincts are weighted by the number of registered voters to prevent the results from being driven by small precincts.

5.5.3 Homicide shocks harm municipal incumbent parties

Building on the survey evidence, Table 5.3 shows that homicide shocks severely hinder the municipal incumbent party’s electoral performance. Column (1) reports that a homicide shock reduces the incumbent party’s vote share by 2.2 percentage points. This equates to almost 0.5 percentage points for each additional homicide per month relative to the baseline levels in the treatment and control group. Column (2) demonstrates that this translates into an 11 percentage point decline in the incumbent party’s probability of being re-elected. These findings show that an inopportune, but not uncommon, temporary increase in violence can have substantial electoral implications for an incumbent if it coincides with the pre-election period when voters are more likely to acquire political information and infer that the incumbent party is relatively incompetent. This substantial decline in incumbent support could primarily reflect the dramatic rise in drug-related homicides since 2006. Columns (3) and (4) of Table 5.3 report estimates interacting an indicator for an election occurring since December 2006 with the homicide shock (as well as the municipality fixed effects). The vote share results in column (3) show that the drop in incumbent support did not differentially occur during Mexico’s drug war. This implies that an uptick in homicides before elections is punished electorally regardless of the broader context of crime, and that while drug-related homicides are perhaps most prominent, less targeted homicides may be equally worrying to voters. Column (4), however, suggests that incumbent parties have only recently become significantly less likely to

212 Table 5.3: The effect of a pre-election homicide shock on municipal incumbent electoral outcomes

Change in Incumbent Change in Incumbent Change in Change in incumbent party incumbent party turnout incumbent party win party win vote share vote share vote share (registered) (1) (2) (3) (4) (5) (6) Homicide shock -0.022** -0.112*** -0.020 -0.031 0.008 -0.013** (0.009) (0.040) (0.013) (0.059) (0.010) (0.006) Homicide shock -0.009 -0.240***

213 × Post-2006 (0.024) (0.085)

Observations 181,408 181,408 181,408 181,408 171,369 171,369 Outcome mean -0.05 0.56 -0.05 0.56 0.00 -0.01 Outcome range [-0.96,0.89] {0,1} [-0.96,0.89] {0,1} [-0.76,0.67] [-0.93,0.94] Homicide shock mean 0.46 0.46 0.46 0.46 0.47 0.47 Post-2006 mean 0.54 0.54

Notes: All specifications include municipal and year fixed effects, and are estimated using OLS. Specifications (3) and (4) also interaction municipality fixed effects with the post-2006 indicator. The lower order post-2006 indicator is subsumed by the year fixed effects. All observations are weighted by the number of registered voters in the electoral precinct. The number of observations in columns (5) and (6) declines because the lagged number of registered voters was unavailable. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. win re-election. Since incumbent vote share losses did not change, the jump in the electoral penalty likely reflects the general increase in political competition—the average incumbent win margin fell from 15.4 percentage points (pre-2007) to 12.8 (since 2007). Furthermore, the decline in electoral support is principally driven by changes in vote choice rather than changes in turnout. Column (5) finds no clear evidence that a homicide shock either mobilizes voters against the incumbent or induces sufficient disenchantment with politics for voters not to turn out in aggregate.51 In conjunction with Bateson(2012), this result suggests that personal involvement may be required to mobilize voters. Nev- ertheless, column (6) reinforces the vote share findings, showing that a homicide shock reduces the incumbent party’s vote share, as a proportion of the registered electorate, by 1.4 percentage points. By not conditioning on the decision to turnout, this estimate clearly indicates that voters shift away from the incumbent party. I demonstrate the robustness of these findings by showing that the results are not sensitive to variable definitions, sample choices, or alternative specifications. First, columns (1) and (2) of Table 5.4 show that homicide shocks defined by one- or three-month windows also reduce the incumbent’s vote share by around two percentage points and the incumbent’s probability of winning by 7-13 percentage points. The smaller one-month window estimates could reflect a smaller sample or weaker signals imparted by such shocks. Sec- ond, by including fixed effects for the total number of homicides over the four-month window (in 10-homicide bins), column (3) provides further evidence that the results are not driven by differing homicide levels.52 Third, I exclude elections most vulnerable to strategic political violence by DTOs: column (4) excludes municipalities that average more than one drug-related murder a month over any pre-election year between 2006 and 2011 (when

51Given my focus on short-term shocks, this result does not necessarily conﬂict with previous ﬁndings suggesting that voters are less likely to mobilize in the face of heightened violence (Ley 2014). 52I obtain similar results using smaller bin sizes, including size 1, at the cost of absorbing large municipalities. 214 Table 5.4: Robustness of the effect of a pre-election homicide shocks on municipal incumbent electoral outcomes

Panel A: Change in One- Three- Violence Few Non- No ﬁxed State- First Main Incumbent Single- Placebo Controls incumbent vote share month month bin drug DTO effects year half of party party party 6 months shock shock effects homicides states effects month incumbent trends incumbent earlier (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) Homicide shock -0.015 (one month) (0.011) Homicide shock -0.028*** (three month) (0.009) Homicide shock -0.026*** -0.019** -0.020** -0.019*** -0.017** -0.018** -0.022** -0.018** -0.032*** -0.016** (0.009) (0.009) (0.010) (0.007) (0.007) (0.009) (0.009) (0.008) (0.012) (0.007) Placebo homicide shock -0.006 (0.010)

Observations 161,377 194,799 181,408 152,216 114,142 181,408 181,408 163,483 173,063 173,063 118,779 178,513 161,697 Outcome mean -0.05 -0.05 -0.05 -0.05 -0.06 -0.05 -0.05 -0.05 -0.05 -0.05 -0.04 -0.05 -0.05 Homicide shock mean 0.52 0.48 0.46 0.48 0.46 0.46 0.46 0.46 0.46 0.46 0.45 0.45 0.47 Panel B: Incumbent One- Three- Violence Few Non- No ﬁxed State- First Main Incumbent Single- Placebo Controls party win month month bin drug DTO effects year half of party party party 6 months shock shock effects homicides states effects month incumbent trends incumbent earlier (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)

215 Homicide shock -0.053 (one month) (0.044) Homicide shock -0.123*** (three month) (0.043) Homicide shock -0.119*** -0.078* -0.089* -0.069** -0.101*** -0.098** -0.112*** -0.103** -0.116** -0.063* (0.037) (0.043) (0.050) (0.034) (0.032) (0.042) (0.041) (0.044) (0.057) (0.035) Placebo homicide shock -0.022 (0.046)

Observations 161,377 194,799 181,408 152,216 114,142 181,408 181,408 163,483 173,063 173,063 118,779 178,513 161,697 Outcome mean 0.57 0.56 0.56 0.55 0.54 0.56 0.56 0.56 0.56 0.56 0.56 0.56 0.56 Homicide shock mean 0.52 0.48 0.46 0.48 0.46 0.46 0.46 0.46 0.46 0.46 0.45 0.45 0.47

Notes: Columns (1) and (2) respectively define homicide shocks over one- and three-month windows. Column (3) includes fixed effects for the total number of homicides over the four-month window in bins of size ten. Column (4) excludes municipalities that average more than one drug-related homicide a month over the 12 months before an election between 2006 and 2011. Column (5) excludes states with high-level of DTO activity (see footnote 53). Column (6) excludes municipality and year fixed effects. Column (7) includes state-year fixed effects. Column (8) includes only elections that take place on or before the 15th of the month. Column (9) includes only election where the PAN, PRD, or PRI are an incumbent. Column (10) includes party-specific incumbent trends. Column (11) includes only observations where the incumbent is not a coalition. Column (12) uses a placebo homicide shock defined six months before the election. Columns (13) includes the controls in Table D.4, with the exception of the variables listed in footnote 54. All specifications are estimated using OLS, and all observations are weighted by the number of registered voters in the electoral precinct. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. such data were collected), while column (5) excludes nine states with high DTO-related drug crime.53 In both cases, the results are robust. Fourth, column (6) shows that the results do not depend upon including municipality and year fixed effects. Conversely, column (7) includes state-year fixed effects to ensure that the results are not driven by election-specific state shocks. Column (8) also shows that the results do not depend upon the inclusion of elections that occurred in the second half of the month. Fifth, column (9) restricts attention to PAN, PRD and PRI incumbents, and thus demonstrates that the results are not driven by the few mayors from smaller parties. Moreover, column (10) includes incumbent-specific trends for the three largest parties to ensure that the results are not driven by differential trends in party-incumbent performance. Similarly, column (11) shows that the results are robust to restricting attention to single-party incumbents. Sixth, I conduct a placebo test where a homicide shock is defined six months earlier according to equation (5.3). The results in column (12) show that prior shocks, which could be indicative of pre-trends, do not affect incumbent electoral performance. Finally, column (13) shows that the results are robust to controlling for the balancing variables.54

5.5.4 Voter updating across elections

Voter responses to pre-election homicide shocks may also depend upon the information they received before previous elections. To examine such updating across elections, Figure 5.5 shows how the effect of a homicide shock on the change in the incumbent party’s vote share varies with the occurrence of a homicide shock before the previous election and whether the current incumbent party differs from the previous incumbent party.

53 Baja California, Chihuahua, Durango, Guerrero, Michoacán, Nuevo León, Sinaloa, Sonora, and Tamaulipas. 54This set of controls excludes the few variables with more than 20% missing data, namely homicide victim characteristics, threatened precincts (federal elections only), municipal spending, and child mortality.

216 otelkl rmn fet fciei h es hs nig niaeta oesalso voters that indicate findings these reports. news, violence local the from in learn crime of effects addition priming In likely shock. the homicide to a a over if presided also incumbents party current different a punish from not incumbent consecutive previous do oversees voters incumbent Conversely, same the shocks. if homicide large also pre-election is shock neg- homicide the Similarly, a party. of incumbent effect different ative a such under experience election not previous did the which before and shock election, a expe- current voters the among before greatest shock is homicide sanctioning a riencing that indicates bar first The parties. incumbent iue55 h feto oiiesoko h hnei nubn at oeshare, vote party incumbent in change the on shock homicide a of effect The 5.5: Figure opeetn h uvylvlrsoss h eut ugs htvtr paeabout update voters that suggest results the responses, survey-level the Complementing lcin rvossoki nidctrfrahmcd hc curn eoeteprevious the before occurring shock homicide a for indicator an election. is shock Previous election. Notes ypeiu oiiesok n nubn dniy(5 ofiec intervals) confidence (95% identity incumbent and shocks homicide previous by aeicmeti nidctrfrtecreticmetprywnigteprevious the winning party incumbent current the for indicator an is incumbent Same :

Marginal effect of homicide shock

no previousshock -.1 -.05 0 .05 .1 Diff. incumbent, Diff. incumbent, previous shock 217 no previousshock Same incumbent, Same incumbent, previous shock 5.5.5 Small effects of long-run homicide rates

My theoretical argument implies that voters are particularly likely to respond to homicides that occur just before elections. Longer-run homicide metrics may not affect responsibility attribution if voters are insufﬁciently politically engaged at other times in the political cycle, even if voters (at least intend to) punish such performance (see Healy and Lenz 2014; Healy and Malhotra 2013). To examine whether homicide shocks before elections indeed induce greater incumbent sanctioning than long-run homicide rates, I use a difference-in-differences (DD) design to estimate the effects of homicides over the prior year and prior (three-year) electoral cycle. Speciﬁcally, I compare changes in support for incumbent parties in municipalities that experienced large increases in their longer-run homicide rate between elections to changes in support for incumbent parties in municipalities that did not. This design entails estimating regressions of the following form:

Ypmt = βAverage monthly homicide ratemt + ηm + µt + εpmt. (5.4)

As a robustness check, I also include municipality-speciﬁc time trends to adjust for differential trends in incumbent support. Without conditioning on municipalities experiencing at least one homicide around the election, the sample size increases.55 The results in Table 5.5 conﬁrm that long-run homicide trends have had limited impact on electoral performance. Columns (1) and (2) of panel A show that ten more homicides a month over the year before an election—almost two-thirds of a standard deviation increase—only translate into a 0.6 percentage point decline in the municipal incumbent

55Similar results are obtained using the subsample included in the homicide shock sample.

218 Table 5.5: The effect of long-run homicide rates on municipal incumbent electoral outcomes

Panel A: 12 months before election Change in incumbent Incumbent party vote share party win (1) (2) (3) (4) Average monthly homicide rate -0.0006** -0.0006** 0.0006 0.0001 (12 months before election) (0.0003) (0.0003) (0.0013) (0.0006)

Observations 261,250 261,250 261,250 261,250 Outcome mean -0.06 -0.06 0.54 0.54 Homicide rate mean 5.59 5.59 5.59 5.59 Homicide rate standard deviation 18.41 18.41 18.41 18.41 Municipality-speciﬁc time trends XX Panel B: Since last election Change in incumbent Incumbent party vote share party win (1) (2) (3) (4) Average monthly homicide rate -0.0002 -0.0001 0.0023* 0.0012 (3 years before election) (0.0002) (0.0006) (0.0013) (0.0019)

Observations 260,768 260,768 260,768 260,768 Outcome mean -0.06 -0.06 0.54 0.54 Homicide rate mean 5.32 5.32 5.32 5.32 Homicide rate standard deviation 15.67 15.67 15.67 15.67 Municipality-speciﬁc time trends XX

Notes: All specifications are estimated using OLS, and include municipality and year fixed effects. All observations are weighted by the number of registered voters in the electoral precinct. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. party’s vote share.56 This effect is around four times smaller than the ten-homicide average change associated with a homicide shock. Columns (3) and (4) show this decline in vote share does not affect the incumbent party’s re-election probability. Panel B shows that voters do not punish poor performance over the full electoral cycle. This difference mirrors the findings of Healy and Lenz(2014), who suggest that U.S. voters wish to punish performance over the full electoral cycle but rely on more recent information in its absence.

56The estimate is not quite statistically signiﬁcant if incumbent vote share is instead used as the outcome.

219 The evidence thus suggests that even if longer-run homicide trends do inform electoral accountability, the effects are substantially smaller than short-term homicide shocks.

5.5.6 Municipal, state, and federal incumbent parties: who is being

blamed?

Although mayors play an important role in local public security, state and federal governments are also important players. It is thus possible that punishment of municipal mayors simply reflects broader punishment of the party controlling higher office. Conversely, if voters believe that mayors are primarily responsible for local crime (or their actions are only weakly correlated with co-partisans at higher levels), they may update less about national parties following local homicide shocks. To disentangle these chains of accountability, I first examine how the effects of homicide shocks vary with the existence of a municipal police force, before comparing the electoral performance of parties at the municipal, state and federal levels using data from simultaneous municipal, state and federal elections. If voters believe that the mayor is responsible for local crime rates because they control the local police, homicide shocks should only be punished in municipalities with their own police force. Using the descriptions of municipal public security forces provided by the municipal governments in the 2000, 2002, 2004, 2011 and 2013 ENGM surveys, and im- puting data for missing years, I count municipalities without a police force or those relying on state, federal, community, private, or other security forces.57 Columns (1) and (2) in Ta- ble 5.6 show that, consistent with voters recognizing that elevated homicides rates may be beyond the control of municipal mayors lacking local police forces, homicide shocks only significantly affect the electoral prospects of incumbent parties commanding a local police

57The Appendix describes the imputation. Given the difﬁculty of assigning responsibility, the few municipalities with inter-municipal and civil association-run police are excluded.

220 Table 5.6: Effect of pre-election homicide shocks on municipal incumbent electoral outcomes, by existence of a municipal-level police force and in the Federal District

Change in Incumbent Change in Incumbent incumbent party incumbent party party win party win vote share vote share (Federal (Federal District) District) (1) (2) (3) (4) Homicide shock -0.026*** -0.132*** -0.011 0.016 (0.010) (0.043) (0.059) (0.086) No municipal police force 0.002 -0.044 (0.018) (0.117) Homicide shock 0.051** 0.272** × No municipal police force (0.024) (0.124)

Observations 173,747 173,747 11,849 11,849 Outcome mean -0.05 0.56 0.09 0.90 Homicide shock mean 0.46 0.46 0.49 0.49 No municipal police force mean 0.09 0.09

Notes: Specifications (1) and (2) include municipal and year fixed effects, and are estimated using OLS. Specifications (3) and (4) use only electoral data from the Federal District (otherwise excluded throughout), and include delegation and year fixed effects, and are estimated using OLS. All observations are weighted by the number of registered voters in the electoral precinct. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

221 force.58 Similarly, columns (3) and (4) show that homicide shocks at the delegation-level have no effect in the Federal District, where the Federal District Police are not specific to different delegations. At least when they consume politically-relevant news, these results echo previous research suggesting that voters in developing contexts differentiate incumbent performance from external forces (Harding and Stasavage 2014; Kronick 2014).59 Combined with evidence that voters learn from homicide shocks, this finding also suggests that voters are not indiscriminately punishing events like American football losses (Healy, Malhotra and Mo 2010). Nevertheless, if the parties of municipal mayors, state deputies, state governors, or the president are correlated, the substantial electoral penalties found above could reflect punishment of other political actors. To test for such spillovers down layers of government, columns (1)-(3) of panel A in Table 5.7 respectively examine the effect of a homicide shock on the municipal vote share of the party of the state deputy, state governor, and the presidency. The results, however, suggest that if anything the governor and president’s parties increase their vote share, and thus provide no support for punishment at higher levels of government spilling over to municipal elections.60 Nevertheless, could still reserve punishment for state and federal politicians. However, columns (1) and (2) of panel B similarly indicate that a homicide shock reduces neither the vote share of state nor federal deputies in concurrent state and federal elections. Although I find no evidence that state and federal incumbent parties are held responsi-

58In neither case is the positive effect in municipalities without police forces significantly different from zero. To address the concern that the effect attributed to police forces reflects the lack of sanctioning in smaller and less developed municipalities, Table D.15 shows that this finding is robust to controlling for the interaction of a homicide shock with various correlates of city size and development. 59Table D.16 provides mixed evidence that voters account for homicide shocks affecting neighboring municipalities. 60Consistent with Table D.17 in the Appendix, the positive effect at the federal level reflects increased support for Calderón’s PAN government.

222 Table 5.7: Simultaneous effects of pre-election homicide shocks at different levels of government

Panel A: Spillovers Change in municipal vote share of... down levels of ...state ...state ...president’s government district governor’s party party party (1) (2) (3) Homicide shock -0.001 0.021 0.031* (0.008) (0.015) (0.017)

Observations 154,783 149,445 181,408 Outcome mean 0.01 0.03 0.04 Homicide shock mean 0.47 0.47 0.46 Panel B: Accountability Change in Change in Change in Change in at higher levels of state federal state vote federal vote government incumbent incumbent share of share of vote share vote share municipal municipal incumbent incumbent (1) (2) (3) (4) Homicide shock 0.006 0.003 0.008 -0.020 (0.009) (0.014) (0.016) (0.016)

Observations 145,431 55,764 138,185 67,373 Outcome mean -0.04 -0.05 0.45 0.38 Homicide shock mean 0.46 0.46 0.46 0.46

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

223 ble for homicide shocks by voters, punishment of the local party could instead spill up to the national level. However, when municipal and higher-level elections are held simultaneously, columns (3) and (4) of panel B show that the state and federal deputy vote shares of the municipal incumbent’s party are not affected by a homicide shock. Table D.17 in the Appendix also shows that voters do not differentially punish different parties for homicide shocks.

5.6 The moderating role of local media

Bringing together the individual- and aggregate-level findings, the final component of my argument is that the electoral sanctioning of local violence before elections depends upon access to a media environment covering local news. To demonstrate that the sanctioning of homicide shocks reflects such media coverage, I exploit variation in media signal coverage to identify the effect of an additional local radio or television station in precincts subject to a municipal homicide shock.

5.6.1 Data

As part of Mexico’s major media reform in 2007 (see Larreguy, Marshall and Snyder 2016; Serra 2012), the IFE required that all radio and television stations in the country submit detailed coverage and technological data. This included the location and power of their antennae and a map detailing their commercial quality coverage—the level of coverage that media stations may base advertising sales on and which the IFE deemed relevant for minimizing cross-state advertising spillovers. The signal inside the commercial quality coverage area is very strong, and should cover virtually all households, while signal quality declines quickly as the distance from the commercial quality coverage boundary

224 increases.61 Figure 5.6 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 stations can travel considerable distances—particularly when aided by stretches of sea water with high electrical ground conductivity. Virtually all of the population is covered by at least one media station, but the number of outlets—both emitting from within a municipality and without—providing commercial quality signals to any given electoral precinct varies considerably. Furthermore, the extent to which a given electoral precinct is covered by a signal varies substantially: in some cases, a signal only covers a tiny fraction of a precinct, while in others the entire precinct is covered. I follow Larreguy, Marshall and Snyder(2015) by deﬁning a precinct as covered by a given media station only if at least 20% of voters fall within the commercial coverage boundary.62

5.6.2 Identiﬁcation strategy

To identify the electoral effects of media stations covering local homicides, I leverage geographic variation in coverage (see also Larreguy, Marshall and Snyder 2015).63 I compare neighboring electoral precincts within the same municipality that differ in the total number of local AM, FM and television stations—deﬁned as media stations whose antennae are located within the electoral precinct’s municipality—by which they are cov-

61The IFE deﬁnes the boundary of the coverage area using a 60 dBµ threshold for signal strength. Ac- cordingly, to the U.S.-based National Communications and Information Administration, this “60 dBµ level is recognized as the area in which a reliable signal can be received using an ordinary radio receiver and antenna.” Outside this area, a high-performance antenna is typically required to avoid receiving a weak signal. This is the threshold commonly used to determine a radio station’s audience and sell advertising space commercially in the United States. 62INEGI provides detailed Census population counts in 2010, for both rural localities and urban blocks. This data is used to identify the proportion of voters affected covered by a commercial quality signal. 63Ansolabehere, Snowberg and Snyder(2006), Enikolopov, Petrova and Zhuravskaya(2011), and Snyder and Strömberg(2010) use similar designs.

225 (a) AM radio stations

(b) FM radio stations

Figure 5.6: Media station commercial quality signal coverage areas

226 Figure 5.7: Identification strategy example Notes: Both precincts are located in the municipality of Villa de Tututepec de Melchor Ocampo in the state of Oaxaca. Precinct 1583 is covered by the local television emitting from within the municipality, but precinct 1571 is not. ered.64 Since signal quality declines, but without disappearing entirely, across the commercial quality boundary, this strategy identifies the “intent to treat” effect of an increase in the probability of exposure to an additional local media station.65 Figure 5.7 illustrates the identification strategy graphically, using the example of electoral precincts 1571 and 1583 in the municipality of Villa de Tututepec de Melchor Ocampo in Oaxaca. The identifying assumption is that these neighboring precincts differ only because precinct 1583 receives a commercial quality signal from a local television station that does not cover precinct 1571. Operationally, for each electoral precinct in the country I identify the set of neighboring precincts from within the same municipality that differ in the number of local media stations

64Although 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. 65Exposure to commercial quality coverage does not constitute a geographic regression discontinuity design because the treatment is not binary (some neighbors differ by more one media station) and where neighbors differ by more than one media station it is not clear how to coherently deﬁne the running variable.

