PUBLICIZING MALFEASANCE: WHEN THE MEDIA STRUCTURE

Transcription

1 PUBLICIZING MALFEASANCE: WHEN THE MEDIA STRUCTURE FACILITATES ELECTORAL ACCOUNTABILITY IN MEXICO HORACIO A. LARREGUY JOHN MARSHALL JAMES M. SNYDER, JR. OCTOBER 2017 Incumbent malfeasance is a major challenge in many developing democracies. Our model suggests that electoral accountability requires a media market structure incentivizing local media to report incumbent malfeasance. We test the model s implications in Mexico, leveraging plausibly exogenous variation in the pre-election release of municipal audits revealing misallocated spending and greater access to broadcast media. We find that, in contrast to non-local media, each additional local media station increases voter sanctioning of incumbent party malfeasance by one percentage point. Local media s pro-accountability effects are greatest when an outlet s audience resides in their municipality, but are crowded out by non-local media. JEL: D72, D78, O17. Key words: electoral accountability; malfeasance; media market structure. We thank Daron Acemoglu, Abhijit Banerjee, Claudio Ferraz, Fred Finan, Miriam Golden, Brian Knight, Rakeen Mabud, Ben Olken, Carlo Prato, Jesse Shapiro, and David Stromberg for helpful comments and suggestions. The paper also benefited from presentations at the 2015 CIRPÉE Political Economy conference, the 2015 Media and Communications Conference, Columbia, Harvard, London Business School and MIT. 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. Department of Government, Harvard University. hlarreguy@fas.harvard.edu. Department of Political Science, Columbia University. jm4401@columbia.edu. Department of Government, Harvard University. jsnyder@fas.harvard.edu. 1

2 1 Introduction An influential body of scholarship asserts that in democracies: (i) elections are a key institution producing political accountability; (ii) in order for elections to support accountability, voters must be adequately informed; and (iii) mass media play an essential role in informing voters (e.g. Ashworth 2012). An important application of this accountability dynamic is the electoral sanctioning of malfeasant behavior, such as corruption and the illegal diversion of funds away from the projects they were intended for. Such malfeasance is particularly prevalent in developing democracies (e.g. Mauro 1995; Treisman 2007), where voters are particularly poorly informed (e.g. Keefer 2007; Pande 2011). Although the media often cover malfeasance revelations (e.g. Puglisi and Snyder 2011), incentives to inform voters about local malfeasance may be weaker. In this article, we examine whether and under what conditions broadcast media enable voters to hold their local government to account by publicizing incumbent party malfeasance. We develop a simple model of media markets to explore how the media market structure influences electoral accountability. Voters choose which media stations to consume and have a preference for both variety and content local to their municipality. Although voters do not explicitly seek information to select between politicians (e.g. Baum 2002; Hamilton 2004; Prior 2007), the local media content that they encounter enables them to learn about and ultimately vote on the basis of the malfeasance of their municipal incumbent party (Prat and Stromberg 2013). Media stations compete a la Melitz (2003) by deciding how much content local to the municipality they are based in to provide, given anticipated voter demand and the share of their audience that is local. 1 In equilibrium, electoral sanctioning is increasing in the number of local stations that an electoral precinct has access to. 2 However, local media s positive sanctioning effect can be limited by both market competition and composition. First, the presence of non-local media stations crowds out consumption of local media. Second, local media stations may supply less local content when only a small fraction of their potential audience is local. This article s main contribution is to identify the extent to which local media markets influence municipal electoral accountability in Mexico. Specifically, we examine the release of audit reports pertaining to the Municipal Fund for Social Infrastructure (FISM), a major social program that provides mayors with federal transfers legally required to be spent on social infrastructure projects benefiting impoverished citizens. Between 2007 and 2012, the audits revealed that the average 1 Like Gentzkow and Shapiro (2006) and Mullainathan and Shleifer (2005), media outlets tailor content to voter tastes. 2 Prat and Stromberg (2013) s similar theoretical result is driven by the media content supply, rather than demand, as in our case. 2