227 that they are covered by. Each such grouping n is defined by a “treated” precinct and the set of neighboring precincts receiving different local media exposure. Focusing on municipalities that are included in the homicide shock sample produced 3,190 neighboring groups, containing an average of 2.2 comparison units per election. The average precinct is covered by 6.9, 9.3 and 4.7 local AM, FM and television stations respectively, while the total number of local media stations covering a precinct ranges from 0 to 44.66 There are good reasons to believe that, among neighboring electoral precincts, local media coverage at the commercial quality boundary is effectively random. First, neighboring precincts often differ in coverage because of physical characteristics such as water, geographic contours and large physical objects that 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 boundary. Second, given that media stations lack the technology to differentiate media markets at fine-grained levels,67 and voters that specifically locate according to the availability of local media are unlikely to choose to live close to a coverage boundary (preferring a location guaranteeing coverage), it is unlikely that coverage reflects strategic sorting by either media stations or voters. Finally, I restrict attention to neighboring precincts with an area of less than 2 km2. This removes larger precincts where media outlets could more plausibly target their signals, and prevents comparisons between urban and suburban, or suburban and rural, precincts that may differ in their electoral behavior. The final sample is shown in Figure D.1c. To assess the plausibility of the claim that differences in local media coverage between neighboring precincts are essentially random, I examine balance across a wide range of demographic, socioeconomic, homicide, and political municipal and precinct characteristics.

66Data from the Secretariat of Communications and Transportation shows that the number of media stations has not changed since 2003. 67The IFE technical 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. 228 Table D.8 shows that only eight of these 102 variables are significantly correlated with the number of local media stations at the 10% level (using the main specification used below). In particular, there are no significant differences in key indicators of rural-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. The few significant differences are small in magnitude. Combining within-neighbor variation in local media coverage with homicide shocks just before an election,68 I estimate the following equation to identify the interaction between homicide shocks and local media coverage:

Ypmnt = β1Homicide shockmt + β2Local mediapt + β3 Homicide shockmt × Local mediapt + ξn + µt + εpmnt, (5.5)

where ξn is a ﬁxed effect for each set of neighboring precincts that ensures I only exploit within-neighbor variation in local media.69 To weight each neighboring group equally, precincts are weighted by electorate size divided by the number of control units per neighbor set in order.

5.6.3 Local media increase punishment of homicide shocks

The estimates in Table 5.8 show that access to local media stations plays a key role in supporting the electoral sanctioning of homicide shocks. The central ﬁnding in column (1) shows that homicide shocks are only punished when an electoral precinct is covered by sufﬁcient local media stations. In the relatively rare case of precincts with fewer than eight local media stations (21% of the sample), the marginal effect plot in Figure 5.8 indicates

68Homicide shocks remain well-balanced across pre-treatment variables in this subsample. 69Since I use within-municipality neighbors, neighbor fixed effects incorporate municipality fixed effects. 229 egbrdfeecsi h ubro oa M Madtlvso ttosseparately. stations television and FM AM, local of number the within- in exploit differences samples—I different neighbor in each—albeit of contribution the cor- identify is to coverage related, television and radio Since parties. incumbent of rewarding and sanctioning election. the before shock homicide that a experience suggests not term do that media incumbents local reward may lower-order voters the on and coefficient Marshall positive Larreguy, the and (2015), (2008) Snyder Finan and Ferraz of Like share party. vote incumbent the municipal reduce the substantially shocks homicide and significant in the kicks sample), coefficient the of interaction exactly (79% almost stations is media local stations eight media than local more for no However, with zero. precinct electoral effect an the in violence, shock pre-election homicide for a party of incumbent neces- the are punish media to local voters that Mexican Implying for shock. sary homicide a of effect significant no is there that media local of number the on conditional shock, homicide a of effect Electoral 5.8: Figure oun 2-4 xmn hc ye flclmdacnrbt ott h electoral the to most contribute media local of types which examine (2)-(4) Columns the Notes x xspostedsrbto ftelclmdavral ntesample. the in variable media local the of distribution the plots axis h rydniypo above plot density gray The 5.8. Table of (1) column from estimates the using Calculated :

Marginal effect of a homicide shock -.2 -.15 -.1 -.05 0 .05 0 ttos(5 oﬁec interval) conﬁdence (95% stations 10 Number oflocalmediastations 230 20 30 40 Table 5.8: Electoral effects of homicide rates, conditional on the presence of local media

Change in incumbent party vote share Media measure: Local media Local AM media Local FM media Local television Non-local media Local media (1) (2) (3) (4) (5) (6) Homicide shock -0.0171 -0.0383 -0.0565** 0.0064 -0.0128 0.0092 (0.0253) (0.0412) (0.0268) (0.0274) (0.0180) (0.0259) Media measure 0.0015 0.0039 0.0010 0.0062 0.001 0.0022** (0.0009) (0.0035) (0.0017) (0.0038) (0.0006) (0.0009) Homicide shock × Media measure -0.0023** -0.0045 -0.0012 -0.0165*** -0.0003 -0.0036*** (0.0011) (0.0041) (0.0025) (0.0052) (0.0004) (0.0012) High higher education 0.0205* (0.0107) 231 Homicide shock × High higher education -0.0543*** (0.0207) Media measure × High higher education -0.0013** (0.0005) Homicide shock × Media measure 0.0026** × High higher education (0.0012)

Observations 38,706 1,531 18,088 19,483 105,001 38,706 Outcome mean -0.03 -0.01 -0.03 -0.04 -0.05 -0.03 Outcome range [-0.63,0.47] [-0.50,0.30] [-0.63,0.41] [-0.62,0.47] [-0.75,0.73] [-0.63,0.47] Homicide shock mean 0.38 0.59 0.34 0.39 0.45 0.38 Media measure mean 20.31 4.13 10.96 4.15 36.90 20.31 Media measure std. dev. 11.19 4.58 4.86 2.45 23.51 11.19

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 in the electoral precinct 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. Given that television is by far the primary media source for Mexicans, it is not surprising to find that local television stations produce the largest effects. As column (4) demonstrates, each additional local television station reduces the vote share of an incumbent facing a homicide shock by 1.7 percentage points. The smaller effects of FM and especially AM radio are relatively similar in magnitude to the average effect reported in column (1), but are not statistically significant in these subsamples. To assess the importance of local media stations, as opposed to media stations emitting from other municipalities, I use the same strategy to identify the effects of non-local media. Consistent with the findings of Larreguy, Marshall and Snyder(2015), column (5) shows that non-local media stations with weaker incentives to report homicides outside their own municipality do not affect how voters sanction their incumbents. This reiterates the importance of voters receiving relevant news about their incumbent party. Finally, I further link the individual-level findings to the aggregate level by examining whether the electoral effect of local media covering a homicide shock also varies with the strength of voter prior beliefs. I proxy for voter priors at the precinct level using an indicator for precincts where more than 40% (the sample median) of adults have some post-secondary level of education, and further interact the homicide shock and local media variables with this proxy. Consistent with the survey findings in Figure 5.3, the significant triple interaction in column (6) suggests that voters with weaker priors indeed respond more at the ballot box to performance information available in the local news.70

70Contrary to the potential confound that the least educated respond most to homicide shocks in the news because local violence afﬂicts such voters most heavily (Díaz-Cayeros et al. 2011), the education triple interaction becomes stronger once triple interactions using precinct-level poverty measures are simultaneously controlled for.

232 5.7 Conclusion

This study demonstrates the importance of the timing of voter news consumption for understanding how voters hold governments to account. Since many voters follow relatively sharp political information cycles, whereby most politically-relevant information is consumed prior to elections, voting behavior can be highly responsive to salient performance indicators, like local homicides, in the news before elections. Leveraging a wealth of fine-grained individual- and precinct-level data, and identification strategies isolating three key components of the theory, I show that voters indeed consume more news before elections, and that pre-election homicide shocks substantially reduce support for the municipal incumbent party among poorly informed voters exposed to local media. Conversely, there is little evidence that long-run performance indicators affect electoral outcomes. News revealed about incumbent performance before elections on other valence issues may similarly impact voting behavior. Ferraz and Finan(2008) and Larreguy, Marshall and Snyder(2015) show respectively that Brazilian and Mexican municipal incumbent parties are punished electorally if audits reports reveal malfeasant behavior in office in the run-up to elections, but only in the presence of local media stations. Brollo(2009) finds that such sanctioning in Brazil decreases with months until the election. Banerjee et al.(2011) show that Indian voters respond to incumbent performance reports cards released in newspapers just before elections. Furthermore, Italian deputies were only punished for alleged criminal wrongdoing following the “Clean Hands” investigation, which was widely reported in the news before the 1992 and 1994 elections (Chang, Golden and Hill 2010). Economic outcomes paint a similar picture. In particular, Latin America’s substantial electoral volatility is correlated with pre-election economic fluctuations (e.g. Roberts and Wibbels 1999). Even in the U.S., Achen and Bartels(2004 b) suggest that economic voting only responds to economic indicators covering the quarters prior to presidential elections. 233 These studies neglect the fundamental role of the timing of voter news consumption. Nevertheless, they suggest that political information cycles may play a central role in explaining why voters often hold governments accountable for salient short-term indicators of performance, other than violence, in a wide variety of contexts. Since a key feature of the Mexican case is that the least informed voters respond most to pre-election information, future research might fruitfully compare these responses to pre-election news with Western Europe’s more politically engaged citizens. Understanding the origins of political information cycles requires further work probing the forces underlying equilibrium information consumption. In particular, voter demand for information remains poorly understood, especially in developing contexts where the marginal effects of information are relatively substantial (Marshall 2016b). Similarly, little is known about when and how media stations report different types of news (although see Lawson 2002). This paper focused on identifying equilibrium cycles and estimating their electoral implications, but these foundational questions demand attention from the increasing number of scholars attributing important effects to political information. A second fundamental question is exactly why voters respond to noisy performance indicators. Such behavior is consistent with rational voters possessing very weak priors using short- term indicators that they believe to be correlated with longer-run performance. Similarly, voters may be observing the government’s response to local homicides, or lack of it, for the first time. Accordingly, how voters process and perceive these indicators merits greater attention. However, voting on the basis of the news before elections may be normatively problematic for democracy. In contrast with models emphasizing the value of information for improving democratic representation (e.g. Besley and Burgess 2002; Besley and Prat 2006; Casey 2015), the case of punishing exogenous homicide shocks prior to elections—which are uncorrelated with longer-run homicide trends—suggests that voters could be commit- 234 ting attribution errors by removing competent incumbents on the basis of noisy short-term performance indicators (see also Achen and Bartels 2004b). Alternatively, voters may be severely punishing parties when they learn that they cannot effectively address crises (Cole, Healy and Werker 2012). Future research should illuminate whether voters indeed elect worse incumbents on the basis of such information, which is itself likely to depend upon how voters view incumbents on average (Ashworth and Bueno de Mesquita 2014b). Never- theless, the benefits of a better-informed electorate depend on the type of information voters acquire and the weight they attach to it. The political information cycles documented here thus add to an increasing body of work challenging the orthodoxy that an informed electorate necessarily enhances electoral accountability (see Ashworth and Bueno de Mesquita 2014a). A key implication is that policy-makers seeking to enhance democratic representation should convey the right performance information at the right time. The results suggest that NGOs and independent media outlets should focus their incumbent performance information around elections—when voters consume political news. If voters substitute short-term performance indices for longer-term measures that they intend to evaluate (Healy and Lenz 2014), it is essential that the news relate to horizons beyond proximate security or economic shocks. Although recent interventions have attempted to achieve this, often by delivering leaflets, broadcast media appears to be the most effective tool for dissemination. Con- sequently, a key challenge is encouraging media stations to overcome sensationalist and political biases to provide longer-term performance indicators.

235 A| Appendix to Chapter 2

Table A.1 presents the results of our balance tests. Table A.2 shows that our survey- level estimates of the effect of the treatment on incumbent competence posterior beliefs are robust to excluding the municipality of Ecatepec de Morelos. Tables A.3-A.8 show our survey-level estimates of the effect of the treatment on challenger competence posterior beliefs, deploying 3 different definitions of challengers, and also comparing the effect of local and comparative performance information.1 Note that good and bad news are excluded because not all coefficients could be estimated due to lack of data. Tables A.9-A.11 similarly show challenger beliefs do not substantially impact incumbent vote share outcomes. In particular, column (4) consistently shows no significant positive interaction with the share of malfeasant spending engaged in by challenger, while column (5) shows that if anything negative updating about the challenger decreases the incumbent’s vote share. Moreover, the positive interaction with the challenger malfeasance prior indicates that treated precincts with the worst priors about the challenger reward the incumbent most, but this therefore cannot be attributed to negative updating about the challenger because such responses already occur among those with the least favorable priors about the challenger. Tables A.12 and A.13 show similar results defining priors and updating from the control group at the municipal level. Table A.14 presents the robustness checks where the incumbent vote share, as a share of registered voters, is the outcome. Finally, Tables A.15 and A.16 report the ag-

1The single block from Tamasopo is dropped for our second challenger deﬁnition because we did not ask about the second-placed party (MC) in that municipality. 236 gregate level estimates distinguishing each of our 4 treatment conﬁgurations.

237 Table A.1: Effect of 46 pre-treatment variables

Information treatment effect Standard error Observations Precinct-level covariates Area -0.637 (0.717) 675 Population -26.237 (36.344) 675 Number of households -6.831 (8.787) 675 Number of private dwellings -9.930 (11.059) 675 Average occupants dwelling 0.014 (0.016) 675 Average occupants per room 0.006 (0.008) 675 Share of homes with 2+ rooms 0.001 (0.006) 675 Share of homes with 3+ rooms 0.001 (0.006) 675 Average years of schooling -0.107* (0.054) 675 Share married 0.001 (0.002) 675 Share working age -0.001 (0.001) 675 Share economically active 0.000 (0.002) 675 Share without health care 0.011** (0.005) 675 Share with state workers health care 0.000 (0.002) 675 Share old 0.001 (0.002) 675 Average children per woman 0.042*** (0.015) 675 Share of households with male head 0.001 (0.003) 675 Share born out of state 0.006 (0.006) 675 Share indigenous speakers 0.008** (0.004) 675 Share of homes without a dirt ﬂoor -0.001 (0.003) 675 Share of homes with a toilet -0.001 (0.005) 675 Share of homes with water 0.002 (0.009) 675 Share of homes with drainage -0.004 (0.006) 675 Share of omes with electricity 0.002 (0.003) 675 Share of homes with water, drainage and electricity -0.008 (0.009) 675 Share of homes with a washing machine 0.004 (0.005) 675 Share of homes with a landline telephone -0.016** (0.007) 675 Share of homes with a radio 0.000 (0.003) 675 Share of homes with a fridge -0.001 (0.006) 675 Share of homes with a cell phone 0.008 (0.005) 675 Share of homes with a television -0.004 (0.003) 675 Share of homes with a car -0.005 (0.006) 675 Share of homes with a computer -0.007 (0.006) 675 Share of homes with internet -0.006 (0.006) 675 Turnout in 2012 0.007** (0.003) 675 Incumbent vote margin in 2012 -0.017*** (0.006) 675 Incumbent vote share in 2012 0.014*** (0.005) 675

Survey-level covariates Female 0.020 (0.018) 4,958 Age -0.528 (0.531) 4,869 Education -0.062 (0.133) 4,948 Income -0.043 (0.081) 4,402 Income (log) -0.010 (0.017) 4,402 Employed -0.006 (0.014) 4,950 Turnout in 2012 0.004 (0.012) 4,958 Incumbent vote in 2012 -0.007 (0.021) 3,122 Political knowledge index 0.006 (0.025) 4,958

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality- treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

238 Table A.2: Effect of information treatment on voter beliefs about incumbent party malfeasance, removing Ecatepec de Morelos

Perceived incumbent party malfeasance (1) (2) (3) (4) (5) (6) Information treatment 0.019 -0.046 0.900** 0.002 -0.161** -0.103 (0.041) (0.057) (0.336) (0.067) (0.073) (0.078) × Incumbent malfeasance prior -0.283*** (0.054) × Strength incumbent prior -0.272** (0.107) × Incumbent malfeasant spending 0.091 (0.197) × Negative incumbent updating 0.173*** (0.042) 239 × Good news -0.295*** (0.100) × Bad news 0.218** (0.101)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Outcome mean -0.21 -0.21 -0.26 -0.21 -0.21 -0.21 Outcome std. dev. 1.46 1.46 1.45 1.46 1.46 1.46 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.20 3.23 0.19 1.00 0.03 / 0.58 Interaction std. dev. 0.85 0.38 0.16 1.05 0.18 / 0.49 Observations 4,181 4,181 4,181 4,181 4,181 4,181

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.3: Effect of information treatment on voter beliefs about challenger party malfeasance, where the challenger is each voter’s second preferred party

Perceived challenger party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) (7) Information treatment -0.029 -0.134*** -0.292 -0.072 -0.024 -0.166*** -0.040 (0.032) (0.042) (0.304) (0.062) (0.042) (0.039) (0.040) × Incumbent malfeasance prior -0.589*** (0.064) × Strength challenger prior 0.085 (0.096) × Incumbent malfeasant spending 0.482 (0.677) × Difference in malfeasant spending -0.038 (0.180)

240 × Negative incumbent updating 0.213*** (0.044) × Good news -0.643*** (0.066) × Bad news 0.083 (0.070)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.19 -0.19 -0.19 -0.19 -0.19 -0.19 -0.19 Control outcome std. dev. 1.30 1.30 1.30 1.30 1.30 1.30 1.30 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.22 0.21 0.21 0.89 0.04 / 0.55 Interaction std. dev. 0.87 0.39 0.17 0.17 1.05 0.19 / 0.50 R2 0.08 0.09 0.08 0.08 0.08 0.09 0.09 Observations 4,958 4,958 4,958 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.4: Effect of information treatment on voter beliefs about challenger party malfeasance, where the challenger is the party that received the largest vote share in the last municipal election

Perceived challenger party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) (7) Information treatment -0.007 -0.109** -0.560 -0.039 -0.005 -0.125*** 0.050 (0.038) (0.053) (0.393) (0.073) (0.051) (0.043) (0.075) × Incumbent malfeasance prior -0.382*** (0.053) × Strength challenger prior 0.177 (0.125) × Incumbent malfeasant spending 0.359 (0.908) × Difference in malfeasant spending -0.010 (0.218)

241 × Negative incumbent updating 0.162*** (0.036) × Good news -0.349*** (0.122) × Bad news 0.000 (0.091)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.30 -0.30 -0.31 -0.30 -0.30 -0.30 -0.30 Control outcome std. dev. 1.36 1.36 1.37 1.36 1.36 1.36 1.36 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.23 0.21 0.21 0.89 0.04 / 0.55 Interaction std. dev. 0.87 0.38 0.17 0.17 1.05 0.19 / 0.50 R2 0.19 0.19 0.19 0.19 0.19 0.19 0.19 Observations 4,958 4,958 4,958 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.5: Effect of information treatment on voter beliefs about challenger party malfeasance, where the challenger is the average perception of PAN, PRI and PRD perceptions where they are not the municipal incumbent

Perceived challenger party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) (7) Information treatment 0.009 -0.074* -0.463 -0.119** 0.049 -0.100*** -0.024 (0.034) (0.042) (0.369) (0.052) (0.042) (0.033) (0.070) × Incumbent malfeasance prior -0.288*** (0.050) × Strength challenger prior 0.147 (0.116) × Incumbent malfeasant spending 1.425** (0.691) × Difference in malfeasant spending -0.320 (0.212)

242 × Negative incumbent updating 0.144*** (0.032) × Good news -0.195 (0.126) × Bad news 0.116 (0.086)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.33 -0.33 -0.33 -0.33 -0.33 -0.33 -0.33 Control outcome std. dev. 1.20 1.20 1.20 1.20 1.20 1.20 1.20 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.22 0.21 0.21 0.89 0.04 / 0.55 Interaction std. dev. 0.87 0.39 0.17 0.17 1.05 0.19 / 0.50 R2 0.30 0.30 0.30 0.30 0.30 0.30 0.30 Observations 4,958 4,958 4,958 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.6: Effect of local and comparative information treatments on voter beliefs about challenger party malfeasance, where the challenger is each voter’s second preferred party

Perceived challenger party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) Local information treatment -0.005 -0.106** -0.178 -0.034 -0.030 -0.122*** (0.037) (0.047) (0.348) (0.073) (0.047) (0.044) Comparative information treatment -0.054 -0.162*** -0.410 -0.110 -0.017 -0.211*** (0.041) (0.047) (0.345) (0.080) (0.056) (0.044) Local × Challenger malfeasance prior -0.555*** (0.077) Comparative × Challenger malfeasance prior -0.619*** (0.072) Local × Strength challenger prior 0.056 (0.109) Comparative × Strength challenger prior 0.115 (0.110) Local × Challenger malfeasant spending 0.319 (0.788) 243 Comparative × Challenger malfeasant spending 0.627 (0.911) Local × Difference in malfeasant spending 0.195 (0.196) Comparative × Difference in malfeasant spending -0.289 (0.222) Local × Negative challenger updating 0.178*** (0.049) Comparative × Negative challenger updating 0.249*** (0.050)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.19 -0.19 -0.19 -0.19 -0.19 -0.19 Control outcome std. dev. 1.30 1.30 1.30 1.30 1.30 1.30 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.22 0.21 0.21 0.89 Interaction std. dev. 0.87 0.39 0.17 0.17 1.05 R2 0.08 0.09 0.08 0.08 0.09 0.09 Observations 4,958 4,958 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.7: Effect of local and comparative information treatments on voter beliefs about challenger party malfeasance, where the challenger is the party that received the largest vote share in the last municipal election

Perceived challenger party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) Local information treatment 0.027 -0.075 -0.527 -0.008 0.007 -0.086* (0.042) (0.057) (0.460) (0.077) (0.052) (0.049) Comparative information treatment -0.042 -0.144** -0.600 -0.069 -0.018 -0.165*** (0.043) (0.056) (0.420) (0.079) (0.060) (0.046) Local × Challenger malfeasance prior -0.370*** (0.054) Comparative × Challenger malfeasance prior -0.391*** (0.062) Local × Strength challenger prior 0.177 (0.145) Comparative × Strength challenger prior 0.179 (0.135) Local × Challenger malfeasant spending 0.388

244 (0.960) Comparative × Challenger malfeasant spending 0.297 (1.024) Local × Difference in malfeasant spending 0.154 (0.228) Comparative × Difference in malfeasant spending -0.186 (0.257) Local × Negative challenger updating 0.151*** (0.036) Comparative × Negative challenger updating 0.171*** (0.043)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.30 -0.30 -0.31 -0.30 -0.30 -0.30 Control outcome std. dev. 1.36 1.36 1.37 1.36 1.36 1.36 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.23 0.21 0.21 0.89 Interaction std. dev. 0.87 0.38 0.17 0.17 1.05 Observations 4,958 4,958 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.8: Effect of local and comparative information treatments on voter beliefs about challenger party malfeasance, where the challenger is the average perception of PAN, PRI and PRD perceptions where they are not the municipal incumbent

Perceived challenger party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) Local information treatment 0.032 -0.051 -0.405 -0.115** 0.064 -0.075** (0.037) (0.045) (0.401) (0.056) (0.044) (0.037) Comparative information treatment -0.014 -0.097** -0.528 -0.121** 0.033 -0.124*** (0.039) (0.046) (0.373) (0.060) (0.051) (0.038) Local × Challenger malfeasance prior -0.280*** (0.052) Comparative × Challenger malfeasance prior -0.294*** (0.056) Local × Strength challenger prior 0.136 (0.125) Comparative × Strength challenger prior 0.159 (0.118) Local × Challenger malfeasant spending 1.631** (0.721) 245 Comparative × Challenger malfeasant spending 1.191 (0.815) Local × Difference in malfeasant spending -0.266 (0.226) Comparative × Difference in malfeasant spending -0.377 (0.236) Local × Negative challenger updating 0.140*** (0.033) Comparative × Negative challenger updating 0.148*** (0.037)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.33 -0.33 -0.33 -0.33 -0.33 -0.33 Control outcome std. dev. 1.20 1.20 1.20 1.20 1.20 1.20 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.22 0.21 0.21 0.89 Interaction std. dev. 0.87 0.39 0.17 0.17 1.05 R2 0.30 0.30 0.30 0.30 0.30 0.30 Observations 4,958 4,958 4,958 4,958 4,958 4,958

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.9: Effect of information treatment on incumbent party vote share, using challenger priors and updating where the challenger is defined by each voter’s second preferred party

Incumbent party vote share (1) (2) (3) (4) (5) (6) (7) Panel A: Incumbent party vote share (share of turnout) Information treatment 0.020*** 0.022*** 0.100** 0.020** 0.025*** 0.023*** 0.024*** (0.004) (0.004) (0.048) (0.009) (0.004) (0.005) (0.006) × Challenger malfeasance prior 0.013 (0.008) × Strength challenger prior -0.026* (0.015) × Challenger malfeasant spending -0.003 (0.092) × Difference in malfeasant spending -0.047** (0.020) × Negative challenger updating -0.006 (0.005) × Good news 0.010 (0.032) × Bad news -0.008 (0.007)

Outcome range [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] Control outcome mean 0.38 0.38 0.38 0.38 0.38 0.38 0.38 Control outcome std. dev. 0.12 0.12 0.12 0.12 0.12 0.12 0.12 R2 0.61 0.61 0.61 0.61 0.61 0.61 0.61 Panel B: Incumbent party vote share (share of registered voters) Information treatment 0.008*** 0.010*** 0.026 0.009* 0.011*** 0.011*** 0.009*** (0.002) (0.002) (0.028) (0.005) (0.002) (0.003) (0.003) × Challenger malfeasance prior 0.012** (0.005) × Strength challenger prior -0.006 (0.009) × Challenger malfeasant spending -0.007 (0.048) × Difference in malfeasant spending -0.026** (0.011) × Negative challenger updating -0.004 (0.003) × Good news 0.008 (0.014) × Bad news -0.004 (0.004)

Outcome range [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] Control outcome mean 0.19 0.19 0.19 0.19 0.19 0.19 0.19 Control outcome std. dev. 0.07 0.07 0.07 0.07 0.07 0.07 0.07 R2 0.62 0.62 0.62 0.62 0.62 0.62 0.62 Treatment mean 0.59 0.59 0.59 0.59 0.59 0.59 0.59 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 0.49 0.49 Interaction mean -0.07 3.23 0.21 0.21 0.87 0.04 / 0.54 Interaction std. dev. 0.87 0.38 0.17 0.17 1.04 0.19 / 0.50 Observations 675 675 675 675 675 675 675

246 Table A.10: Effect of information treatment on incumbent party vote share, using challenger priors and updating where the challenger is the party that received the largest vote share in the last municipal election

Incumbent party vote share (1) (2) (3) (4) (5) (6) (7) Panel A: Incumbent party vote share (share of turnout) Information treatment 0.020*** 0.023*** 0.091* 0.020** 0.025*** 0.025*** 0.021** (0.004) (0.004) (0.052) (0.009) (0.004) (0.004) (0.010) × Challenger malfeasance prior 0.015*** (0.004) × Strength challenger prior -0.023 (0.016) × Challenger malfeasant spending -0.003 (0.092) × Difference in malfeasant spending -0.047** (0.020) × Negative challenger updating -0.007** (0.003) × Good news 0.014 (0.011) × Bad news -0.006 (0.011)

Outcome range [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] Control outcome mean 0.38 0.38 0.38 0.38 0.38 0.38 0.38 Control outcome std. dev. 0.12 0.12 0.12 0.12 0.12 0.12 0.12 R2 0.61 0.61 0.61 0.61 0.61 0.61 0.61 Panel B: Incumbent party vote share (share of registered voters) Information treatment 0.008*** 0.010*** 0.024 0.009* 0.011*** 0.011*** 0.007 (0.002) (0.002) (0.030) (0.005) (0.002) (0.003) (0.005) × Challenger malfeasance prior 0.008** (0.004) × Strength challenger prior -0.005 (0.010) × Challenger malfeasant spending -0.007 (0.048) × Difference in malfeasant spending -0.026** (0.011) × Negative challenger updating -0.004 (0.002) × Good news 0.008 (0.008) × Bad news -0.001 (0.006)

Outcome range [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] Control outcome mean 0.19 0.19 0.19 0.19 0.19 0.19 0.19 Control outcome std. dev. 0.07 0.07 0.07 0.07 0.07 0.07 0.07 R2 0.62 0.62 0.61 0.62 0.62 0.62 0.62 Treatment mean 0.59 0.59 0.59 0.59 0.59 0.59 0.59 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 0.49 0.49 Interaction mean -0.08 3.24 0.21 0.21 0.89 0.04 /0.52 Interaction std. dev. 0.89 0.36 0.17 0.17 1.08 0.20 / 0.50 Observations 675 675 668 675 675 675 675

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in column (3) reflects a lack of challenger prior data in Tamasopo. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

247 Table A.11: Effect of information treatment on incumbent party vote share, using challenger priors and updating where the challenger is the average perception of PAN, PRI and PRD perceptions where they are not the municipal incumbent

Incumbent party vote share (1) (2) (3) (4) (5) (6) (7) Panel A: Incumbent party vote share (share of turnout) Information treatment 0.020*** 0.023*** 0.093* 0.020** 0.025*** 0.024*** 0.023** (0.004) (0.004) (0.052) (0.009) (0.004) (0.004) (0.010) × Challenger malfeasance prior 0.012*** (0.003) × Strength challenger prior -0.023 (0.016) × Challenger malfeasant spending -0.003 (0.092) × Difference in malfeasant spending -0.047** (0.020) × Negative challenger updating -0.006** (0.003) × Good news 0.009 (0.013) × Bad news -0.008 (0.011)

Outcome range [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] Control outcome mean 0.38 0.38 0.38 0.38 0.38 0.38 0.38 Control outcome std. dev. 0.12 0.12 0.12 0.12 0.12 0.12 0.12 R2 0.61 0.61 0.61 0.61 0.61 0.61 0.61 Panel B: Incumbent party vote share (share of registered voters) Information treatment 0.008*** 0.010*** 0.024 0.009* 0.011*** 0.010*** 0.009* (0.002) (0.002) (0.029) (0.005) (0.002) (0.003) (0.005) × Challenger malfeasance prior 0.006** (0.002) × Strength challenger prior -0.005 (0.009) × Challenger malfeasant spending -0.007 (0.048) × Difference in malfeasant spending -0.026** (0.011) × Negative challenger updating -0.003 (0.002) × Good news 0.003 (0.008) × Bad news -0.002 (0.006)

248 Table A.12: Effect of information treatment on voter beliefs about incumbent party malfeasance, using priors and updating deﬁned by control precincts at the municipal level

Perceived incumbent party malfeasance (very low - very high) (1) (2) (3) (4) (5) (6) Information treatment -0.001 -0.015 0.429 0.016 -0.096** -0.054 (0.040) (0.037) (0.482) (0.067) (0.047) (0.044) × Incumbent malfeasance prior -0.126*** (0.035) × Strength incumbent prior -0.133 (0.151) × Incumbent malfeasant spending -0.083 (0.214) × Negative incumbent updating 0.102***

249 (0.030) × Good news 0.132*** (0.045) × Bad news 0.087 (0.075)

Outcome range {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} {-2,-1,0,1,2} Control outcome mean -0.14 -0.14 -0.14 -0.14 -0.14 -0.14 Control outcome std. dev. 1.48 1.48 1.48 1.48 1.48 1.48 Treatment mean 0.77 0.77 0.77 0.77 0.77 0.77 Treatment std. dev. 0.42 0.42 0.42 0.42 0.42 0.42 Interaction mean -0.08 3.24 0.21 0.89 0.04 / 0.52 Interaction std. dev. 0.89 0.37 0.17 1.07 0.20 / 0.50 R2 0.29 0.29 0.29 0.29 0.29 0.29 Observations 4,624 4,624 4,624 4,624 4,624 4,624

Notes: All specifications include block fixed effects, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A.13: Effect of information treatment on incumbent party vote share, using priors and updating defined by control precincts at the municipal level

Incumbent party vote share (1) (2) (3) (4) (5) (6) Panel A: Incumbent party vote share (share of turnout) Information treatment 0.029*** 0.027*** 0.203*** 0.046*** 0.036*** 0.020*** (0.006) (0.006) (0.073) (0.008) (0.007) (0.007) × Incumbent malfeasance prior 0.011* (0.006) × Strength incumbent prior -0.055** (0.023) × Incumbent malfeasant spending -0.080*** (0.027) × Negative incumbent updating -0.011** (0.005) × Good news 0.084** (0.034) × Bad news 0.005 (0.014)

Outcome range [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] Control outcome mean 0.38 0.39 0.39 0.38 0.39 0.39 Control outcome std. dev. 0.12 0.12 0.12 0.12 0.12 0.12 R2 0.63 0.62 0.62 0.63 0.62 0.62 Panel B: Incumbent party vote share (share of registered voters) Information treatment 0.016*** 0.015*** 0.119** 0.027*** 0.021*** 0.014*** (0.004) (0.004) (0.052) (0.006) (0.004) (0.004) × Incumbent malfeasance prior 0.007** (0.003) × Strength incumbent prior -0.033** (0.016) × Incumbent malfeasant spending -0.050*** (0.016) × Negative incumbent updating -0.007*** (0.002) × Good news 0.038** (0.015) × Bad news -0.002 (0.009)

Outcome range [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] Control outcome mean 0.19 0.20 0.20 0.19 0.20 0.20 Control outcome std. dev. 0.07 0.07 0.07 0.07 0.07 0.07 R2 0.66 0.65 0.66 0.66 0.66 0.66 Treatment mean 0.59 0.59 0.59 0.59 0.59 0.59 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 0.49 Interaction mean -0.10 3.23 0.21 0.91 0.04 / 0.52 Interaction std. dev. 0.83 0.26 0.17 1.00 0.20 / 0.50 Observations 675 651 651 675 651 651

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in columns (2), (3), (5) and (6) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

250 Table A.14: Robustness of information treatment on incumbent party vote share (registered voters)

Incumbent party vote share (registered voters) (1) (2) (3) (4) (5) (6) Panel A: Removing Ecatepec de Morelos’ 64 precincts Information treatment 0.016*** 0.016*** 0.002 0.028*** 0.021*** 0.013*** (0.004) (0.004) (0.040) (0.006) (0.005) (0.005) × Incumbent malfeasance prior 0.008* (0.004) × Strength incumbent prior 0.004 (0.013) × Incumbent malfeasant spending -0.065*** (0.015) × Negative incumbent updating -0.007*** (0.003) × Good news 0.058*** (0.008) × Bad news -0.001 (0.010) Panel B: Controlling for covariates Information treatment 0.011*** 0.011*** 0.006 0.020*** 0.016*** 0.011** (0.003) (0.003) (0.038) (0.005) (0.004) (0.004) × Incumbent malfeasance prior 0.008*** (0.003) × Strength incumbent prior 0.003 (0.012) × Incumbent malfeasant spending -0.042*** (0.015) × Negative incumbent updating -0.007*** (0.002) × Good news 0.046* (0.026) × Bad news -0.004 (0.006) Panel C: Unweighted precinct estimates Information treatment 0.008*** 0.008*** 0.021 0.014*** 0.012*** 0.010*** (0.002) (0.002) (0.025) (0.003) (0.003) (0.003) × Incumbent malfeasance prior 0.005** (0.002) × Strength incumbent prior -0.004 (0.007) × Incumbent malfeasant spending -0.029** (0.013) × Negative incumbent updating -0.005** (0.002) × Good news 0.036*** (0.012) × Bad news -0.008 (0.005)

251 Table A.15: Effect of information treatment variants on incumbent party vote share (turnout)

Incumbent party vote share (share of turnout) (1) (2) (3) (4) (5) Private local treatment 0.046*** 0.047*** -0.032 0.072*** 0.047*** (0.012) (0.012) (0.097) (0.020) (0.013) Public localinformation treatment 0.006 0.005 0.183 0.023 0.026* (0.014) (0.012) (0.129) (0.025) (0.014) Private comparativeinformation treatment 0.033*** 0.027** -0.011 0.062*** 0.041*** (0.012) (0.011) (0.064) (0.016) (0.012) Public comparative information treatment 0.031*** 0.030*** 0.160 0.025 0.023* (0.010) (0.010) (0.115) (0.016) (0.012) Private local × Incumbent malfeasance prior 0.005 (0.010) Public local× Incumbent malfeasance prior 0.029*** (0.011) Private comparative× Incumbent malfeasance prior 0.017* (0.010) Public comparative × Incumbent malfeasance prior -0.012 (0.009) Private local × Strength incumbent prior 0.024 (0.032) Public local× Strength incumbent prior -0.056 (0.040) Private comparative× Strength incumbent prior 0.011 (0.021) Public comparative × Strength incumbent prior -0.040 (0.035) Private local × Incumbent malfeasant spending -0.125** (0.056) Public local× Incumbent malfeasant spending -0.078 (0.088) Private comparative× Incumbent malfeasant spending -0.131*** (0.041) Public comparative × Incumbent malfeasant spending 0.024 (0.063) Private local × Negative incumbent updating -0.001 (0.008) Public local× Negative incumbent updating -0.027** (0.011) Private comparative× Negative incumbent updating -0.017** (0.008) Public comparative × Negative incumbent updating 0.009 (0.008)

Outcome range [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] [0.07,0.85] Control outcome mean 0.19 0.20 0.20 0.19 0.20 Control outcome std. dev. 0.07 0.07 0.07 0.07 0.07 Treatment mean 0.59 0.59 0.59 0.59 0.59 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 Interaction mean -0.08 3.24 0.21 0.89 Interaction std. dev. 0.89 0.37 0.17 1.08 R2 0.66 0.67 0.66 0.67 0.67 Observations 675 651 651 675 651

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in columns (2), (3) and (5) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

252 Table A.16: Effect of information treatment variants on incumbent party vote share (registered voters)

Incumbent party vote share (registered voters) (1) (2) (3) (4) (5) Private local treatment 0.028*** 0.029*** -0.018 0.045*** 0.031*** (0.010) (0.010) (0.054) (0.016) (0.010) Public localinformation treatment 0.002 0.003 0.042 0.010 0.015 (0.008) (0.007) (0.075) (0.013) (0.009) Private comparativeinformation treatment 0.019** 0.014** -0.056 0.033*** 0.024*** (0.007) (0.006) (0.042) (0.010) (0.007) Public comparative information treatment 0.014** 0.014** 0.051 0.016* 0.010 (0.006) (0.005) (0.069) (0.009) (0.007) Private local × Incumbent malfeasance prior 0.007 (0.007) Public local× Incumbent malfeasance prior 0.018*** (0.006) Private comparative× Incumbent malfeasance prior 0.013** (0.005) Public comparative × Incumbent malfeasance prior -0.007 (0.005) Private local × Strength incumbent prior 0.014 (0.019) Public local× Strength incumbent prior -0.012 (0.023) Private comparative× Strength incumbent prior 0.021 (0.013) Public comparative × Strength incumbent prior -0.011 (0.021) Private local × Incumbent malfeasant spending -0.081** (0.038) Public local× Incumbent malfeasant spending -0.035 (0.046) Private comparative× Incumbent malfeasant spending -0.067** (0.025) Public comparative × Incumbent malfeasant spending -0.014 (0.033) Private local × Negative incumbent updating -0.003 (0.005) Public local× Negative incumbent updating -0.015*** (0.005) Private comparative× Negative incumbent updating -0.012*** (0.004) Public comparative × Negative incumbent updating 0.005 (0.004)

Outcome range [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] [0.03,0.47] Control outcome mean 0.19 0.20 0.20 0.19 0.20 Control outcome std. dev. 0.07 0.07 0.07 0.07 0.07 Treatment mean 0.59 0.59 0.59 0.59 0.59 Treatment std. dev. 0.49 0.49 0.49 0.49 0.49 Interaction mean -0.08 3.24 0.21 0.89 Interaction std. dev. 0.89 0.37 0.17 1.08 R2 0.66 0.67 0.66 0.67 0.67 Observations 675 651 651 675 651

Notes: All specifications include block fixed effects, weight by the inverse of the share of the precinct that was treated, and are estimated using OLS. Lower-order interaction terms are absorbed by the block fixed effects. The smaller sample in columns (2), (3) and (5) reflect lack of prior data in Apaseo el Alto. Standard errors clustered by municipality-treatment are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

253 B| Appendix to Chapter 3

B.1 Proofs

Proof of Proposition1 . First, consider the case of a fully separating equilibrium, where

β (n∗(θ)) = θ,∀θ given the equilibrium strategy n∗(θ). Given the maximand in equation

2 (3.2), call it u(θ,β,n), is twice-differentiable on [θ,θ] × [0,∞), uβ > 0, uθn < 0, the con-

vexity of c(n,θ) ensures a unique (bounded) maximizer solving un(θ,θ,n) = 0, and uβ > 0 implies that θ freely maximizes (akin to a Riley equilibrium), I invoke Theorems 1 and 2 of Mailath(1987). Consequently, the equation (3.3)—which follows from the derivative of an

inverse function and n = n∗(θ)—exists and deﬁnes the incentive compatibility constraint for all types θ, which is continuously differentiable on [θ,θ]. Integrating over θ yields equilibrium information acquisition (as shown in equation (D.1)). Incentive compatibility and the existence of a unique solution for θ imply the existence of a separating equilibrium, and its uniqueness among separating equilibria. Second, consider the case of a semi-separating equilibria where there exist some low

θ types that do not acquire information because pwβ (n) − c(n,θ) < 0,∀n (i.e. a corner solution, given n ∈ [0,∞)). For such types, n∗(θ; p,w) = 0. Given that u is monotoni- cally increasing in θ, and provided that n∗(θ; p,w) > 0 (a necessary condition for a semi- separating equilibrium), the intermediate value theorem implies that there exists an θ˜(p,w) and n∗(θ˜(p,w); p,w) where the θ˜ is indifferent between acquiring information to separate

254 from low types and pooling with low types:

pw Z θ˜ pwθ˜ − c(n∗(θ˜; p,w),θ˜) = θdF(θ). (B.1) F(θ˜) θ

Monotonicity of the maximization problem in θ implies that all types θ ≤ θ˜(p,w) will prefer to acquire n∗ = 0 news rather than n∗(θ˜(p,w); p,w). For such types, β (n) = ˜(p w) 1 R θ , θdF(θ),∀n < n∗(θ˜(p,w); p,w), because they cannot be differentiated in F(θ˜(p,w)) θ equilibrium. For θ > θ˜(p,w), exactly the characteristics of the fully separating equilibrium apply where n∗(θ) = 0 and they integrate from θ˜(p,w) to θ in order to calculate

∗ n (θ; p,w).

Proof of Proposition5 . I examine the comparative static predictions in order, differentiating the equilibrium outcomes by the respective parameters. First, the results at the exten-

sive margin follow by differentiating S ≡ 1 − F(θ˜(p,w)). I ﬁrst use the implicit function theorem to derive:

˜(p,w) R θ θdF(θ) w θ˜(p,w) − θ dθ˜(p,w) F(θ˜(p,w)) = − < 0, (B.2) dp D ˜(p,w) R θ θdF(θ) p θ˜(p,w) − θ dθ˜(p,w) F(θ˜(p,w)) = − < 0, (B.3) dw D dθ˜(p,w)∂θ˜(p,w) dθ˜(p,w)∂θ˜(p,w) = dp∂w dw∂ p ˜(p,w) R θ θdF(θ) c (n∗(θ˜(p,w); p,w),θ˜(p,w)) θ˜(p,w) − θ θ F(θ˜(p,w)) = D2 < 0, (B.4)

˜ −1 R θ˜(p,w) where the signs follow from noting that cθ < 0 and θ(p,w) > [F(θ)] θ θdF(θ),

255 and given that the denominator is positive:

˜(p w) F(θ˜(p,w))F0(θ˜(p,w))θ˜(p,w) − F0(θ˜(p,w)) R θ , θdF(θ) D ≡ pw 1 − θ [F(θ˜(p,w))]2 ∗ −cθ (n (θ˜(p,w); p,w),θ˜(p,w)) > 0, (B.5)

0 ˜ ˜ R θ˜(p,w) 0 ˜ 0 ˜ given F > 0, F(θ(p,w)) < 1 (provided θ(p,w)) 6= θ), θ θF (θ)dθ > θ(p,w)F (θ(p,w)), ∗ ˜ and the envelope theorem implies that ∂n (θ(p,w)) = 0 because the indifference condition is ∂θ˜(p,w) evaluated at the value function maximized at n∗(θ˜(p,w), p,w). Then,

∂S dθ˜(p,w) = −F0(θ˜(p,w)) > 0, (B.6) ∂ p dp ∂S dθ˜(p,w) = −F0(θ˜(p,w)) > 0, (B.7) ∂w dw ∂ 2S dθ˜(p,w) dθ˜(p,w) d2θ˜(p,w) = − F00(θ˜(p,w)) + F0(θ˜(p,w)) > 0.(B.8) ∂ p∂w dp dw dpdw

The ﬁnal condition holds provided that F00 is not too large.