3 municipality spent 5.8% of FISM funds on unauthorized (i.e. non-social infrastructure) projects, which usually constitute corruption, while 6.6% was misallocated to projects that did not benefit the poor. Voters are poorly informed about mayoral responsibilities to provide public services (Chong et al. 2015), and principally rely on local broadcast media to learn about mayoral malfeasance (e.g. Castañeda Sabido 2011). Although mayors could not seek re-election (due to term limits), political competition and voter knowledge of politics are oriented around political parties rather than individual candidates in Mexico s party-centric system (e.g. Langston 2003; Larreguy, Marshall and Snyder forthcoming; Marshall 2017). 3 We thus examine the extent to which voters reward or punish incumbent parties for mayoral distribution of FISM funds, as in recent studies informing voters of audit report outcomes through non-media channels (e.g. Chong et al. 2015). Guided by our theoretical model, we exploit two sources of plausibly exogenous variation to estimate the effect of access to additional local radio and television stations reporting mayoral malfeasance revelations in Mexico. To estimate the effects of malfeasance revelations, we follow Ferraz and Finan (2008) in leveraging the timing of the release of audit reports before and after municipal elections. We thus compare incumbent parties whose mayor was revealed to have engaged in malfeasant behavior either unauthorized spending or diverting funds to projects not benefiting the poor before an election to similarly malfeasant mayors whose audit reports were not released until after the election. To estimate the effects of access to local media, we use fine-grained spatial data to leverage variation in access to local radio and television signals between neighboring electoral precincts within the same municipality. Access to such signals differs due to plausibly exogenous factors such as antenna power and geographic features lying between the antenna and particular precincts. 4 Extensive balance, placebo, and sensitivity tests support both identification strategies. We find that each additional local media station a station broadcasting from within the same municipality in which an electoral precinct is located significantly increases voter sanctioning of incumbent parties revealed to be malfeasant. For both unauthorized spending and misallocating funds earmarked for projects benefiting the poor, our estimates indicate that each additional local radio or television station reduces the vote share of incumbent political parties whose mayor was revealed to be malfeasant by up to one percentage point. The extent of such sanctioning increases with the severity of malfeasance. Moreover, the effect of revealing an incumbent to have neglected the poor is greater in less developed precincts where voters stand to benefit most from FISM 3 Mayors elected in 2015 will become eligible for re-election in most states for the first time in As discussed below, we estimate the intent-to-treat effect on voters with access to media, given that commercial quality coverage boundaries reduce the likelihood that voters receive a media signal but cannot preclude coverage entirely. We lack the data required to compute a first stage, so focus on reduced form estimates. 3

4 projects. We also find that voters reward good performance in office: when the incumbent mayor did not engage in malfeasant behavior in office, an audit report released before an election increases their party s vote share by around 0.5 percentage points for each additional local media station. In contrast, we find no evidence that non-local media support electoral sanctioning. Consistent with our theoretical model, local media s capacity to support electoral accountability is moderated by the municipal media market structure. First, we show that access to an additional non-local media station crowds out the sanctioning effect of local media. Second, suggesting that the extent of local news coverage of mayoral performance increases with the share of a station s audience for whom the news is relevant, we find evidence that the effect of a local media station is larger when the media station s audience principally resides inside the municipality. These results suggest that electoral accountability requires that local media markets are structured to ensure that stations have incentives to supply politically-relevant information to their audiences. Our findings extend our understanding of when voters hold local governments to account in several ways. First, while the extant literature has used innovative strategies to estimate the electoral effects of audit reports and of media coverage in isolation, we identify the interaction of the two by exploiting plausibly exogenous variation in both the release of audits and access to local media. As Ferraz and Finan (2008) note, their path-breaking study cannot definitively separate the amplifying effects of radio from radio s correlates such as education and demand for political news that could also increase voter sanctioning of politicians revealed to be corrupt. 5 Similarly, although studies randomizing access to incumbent performance information delivered via leaflets show that voters can be induced to sanction poor performance in office (Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013), 6 a second source of variation is required to distinguish the effects of the information provided from the attention drawn to the information during leaflet delivery or the timing of the delivery. Our findings thus demonstrate that malfeasance revelations and media coverage are indeed complements in supporting electoral accountability, and can both harm malfeasant and benefit clean incumbent parties (e.g. Banerjee et al. 2011; Ferraz and Finan 2008). Moreover, local media s sanctioning effects exceed those associated with delivering leaflets, suggesting that media s priming and framing capacity may be substantial. 5 Appendix Table A3 shows that media access is substantially greater in more urban, literate, and economically developed areas in Mexico. This could induce bias if such variables themselves have important political consequences. For example, Klašnja (forthcoming) finds 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) similarly observe greater punishment of corruption among literate voters in Brazil. 6 Observational 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). 4