Second, I examine the effects of p and w on n∗(θ; p,w). First note that equation (3.3) in the main text demonstrates that n∗(θ; p,w) is increasing in θ. Given this, for θ ≥ θ˜(p,w) > θ, the implicit function theorem yields:

dn∗(θ; p,w) Z θ w pw ∂θ˜(p,w) = d − ∗ θ ∗ dp θ˜(p,w) cn(n (θ; p,w),θ) cn(n (θ˜(p,w); p,w),θ˜(p,w)) ∂ p ∂n∗(θ˜(p,w)); p,w) + , (B.9) ∂ p dn∗(θ; p,w) Z θ p pw ∂θ˜(p,w) = d − ∗ θ ∗ dw θ˜(p,w) cn(n (θ; p,w),θ) cn(n (θ˜(p,w); p,w),θ˜(p,w)) ∂w ∂n∗(θ˜(p,w)); p,w) + , (B.10) ∂w

where the ﬁrst two terms for each expression reﬂect the differentiation effect, while the

256 third term reﬂects the level effect coming from a shift in n∗(θ˜(p,w); p,w). The proposition ∂θ˜(p,w) ∂θ˜(p,w) focuses only on the former effect, which is clearly positive given that ∂w < 0, ∂w <

0, and cn > 0. The level effect is also positive from inspection of equation (B.6). Furthermore,

dn∗(θ; p,w)∂n∗(θ; p,w) Z θ 1 = d ∗ θ dp∂w θ˜(n,W ) cn(n (θ˜(p,w); p,w),θ˜(p,w)) θ˜(p,w) ∂θ˜(p,w) wc (n∗(θ˜(p,w); p,w),θ˜(p,w)) − 1 − nn ∗ ∗ cn(n (θ˜(p,w); p,w),θ˜(p,w)) ∂ p cn(n (θ˜(p,w); p,w),θ˜(p,w)) ∂θ˜(p,w) wc (n∗(θ˜(p,w); p,w),θ˜(p,w)) 1 − nθ ∗ ∂w cn(n (θ˜(p,w); p,w),θ˜(p,w)) ∂ 2θ˜(p,w) ∂ 2n∗(θ˜(p,w)); p,w) +w + , (B.11) ∂ p∂w ∂w∂ p

which is positive given the differentials for θ˜(p,w) above, and a sufﬁcient condition that:

c (n∗(θ˜(p,w); p,w),θ˜(p,w)) c (n∗(θ˜(p,w); p,w),θ˜(p,w)) w < max nn , nθ . (B.12) ∗ ∗ cn(n (θ˜(p,w); p,w),θ˜(p,w)) cn(n (θ˜(p,w); p,w),θ˜(p,w))

B.2 Variable deﬁnitions

B.2.1 Experimental data

Political quiz score. Number of political quiz questions, out of 10, that the respondent correctly answered. The quiz included the following questions:

1. What party obtained the fourth largest vote share in the elections for Federal Deputy? [A: MC; B: MORENA; C: PRD; D: PVEM]

2. On the 29th May 2015, the Tribunal Electoral del Poder Judicial de la Federación revoked the 3-day suspension of campaign advertising on radio and television of which party? [A: MORENA; B: PAN; C: PRD; D: PVEM] 257 3. According to the National Electoral Institute, how many polling stations in Chiapas, Guerrero and Oaxaca were not installed due to social conﬂict? [A: 36; B: 88; C: 125; D: 200]

4. The candidate from which party won the Governor’s election in Queretaro? [A: PAN; B: PRD; C: PRI; D: Independent]

5. Which of the following states did no hold an election for Governor on the 7th June 2015? [A: Colima; B: Morelos; C: Nueva León; D: San Luis Potosí]

6. Who has responsible for setting ﬁre to ballots in Oaxaca? [A: the CNTE; B: people from the PRD; C: people from the PRI; the SNTE]

7. On 3rd June 2015, the candidate for Federal Deputy, Ángel Luna Munguía, was assassinated in his campaign ofﬁce. From what party was he? [A: MORENA: B: PAN; C: PRD; D: PRI]

8. The candidate from which party won the election to become delegational head of Álvaro Obregón? [A: MORENA: B: PAN; C: PRD; D: PRI]

9. What institution is responsible for verifying the Programa de Resultados Electorales Preliminares (PREP) information system used to verify the Federal Deputy elections in 2015? [A: the U.S. government; B: the federal government; C: Tribunal Electoral del Poder Judicial de la Federación (TEPJF); D: Universidad Nacional AutÃ¸snoma de México (UNAM)]

10. For which party was ex-footballer Cuauhtémoc Blanco a mayoral candidate in Cuer- navaca, Morelos? [A: MC; B: MORENA; C: PRD; D: PSD]

The ordering of multiple choices was randomized.

258 Social treatment. Indicator coded 1 for respondents that were randomly assigned to receive the social treatment informing voters that their performance on the quiz will be sent to the email addresses of the three friends that they enumerated earlier in the baseline survey. Believe received treatment. Indicator coded 1 for respondents that said after the political quiz that they believed that the three friends whose email addresses they listed on the baseline survey would be informed of the results of the quiz. Respondents coded zero include those that said 0, all students, or don’t know. Sophisticated. Indicator coded 1 for respondents that answered all three questions on the baseline survey correctly. Hours of internet news a week (baseline). The number of hours of news consumed on the internet week, according to the baseline survey. Follow national news. Indicator coded 1 for respondents that report following the national news, according to the baseline survey. Demonstrate knowledge. Indicator coded 1 for respondents that stated on the baseline survey that they acquire information in order to demonstrate to their friends and family that they are informed about politics. Know more than friends. Five-point scale rating how much more respondents believe that they know about politics than their friends, according to the baseline survey. High interest friends. Indicator coded 1 for respondents that in the baseline survey rated their three friends’ political interest at 5 or greater on a scale from 0 to 10. Male. Indicator coded 1 for being male. Year of birth. Year of birth in years. Political interest. A scale ranging from 0 to 10 rating an individual’s self-reported interest in politics.

259 Items listed. Count of the number of items that respondents say that they engaged in during recent weeks: attend a campaign activity; watch the news on television; write an article about politics on the internet; and, in the case of a random subset of respondents, talk about the questions on the quiz with a friend. List experiment treatment. Indicator coded 1 for respondents that were randomly assigned to receive the ﬁnal item listed in the previous variable. Interest in politics. Eleven-point scale ranging from 0 to 10 denoting the respondent’s stated level of interest in politics in the second survey (before the quiz). Acquire to speak with friends/to demonstrate knowledge/to choose best candidate/due to interest/for work/due to duty. Indicator coded 1 for respondents that, in the second survey, report acquiring political information for these respective reasons. Estimated friend score. The respondent’s estimate of the number of questions that they believe that the three friends they listed on the ﬁrst survey would have answered correctly. Political interest of friends. A scale ranging from 0 to 10 rating an individual’s perceived interest in politics of their friends in the second survey (after the quiz).

B.2.2 ENCUP survey data

Upcoming local election. Indicator coded 1 for respondents living in a state/municipality with an upcoming local election occurring within the year of the survey. States/municipalities where an election has already occurred within the year of the survey are coded 0. Months until local election. Number of months until the next local election. Watch and listen to news and political programs ever/monthly/weekly/daily. Indicator coded 1 for a respondent that answers that they watch political programs or listen to news at least ever/once a month/at least once a week/daily. (“¿Qué tan seguido escucha noticias o ve programas sobre política?”)

260 Watch and listen to news and political programs scale. 5-point scale from 0 to 4, with values corresponding to levels of watching and listening to new and political programs (in ascending order). Topical political knowledge. First factor from a factor analysis containing the following topical questions: What is the name of the youth movement that recently started in Mexico? (2012) Where was 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) Which party holds your state governorship? (2001, 2003, 2005, 2012) What is the name of your state governor? (2001) Institutional political knowledge. First factor from a factor analysis containing the following topical questions: How many years do federal representatives serve for? (2001, 2003, 2005, 2012) What are the three separated powers of government? (2012) Who has the authority to approve changes to the constitution? (2001) Political network scale. A standardized summative rating scale combining the following three variables: the number of political organizations (general political, party, or cooperative) that a respondent reports being a member of, or previously being a member of; the number of political organizations at which an individual has attended a meeting during the last year; and a scale measuring the regularity with which respondents discuss problems in the community with friends and neighbors, ranging through never (coded 0), occasionally (coded 1) and frequently (coded 2). The final component of the scale and the cooperative organization indicator were not asked about in the 2012 survey; accordingly, I multiply imputed these responses over ten datasets using the following pre-treatment variables: survey and municipality fixed effects, gender, age, indigenous language speaker, and Catholic. As noted in the main text, the scale has a Cronbach’s alpha of 0.57. Civic network scale. A standardized summative rating scale combining the following three variables: the number of civic organizations (pensioner, professional, labor, social, 261 voluntary, religious, neighbor, cultural, sporting, parents and citizen organizations) that a respondent reports being a member of, or previously being a member of; and the number of civic organizations at which an individual has attended a meeting during the last year. The scale has a Cronbach’s alpha of 0.66. Municipal incumbent win margin (last election). The difference in vote share between the incumbent and second-placed finisher at the previous municipal mayoral election. In Usos y Custombres in Oaxaca, the incumbent win margin is set to the maximum of 1. Municipal ENPV (last election). The effective number of political parties (by vote share) at the previous municipal mayoral election. In Usos y Custombres in Oaxaca, ENPV is set to the maximum of 1. Voted for mayor since 2000. Indicator coded 1 for individuals that have reported voting in their municipal election in 2000. Media stations within municipality. Total number of AM, FM and television stations with antennae located within the municipality.

B.3 Additional experimental results

Tables B.1 and B.2 respectively show balance across 49 pre-treatment variables in the final sample and the initial assignment.1 The control mean provides the baseline summary statistics for the sample, while the treatment effect column and associated standard error examine differences between treatment and control participants. Consistent with chance, only 5 of the variables show a significant difference (at the 10% level) in the final sample, while only 4 differ significantly in the entire sample. The fact that these are the same variables lends further suggests that there is no differential attrition.

1I use the ﬁnal sample of 539 respondents used for the main analysis, although the results are unchanged if the single dropped observation with a missing value for interest in politics is included.

262 Table B.1: Balance tests for the ﬁnal sample

Pre-treatment variable Control Treatment Standard Pre-treatment variable Control Treatment Standard mean effect error mean effect error Latitude 19.995 0.061 (0.312) Hours of radio news a week 2.558 -0.136 (0.270) Longitude -98.72 -0.395 (0.433) Hours of television news a week 2.462 -0.047 (0.269) Year of birth 1992.9 0.322* (0.176) Follow local news 0.315 -0.062 (0.039) Male 0.695 -0.061 (0.041) Follow national news 0.882 -0.009 (0.028) Politics student 0.201 0.026 (0.035) Follow international news 0.713 0.029 (0.038) Total correct answers 2.477 -0.011 (0.136) Follow state news 0.527 -0.035 (0.043) Sophisticated 0.688 0.000 (0.040) Follow no news 0.000 0.012* (0.007) Interest in politics 7.832 -0.358* (0.191) Don’t know if follow news 0.004 -0.004 (0.004) Political interest of friends 7.051 -0.126 (0.175) Student organization 0.595 0.043 (0.042) Frequency of political discussion with friends 3.358 -0.058 (0.082) Voluntary organization 0.491 0.020 (0.043)

263 Know more than friends 3.713 0.021 (0.099) Sindicate 0.004 -0.004 (0.004) Respect for knowledge about politics 8.100 -0.218 (0.183) Religious organization 0.179 0.036 (0.034) Participate in political conversations 3.466 -0.054 (0.057) Citizen organization 0.133 -0.025 (0.028) Comfortable when don’t know about politics 3.018 -0.030 (0.090) Neighbor organization 0.168 -0.045 (0.030) Acquire to vote for best candidate 0.824 -0.021 (0.034) Cultural organization 0.240 0.029 (0.038) Acquire to speak with friends 0.258 -0.016 (0.037) Party organization 0.154 -0.066** (0.028) Acquire due to interest 0.595 -0.057 (0.043) Sports organization 0.477 0.000 (0.043) Acquire for work 0.204 0.057 (0.036) Other organization 0.036 -0.001 (0.016) Acquire out of civic duty 0.713 0.006 (0.039) Total organizations 2.584 0.008 (0.060) Acquire to demonstrate knowledge 0.061 0.016 (0.022) PAN partisan 0.437 -0.010 (0.043) Acquiring information not important 0.011 0.024* (0.013) PRI partisan 0.140 -0.009 (0.030) Don’t know why acquire 0.000 0.008 (0.005) PRD partisan 0.039 -0.009 (0.016) Attrition incentive 0.516 0.011 (0.043) MORENA partisan 0.007 0.012 (0.010) Hours of newspaper news a week 2.401 0.207 (0.260) Non-partisan 0.323 -0.019 (0.040) Hours of internet news a week 5.655 -0.140 (0.405)

Notes: All speciﬁcations are difference in means between the respondents treated by the social treatment and those that were not. Each regression includes 539 observations. Robust standard errors are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table B.2: Balance tests from the initial assignment in the baseline survey

Pre-treatment variable Control Treatment Standard Pre-treatment variable Control Treatment Standard mean effect error mean effect error Latitude 19.868 0.262 (0.266) Hours of radio news a week 2.561 -0.145 (0.221) Longitude -98.539 -0.396 (0.500) Hours of television news a week 2.563 -0.109 (0.230) Year of birth 1993.1 0.193 (0.150) Follow local news 0.305 -0.041 (0.033) Male 0.679 -0.045 (0.035) Follow national news 0.892 -0.031 (0.024) Politics student 0.182 0.027 (0.029) Follow international news 0.697 0.062* (0.032) Total correct answers 2.495 -0.035 (0.115) Follow state news 0.508 -0.005 (0.036) Sophisticated 0.647 0.021 (0.035) Follow no news 0.003 0.008 (0.006) Interest in politics 7.778 -0.299* (0.158) Don’t know if follow news 0.003 -0.003 (0.003) Political interest of friends 7.032 -0.214 (0.148) Student organization 0.587 0.050 (0.035) Frequency of political discussion with friends 3.350 -0.059 (0.069) Voluntary organization 0.513 -0.002 (0.036)

264 Know more than friends 3.691 0.087 (0.084) Sindicate 0.003 -0.003 (0.003) Respect for knowledge about politics 8.079 -0.122 (0.151) Religious organization 0.184 0.024 (0.029) Participate in political conversations 3.453 -0.057 (0.049) Citizen organization 0.124 -0.030 (0.023) Comfortable when don’t know about politics 3.019 -0.043 (0.075) Neighbor organization 0.166 -0.035 (0.026) Acquire to vote for best candidate 0.826 -0.035 (0.029) Cultural organization 0.266 0.023 (0.033) Acquire to speak with friends 0.284 -0.03 (0.032) Party organization 0.137 -0.051** (0.023) Acquire due to interest 0.587 -0.017 (0.036) Sports organization 0.476 -0.003 (0.036) Acquire for work 0.203 0.062** (0.031) Other organization 0.039 -0.007 (0.014) Acquire out of civic duty 0.711 -0.013 (0.033) Total organizations 2.542 -0.010 (0.054) Acquire to demonstrate knowledge 0.068 0.001 (0.018) PAN partisan 0.413 0.012 (0.036) Acquiring information not important 0.011 0.014 (0.010) PRI partisan 0.142 -0.032 (0.024) Don’t know why acquire 0.000 0.005 (0.004) PRD partisan 0.039 -0.007 (0.014) Attrition incentive 0.497 0.005 (0.036) MORENA partisan 0.008 0.013 (0.009) Hours of newspaper news a week 2.502 0.074 (0.222) Non-partisan 0.326 -0.008 (0.034) Hours of internet news a week 5.555 -0.148 (0.329)

Notes: All speciﬁcations are difference in means between the respondents treated by the social treatment and those that were not. Numbers of observations vary slightly with non-responses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. B.4 Additional observational results

Table B.3 shows that upcoming local elections are well balanced across 22 individual and municipal level characteristics in the ENCUP surveys. Consistent with chance, only two differences are statistically significant. Sample size differences reflect the unavailability of some questions in particular survey waves. Moreover, Table B.4 demonstrates that neither changes in upcoming local elections nor changes in measures of local violence predict whether a municipality is included in any given survey wave. The sample includes only municipalities where surveys are conducted during at least one survey wave. Table B.5 shows how the effect of local election by politically-oriented networks varies by intensity of political news consumption. The results show that news consumption prior to elections is sigincreased among voters in politically engaged networks at every news consumption intensity. Table B.6 examines how the effects of an upcoming local election vary with the definition of such an election. In particular, for the three main outcomes, columns (1)-(10) show that the results are robust to using any number of months between 1 and 10 to define an upcoming local election. Categories are groups in some cases because the coefficients are identical. In combination with panel G of Table 4.7, these results indicate that the findings are not sensitive to the definition of an upcoming local election.

265 Table B.3: Balance tests for upcoming local elections over 22 individual and municipal-level variables

Female Speaks Catholic Age Education Voted for Political Media PAN PRI PRD Indigenous mayor network stations within governor governor governor language since 2000 scale municipality (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Upcoming local election 0.009 -0.017 -0.032* -0.009 0.045 0.037 -0.045 -0.330 -0.017 0.000 -0.006 (0.011) (0.021) (0.018) (0.387) (0.034) (0.024) (0.063) (0.408) (0.075) (0.080) (0.033)

Observations 17,213 13,030 13,030 17,213 12,513 13,030 17, 213 17,213 17,133 17,133 17,133 Outcome mean 0.55 0.08 0.81 40.70 1.74 0.73 0.00 10.13 0.27 0.56 0.14 266 PAN PRI PRD Municipal Municipal Municipal Municipal Municipal Total Municipal Homicides municipal municipal municipal incumbent incumbent incumbent ENPV registered municipal police per month incumbent incumbent incumbent win margin win indicator vote share (last voters spending per voter (last year) (last election) (last election) (last election) election) (last election) (last year) (12) (13) (14) (15) (16) (17) (18) (19) (20) (21) (22) Upcoming local election 0.097** -0.047 -0.019 0.015 0.109 0.023 -0.214* 5373.022 96.866 -0.123 0.902 (0.039) (0.067) (0.036) (0.018) (0.066) (0.016) (0.114) (15211.094) (91.616) (0.157) (0.954)

Observations 17,015 17,015 17,015 17,213 17,213 17,015 17,213 17,015 10,530 14,016 17,178 Outcome mean 0.36 0.45 0.15 0.15 0.63 0.47 2.62 281791.39 841.91 2.24 4.73 Notes: All speciﬁcations include state and survey ﬁxed effects, and are estimated using OLS. Several underlying elements of the political network scale were imputed in 2012 (see Online Appendix for details), and the estimates in column (7) are across ten multiply imputed datasets. Standard errors clustered by state are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table B.4: Predictors of municipalities included in each survey wave

Panel A: Linear predictors Surveyed municipality indicator (1) (2) (3) (4) Local election 0.024 (0.046) Homicide in last month -0.023 (0.046) Homicides last year -0.001 (0.001) Homicides last 3 years -0.000 (0.001)

Observations 2,156 2,152 2,152 2,152 Surveyed municipality mean 0.48 0.48 0.48 0.48

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

267 Table B.5: The effect of upcoming local elections on news consumption intensity, by political network engagement

Watch and listen to news and political programs...... ever ...monthly ...weekly ...daily ...scale (1) (2) (3) (4) (5) Upcoming local election 0.009 0.053 0.063** 0.050 0.174 (0.025) (0.032) (0.034) (0.032) (0.112) Political network scale 0.022** 0.031** 0.037** 0.032** 0.122** (0.008) (0.010) (0.012) (0.010) (0.039) Upcoming local election 0.033** 0.055*** 0.054*** 0.050*** 0.192***

268 × Political network scale (0.015) (0.018) (0.015) (0.017) (0.055)

Observations 13,030 13,030 13,030 13,030 13,030 Outcome mean 0.87 0.69 0.63 0.39 2.58 Outcome std. dev. 0.34 0.46 0.48 0.49 1.47 Upcoming local election mean 0.19 0.19 0.19 0.19 0.19 Political network scale mean 0.00 0.00 0.00 0.00 0.00 Political network scale std. dev. 1.00 1.00 1.00 1.00 1.00 Survey year not asked 2001 2001 2001 2001 2001

Panel A: Watch and Upcoming local election deﬁned as an indicator elections within ... of survey listen to news ever 1/2 months 3/4 months 5/6 months 7 months 8/9 months 10 months (1) (2) (3) (4) (5) (6) Upcoming local election 0.050*** 0.021 0.009 0.059** 0.055** 0.054* (0.018) (0.019) (0.025) (0.029) (0.027) (0.027) Upcoming local election 0.013 0.001 0.033** 0.026** 0.023* 0.024* × Political network scale (0.008) (0.001) (0.015) (0.012) (0.012) (0.012) Panel B: Watch and Upcoming local election deﬁned as an indicator elections within ... of survey listen to news scale 1/2 months 3/4 months 5/6 months 7 months 8/9 months 10 months (1) (2) (3) (4) (5) (6)

269 Upcoming local election 0.185** 0.196 0.174 0.501*** 0.450*** 0.465*** (0.082) (0.117) (0.112) (0.161) (0.144) (0.148) Upcoming local election 0.187*** 0.101* 0.192*** 0.160*** 0.146** 0.151** × Political network scale (0.041) (0.042) (0.055) (0.050) (0.050) (0.049) Panel C: Topical Upcoming local election defined as an indicator elections within ... of survey political knowledge 1 month 2 months 3/4 months 5/6 months 7 months 8 months 9 months 10 months (1) (2) (3) (4) (5) (6) (7) (8) Upcoming local election 0.204*** 0.051 0.363 0.188 0.110 0.122 0.104 0.100 (0.072) (0.146) (0.216) (0.133) (0.144) (0.137) (0.125) (0.122) Upcoming local election 0.070 0.228* 0.147 0.175** 0.142** 0.146** 0.151** 0.156** × Political network scale (0.047) (0.130) (0.086) (0.063) (0.066) (0.064) (0.063) (0.062) Notes: All specifications include state and survey fixed effects, and are estimated using OLS. Several underlying elements of the political network scale were imputed in 2012 (see Online Appendix for details), and the estimates in column (7) are across ten multiply imputed datasets. Standard errors clustered by state are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. C| Appendix to Chapter 4

C.1 Variable deﬁnitions

Change in incumbent party vote share. The change in incumbent vote share, as a proportion of total votes cast, in a given electoral precinct. Where the incumbent is part of a coalition, we count the vote share of the affiliation of the mayor at the next election. For around a quarter of municipalities in the sample we hand-coded the mayor’s affiliation. Source: IFE and State Electoral Institutes. Incumbent re-elected. Indicator coded one where the incumbent party is re-elected (municipal level variation). Incumbents are defined as above. Source: IFE and State Electoral Institutes. Unauthorized spending. Percentage of FISM funds spent in an unauthorized manner. See text for further discussion. The variables Corrupt Q3 and Corrupt Q4 are separately defined by the third and fourth quartiles of our full sample and GRD sample. Source: ASF audit reports. Spending not on the poor. Percentage of FISM funds spent not spent on the poor. See text for further discussion. The variables Not poor Q3 and Not poor Q4 are separately defined by the third and fourth quartiles of our full sample and GRD sample. Source: ASF audit reports. PAN/PRI/PRD incumbent. Indicator coded one for the incumbent mayor represents a coali-

270 tion containing the PAN, PRI or PRD. Source: IFE and State Electoral Institutes. Coalition partners. The number of parties in the incumbent mayor’s coalition. Source: computed from IFE and State Electoral Institutes. Municipal incumbent victory margin (lag). The difference between the largest and the second large party in a given electoral precinct at the last election. Source: computed from IFE and State Electoral Institutes. Municipal effective number of political parties (lag). The effective number of political parties in the municipal, deﬁned by the following formula:

1 , P 2 ∑p=1 vp

where vp is the municipal vote share of party p = 1,...,P. Source: computed from IFE and State Electoral Institutes. Incumbent party vote share (lag). The vote share of the incumbent party in a given electoral precinct at the previous election. Source: IFE and State Electoral Institutes. Effective number of political parties (lag). The effective number of political parties, deﬁned at the precinct level. Source: computed from IFE and State Electoral Institutes. Registered voters. The number of voters registered to vote in the electoral precinct. Source: IFE and State Electoral Institutes. Turnout (lag). Precinct-level electoral turnout at the previous election. Source: IFE and State Electoral Institutes. Distance to municipal head from precinct border (log). The nearest (logged) distance in meters from the municipal head to the border of the electoral precinct. Source: computed from INEGI data. Distance to municipal head from precinct centroid (log). The (logged) distance in meters from the municipal head to the centroid of the electoral precinct. Source: computed from

271 INEGI data. Area (km2). Electoral precinct area in square kilometers. Source: computed from IFE data. Population (log). Electoral precinct population (logged). Source: Mexican 2010 Census. Population density (log). Electoral precinct population density (logged). Source: computed from the two variables above. Local media. The total number of AM, FM and TV stations emitting from within an electoral precinct’s municipality. Source: computed from IFE data. Local AM/FM/TV. The total number of AM/FM/TV stations emitting from within the municipality. Source: computed from IFE data. Non-local media. The total number of AM, FM and TV stations emitting from outside an electoral precinct’s municipality. Source: computed from IFE data. Average children per woman. Average number of children per woman in the electoral precinct. Source: Mexican 2010 Census. Share households with male head. Share of households in the electoral precinct with a male head of household. Source: Mexican 2010 Census. Share indigenous speakers. Share of the electoral precinct population aged 3+ that speaks an indigenous language. Source: Mexican 2010 Census. Average years of schooling. Average number of completed grades of schooling among the population aged above 15. Source: Mexican 2010 Census. Share illiterate. Share of the electoral precinct population aged above 15 that is illiterate in 2010. Source: Mexican 2010 Census. Share no schooling/incomplete primary schooling/complete primary schooling/incomplete secondary schooling/complete secondary schooling/higher education. Share of the electoral precinct population aged above 15 for whom no schooling/incomplete primary schooling/complete primary schooling/incomplete secondary schooling/complete secondary schooling/higher education is the highest level of education that they possess. Source: Mexican 272 2010 Census. Share economically active. Share of the electoral precinct population aged above 12 that is economically active (i.e. (had job, had job but not working, looking for job)). Source: Mexican 2010 Census. Share without health insurance. Share of the electoral precinct population without public or private health care. Source: Mexican 2010 Census. Share state workers health care. Share of the electoral precinct population with state workers health care. Source: Mexican 2010 Census. Average occupants per dwelling/room. Average number of occupants per dwelling/room in the electoral precinct. Source: Mexican 2010 Census. Share non-dirt ﬂoor. Share of the electoral precinct population in a private dwelling without a dirt ﬂoor. Source: Mexican 2010 Census. Share toilet at home. Share of the electoral precinct population with a toilet. Source: Mexican 2010 Census. Share running water. Share of the electoral precinct population in a private dwelling with running water. Source: Mexican 2010 Census. Share drainage. Share of the electoral precinct population in a private dwelling with drainage. Source: Mexican 2010 Census. Share electricity. Share of the electoral precinct population in a private dwelling with electricity. Source: Mexican 2010 Census. Share fridge. Share of the electoral precinct population in a private dwelling with a refrig- erator. Source: Mexican 2010 Census. Share washing machine. Share of the electoral precinct population in a private dwelling with a washing machine. Source: Mexican 2010 Census. Share cell phone. Share of the electoral precinct population in a private dwelling with a cell phone. Source: Mexican 2010 Census. 273 Share cell phone. Share of the electoral precinct population in a private dwelling with a car or truck. Source: Mexican 2010 Census. Share cell phone. Share of the electoral precinct population in a private dwelling with a computer. Source: Mexican 2010 Census.

C.2 Audit reports

Figures C.1 and C.2 provide an example of an audit report from 2008 for the municipality of Ajalpan in the state of Puebla.

C.3 Additional local media stations and news consump-

tion

As noted in the main text, we provide survey evidence that the presence of an additional media station is positively correlated with news consumption. In particular, we use data from the 2009 CIDE-CSES post-election survey which asked 2,400 respondents in detail about their media consumption. Such detail is rare in such surveys and reflects the major media reform that occurred in 2007 (see Larreguy, Marshall and Snyder 2016). All respondents were asked the identities of any news programs that they regularly watch on TV (and could name up to two), while half the sample was asked for the identities of any news program that they regularly listen to on the radio. To examine the relationship between media access and consumption, we look at how the number of local media stations affects stated consumption. Controlling for state fixed effects, the estimates in Table C.1 demonstrate a clear positive correlation. Columns (1) and (2) of panel A show that each additional local radio station significantly increases the probability that an individual listens to a news program on the radio by 0.4 percentage 274 Figure C.1: Sample ASF audit report (page 1)

Notes: Extracted from the ASF audit report on the use of FIMS funds by the municipal government of the municipality of Ajalpan in the state of Puebla in 2008. The red squares indicate the lines where the ASF reports the FISM funds spent “in an unauthorized manner” and the share spent on projects “not beneﬁting the poor.”

275 Figure C.2: Sample ASF audit report (page 2)

Note: See Figure C.1.

276 Table C.1: Correlation between access to broadcast media and news consumption

Panel A: Local media Any Number Any Number Any Total radio of radio television of television news news news news news news program programs program programs program programs Local radio stations 0.004* 0.005** (0.002) (0.002) Local television stations 0.014** 0.021* (0.005) (0.011) Local media 0.004*** 0.014*** (0.001) (0.004)

Observations 1,055 1,055 2,110 2,110 2,110 1,055 Outcome mean 0.29 0.34 0.80 1.10 0.81 1.44 Outcome std. dev. 0.45 0.56 0.40 0.71 0.39 0.98 Media mean 8.40 8.40 8.40 8.40 8.40 8.40 Media std. dev. 8.96 8.96 8.95 8.95 8.95 8.96 Panel B: Non-local media Any Number Any Number Any Total radio of radio television of television news news news news news news program programs program programs program programs Non-local radio stations 0.002 0.002 (0.001) (0.002) Non-local television stations 0.008 0.013 (0.007) (0.011) Non-local media 0.002 0.004 (0.001) (0.003)

Observations 1,055 1,055 2,110 2,110 2,110 1,055 Outcome mean 0.29 0.34 0.80 1.10 0.81 1.44 Outcome std. dev. 0.45 0.56 0.40 0.71 0.39 0.98 Media mean 22.70 22.70 22.70 22.70 22.70 22.70 Media std. dev. 21.68 21.68 21.67 21.67 21.67 21.68

Notes: All speciﬁcations include state ﬁxed effects, and are estimated using OLS. Standard errors are clustered by municipality. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

277 points and increases the average number of programs listened to by 0.005. Conversely, panel B shows that non-local media stations do not significantly increase news consumption. Similarly, columns (3) and (4) of panel A show that each additional local television station significantly increases the probability that an individual watches a news program on television by 1.4 percentage points and increases the average number of new programs watched by 0.02. Again, panel B shows no significant effect for non-local media stations. Finally, columns (5) and (6) aggregate across radio and television, and similarly show that local, but not non-local, media stations significantly increase news consumption.

C.4 Lack of balance across media stations

Table C.2 shows that the number of local media stations that cover an electoral precinct is strongly correlated with virtually every precinct-level covariate. In particular, we ﬁnd that precincts with more local media stations are inhabited by more urban and pro-incumbent households, that are also far more likely to be literate, possess more household necessities and amenities (see variable deﬁnitions above), and have fewer children. In short, precincts with more media stations are more developed.

C.5 Additional results

Table C.3 presents our DD estimates for the neighbor sample. Due to the signiﬁcant decline in sample size, and the fact that the GRD estimates only identify off within-match variation (and thus only from matches cross municipality borders), our estimates are considerably noisier. Nevertheless, the point estimates are broadly similar to those in the Table 4 of the main paper. Table C.4 estimates the triple interaction between revealing corruption or not spending

278 Table C.2: Balance across the number of local media stations (DD sample)

Correlation with local media Incumbent party vote share (lag) 0.001* (0.001) Effective number of political parties (lag) -0.009*** (0.003) Registered voters (log) -0.001 (0.002) Turnout (lag) 0.000 (0.001) Distance to municipal head from precinct border (log) 0.003 (0.003) Distance to municipal head from precinct centroid (log) -0.005* (0.003) Area (log) -0.035*** (0.005) Population (log) -0.002 (0.002) Population density (log) 0.053*** (0.009) Non-local media -0.891*** (0.148) Average children per woman -0.009*** (0.001) Share households with male head -0.002*** (0.000) Share indigenous speakers -0.003*** (0.000) Average years of schooling 0.059*** (0.007) Share illiterate -0.002*** (0.000) Share no schooling -0.002*** (0.000) Share incomplete primary schooling 0.002*** (0.000) Share complete primary schooling 0.004*** (0.000) Share incomplete secondary schooling 0.004*** (0.001) Share complete secondary schooling 0.005*** (0.001) Share higher education 0.005*** (0.001) Share economically active 0.002*** (0.000) Share without health care -0.002*** (0.001) Share state workers health care 0.000* (0.000) Average occupants per dwelling -0.009*** (0.002) Average occupants per room -0.009*** (0.001) Share non-dirt ﬂoor 0.001*** (0.000) Share toilet at home 0.002*** (0.000) Share running water 0.004*** (0.001) Share drainage 0.004*** (0.000) Share electricity 0.001*** (0.000) Share washing machine 0.007*** (0.001) Share fridge 0.005*** (0.000) Share cell phone 0.007*** (0.001) Share car 0.007*** (0.001) Share computer 0.007*** (0.001)

279 Table C.3: The effect of audits revealing corruption (neighbor sample)

Change in incumbent party vote share Incumbent party re-elected (1) (2) (3) (4) Constant -0.106*** -0.137*** 0.160 0.335* (0.024) (0.025) (0.161) (0.174) Audit 0.101*** 0.152 0.238 0.476** (0.037) (0.033) (0.226) (0.222) Corrupt Q3 0.040 0.304 (0.035) (0.195) Audit × Corrupt Q3 -0.003 0.520** (0.042) (0.262) Audit 0.000 (.) Corrupt Q4 -0.020 0.008 (0.053) (0.249) Audit × Corrupt Q4 -0.052 0.344 (0.056) (0.356) Not poor Q3 0.059 0.023 (0.056) (0.218) Audit × Not poor Q3 -0.099 -0.091 (0.086) (0.288) Not poor Q4 0.162*** 0.184 (0.040) (0.251) Audit × Not poor Q4 -0.231*** -0.806*** (0.056) (0.308)

Observations 17,312 17,312 17,312 17,312

Notes: All speciﬁcations include pre-treatment controls and match and year ﬁxed effects, use up to two possible matches, and are estimated using OLS. All regressions have 19,230 observations except Panel A which contains 16,947 observations. Standard errors are clustered by municipality-year. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

280 on the poor by the number of local media stations in the DD sample. As noted in the main text, we ﬁnd broadly similar results to the results exploiting exogenous variation in local media.

281 Table C.4: Effects of local media publicizing audits reports revealing malfeasance before an election (DD sample)

Change in incumbent party vote share (1) (2) Audit -0.015 -0.008 (0.021) (0.022) Audit × Local media 0.002 0.003 (0.001) (0.002) Audit × Corrupt Q3 -0.004 (0.037) Audit × Local media × Corrupt Q3 0.002 (0.003) Audit × Corrupt Q4 0.041 (0.062) Audit × Local media × Corrupt Q4 -0.009** (0.004) Audit × Not poor Q3 -0.028 (0.036) Audit × Local media × Not poor Q3 -0.001 (0.003) Audit × Not poor Q4 0.044 (0.060) Audit × Local media × Not poor Q4 -0.011*** (0.004)

Observations 47,938 47,938 Outcome mean -0.05 -0.05 Outcome std. dev. 0.15 0.15 Local media mean 12.76 12.76 Local media std. dev. 11.68 11.68

Notes: All speciﬁcations include demographic and socioeconomic controls and year ﬁxed effects, and are estimated using OLS. The omitted category for corruption and not spending on the poor is Q1 and Q2. Standard errors are clustered by municipal election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

282 D| Appendix to Chapter 5

D.1 Formal model illustrating the voter updating process

This section develops a simple decision-theoretic selection model (alla Fearon 1999) to clarify formally the theoretical implications discussed in the main paper. In particular, the model incorporates salience and learning considerations to examine how information about incumbent performance affects the voting behavior of forward-looking voters seeking to choose the best candidate. By consuming information, forward-looking voters learn about the contemporary political world. Specifically, let a voter receive n signals about incumbent performance in office, and thus the suitability of the incumbent party for continuing in office (or “competence”).1 For simplicity, each signal s of incumbent performance is an independent draw from a Nor-

mal distribution N(µ,σ 2) with known variance σ 2 > 0. The mean of the distribution µ represents the unknown true level of incumbent competence, while a larger variance re- ﬂects the possibility that different media sources distort or under-report certain types of news (e.g. Besley and Prat 2006; Gentzkow and Shapiro 2006; Mullainathan and Shleifer 2005), fail to serve all segments of the market (Prat and Strömberg 2005), or are dwarfed by news that does not pertain to the voter’s incumbent party (Larreguy, Marshall and Snyder

2015). The voter also possesses a prior belief N(δ,τ2) over incumbent competence. For

1Politicians are assumed to perform according to their underlying competence level µ or 0 when in ofﬁce.

283 simplicity the challenger’s level of competence is normalized to 0, and I assume that information about the incumbent does not cause voters to also update about the challenger.2 The voter updates their belief about the incumbent’s competence downwards (upwards) when

n δ > (<)s¯, where s¯ = ∑t=1 st /n is the average signal received by a voter. The voter also receives a shock b ∈ R toward the incumbent. This shock b may be positive or negative, and could derive from partisanship, clientelistic ties, or other valence factors that helped the incumbent party win the last election. Ultimately, individuals vote according to both their bias and expectations of politician competence in ofﬁce. To capture issue salience, the voter weights the importance of competence by the weighting function w(n) > 0, which is increasing in n.3 For simplicity, I assume that the voter is risk-averse and simply maximizes the sum of (expected) competence and bias.4 Upon receiving n signals, a voter learns from the news about the competence of the incumbent party. Updating their prior belief distribution, the voter’s posterior belief distribution is given by the following standard Normal learning result:

δ ns −1! 2 + 2 1 n N τ σ , + (D.1) 1 + n τ2 σ 2 τ2 σ 2 where s is the mean signal received by the voter. The extent to which signals about in-

2The results could be easily extended to allow for voters to have a distribution of beliefs over the challenger. If I allowed beliefs about the incumbent and challenger to be correlated, then the results would be unaffected for a sufﬁciently small correlation.

3Similar results would hold if the weighting also depended on the content of the news, i.e. w(n,s¯), where most plausibly ws¯ < 0 and wns¯ < 0. Changes in the weight attached to beliefs about incumbent competence could also reﬂect a voter’s certainty about the information; in such a setting, more signals would then cause voters to increasingly voter according with the signal because it relatively increases the precision of their belief about the incumbent’s competence (relative to other factors affecting vote choice). 4Abstracting from risk-aversion presents the results particularly clear. If the voter were risk-averse, information would help the incumbent by reducing uncertainty about the utility voters would expect to receive from electing them. This could easily be achieved by adopting a risk-averse utility function (for tractabil- ity a constant absolute risk-aversion functional form could easily map Normal posterior beliefs into a utility function that separates the utility associated with the mean and variance).

284 cumbent performance, particularly when noisy or relatively uninformative (i.e. high σ 2), affect voter posterior beliefs depends upon the strength of their prior beliefs. A voter with

weak priors (i.e. large τ2) are mostly likely to update their beliefs. While a noisy signal should have limited effect on well-informed voters who recognize that a short-term shock is not a good reﬂection of incumbent competence, a voter with a weak prior does not know whether a short-term shock reﬂects long-term performance. To the extent that uninformed voters believe that such information may be informative about incumbent competence, their posterior beliefs—both the mean and variance—about incumbent competence change. The expected utility associated with the incumbent politician is then given by integrating over this posterior distribution. Combining this with the bias b toward the incumbent, a voter with bias b votes for the incumbent when:

δσ 2 + nsτ2 B ≡ b + w(n) ≥ 0 (D.2) σ 2 + nτ2

Intuitively, a voter is more likely to re-elect the incumbent when the bias b toward the incumbent is high and the expected competence of the incumbent is high. The relative importance attached to these factors is reﬂected by w(n). The individual-level implications could be aggregated into party-level vote shares by integrating over the distribution of voter priors and biases. To understand how information impacts the voters propensity of voting for the incumbent, I differentiate B with respect to the number of signals, n:

∂B σ 2τ2(s¯− δ ) δσ 2 + nsτ2 = w(n) +w0(n) . (D.3) ∂n (σ 2 + nτ2)2 σ 2 + nτ2 | {z } | {z } Learning e f f ect Salience e f f ect

The ﬁrst term represents the learning effect: how the impact of new information, by changing the voter’s posterior belief, affects their likelihood of supporting the incumbent. The 285 following proposition analyzes the comparative static implications of the learning effect:

Proposition 3. If s < δ, such that information does not help the incumbent, then the learning effect impacts voting behavior as follows:

1. The likelihood that the voter votes for the incumbent decreases with the number of signals received (n).

2. When 1 < n , the learning increases in the precision of the signal ( 1 ), decreases τ2 σ 2 σ 2 √ in the precision of the prior ( 1 ), and (if 2 − n 2 < 2n 2 2) increases in their τ2 σ τ τ σ interaction.

The reverse results hold when s > δ.

Proof: Focusing on the learning effect sets the salience effect to zero. For part 1, given w(n) > 0 and all variances are positive, it is easy to see that the learning effect, L ≡ 2 2 w(n) σ τ (s¯−δ ) , is positive when s < . Part 2 is established by taking subsequent deriva- (σ 2+nτ2)2 δ tives. The cross-partial (i.e. the partial of the learning effect) with respect to σ 2 is given by:

∂L w(n)(s − δ )τ2[nτ2 − σ 2] = > 0, (D.4) ∂σ 2 (σ 2 + nτ2)3

which, given s < δ, is negative when nτ2 < σ 2. Similarly,

∂L w(n)(s − δ )σ 2[nτ2 − σ 2] = − < 0, (D.5) ∂τ2 (σ 2 + nτ2)3 by the same conditions. Finally,

∂ 2L w(n)(s − δ )σ 2[2nτ2σ 2 − (σ 2 − (nτ2)2)] = − > 0, (D.6) ∂τ2∂σ 2 (σ 2 + nτ2)3

286 √ provided that σ 2 − nτ2 < 2nτ2σ 2. The multiplicative nature of the comparative statics

ensures that the reverse results hold when s > δ. When s < (>)δ, the voter updates negatively (positively) about the competence of their incumbent. Consider the case where s < δ. The average signal of incumbent performance that they receive from the news suggests that the incumbent’s true level of competence is

below their prior δ. Consequently, part 1 unsurprisingly states that when the news suggests that incumbent performance is low, a voter is less likely to re-elect the incumbent. This reflects the decrease in the posterior belief over the competence of the incumbent, while the posterior belief about the challenger is unaffected. Part 2 further shows that the punishment of the incumbent is increasing in the precision of the signals and decreasing in the precision of the prior—provided that the prior is sufficiently weak ( 1 < n ). Furthermore, τ2 σ 2 under an additional condition, the impact of a precise signal is strongest when voter priors are weakest. Intuitively, when news coverage is particularly credible or informative relative to their prior beliefs, voters are less likely to believe that the incumbent has sufficient competence to merit re-election. The second term in equation (D.3) represents the salience effect: the impact of information increasing the relative importance of incumbent performance to the voter. This increases support for the incumbent if the posterior belief is positive (and thus exceeds the normalized belief about the challenger’s competence). This opens the possibility that even

if voters receive a negative signal, such that s < δ, the beneﬁts of priming an issue on which the incumbent initially scored well could outweigh the negative learning effect. The following result clariﬁes this insight, and describes several key comparative static predictions in terms of vote choice:

Proposition 4. If s < δ, such that information does not help the incumbent, then the salience effect impacts voting behavior as follows:

287 1. The likelihood that the voter votes for the incumbent increases (decreases) with the number of signals received (n) when the posterior belief about incumbent compe-

2 2 tence, δσ +nsτ , is positive (negative). σ 2+nτ2

2. A negative salience effect increases in the precision of the signal ( 1 ), decreases in σ 2 the precision of the prior ( 1 ), and (if 1 < n ) decreases in their interaction. τ2 τ2 σ 2

Proof: Focusing on the salience effect sets the learning effect to zero. For part 1, given w0(n) > 0 and all variances are positive, it is easy to see that the salience effect, S ≡

2 2 2 2 w0(n) δσ +nsτ , is positive when the posterior belief δσ +nsτ > 0 (or when 2 +ns 2 > 0, σ 2+nτ2 σ 2+nτ2 δσ τ given the denominator is positive). Part 2 is again established by taking subsequent deriva-

tives. The cross-partial (i.e. the partial of the salience effect) with respect to σ 2 is given by:

∂S w0(n)nτ2[δ − s¯] = > 0. (D.7) ∂σ 2 (σ 2 + nτ2)2

Similarly,

∂S w0(n)nσ 2[δ − s¯] = − < 0. (D.8) ∂τ2 (σ 2 + nτ2)2

Finally,

∂ 2S w0(n)n[δ − s¯](σ 2 − nτ2) = < 0, (D.9) ∂τ2∂σ 2 (σ 2 + nτ2)3 provided that σ 2 − nτ2 < 0. The multiplicative nature of the comparative statics again ensures that the reverse results hold when s > δ. The ﬁrst part of the proposition reiterates that the direction of the salience effect depends upon the voter’s posterior belief—namely if it is above that of the challenger, i.e.