5 Beyond demonstrating that access to media increases voter sanctioning of malfeasance revelations, our second main contribution is to demonstrate the particular importance of local media and the moderating role of the local media market structure for electoral accountability. Our finding that non-local media stations can crowd out access to relevant local content, particularly as the market segments along the local/non-local dimension (Gentzkow and Shapiro 2006), reinforces the findings of Durante and Zhuravskaya (forthcoming) and Eisensee and Strömberg (2007). They respectively show that major news events can crowd out coverage of political issues such as small-scale military attacks and disaster relief. By further finding that local media s effects are greatest where the local audience is large, our results complement Snyder and Strömberg (2010). We first extend their findings by showing that incentives for media outlets to tailor local news to their audience can also enhance political accountability in a substantially less developed context where accountability is undermined by clientelistic practices. Moreover, in contrast with Snyder and Strömberg s (2010) focus on how the prospect of press coverage within a politician s district preemptively alters their legislative effort, we show that congruence between media markets and municipality boundaries is essential for voter sanctioning once political performance is revealed. Since 39% of Mexico s electoral precincts are not covered by a single local AM radio, FM radio, or television station, much of the population may have limited capacity to sanction incumbent party performance. This may help explain the varying effects of malfeasance revelations across the world (e.g. Chang, Golden and Hill 2010; Eggers 2014; Ferraz and Finan 2008), and suggests that declining local markets in many countries may have negative ramifications for democratic representation (see also Petrova 2011). 7 Third, we identify the role of additional media outlets covering relatively similar content. Our analysis thus complements innovative recent research focusing on the electoral implications of access to media outlets providing distinct news content (DellaVigna and Kaplan 2007; Durante, Pinotti and Tesei 2015; Enikolopov, Petrova and Zhuravskaya 2011), and implies that both types of content can influence voting behavior. In this regard, our design and non-consolidated democratic context relate most closely to Enikolopov, Petrova and Zhuravskaya (2011), who show that criticism of the Russian federal government decreased United Russia s vote share. Given a conducive media market structure, our findings offer hope that access to information could incentivize politicians to serve their constituents (Besley and Burgess 2002; Casey 2015; Snyder and Strömberg 2010). However, without universal access to local content, the levels of malfeasance that we ob- 7 Over 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 in 1996, compared to 36% in 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