288 2 2 δσ +nsτ > 0. The magnitude of the negative impact of the salience effect is increasing σ 2+nτ2 in the precision of the signal (1/σ 2) and increasing in the weakness of the voter’s prior (1/τ2). The interaction shows that such results trade-off. Finally, I consider when the salience effect overpowers the learning effect. In particular,

Proposition 5. If s < δ, the salience effect compounds the negative learning effect when the

2 2 posterior belief about incumbent competence, δσ +nsτ , is positive (negative). The learning σ 2+nτ2 0 2 2 effect dominates a confounding salience effect when w (n) < σ τ (δ−s¯) w(n) (δσ 2+ns¯τ2)(σ 2+nτ2)

Proof: The results follow simply from part 1 of Proposition4, and setting L < S. This result unsurprisingly shows that when the salience effect is negative, it compounds the negative learning effect—and that this again requires that the posterior belief about the incumbent is below that of the challenger. Even if this does not hold, the second element of the proposition notes new information may still have a negative effect if the learning effect dominates the salience effect, which requires that the relative increase in salience is sufﬁciently small or the signal signiﬁcantly departs from the prior.

D.2 Months and years of municipal elections

Table D.1 reports the municipal elections potentially covered by the survey and aggregate elections in the main analysis.

D.3 Data description

D.3.1 ENCUP survey data

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 México July 2000, March 2003, March 2006, July 2009, July 2012. Michoacán 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 León 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. Querétaro 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 Potosí 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. Yucatán 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 elections were held in Hidalgo in February 2002 without concurrent municipal elections, and are counted as upcoming local elections. Italicized elections are not included in the sample for the homicide shocks analysis due to data unavailability (or exclusion in the case of the Federal District). Except in the cases of Baja California 2001 and 2004 and Oaxaca 2013, missingness reﬂects the fact that data from the preceding election required to deﬁne the change in vote share was not available. For Baja California 2001 and 2004, the precinct numbering changed across elections and thus cannot be matched. For Oaxaca 2013, precinct level data was unavailable.

290 where an election 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 1 for a respondent that answers that they watch political programs or listen to news at least ever/once a month/at least once a week/daily. (“¿Qué tan seguido escucha noticias o ve programas sobre política?”) Watch and listen to news and political programs scale. 5-point scale from 0 to 4, with values corresponding to levels of watching and listening to new and political programs (in ascending order). Political knowledge quiz. First factor from a factor analysis containing the following questions: What is the name of the youth movement that recently started in Mexico? (2012) Where was 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) Which party holds your state governorship? (2001, 2003, 2005, 2012) What is the name of your state governor? (2001) How many years do federal representatives serve for? (2001, 2003, 2005, 2012) What are the three separated powers of government? (2012) Who has the authority to approve changes to the constitution? (2001) Topical knowledge questions. First factor from a factor analysis containing the following questions: What is the name of the youth movement that recently started in Mexico? (2012) Where was 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) Which party holds your state governorship? (2001, 2003, 2005, 2012) What is the name of your state governor? (2001) Institutional knowledge questions. First factor from a factor analysis containing the following questions: How many years do federal representatives serve for? (2001, 2003, 2005, 2012) What are the three separated powers of government? (2012) Who has the authority to approve changes to the constitution? (2001) 291 Interest in politics. Indicator coded 1 for respondents reporting some or a lot of interest in politics (2003, 2005, 2012) or public affairs (2001). Actively engage in political discussion. Indicator coded 1 for respondents that answer that, when conversation turns to politics, they generally or sometimes participate in discussion and speak their opinion (“Por lo general, cuando usted está conversando con algunas personas y éstas empiezan a hablar de política. ¿Qué hace usted?”). Education. 4-point variable, where 0 is less than completed primary education, 1 is a maximum education level of completing high school, 2 is a maximum level of completing lower secondary education, 3 is a maximum level of complete secondary education (preparatoria), and 4 is at least a university degree. Homicide last month. Indicator coded 1 for respondents from a municipality where INEGI registers the occurrence of a homicide to a resident of the municipality within the month of the survey or the preceding month. The preceding month is used because some interviews were carried out at the beginning of the month, but respondent-specific interview times are unavailable. (Note that homicides figures are subject to reclassification across time.) Homicide shock. This indicator is coded 1 if either the number of homicides in the two months prior to the month of the survey (including the survey month) or the two months prior to the survey month exceed those in the two months immediately after the month of the survey, based on the INEGI monthly homicide statistics for the occurrence of homicides among a municipality’s residents. In 2005, the indicator is coded using the current month if the day of the month is greater than 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 the municipality in the preceding 12 months (excluding the current month). (Note that homicides figures are subject to reclassification across time.) 292 Homicides last 3 years. Average number of residents suffering a homicide per month within the municipality in the preceding 36 months (excluding the current month). (Note that homicides figures are subject to reclassification across time.) Public insecurity the major problem in the community. Indicator coded 1 if, in an open response, a respondent lists violence, crime or public security as the main problem facing their community (including as 0s respondents that listed no problem). Low confidence in police/mayor. Indicator coded 1 for respondents answering that their confidence in the police or municipal mayor is 5 or below on a scale from 0 to 10 (for the 2003, 2005, and 2012 surveys) or expressing no or almost no confidence in the police or municipal mayor (in 2001), in response to a question asking 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 member of, or previously being a member of. The number of possible organizations slightly varies across survey. Organizations talk about politics. Indicator coded 1 for respondents that answer that politics is discussed at the organizations they are a member of. Number of group meetings. The number of political organizations at which an individual has attended a meeting during the last year. Discuss community problems. A scale measuring the regularity with which respondents discuss 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. In Usos y Custombres in Oaxaca, the incumbent win margin is set to the maximum of 1. ENPV (lag). The effective number of political parties (by vote share) at the previous municipal mayoral election. In Usos y Custombres in Oaxaca, ENPV is set to the maximum 293 of 1. Incumbent won (lag). Indicator coded 1 for municipalities where the incumbent party was re-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 recent election (or an election held later in the year of the survey). Registered voters (lag). Number of registered voters in the municipality in the most recent election (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 (in thousands), divided by the total number of registered voters.

D.3.2 Homicide and electoral data

Change in incumbent party vote share. Change in the precinct-level share of all votes cast for the incumbent between the current municipal election and the prior municipal election (3 or 4 years earlier). When multiple parties form an incumbent coalition, the incumbent vote share is determined by the vote share of the largest party/coalition containing an incumbent party at the next election in terms of vote share. Incumbent party win. Indicator coded 1 if the incumbent party wins the municipal election. In the case of coalitions, is deﬁned similarly to the above. Change in turnout. Change in the precinct-level turnout rate between the current municipal election 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 a share of all registered voters, cast for the incumbent between the current municipal election and the prior municipal election (3 or 4 years earlier). Homicide shock. Deﬁned in equation (5.2) of the main paper, using INEGI homicide statistics for intentional homicides that occurred in each month to residents of a given

294 municipality. One- and three-month versions are similarly defined. (Note that homicides figures are subject to reclassification across time.) Placebo homicide shock (6 months earlier). Defined as in equation (5.2) of the main paper, with the exception that all months are shifted 6 months forward in time. (Note that homicides figures are subject to reclassification across time.) Average monthly homicide rate (12 months/3 years before election). Average number of residents suffering a homicide per month within the municipality in the 12 months/3 years prior to the municipal election, again based on INEGI homicide data. (Note that homicides figures are subject to 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 state police 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 were excluded because channels of accountability are hard to discern. These categorizations were homogenized across the 2000, 2002, 2004, 2011 and 2013 ENGM surveys. Missing years were imputed according to the following rules: I first used the last available data, and if no previous coding was available took the nearest year in the future. Calderón Presidency. Indicator coded one for elections in the years 2007-2012. PAN/PRD. Indicator coded 1 for PAN/PRD municipal incumbents.

D.3.3 Local media coverage data

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

295 cality locations). Local AM radio/FM radio/television. Number of AM radio/FM radio/television stations, with an 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 locality locations). Non-local media. Number of AM radio, FM radio or television stations, with an emitter based outside the precinct’s municipality, covering at least 20% of the precinct population (as deﬁned by detailed 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% of residents had experienced higher education in 2010, according to the 2010 Census.

D.4 Map of municipalities included in different samples

In separate ﬁgures, Figure D.1 shades in red the municipalities that appear at least once in each of the main empirical analyses—the survey sample, the homicide shock sample, and the local media sample.

D.5 Additional analyses

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

D.5.1 Assessing the identiﬁcation assumptions

Tables D.2-D.8 show the results of balance tests for the three main sets of empirical ﬁndings in the paper: Table D.2 shows that upcoming local elections are well balanced

296 (a) Municipalities in the ENCUP survey samples

(b) Municipalities in the homicide shock sample

Figure D.1: Municipalities included in each empirical analysis (shaded in red)

297 across individual and municipal characteristics in the ENCUP survey data; Table D.4 shows that homicide shocks are well balanced across a wide variety of covariates, and Table D.5 shows that such balance holds once fixed effects are removed for the time-invariant Census and geographic variables; and Table D.8 shows that the number of local media stations is well-balanced across these same covariates. In addition, Table D.3 demonstrates that neither changes in upcoming local elections nor changes in measures of local violence predict whether a municipality is included in any given year. In each panel the outcome is an indicator for whether a municipality is included in that survey wave, conditional on a municipality appearing in at least one of the four ENCUP surveys. I also provide additional tests to support the exogeneity of homicide shocks. First, Fig- ure D.2 shows that the distribution of homicides does not vary across the period prior to that used to define the treatment, in the two months before an election, or the two months after an election. Such evidence is consistent with sampling variability, rather than strategic behavior. Second, using a simple specification containing municipality and year fixed effects, Table D.6 shows that the number of police per voter does not systematically vary across election and non-election years. Third, Table D.7 shows that election outcomes are uncorrelated with homicide counts in the two months after elections. In particular, columns (3) and (4) find no correlation between the identity of the winning party and post-election homicides either throughout the sample or during the Calderón administration. This does 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.

298 Table D.2: Balance of upcoming local elections in the ENCUP surveys over 20 individual and municipal level variables

Female Catholic Age Education Employed Own econ. Number Org.s Number Discuss last year in month of talk about of group community org.s politics meetings problems (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Upcoming local election -0.007 -0.011 -0.100 0.202*** -0.015 -0.029 -0.066 -0.013 -0.084 0.005 (0.011) (0.020) (0.394) (0.075) (0.016) (0.021) (0.077) (0.021) (0.103) (0.029)

Observations 15,976 11,983 15,976 11,756 11,756 12,322 15,976 12,576 15,976 12,576 Outcome mean 0.55 0.81 40.76 1.70 0.47 1.57 1.06 0.26 1.52 0.70

299 Local election mean 0.16 0.20 0.16 0.18 0.18 0.20 0.16 0.20 0.16 0.20 Homicides Homicide Homicide Homicide Incumbent ENPV Incumbent Incumbent Registered Police last month shock last year last 3 years win margin (lag) won vote share voters per voter (lag) (lag) (lag) (lag) (last year) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) Upcoming local election 0.060 -0.022 1.607 1.607 0.021 0.011 0.074 0.012 107,854.002* -0.063 (0.058) (0.081) (1.080) (1.090) (0.016) (0.066) (0.059) (0.011) (59,015.253) (0.178)

Observations 15,941 12,664 15,941 15,941 15,976 15,976 15,976 15,778 15,778 13,666 Outcome mean 0.69 0.44 4.50 4.82 0.15 2.57 0.61 0.48 252,422 2.30 Local election mean 0.16 0.17 0.16 0.16 0.16 0.16 0.16 0.17 0.17 0.16

Notes: All speciﬁcations include survey year ﬁxed effects, and are estimated using OLS. Standard errors clustered by municipality in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table D.3: 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 2,088 Outcome 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 2,088 Outcome mean 0.46 0.52 0.46 0.46

Notes: All speciﬁcations include municipality and survey-year ﬁxed effects, 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. 300 Table D.4: Balance over 104 variables over pre-election homicide shocks

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

Area (km2) 0.633 (1.038) Single victims 0.035 (0.024) Electorate 72.164* (43.837) Job-related homicides 0.014 (0.021) Electorate density 49.090 (38.848) Organs examined after homicide 0.002 (0.009) Municipal electorate 2968.9 (5635.7) Family-incident homicide -0.001 (0.010) Local media -0.003 (0.007) Urban victims 0.002 (0.017) Non-local media -0.001 (0.009) Non-Mexican victims 0.003 (0.006) Incumbent vote share (lag) 0.002 (0.007) Neighbor average homicide shock 0.023 (0.024) Incumbent win (lag) 0.049 (0.036) Electorate-weighted neigh. ave. hom. shock 0.016 (0.025) Win margin (lag) -0.001 (0.008) Average non-homicide deaths (prior 3 years) -1.354 (2.341) Municipality win margin (lag) 0.002 (0.011) Average non-homicide deaths (prior year) -3.042 (2.487) ENPV (lag) -0.023 (0.031) Average child mortalities (prior 3 years) -0.202 (0.283) Municipality ENPV (lag) -0.031 (0.035) Average child mortalities (prior year) -0.040 (0.253) Turnout (lag) -0.011 (0.008) Total municipal spending 8.730 (70.125) PRI incumbent -0.047 (0.033) No municipal police force -0.040** (0.017) PAN incumbent 0.027 (0.031) Police per voter 0.024 (0.068) PRD incumbent 0.025 (0.029) Occupants per dwelling -0.001 (0.001) Mayor’s party aligned with president 0.027 (0.033) Occupants per room -0.001 (0.001) Mayor’s party aligned with governor -0.035 (0.036) Share with 2 bedrooms 0.000 (0.001) Homicides 12 months before election -2.686 (2.418) Share 3+ bedrooms 0.000 (0.001) Homicides 11 months before election -2.376 (3.037) Share female 0.000 (0.000) Homicides 10 months before election -3.165 (2.850) Share working age 0.000 (0.000) Homicides 9 months before election -2.687 (2.619) Children per woman 0.002 (0.002) Homicides 8 months before election -3.611 (2.262) Share born out of state -0.001** (0.000) Homicides 7 months before election -2.874 (2.354) Share Catholic 0.000 (0.000) Homicides 6 months before election -2.752 (2.499) Share indigenous speakers 0.000 (0.000) Homicides 5 months before election -2.238 (1.886) Years of schooling -0.017 (0.015) Homicides 4 months before election -2.009 (2.631) Female years of schooling -0.007 (0.007) Homicides 3 months before election -2.212 (2.514) Male years of schooling -0.007 (0.008) Average homicides (prior 3 years) -2.985 (2.706) Share illiterate 0.000 (0.000) Average homicides (prior year) -2.661 (2.463) Share with no schooling 0.000 (0.000) Top homicide quartile (prior year) -0.017 (0.032) Share incomplete primary school 0.000 (0.000) Top homicide decile (prior year) -0.008 (0.019) Share complete primary school 0.000 (0.000) Average drug-related homicides (prior year) 0.811 (0.755) Share incomplete secondary school -0.001 (0.001) High-risk electoral precinct 0.002 (0.004) Share complete secondary school -0.001 (0.001) Gun-related homicides -2.579 (3.173) Share higher education 0.000 (0.001) Chemical substance-related homicides 0.001 (0.012) Share economically active 0.000 (0.000) Hanging-related homicides 0.089 (0.097) Share without health care 0.001** (0.000) Drowning-related homicides 0.031 (0.019) Share state workers health care 0.000 (0.000) Explosives-related homicides -0.099 (0.090) Share running water -0.001 (0.000) Smoke/ﬁre-related homicides 0.014 (0.019) Share drainage 0.000 (0.000) Cutting object-related homicides -0.008 (0.134) Share washing machine -0.001* (0.000) Blunt object-related homicides -0.011 (0.031) Share landline telephone 0.000 (0.001) Male homicide victims 0.010 (0.019) Share radio 0.000 (0.000) Delayed homicide registration -2.692 (3.532) Share fridge -0.001 (0.000) Age of homicide victims 1.710** (0.807) Share cell phone -0.001 (0.001) Victims with no schooling 0.011 (0.014) Share television 0.000 (0.000) Victims with incomplete primary schooling -0.011 (0.014) Share car or truck 0.000 (0.000) Victims with complete primary schooling -0.012 (0.022) Share computer 0.000 (0.001) Victims with secondary schooling 0.019 (0.030) Share internet 0.000 (0.001) Pre-election between-cartel violence 0.639 (0.535) Pre-election drug seizures 0.536 (1.106) Pre-election violent enforcement 0.001 (0.098) Pre-election asset seizures -0.274 (0.239) Pre-election drug-related arrests 0.037 (0.703) Pre-election gun seizures 0.337 (0.410)

Notes: All speciﬁcations include municipality and year ﬁxed effects, and are estimated using OLS. Most variables cover almost the entire main sample. Standard errors clustered by municipality in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

301 Table D.5: Balance over 37 time-invariant variables over pre-election homicide shocks, excluding ﬁxed effects

Outcome Effect of homicide shock Coefﬁcient estimate Standard error Area (km2) 5.096** (2.191) Local media -1.495 (1.276) Non-local media -1.066 (1.379) Occupants per dwelling -0.013 (0.016) Occupants per room 0.012 (0.013) Share with 2 bedrooms -0.012 (0.007) Share 3+ bedrooms -0.011 (0.008) Share female 0.000 (0.001) Share working age -0.001 (0.002) Children per woman 0.016 (0.018) Share born out of state -0.015 (0.013) Share Catholic -0.005 (0.006) Share indigenous speakers 0.004 (0.004) Years of schooling -0.093 (0.09) Female years of schooling -0.066 (0.088) Male years of schooling -0.087 (0.092) Share illiterate 0.003 (0.003) Share with no schooling 0.002 (0.003) Share incomplete primary school -0.002 (0.003) Share complete primary school -0.007 (0.005) Share incomplete secondary school -0.007 (0.007) Share complete secondary school -0.007 (0.007) Share higher education -0.005 (0.008) Share economically active -0.005* (0.003) Share without health care -0.003 (0.006) Share state workers health care 0.002 (0.002) Share running water -0.008 (0.006) Share drainage -0.008 (0.005) Share washing machine -0.013 (0.01) Share landline telephone -0.027* (0.014) Share radio -0.011* (0.006) Share fridge -0.011 (0.008) Share cell phone -0.009 (0.01) Share television -0.006* (0.003) Share car or truck -0.011 (0.01) Share computer -0.014 (0.011) Share internet -0.013 (0.01)

Notes: All speciﬁcations are unadjusted differences-in-means, and are estimated using OLS. Most variables cover almost the entire main sample. Standard errors clustered by municipality in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

302 L.Mncplte ihn mlye r xldd tnaderr lsee ymncplt r in are municipality by clustered errors Standard denotes excluded. * are parentheses. employees no with Municipalities OLS. Notes Density al .:Efc flcleetoso uiia ulcscrt employment security public municipal on elections local of Effect D.6: Table

l pcfiain nld uiiaiyadsre-erfie fet,adaeetmtdusing estimated are and effects, fixed survey-year and municipality include specifications All : 0 .1 .2 .3 .4 0 iueD2 itiuino oiie yrlto oelection to relation by homicides of Distribution D.2: Figure ucm en42 4.23 0.31 9,655 [.01,201.86] 4.23 [.01,201.86] 9,655 0.31 -0.038 trends time Municipality-specific 0.006 mean year election Local range Outcome mean Outcome Observations year election Local Two monthsbeforeelection Months 5and6beforeelection p < 5 .,* denotes ** 0.1, Average numberofhomicidespermonth 10 p < 303 .5 * denotes *** 0.05, mlye e 00voters 1000 per employees uiia ulcsecurity public Municipal 009 (0.082) (0.069) 1 (2) (1) Two monthsafterelection Months 3and4beforeelection 15 p < 0.01. X 20 25 Table D.7: Correlation between election outcomes and post-election homicides

Average homicides in the two months after an election (1) (2) (3) (4) Change in incumbent party vote share -13.103 (9.784) Incumbent party win 2.326 (5.230) PAN win -0.093 -17.484 (5.456) (14.481) PRI win 3.604 -10.075 (3.550) (16.397) PRD win -4.166 -13.681 (6.124) (29.453)

Observations 181,408 181,408 181,408 80,221 Outcome mean 18.52 18.52 18.52 28.23 Election outcome mean -0.05 0.56

304 Table D.8: Balance over 102 variables over the number of local media stations

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

Area (km2) -0.002 (0.006) Victims with complete primary schooling 0.000 (0.000) Electorate -20.327 (15.440) Victims with secondary schooling -0.000 (0.000) Electorate density -30.091 (36.848) Single victims 0.000 (0.000) Municipal electorate 6.945 (24.007) Job-related homicides 0.000 (0.000) Distance from centroid to municipal head (log) -0.003 (0.003) Organs examined after homicide -0.000 (0.000) Non-local media -0.035 (0.162) Family-incident homicide 0.000 (0.000) Incumbent vote share (lag) -0.000 (0.000) Urban victims 0.000 (0.000) Incumbent win (lag) -0.000 (0.000) Non-Mexican victims -0.000 (0.000) Win margin (lag) 0.000 (0.000) Neighbor average homicide shock 0.000 (0.000) Municipality win margin (lag) 0.000 (0.000) Electorate-weighted neigh. ave. hom. shock 0.000 (0.000) ENPV (lag) 0.002 (0.002) Average non-homicide deaths (prior 3 years) -0.012 (0.018) Municipality ENPV (lag) 0.000 (0.000) Average non-homicide deaths (prior year) -0.012 (0.018) Turnout (lag) 0.000 (0.001) Average child mortalities (prior 3 years) 0.001 (0.001) PRI incumbent 0.000 (0.000) Average child mortalities (prior year) 0.000 (0.001) PAN incumbent -0.000 (0.000) Total municipal spending -0.094 (0.315) PRD incumbent 0.000 (0.000) No municipal police force 0.000 (0.000) Mayor’s party aligned with president -0.000 (0.000) Police per voter 0.000 (0.000) Mayor’s party aligned with governor -0.000 (0.000) Occupants per dwelling -0.003 (0.002) Homicide shock -0.000 (0.000) Occupants per room -0.003** (0.001) Homicides 12 months before election 0.003 (0.007) Share with 2 bedrooms 0.000 (0.001) Homicides 11 months before election 0.002 (0.010) Share 3+ bedrooms 0.001* (0.001) Homicides 10 months before election -0.002 (0.006) Share female 0.000 (0.000) Homicides 9 months before election 0.001 (0.007) Share working age -0.000 (0.000) Homicides 8 months before election 0.001 (0.005) Children per woman -0.001 (0.002) Homicides 7 months before election -0.008 (0.008) Share born out of state 0.000 (0.000) Homicides 6 months before election -0.001 (0.009) Share Catholic 0.000 (0.001) Homicides 5 months before election -0.003 (0.009) Share indigenous speakers 0.000 (0.000) Homicides 4 months before election -0.000 (0.006) Years of schooling 0.010 (0.009) Homicides 3 months before election 0.001 (0.005) Female years of schooling 0.005 (0.009) Average homicides (prior 3 years) 0.003 (0.004) Male years of schooling 0.014 (0.009) Average homicides (prior year) -0.001 (0.007) Share illiterate -0.000* (0.000) Top homicide quartile (prior year) -0.000 (0.000) Share with no schooling -0.000 (0.000) Top homicide decile (prior year) -0.000 (0.000) Share incomplete primary school 0.000 (0.000) Average drug-related homicides (prior year) 0.000 (0.001) Share complete primary school 0.000 (0.000) High-risk electoral precinct 0.001 (0.001) Share incomplete secondary school 0.001 (0.001) Gun-related homicides -0.000 (0.005) Share complete secondary school 0.001 (0.001) Chemical substance-related homicides 0.000 (0.000) Share higher education 0.001 (0.001) Hanging-related homicides 0.000* (0.000) Share economically active 0.000 (0.000) Drowning-related homicides -0.000 (0.000) Share without health care -0.000 (0.000) Explosives-related homicides -0.000 (0.000) Share state workers health care 0.000 (0.000) Smoke/ﬁre-related homicides -0.000 (0.000) Share running water 0.001** (0.001) Cutting object-related homicides -0.000 (0.000) Share drainage 0.001** (0.000) Blunt object-related homicides -0.000 (0.000) Share washing machine 0.001 (0.001) Delayed homicide registration -0.000 (0.005) Share fridge 0.000 (0.001) Male victims -0.000 (0.000) Share cell phone 0.000 (0.001) Age of victims -0.008 (0.006) Share car or truck 0.000 (0.001) Victims with no schooling -0.000 (0.000) Share computer 0.001 (0.001) Victims with incomplete primary schooling 0.000 (0.000) Share internet 0.002* (0.001) Pre-election between-cartel violence 0.000 (0.002) Pre-election drug seizures 0.004 (0.004) Pre-election violent enforcement 0.000 (0.000) Pre-election asset seizures 0.001 (0.001) Pre-election drug-related arrests 0.004* (0.002) Pre-election gun seizures 0.000 (0.001)

305 Table D.9: The effect of upcoming local elections on political news consumption and political engagement, controlling ﬂexibly for education

Watch Watch Topical and listen and listen knowledge ever scale questions (1) (2) (3) Upcoming local election 0.033* 0.154* 0.203*** (0.018) (0.080) (0.060)

Observations 7,763 7,763 11,756 Outcome mean 0.86 2.56 -0.04 Outcome range {0,1} {0,1,2,3,4} [-1.5,2.2] Upcoming local election mean 0.23 0.23 0.17 Survey year without data 2001, 2005 2001, 2005 2005

D.5.2 Robustness of survey results

Table D.9 shows that the main impacts of a local election on information consumption are robust to controlling flexibly for education. In particular, at the cost of losing data from 2005 (when education was not measured), I include indicators for four different levels of schooling. At a considerable efficiency cost arising from effectively removing half the states in the sample that never experience an upcoming election, Table D.10 shows that I obtain similar results when I include state fixed effects. The results are unsurprisingly more precise in column (3), given that the 2001 wave can also be used. Table D.11 examines the robustness of my proximate election indicator by examining the effect of the number of months until the next election. Panel A shows substantively similar results: the negative coefficient on the number of months until the next election indicates that voters consume significantly more information as elections approach, and that such consumption continues to translate into political knowledge and discussion. Panel

306 Table D.10: The effect of upcoming local elections on political news consumption and political engagement, controlling for state ﬁxed effects

Watch Watch Topical and listen and listen knowledge ever scale questions (1) (2) (3) Upcoming local election 0.013 0.189* 0.192** (0.019) (0.101) (0.076)

Observations 11,983 11,983 15,976 Outcome mean 0.86 2.55 -0.05 Outcome range {0,1} {0,1,2,3,4} [-1.5,2.2] Upcoming local election mean 0.19 0.19 0.16 Survey year without data 2001 2001

Notes: All speciﬁcations include state 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, *** denotes p < 0.01.

B includes a quadratic term to explore the anticipated non-linearity of the effect: consistent with the 5-month indicator used in the main paper, the positive quadratic term shows that the highest marginal effect of proximity an the election is just before an election; in the years before an election, moving a month closer is relatively unimportant. To further characterize the political information cycle, I interact the upcoming local election indicator with the number of months until the election in panel C. The negative interaction indicates that even within the months just before an election, proximity continues to increase consumption. Together these findings suggest that voters rarely consume information about politics outside of electoral campaign, but during campaigns—and especially right before elections—consumption increases significantly. Table D.12 shows that homicides coinciding with upcoming local elections also cause voters to reduce their confidence in the police.5 Columns (1)-(4) in demonstrate that local elections only increase the likelihood that a respondent expresses low confidence in the

5The questions do not distinguish between municipal, state, and federal forces.

307 Table D.11: The effect of months until local elections on political news consumption and political engagement

Watch Watch Political and listen and listen knowledge ever scale quiz (1) (2) (3) Panel A: Linear relationship Months until next election -0.0017*** -0.0146*** -0.0039*** (0.0006) (0.0026) (0.0015)

Observations 11,983 11,983 15,976 Outcome mean 0.86 2.55 0.00 Local election mean 17.66 17.66 18.75 Survey year without data 2001, 2005 2001, 2005 2005 Panel B: Quadratic relationship Months until next election -0.0070*** -0.0256** -0.0376*** (0.0023) (0.0110) (0.0066) Month until next election squared 0.0001** 0.0003 0.0009*** (0.0001) (0.0003) (0.0002)

Observations 11,983 11,983 15,976 Outcome mean 0.86 2.55 0.00 Local election mean 17.66 17.66 18.75 Survey year without data 2001, 2005 2001, 2005 2005 Panel C: Interaction with campaign indicator Upcoming local election 0.085*** 0.307* 0.491*** (0.022) (0.168) (0.151) Local election 0.107*** 0.426** 0.619*** (0.023) (0.169) (0.150) Months until next election -0.001 -0.013*** 0.002 (0.001) (0.003) (0.001) Upcoming local election × Months until -0.018*** -0.085** -0.081** next election (0.005) (0.036) (0.033)

Observations 11,983 11,983 15,976 Outcome mean 0.86 2.55 0.00 Local election mean 0.19 0.19 0.16 Survey year without data 2001, 2005 2001, 2005 2005

Notes: All specifications include survey-year fixed effects, dummies for completing primary education, lower secondary education, full secondary education, and university, and are estimated using OLS. Stan- dard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. 308 Table D.12: Heterogeneous effects of upcoming local elections on confidence in the police, by municipal homicide measures Homicide measure: Homicide Homicide Homicides Homicides last month shock last year last 3 years Low confidence in the police (1) (2) (3) (4) Upcoming local election -0.042 0.002 0.004 0.002 (0.033) (0.032) (0.026) (0.026) Homicide measure 0.017 -0.032** 0.002*** 0.001*** (0.015) (0.016) (0.001) (0.000) Upcoming local election 0.125*** 0.142*** 0.009*** 0.010*** × Homicide measure (0.040) (0.042) (0.002) (0.002)

Observations 15,530 12,382 15,530 15,530 Outcome mean 0.48 0.48 0.48 0.48 Outcome range {0,1} {0,1} {0,1} {0,1} Upcoming local election mean 0.16 0.17 0.16 0.16 Homicide measure mean 0.69 0.44 4.57 4.90 Notes: All specifications include survey year fixed effects, 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. police in violent municipalities. In each column, the coefficient on local elections is statistically insignificant, while the interaction with both short-term and longer-term homicide measures is positive and highly statistically significant. Although all types of homicide shock affect confidence in the police, the effects of the short-term measures are notably larger than for the means of the longer-term measures: the effect of an additional homicide using the short-term measures doubles the effect of an additional homicide over the prior year or 3 years.6 Such declines in confidence in the police complement studies documenting a strong negative correlation between perceptions of public insecurity and lower institutional trust (see Blanco 2013).

6Based on an F test comparing the short-term interactive coefficient, multiplied with the average number of homicides or change in average number of homicides, with the long-term coefficients. Moreover, controlling for all measures simultaneously, the short-term measures remain large and significant.

309 Table D.13: Effect of pre-election drug-related homicide shocks on municipal incumbent electoral outcomes

Change in Incumbent incumbent party party win vote share (1) (2) Drug-related homicide shock -0.011 -0.040 (0.016) (0.067)

Observations 41,405 41,405 Outcome mean -0.04 0.56 Homicide shock mean 0.47 0.47

D.5.3 Drug-related homicides

As noted in the main text, the Calderón government released monthly municipal data on the number of drug-related deaths between 2007 and 2011. However, there are data issues with such data. First, these numbers are contentious (see Heinle, Rodríguez Ferreira and Shirk 2014), and do not necessarily follow the homogeneous coroner’s report criteria used by INEGI. Second, the limited availability of this data, combined with the deﬁnition of a homicide shock that requires at least one drug-related death over the four-month window around elections, substantially reduces the sample size by around 75%. More generally, it is not clear theoretically whether voters should respond more or less to drug-related homicides, as opposed to other homicides. In fact, voters might think that these are less relevant to them than more arbitrary murders which constitute around 50% of totals 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.

310 The results in Table D.13 report the effect of drug-related homicide shocks just before an election, defined according to equation (5.2) but instead using drug-related homicides. Given that only five years of data exist, and thus many municipalities only appear once, I use state fixed effects instead of municipality fixed effects. Although the large drop in sample size unsurprisingly reduces precision substantially, the point estimates are relatively similar—negative, and large—to those reported in Table 5.3. These findings thus further reinforce the claim in the main text that voters respond similarly to different types of homicide.

D.5.4 Alternative approaches to capturing homicide shocks

The results in the main text focus on homicide shocks coded as a binary variable. A key advantage of this approach is that it is highly short-term, and thus accurately maps political information cycles. By exploiting idiosyncratic shifts in monthly homicide counts, I am able to generate plausibly exogenous variation. However, I also now consider two different approaches to capturing the effects of short-run shocks around local elections. First, I utilize a difference-in-differences design to identify how changes in the number of homicides around elections, relative to previous elections, affect vote choice. Since the homicide level, as opposed to the change exploited in the main paper, just before an election is highly correlated with the general homicide level, a meaningful test of increased homicides around elections requires a subtler design.7 Accordingly, I use the difference between the average number of homicides in the two months prior to an election and the the average number of homicides over the entire electoral cycle. This captures differences in the magnitude of pre-election shocks across elections, and thus exploits variation in the intensity of these shocks across municipalities. The identifying assumption is that municipalities

7The results of such an approach are thus similar to the small long-run estimates reported in Table 5.5.

311 Table D.14: Alternative approaches to measuring the effect of short-run homicides rates on municipal incumbent electoral outcomes

Panel A: Deviations in pre-election Change in incumbent Incumbent homicide levels party vote share party win (1) (2) (3) (4) Deviation from average monthly homicide -0.0012** -0.0011** -0.0020 -0.0013 (0.0006) (0.0005) (0.0020) (0.0018)

Observations 181,408 181,408 181,408 181,408 Outcome mean -0.05 -0.05 0.56 0.56 Deviation mean 0.99 0.99 0.99 0.99 Deviation standard deviation 15.42 15.42 15.42 15.42 Municipality-speciﬁc time trends XX Panel B: Proportional change in Change in incumbent Incumbent homicide level party vote share party win (1) (2) Proportional change in homicides -0.0076* -0.0335** (2 months before v. after election) (0.0042) (0.0167)

Observations 151,074 151,074 Outcome mean -0.05 0.57 Proportional change mean -0.10 -0.10 Proportional change standard deviation 1.09 1.09

Notes: All specifications are estimated using OLS, and include municipality and year fixed effects. All observations are weighted by the number of registered voters in the electoral precinct. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. experiencing different homicide deviations from cycle trends before elections otherwise follow parallel trends in incumbent support. Panel A of Table D.14 reports estimates from the analog of equation (5.4), and demonstrates that this measure of homicide shocks also robustly decreases the incumbent party’s vote share. The estimates are substantially larger than the long-run estimates for the monthly homicide rate over the entire electoral cycle reported in panel B of Table 5.5. Although the effect on the incumbent’s win probability is negative in columns (3) and (4), it is not statistically significant. This is likely because small changes in the shock are rarely sufficient to push incumbent to lose.

312 Second, to capture the intensity of the shock I refine the pre-election homicide shock used in the main paper. Instead, I use the proportional change in the number of homicides before as opposed to after the election, i.e. the difference between homicides in the two months before the election relative to the two months after the election divided by the number of homicides in the two months before the election. Panel B of Table D.14 reports large negative coefficients on the proportional measure. These findings support the main findings of the paper by showing that larger pre-election homicide shocks produce bigger electoral responses.

D.5.5 Robustness of no municipal police results

Table D.15 examines how the interaction between a homicide shock and the existence of a municipal police force changes when other interactions are simultaneously controlled for. In particular, to address the concern that the effect in municipalities without their own police force is simply driven by other differences, I simultaneously control for the interaction between a homicide shock and the number of registered voters in the municipality, the number of local media stations covering a precinct, and the lagged incumbent vote share. The results indicate that these potentially confounding explanations do not appear to explain the signiﬁcant interaction between the existence of municipal police and a homicide shock.

D.5.6 Learning from the experiences of neighboring municipalities

To identify how comparative performance information impacts voter behavior, I compute the average homicide shock across neighboring municipalities.8 The exogeneity of homicide shocks within municipalities facilitates a causal interpretation of the interaction

8The results are robust to weighting neighboring municipalities by the size of their registered electorates.

313 Table D.15: Effect of pre-election homicide shocks on municipal incumbent electoral outcomes, by existence of a municipal-level police force

Change in Incumbent incumbent party party win vote share (1) (2) Homicide shock × No municipal police force 0.039* 0.295** (0.023) (0.127)

Observations 171,482 171,482 Outcome mean -0.05 0.56 Homicide shock mean 0.46 0.46 No municipal police force mean 0.09 0.09

Notes: All specifications are estimated using OLS, include municipality and year fixed effects, and control for the number of registered voters in the municipality, the number of local media stations covering a precinct, the lagged incumbent vote share, and the interaction of each of these variables with a homicide shock. The coefficients of these variables are omitted. All observations are weighted by the number of registered voters in the electoral precinct. Standard errors clustered by municipality are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. between homicide shocks within a voter’s municipality and shocks to their neighbors. Con- sistent with voters benchmarking their incumbent’s performance against that of their neighbors, column (2) shows that the probability that an incumbent party is re-elected increases with the average shock experienced by neighboring municipalities. However, the negative interaction between the homicide shock and the neighbor average implies that relative performance considerations do not apply when a voter’s own municipality is afflicted by a shock.9 Plausibly consistent with increased salience, mayoral punishment instead increases with the proportion of shocks experienced by neighbors. The change is vote share follows a similar pattern but is not statistically significant, while columns (3) and (4) report similar results when an electorate-weighted neighboring municipal shock measure is used instead.

9Unreported results separating neighboring incumbents from the same and different parties, but show no meaningful differences.

314 Table D.16: Effect of pre-election homicide shocks on municipal incumbent electoral outcomes, by neighbor average homicide shock

Change in Incumbent Change in Incumbent incumbent party incumbent party party win party win vote share vote share (1) (2) (3) (4) Homicide shock -0.007 0.070 -0.010 0.039 (0.017) (0.068) (0.017) (0.067) Neighbor average homicide shock 0.009 0.273*** (0.020) (0.097) Homicide shock × Neighbor -0.032 -0.376*** average homicide shock (0.029) (0.129) Electorate-weighted neighbor 0.012 0.249** average homicide shock (0.019) (0.095) Homicide shock × Electorate-weighted -0.027 -0.327** neighbor average homicide shock (0.027) (0.129)

Observations 176,982 176,982 181,358 181,358 Outcome mean -0.05 0.56 -0.05 0.56 Homicide shock mean 0.46 0.46 0.46 0.46 Interaction mean 0.48 0.48 0.47 0.47

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

315 Table D.17: Effects of pre-election homicide shocks on municipal incumbent electoral outcomes, by incumbent party

Change in Change in Change in Change in Change in incumbent PAN PRD PRI PAN party incumbent incumbent incumbent incumbent vote share vote share vote share vote share vote share (1) (2) (3) (4) (5) Homicide shock -0.008 -0.028** -0.059*** -0.012 0.029 (0.013) (0.012) (0.021) (0.015) (0.020) Homicide shock × PAN -0.021 (0.021) Homicide shock × PRD -0.037 (0.035) Homicide shock × Calderón 0.042 Presidency (0.042)

Observations 173,063 61,388 20,718 90,957 181,408 Outcome mean -0.05 -0.06 -0.09 -0.03 -0.05 Homicide shock mean 0.46 0.46 0.45 0.47 0.46

D.5.7 Differential impacts of homicide shocks across parties

I also examine differential punishment across parties. Due to the PRI’s more extensive clientelistic ties (e.g. Cornelius 1996; Fox 1994; Greene 2007; Magaloni 2006), PRI voters may be less susceptible to performance information, and thus less inclined to punish PRI incumbents for homicide shocks. Table D.17 provides tentative evidence consistent with this expectation. Although the interaction terms are not statistically signiﬁcant, column (1) shows that the PAN and PRD are punished relatively more than the PRI. Although punishment of PRI incumbents is not statistically, columns (2)-(4) register signiﬁcant punishment of PAN and PRD mayors. This is likely because the PAN and especially PRD control more urban areas where the effects of homicide shocks are larger.

316 Finally, who do voters turn to to reduce local violence? If reducing violence is a major concern, and voters believe that Calderón’s tough stance on drug-related crime may help their municipality (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 vote share by 4.2 percentage points during Calderón’s presidency, this interaction is not quite statistically significant.

317 | Bibliography

Abrams, Samuel, Torben Iversen and David Soskice. 2011. “Informal social networks and rational voting.” British Journal of Political Science 41(2):229–257.

Achen, Christopher H. and Larry M. Bartels. 2004a. “Blind retrospection: Electoral responses to drought, ﬂu, and shark attacks.” Working paper.

Achen, Christopher H. and Larry M. Bartels. 2004b. “Musical chairs: Pocketbook voting and the limits of democratic accountability.” Working paper.

Adams, James and Lawrence Ezrow. 2009. “Who do European parties represent? How Western European parties represent the policy preferences of opinion leaders.” Journal of Politics 71(1):206–223.

Adena, Maja, Ruben Enikolopov, Maria Petrova, Veronica Santarosa and Ekaterina Zhu- ravskaya. 2015. “Radio and the Rise of The Nazis in Prewar Germany.” Quarterly Jour- nal of Economics 130(4):1885–1939.

Aldashev, Gani. 2010. “Political Information Acquisition for Social Exchange.” Quarterly Journal of Political Science 5(1):1–25.

Alt, James E., Amalie Jensen, Horacio A. Larreguy, David D. Lassen and John Marshall. 2016. “Contagious political concerns: Identifying unemployment concern peer effects using the Danish population network.” Working paper.

Alt, James E., David D. Lassen and John Marshall. 2016. “Information sources, belief updating, and the politics of economic expectations: Evidence from a Danish survey experiment.” Journal of Politics 78(2):327–343.

Alvarez, Michael E. 2001. Information and Elections. Ann Arbor: University of Michigan Press.

Anderson, Cameron D. 2006. “Economic voting and multilevel governance: a comparative individual-level analysis.” American Journal of Political Science 50(2):449–463.

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

318 Anderson, Simon P. and John McLaren. 2012. “Media mergers and media bias with rational consumers.” Journal of the European Economic Association 10(4):831–859. Ansolabehere, Stephen, Erik C. Snowberg and James M. Snyder. 2006. “Television and the incumbency advantage in US elections.” Legislative Studies Quarterly 31(4):469–490. Ansolabehere, Stephen and Shanto Iyengar. 1995. Going Negative. New York: Free Press. Arias, Eric. 2015. “The Diffusion of International Norms: A Field Experiment on the Role of Common Knowledge.” Working paper. Arias, Eric, Horacio A. Larreguy, John Marshall and Pablo Querubín. 2016. “Priors rule: When do malfeasance revelations help and hurt incumbent parties?” Working paper. 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. 2014a. “Is Voter Competence Good for Voters? Information, Rationality, and Democratic Performance.” American Political Sci- ence Review 108(3):565–587. Ashworth, Scott and Ethan Bueno de Mesquita. 2014b. “Learning About Voter Rational- ity.” Working paper. Auditoría Superior de la Federación. 2014. “Informe del Resultado de la Fiscalización Superior de la Cuenta Pública 2012.” Audit Summary Report. Austen-Smith, David and Roland G. Fryer. 2005. “An Economic Analysis of “Acting White”.” Quarterly Journal of Economics 120(2):551–583. Baker, Andy, Barry Ames and Lucio R. Renno. 2006. “Social context and campaign volatility in new democracies: networks and neighborhoods in Brazil’s 2002 Elections.” Amer- ican Journal of Political Science 50(2):382–399. Banerjee, Abhijit, Donald P. Green, Jeffery McManus and Rohini Pande. 2014. “Are poor voters indifferent to whether elected leaders are criminal or corrupt? A vignette experiment in rural India.” Political Communication 31(3):391–407. Banerjee, Abhijit V., Selvan Kumar, Rohini Pande and Felix Su. 2011. “Do Informed Voters Make Better Choices? Experimental Evidence from Urban India.” Working paper. Barabas, Jason and Jennifer Jerit. 2009. “Estimating the Causal Effects of Media Coverage on Policy-Speciﬁc Knowledge.” American Journal of Political Science 53(1):73–89. Baron, David P. 2006. “Persistent media bias.” Journal of Public Economics 90(1):1–36. Barro, Robert J. 1973. “The control of politicians: An economic model.” Public Choice 14(1):19–42. 319 Bartels, Larry M. 2008. Unequal democracy: The political economy of the New Gilded Age. Princeton: Princeton University Press.