6 serve suggest that a positive probability of being audited is insufficient to prevent mayoral malfeasance. This is consistent with evidence that audits are most effective at reducing corruption when politicians know in advance or from recent experience that the reports could result in criminal prosecution (Avis, Ferraz and Finan forthcoming; Olken 2007) or will be released before an election (Bobonis, Fuertes and Schwabe 2016). Without extensive local coverage, media-induced electoral accountability may then support democracy by reducing the probability that the parties of malfeasant mayors are re-elected. 2 Local media, market structure and electoral sanctioning In theory, a single news-reporting media outlet could be sufficient to support electoral accountability. For one news-reporting media station to be able to generate significant electoral sanctioning of malfeasance revelations, many voters would have to receive credible information from such an outlet, update their beliefs about their incumbent party s malfeasance, and ultimately vote according to such beliefs. In practice, however, the media s role in supporting this mechanism of electoral accountability faces two obstacles. First, media outlets may choose not to report revelations of incumbent malfeasance because of perceived lack of audience interest or weak advertising revenue incentives (e.g. Ansolabehere, Snowberg and Snyder 2006; Prat and Stromberg 2013; Strömberg 2004; Snyder and Strömberg 2010). 8 Second, voters may not choose to consistently consume local political information, even when it is available to them (e.g. Delli Carpini and Keeter 1996; Hamilton 2004; Pande 2011). Our model captures these constraints by examining consumption of local content reporting on mayoral malfeasance when voters have a preference for variety and media outlets maximize their audience across the municipalities that they serve. This simple approach to modeling media consumption shows that each additional media station reporting local news content can alter voting behavior by increasing the number of signals that voters receive about incumbent malfeasance. The model demonstrates that, in equilibrium, local media s effect on electoral sanctioning is generally decreasing in the number of non-local media stations and increasing in the proportion of a local media s station audience that resides within the station s municipality. The model s primary goal is to guide our empirical analysis. 8 While our model abstracts from this possibility, incumbent control of the airwaves could be another force driving a lack of reporting (e.g. Besley and Prat 2006). 6

7 2.1 Media market model Each municipality q = 1,..., Q contains a continuum of electoral precincts p q. In turn, each precinct contains a continuum of voters of mass f(p q ). Voters in any given precinct must choose which media stations they want to consume, and may consequently learn about the municipal incumbent party s malfeasance in office. Without loss of generality, we focus on a generic precinct p q in municipality q. Media stations. Our generic precinct p q is covered by masses M L (p q ) 0 and M N (p q ) > 0 of local media stations (L) and non-local media stations (N) respectively, which compete for voter audience. Within each category, there is a continuum of identical individual stations; one such station covering precinct p q is denoted by θ(p q ), and may be of type L or N. 9 The difference between local and non-local media stations is that only local stations have the capacity to provide content specific to the municipality in which they are located, e.g. because their office and antennae are located in that municipality and they thus possess a substantial comparative advantage in producing local news. We denote l(θ) as the content specific to the municipality where a media station is located. By definition, l(θ) = 0 when θ(p q ) = N but l(θ) [0, 1] when θ(p q ) = L. The total audience of a media station θ located in municipality q is then given by: ( ) ( ) A l(θ) := h p q, l(θ) f(p q )dp q + ( ) h p q, 0 f(p q )dp q, (1) p q P q(θ) q q p q P q(θ) ( ) where P q (θ) is the set of precincts in municipality q that station θ covers, h ( ) p q, l(θ(p q )) 0 denotes hours of station θ consumed, and h p q, l(θ) f(p q ) is θ s audience in precinct p q for a given l(θ). The first term in equation (1) represents θ s audience in the municipality q that it is local to, while the second represents its audience in all other municipalities. Media stations maximize their total audience by providing general and local content. General content includes entertainment, national news, and content only relevant for other municipalities (from the perspective of our generic precinct). We conceive of local content as news, including but not restricted to political news like mayoral audit reports. Due to their larger local audience share, local media stations have stronger incentives to produce local content relevant to our particular precinct s municipality (e.g. Snyder and Strömberg 2010). 10 Local stations choose to devote 9 A continuum with fixed mass is a popular modeling device, but reasonably approximates our context where there are many media stations. Additional layers of station heterogeneity, such as quality or variation in the proportion of their audience that is local, could be included at the cost of mathematical complexity. 10 Given the large mass of local media stations, as in our empirical context, we assume that competitors will expose factual inaccuracies, and thus limit the slanting of news content (Gentzkow and Shapiro 2006; Lawson 2002; Puglisi 7

8 fraction l(θ) [0, 1] of their broadcasting to local content, at cost l(θ) γ /γ where γ > 1. For simplicity, we assume that general content can be provided costlessly (e.g. through retransmission). Recall also that non-local media stations are unable to provide local content, e.g. because their reach is insufficiently large to warrant building a local office. These simplifying assumptions do not drive our results. Voters. Within precinct p q, we assume that the media-content consumption of all voters is identical. We thus focus on the media consumption of a representative voter who must choose among the M L (p q ) + M N (p q ) mass of media outlets that they have access to. Their choices reflect two motivations: (i) a preference for variety across stations; and (ii) a preference for greater local content. 11 The preference for variety, which implies that voters would prefer to consume a fixed amount of content from many different stations than the same amount from a single station, is the standard assumption underpinning consumer choice in new trade theory (e.g. Melitz 2003). It also represents a reasonable approximation to competitive media markets where stations provide differentiated content (and no station can provide only high-quality content) and new outlets are able to attract listeners and viewers (e.g. DellaVigna and Kaplan 2007; Durante and Knight 2012; Durante, Pinotti and Tesei 2015; Enikolopov, Petrova and Zhuravskaya 2011; Strömberg 2004). Given that no single voter is likely to affect election outcomes, the preference for local content reflects individual benefits arising from the greater relevance or interest of localized content. In Table 1 below we provide evidence indicating such a preference for local news in Mexico. Formally, we characterize such preferences by the following constant elasticity of substitution utility function: U(p q ) := ( ML (p q)+m N (p q) 0 [ ) 1 ( ) ρ ρ h p q, l(θ(p q )) [1 + l(θ(p q ))]] dθ(p q ) (2) where ρ (0, 1) represents substitutability across stations. The multiplier [1 + l(θ(p q ))] captures differential interest in local content by allowing voters to obtain greater utility from consuming an hour of media from a station where fraction l(θ(p q )) is devoted to locally-relevant content. Voter choice is subject to the time constraint that they cannot consume more than T hours of content in total: ML (p q)+m N (p q) 0 ( ) h p q, l(θ(p q )) dθ(p q ) T. (3) and Snyder 2011). 11 To simplify exposition, we assume that voters cannot specifically tune in to watch scheduled local content. 8

9 Electoral sanctioning. Because our principal focus is on the media environment, we adopt a reduced form sanctioning model where Bayesian voters decide whether to re-elect an incumbent party by comparing it to a known challenger. 12 Upon consuming media content, voters receive informative signals indicating that the incumbent party I q in municipality q misallocated share s Iq of funds. Each signal s Iq is independently drawn from normal distribution N (x Iq, ψ 2 I q ), where the expectation x Iq is centered on the true level of malfeasant misallocation with variance ψ 2 I q > The number of signals is given by the total number of hours of local content consumed: l(p q ) := ML (p q)+m N (p q) 0 ( ) h p q, l(θ(p q )) l(θ(p q ))dθ(p q ). (4) Exposure to malfeasance revelations is thus increasing in the total number of hours of local content consumed. Although voters do not explicitly acquire information to sanctions politicians in our model, such a motivation could be incorporated as an additional utility received from consuming local news. Alternatively, such information may just be an externality of preferring local content. Voters possess common prior beliefs regarding the malfeasance of the incumbent party I q and challenger party C q in municipality q. Specifically, prior beliefs are normally distributed according to N (κ jq, τ 2 j q ), where κ jq (0, 1) is expected malfeasance and τ 2 j q > 0 is the prior s variance for j {I q, C q }. In addition to preferring less malfeasant incumbent parties, each voter i(p q ) in precinct p q possesses a partisan bias δ(i, p q ) in favor of I q. This bias is independently and uniformly distributed across each precinct s electorate over the interval [ 1 2φ, 1 2φ ]. Voter i(p q) s voting decision is v(i, p q ) {I q, C q }, where i(p q ) re-elects I q (i.e. v (i, p q ) = I q ) if: 14 ] δ(i, p q ) [ω(l(p q ))x Iq + [1 ω(l(p q ))]κ Iq κ Cq, (5) where the second term on the left hand side is voters common posterior expectation κ Iq of the incumbent party s malfeasance. This term assigns weight ω(l(p q )) := l(pq)τ 2 Iq ψ 2 Iq +l(pq)τ 2 Iq to the mean signal received, x Iq, and weight 1 ω(l(p q )) to the prior expectation. Naturally, the weight ω(l(p q )) is increasing in the number of signals received l(p q ) and the variance of the prior τ 2 I q, and decreasing in the variance of the signal ψ 2 I q. 12 Allowing voters to also learn about the challenger (if challenger types are believed to be correlated with incumbent party types) would impose significant mathematical complexity. However, the guiding intuitions apply to contexts where voters believe that the correlation of types is not too high. 13 For tractability, we abuse the unbounded normal distribution in representing the prior beliefs about and signals of the share of malfeasant spending. For sufficiently tight distributions with means between 0 and 1, beliefs and signal will generally remain between 0 and For simplicity, we assume that voters are risk-neutral. 9