Bates, Robert H. 1981. Markets and States in Tropical Africa: The Political Basis of Agricultural Policies. Berkeley: University of California Press.

Bateson, Regina. 2012. “Crime victimization and political participation.” American Politi- cal Science Review 106(3):570–587.

Baumeister, Roy F. and Mark R. Leary. 1995. “The need to belong: desire for interpersonal attachments as a fundamental human motivation.” Psychological Bulletin 117(3):497– 529.

Bellows, John and Edward Miguel. 2009. “War and local collective action in Sierra Leone.” Journal of Public Economics 93(11):1144–1157.

Bénabou, Roland and Jean Tirole. 2006. “Incentives and Prosocial Behavior.” American Economic Review 96(5):1652–1678.

Berrebi, Claude and Esteban F. Klor. 2006. “On terrorism and electoral outcomes theory and evidence 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 and government 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.

Blair, Graeme and Kosuke Imai. 2012. “Statistical analysis of list experiments.” Political Analysis 20(1):47–77.

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

Blanco, Luisa R. 2013. “The impact of crime on trust in institutions in Mexico.” European Journal of Political Economy 32:38–55.

Boas, Taylor C. and F. Daniel Hidalgo. 2011. “Controlling the airwaves: Incumbency advantage and community radio in Brazil.” American Journal of Political Science 55(4):869–885.

Bobonis, Gustavo J., Luis R. Cámara Fuertes and Rainer Schwabe. 2014. “Monitoring Corruptible Politicians.” Working paper.

320 Botero, Sandra, Rodrigo Castro Cornejo, Laura Gamboa, Nara Pavao and David W. Nick- erson. 2015. “Says Who? An Experiment on Allegations of Corruption and Credibility of Sources.” Political Research Quarterly 68(3):493–504.

Brader, Ted and Joshua Tucker. 2012. “Follow the Leader: Party Cues, Policy Opin- ion, and the Power of Partisanship in Three Multiparty Systems.” Comparative Politics 44(4):403–420.

Brennan, Geoffrey and Alan Hamlin. 1998. “Expressive voting and electoral equilibrium.” Public Choice 95(1-2):149–175.

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

Cagé, Julia. 2014. “Media competition, information provision and political participation.” Working Paper.

Cantú, Francisco. 2014. “Identifying Electoral Irregularities in Mexican Local Elections.” American Journal of Political Science 58(4):936–951.

Casey, Katherine. 2015. “Crossing Party Lines: The Effects of Information on Redistribu- tive Politics.” American Economic Review 105(8):2410–2428.

Castañeda Sabido, Fernando. 2011. “Investigación sobre la percepción ciudadana acerca de la transparencia, rendición de cuentas y ﬁscalización en el uso de los recursos públicos en México.” Reporte para la Auditoría Superior de la Federación.

Chandra, Kanchan. 2007. Why ethnic parties succeed: Patronage and ethnic head counts in India. Cambridge: Cambridge University Press.

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

Chattopadhyay, Raghabendra and Esther Duﬂo. 2004. “Women as Policy Makers: Evi- dence from a Randomized Policy Experiment in India.” Econometrica 72(5):1409–1443.

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

Chong, Alberto, Ana De La O, Dean Karlan and Leonard Wantchekon. 2015. “Does Cor- ruption Information Inspire the Fight or Quash the Hope? A Field Experiment in Mexico on Voter Turnout, 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.

321 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.

Costas, Elena, Albert Solé-Ollé and Pilar Sorribas-Navarro. 2011. “Do voters really tolerate corruption? Evidence from Spanish Mayors.” Working paper.

Cruz, Cesi, Philip Keefer and Julien Labonne. 2015. “Incumbent Advantage, Voter Infor- mation and Vote Buying.” Working paper.

Cummins, Jeff. 2009. “Issue Voting and Crime in Gubernatorial Elections.” Social Science Quarterly 90(3):632–651.

Da Silveira, Bernardo S. and João M.P. De Mello. 2011. “Campaign advertising and election outcomes: Quasi-natural experiment evidence from gubernatorial elections in Brazil.” Review of Economic Studies 78(2):590–612.

Dalton, Russell J. and Martin P. Wattenberg. 2000. Parties Without Partisans: Political Change in Advanced Industrial Democracies. Comparative Politics. New York: Oxford University Press.

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

Dell, Melissa. 2015. “Trafﬁcking Networks and the Mexican Drug War.” American Eco- nomic 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.

DellaVigna, Stefano, John A. List, Ulrike Malmendier and Gautam Rao. 2014. “Voting to tell others.” NBER working paper No. 19832.

Delli Carpini, Michael X. and Scott Keeter. 1996. “What Americans Know about Politics and Why It Matters.” New Haven: Yale University Press .

Díaz-Cayeros, Alberto, Beatriz Magaloni, Aila Matanock and Vidal Romero. 2011. “Liv- ing in fear: Mapping the social embeddedness of drug gangs and violence in Mexico.” Working paper.

Diaz-Cayeros, Alberto, José Antonio González and Fernando Rojas. 2006. Mexico’s De- centralization at a Crossroads. In Federal and Economic Reform: International Perspec- tives, ed. Jessica S. Wallack and T. N. Srinivasan. New York: Cambridge University Press pp. 364—406.

Djankov, Simeon, Caralee McLiesh, Tatiana Nenova and Andrei Shleifer. 2003. “Who Owns the Media?” Journal of Law and Economics 46(2):341–382. 322 Downs, Anthony. 1957. An Economic Theory of Democracy. Boston: Addison Wesley.

Druckman, James D. 2005. “Media Matter: How Newspapers and Television News Cover Campaigns and Inﬂuence Voters.” Political Communication 22(4):463–481.

Durante, Ruben and Brian Knight. 2012. “Partisan control, media bias, and viewer responses: Evidence from Berlusconi’s Italy.” Journal of the European Economic Associ- ation 10(3):451–481.

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.

Eggers, Andrew. 2014. “Partisanship and Electoral Accountability: Evidence from the UK Expenses Scandal.” Quarterly Journal of Political Science 9(4):441482.

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

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

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

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

Feddersen, Timothy and Alvaro Sandroni. 2006b. “A Theory of Participation in Elections.” American Economic Review 96(4):1271–1282.

Ferejohn, John. 1986. “Incumbent performance and electoral control.” Public Choice 50(1):5–25.

Fergusson, Leopoldo. 2014. “Media markets, special interests, and voters.” Journal of Public Economics 109(1):13–26.

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

Ferraz, Claudio and Frederico Finan. 2011. “Motivating Politicians: The Impacts of Mon- etary Incentives on Quality and Performance.” Working paper.

323 Fox, Jonathan. 1994. “The difﬁcult transition from clientelism to citizenship: Lessons from Mexico.” World Politics 46(2):151–184.

García-Ponce, Omar, Leonard Wantchekon and Thomas Zeitzoff. 2014. “Are Voters too Afraid to Tackle Corruption? Survey and Experimental Evidence from Mexico.” Work- ing paper.

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 US daily newspapers.” Econometrica 78(1):35–71.

Gentzkow, Matthew, Jesse M. Shapiro and Michael Sinkinson. 2014. “Competition and ideological diversity: Historical evidence from U.S. newspapers.” American Economic Review 104(10):3073–3114.

Gerber, Alan S., Dean S. Karlan and Daniel Bergan. 2009. “Does the Media Matter? A Field Experiment Measuring the Effect of Newspapers on Voting Behavior and Political Opinions.” American Economic Journal: Applied Economics 1(2):35–52.

Gerber, Alan S., Donald P. Green and Christopher W. Larimer. 2008. “Social pressure and voter turnout: Evidence from a large-scale ﬁeld experiment.” American Political Science Review 102(1):33–48.

Gerber, Alan S., James G. Gimpel, Donald P. Green and Daron R. Shaw. 2011. “How large and long-lasting are the persuasive effects of televised campaign ads? Results from a randomized ﬁeld experiment.” American Political Science Review 105(1):135–150.

Gilens, Martin. 2005. “Inequality and democratic responsiveness.” Public Opinion Quar- terly 69(5):778–796.

Gilliam, Franklin D., Jr. and Shanto Iyengar. 2000. “Prime suspects: The inﬂuence of local television news on the viewing public.” American Journal of Political Science 44(3):560– 573.

Glazer, Amihai and Kai A. Konrad. 1996. “A signaling explanation for charity.” American Economic Review 86(4):1019–1028.

Golden, Miriam. 2010. Puzzles of Political Corruption in Modern Advanced Democra- cies. In Democracy and Accountability: Globalized Political Responsibility, ed. Hideko Magara. Toyko: Fukosha pp. 184–199.

Gomez, Brad T. and J. Matthew Wilson. 2006. “Cognitive heterogeneity and economic voting: A comparative analysis of four democratic electorates.” American Journal of Political Science 50(1):127–145. 324 Greene, Kenneth F. 2007. Why Dominant Parties Lose: Mexico’s Democratization in Com- parative Perspective. Cambridge: Cambridge University Press.

Greene, Kenneth F. 2011. “Campaign Persuasion and Nascent Partisanship in Mexico’s New Democracy.” 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.

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

Harding, Robin and David Stasavage. 2014. “What democracy does (and doesnâA˘ Zt´ do) for basic services: School fees, school inputs, and African elections.” Journal of Politics 76(1):229–245.

Healy, Andrew and Gabriel S. Lenz. 2014. “Substituting the End for the Whole: Why Voters 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 voters’ 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 Rodríguez Ferreira and David A. Shirk. 2014. “Drug Violence in Mexico: Data and Analysis Through 2013.” Justice in Mexico Project Special Report.

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

Hsieh, Chang-Tai and Enrico Moretti. 2006. “Did Iraq Cheat the United Nations? Un- derpricing, Bribes, and the Oil for Food Program.” Quarterly Journal of Economics 121(4):1211–1248.

Huckfeldt, R. Robert and John Sprague. 1995. Citizens, politics and social communication: Information and inﬂuence in an election campaign. New York: Cambridge University Press.

Huckfeldt, Robert, Paul E. Johnson and John Sprague. 2004. Political Disagreement: The Survival of Diverse Opinions within Communication Networks. New York: Cambridge University Press. 325 Hughes, Sallie and Chappell H. Lawson. 2004. “Propaganda and crony capitalism: Partisan bias in Mexican television news.” Latin American Research Review 39(3):81–105.

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

Iversen, Torben and David Soskice. 2013. “Biased Democracies: The Social and Economic Logic of Interest-Based Voting.” Working paper.

Iversen, Torben and David Soskice. 2015. “Information, Inequality, and Mass Polarization: Ideology in Advanced Democracies.” Comparative Political Studies 48(13):1781–1813.

Iyengar, Shanto. 1994. Is anyone responsible? How television frames political issues. Chicago: University of Chicago Press.

Iyengar, Shanto, Mark D. Peters and Donald R. Kinder. 1982. “Experimental demon- strations of the “not-so-minimal” consequences of television news programs.” American Political Science Review 76(4):848–858.

Jennings, M. Kent, Laura Stoker and Jake Bowers. 2009. “Politics across generations: Family transmission reexamined.” Journal of Politics 71(3):782–799.

Kaufmann, Daniel, Aart Kraay and Massimo Mastruzzi. 2009. “Governance matters VIII: aggregate and individual governance indicators, 1996-2008.” World Bank Policy Re- search Working Paper 4978.

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

Kendall, Chad, Tommaso Nannicini and Francesco Trebbi. 2015. “How Do Voters Respond to Information? Evidence from a Randomized Campaign.” American Economic Review 105(1):322–53.

King, Gary, James Honaker, Anne Joseph and Kenneth Scheve. 2001. “Analyzing incomplete political science data: An alternative algorithm for multiple imputation.” 95(1):49– 69.

Klašnja, Marko. 2011. “Why Do Malfeasant Politicians Maintain Political Support? Test- ing the “Uninformed Voter” Argument.” Working paper.

Klašnja, Marko and Rocío Titiunik. forthcoming. “The Incumbency Curse: Weak Parties, Term Limits, and Unfulﬁlled Accountability.” American Political Science Review .

Kronick, Dorothy. 2014. “Crime and Electoral Punishment.” Working paper.

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

Kuran, Timur. 1991. “Now out of never: The element of surprise in the East European revolution of 1989.” World Politics 44(1):7–48.

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

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

Larcinese, Valentino. 2007. “The instrumental voter goes to the newsagent demand for information, marginality and the media.” Journal of Theoretical Politics 19(3):249–276.

Larreguy, Horacio A. and Joana C. M. Monteiro. 2014. “Media Networks and Political Accountability: Evidence from Radio Networks in Brazil.” Working paper.

Larreguy, Horacio A., John Marshall and James M. Snyder, Jr. 2015. “Revealing Malfea- sance: How Local Media Faciliates Electoral Sanctioning of Mayors in Mexico.” NBER Working Paper 20697.

Larreguy, Horacio A., John Marshall and James M. Snyder, Jr. 2016. “Leveling the playing ﬁeld: How equalizing access to campaign advertising helps locally non-dominant parties in consolidating democracies.” Working paper.

Larreguy, Horacio A., John Marshall and Laura Trucco. 2015. “Breaking clientelism or rewarding incumbents? Evidence from an urban titling program in Mexico.” Working paper.

Larreguy, Horacio A., John Marshall and Pablo Querubín. 2016. “Parties, Brokers and VoterMobilization: How Turnout Buying Depends Upon the Party’s Capacity to Monitor Brokers.” American Political Science Review 110(1):160–179.

Lavielle, Briseida, Kristina Pirker and Alberto Serdán. 2006. “Monitoreo de Programas Sociales en Contextos Electorales.” Fundar Working paper.

Lawson, Chappell H. 2002. Building the fourth estate: Democratization and the rise of a free press in Mexico. Berkeley: University of California Press.

Lawson, Chappell H. 2004. Television Coverage, Vote Choice, and the Campaign. In Mex- ico’s Pivotal Democratic Election: Candidates, Voters, and the Presidenial Campaign of 2000, ed. Jorge I. Domínguez and Chappell H. Lawson. Stanford: Stanford University Press pp. 187–211.

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

Lawson, Chappell, Jorge Domínguez, Kenneth Greene and Alejandro Moreno. 2013. “The Mexico 2012 Panel Study.” Dataset.

Lee, David S., Enrico Moretti and Matthew J. Butler. 2004. “Do voters affect or elect policies? Evidence from the U.S. House.” Quarterly Journal of Economics pp. 807–859.

Lenz, Gabriel S. 2009. “Learning and opinion change, not priming: Reconsidering the priming hypothesis.” American Journal of Political Science 53(4):821–837.

Lenz, Gabriel S. 2013. Follow the leader? How voters respond to politicians’ policies and performance. Chicago: University of Chicago Press.

Lenz, Gabriel S. and Jonathan M. Ladd. 2009. “Exploiting a Rare Communication Shift to Document the Persuasive Power of the News Media.” American Journal of Political Science 53(2):394–410.

Levitt, Steven D. 1997. “Using Electoral Cycles in Police Hiring to Estimate the Effect of Police on Crime.” American Economic Review 87(3):270–90.

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

Lieberman, Evan S., Daniel N. Posner and Lily L. Tsai. 2014. “Does information lead to more active citizenship? Evidence from an education intervention in rural Kenya.” World Development 60:69–83.

Loewen, Peter John, Henry Milner and Bruce M. Hicks. 2008. “Does compulsory voting lead to more informed and engaged citizens? An experimental test.” Canadian Journal of Political Science 41(3):655–672.

Magaloni, Beatriz. 2006. Voting for Autocracy: Hegemonic Party Survival and its Demise in Mexico. New York: Cambridge University Press.

Mailath, George J. 1987. “Incentive Compatibility in Signaling Games with a Continuum of Types.” Econometrica 55(6):1349–65.

Manin, Bernard, Adam Przeworski and Susan C. Stokes. 1999. Elections and Represen- tation. In Democracy, Accountability, and Representation, ed. Bernard Manin, Adam Przeworski and Susan C. Stokes. New York: Cambridge University Press pp. 29–54.

Marshall, John. 2016a. “Political information cycles: When do voters sanction incumbent parties for high homicide rates?” Working paper.

328 Marshall, John. 2016b. “Signaling sophistication: Experimental evidence that social expectations can increase political information acquisition.” Working paper. Mauro, Paolo. 1995. “Corruption and Growth.” Quarterly Journal of Economics 110(3):681–712. McCann, James A. and Chappell H. Lawson. 2003. “An Electorate Adrift?: Public Opinion and the Quality of Democracy in Mexico.” Latin American Research Review 38(3):60– 81. McClendon, Gwyneth H. 2014. “Social Esteem and Participation in Contentious Politics: A Field Experiment at an LGBT Pride Rally.” American Journal of Political Science 58(2):279–290. McKelvey, Richard D. and Peter C. Ordeshook. 1985. “Elections with limited information: A fulfilled expectations model using contemporaneous poll and endorsement data as information sources.” Journal of Economic Theory 36(1):55–85. México Evalúa. 2012. “Seguridad y Justicia Penal en los estados.” Report. Mullainathan, Sendhil and Andrei Shleifer. 2005. “The Market for News.” American Eco- nomic Review 95(4):1031–1053. Newman, Benjamin J. 2014. “My Poor Friend: Financial Distress in One’s Social Network, the Perceived Power of the Rich, and Support for Redistribution.” Journal of Politics 76(1):126–138. Nickerson, David W. 2008. “Is voting contagious? Evidence from two field experiments.” American Political Science Review 102(1):49–57. Niehaus, Paul and Sandip Sukhtankar. 2013. “Corruption Dynamics: The Golden Goose Effect.” American Economic Journal: Economic Policy 5(4):230–269. Nordhaus, William D. 1975. “The Political Business Cycle.” Review of Economic Studies 42(2):169–190. Olken, Benjamin A. 2007. “Monitoring Corruption: Evidence from a Field Experiment in Indonesia.” Journal of Political Economy 115(2):200–249. Osorio, Javier. 2015. “The Contagion of Drug Violence: Spatiotemporal Dynamics of the Mexican War 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. Powell, G. Bingham, Jr and Guy D. Whitten. 1993. “A Cross-National Analysis of Eco- nomic Voting: Taking Account of the Political Context.” American Journal of Political Science 37(2):391–414. 329 Prat, Andrea and David Strömberg. 2005. “Commercial Television and Voter Information.”. CEPR Discussion Paper No. 4989.

Prior, Markus. 2007. Post-broadcast democracy: How media choice increases inequality in political involvement and polarizes elections. New York: Cambridge University Press.

Puglisi, Riccardo and James M. Snyder. 2011. “Newspaper coverage of political scandals.” Journal of Politics 73(3):931–950.

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 Latin America: a test of economic, institutional, and structural explanations.” American Political Science Review 93(3):575–590.

Rodden, Jonathan and Erik Wibbels. 2011. “Dual accountability and the nationalization of party competition: Evidence from four federations.” Party Politics 17(5):629–653.

Rogoff, Kenneth. 1990. “Equilibrium Political Budget Cycles.” American Economic Re- view 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. Washington: Woodrow Wilson International Center for Scholars Mexico Institute pp. 247–269.

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

Serra, Gilles. 2012. “The Risk of Partyarchy and Democratic Backsliding Mexico.” Taiwan Journal of Democracy 8(1):93–118.

Serra, Gilles. 2013. “Was Mexico’s 2012 Election Undemocratic? Assessing the Fraud Accusations.” CIDE Working Papers, DTEP-259.

Serra, Gilles. 2014. “The 2012 elections in Mexico: Return of the dominant party.” Elec- toral Studies 34:349–353.

Shineman, Victoria. 2013. “If You Mobilize Them, They Will Become Informed: Experi- mental Evidence that Information Acquisition is Endogenous to Costs and Incentives to Participate.” Working paper.

330 Sinclair, Betsy. 2012. The social citizen: Peer networks and political behavior. Chicago: University of Chicago Press.

Singer, Matthew M. and Ryan E. Carlin. 2013. “Context counts: The election cycle, development, and the nature of economic voting.” Journal of Politics 75(3):730–742.

Snyder, James M. and David Strömberg. 2010. “Press Coverage and Political Accountabil- ity.” Journal of Political Economy 118(2):355–408.

Stanig, Piero. 2015. “Regulation of Speech and Media Coverage of Corruption: An Empir- ical Analysis of the Mexican Press.” American Journal of Political Science 59(1):175– 193.

Strömberg, David. 2004. “Radio’s impact on public spending.” Quarterly Journal of Eco- nomics 119(1):189–221.

Tarrow, Sidney. 1994. Power in movement: Social movements, collective action and politics. New York: Cambridge University Press.

Tran, Anh. 2009. “Can procurement auctions reduce corruption? Evidence from the internal records of a bribe-paying ﬁrm.” Working paper.

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

UNODC. 2013. “Global Study on Homicide.” Report, available here.

Vivanco, Edgar F., Jorge Olarte, Alberto Díaz-Cayeros and Beatriz Magaloni. 2015. Drugs, Bullets, and Ballots: The Impact of Violence on the 2012 Presidential Election. In Mex- ico’s Evolving Democracy: A Comparative Study of the 2012 Elections, ed. Jorge I. Domínguez, Kenneth F. Greene, Chappell H. Lawson and Alejandro Moreno. Balti- more: Johns Hopkins University Press pp. 153–181.

Walsh, Katherine Cramer. 2004. Talking about politics: Informal groups and social identity in American life. Chicago: University of Chicago Press.

Weitz-Shapiro, Rebecca and Matthew S. Winters. forthcoming. “Can Citizens Discern? Information Credibility, Political Sophistication, and the Punishment of Corruption in Brazil.” Journal of Politics .

Wellenstein, Anna, Angélica Núñez and Luis Andrés. 2006. Social Infrastructure: Fondo de Aportaciones para la Infraestructura Social (FAIS). In Decentralized service delivery for the poor, Volume II: Background papers, ed. The World Bank. Mexico City: The World Bank pp. 167–222.

Yanagizawa-Drott, David. 2014. “Propaganda and Conﬂict: Evidence from the Rwandan Genocide.” Quarterly Journal of Economics 129(4):1947–1994. 331 Zaller, John R. 1992. The Nature and Origins of Mass Opinion. New York: Cambridge University Press.

Zamboni, Yves and Stephan Litschig. 2014. “Audit Risk and Rent Extraction: Evidence from a Randomized Evaluation in Brazil.” Working paper.

332