10 The game s timing can be summarized as: 1. Media stations simultaneously choose their fraction of local content l(θ). ( ) 2. Voters chooses to consume h p q, l(θ(p q )) hours of each media station θ. 3. Voters receive l(p q ) signals regarding incumbent misallocation. 4. Voters make their vote choice v(i, p q ). 2.2 Equilibrium media consumption and sanctioning We work backwards to solve for the subgame perfect equilibrium (SPNE). Given total local media consumption l(p q ) in equilibrium, the optimal voting rule in equation (5), and integrating over voter partisan biases, the incumbent party vote share V Iq (p q ) in our generic precinct is given by: 15 V Iq (p q ) = 1 2 φ [ ω(l(p q ))[x Iq κ Iq ] + κ Iq κ Cq ]. (6) This vote share makes clear that voters become more likely to vote against an incumbent party revealed to have engaged in more malfeasance than expected (i.e. x Iq revealed to have engaged in less malfeasance than expected (x Iq > κ Iq ), and for a party < κ Iq ). Without receiving any local content, the incumbent party s probability of election is decreasing in the difference in prior expectations of malfeasance, κ Iq κ Cq The key insights of the model concern media consumption, and thus the level ) of l(p q ). In particular, voters optimize their media consumption by choosing h (p q, l(θ(p q )) to maximize equation (2) subject to equation (3), taking the proportion of local content l(θ(p q )) for each station as given. Given that l(n) = 0, we show in the Appendix that media consumption for each station is: ) h (p q, l(θ(p q )) = T [1+l(L)] σ M L (p q)[1+l(l)] σ +M N (p q) T M L (p q)[1+l(l)] σ +M N (p q) if θ(p q ) = L if θ(p q ) = N where σ = 1 is the constant elasticity of substitution between media outlets. Demand for local 1 ρ media stations is thus greater than for non-local media if l(l) > 0, and is increasing in l(l). Given ] 15 Given that vote share is bounded on [0, 1] we assume that φ [ω(l(p q ))[x Iq κ Iq ] + κ Iq κ Cq thus an interior solution exists. (7) ( 1 2, 1 2 ), and 10

11 their preference for variety, voters consume content from every media station to which they have access. The final step of our backward induction sees media stations choose fraction l(θ) of content local to their municipality q. Since there is a continuum of media stations facing) identical maximization problems, and given anticipated voter media consumption h (p q, l(θ(p q )) in each precinct, each local station θ solves the following optimization problem: 16 { ( max l(θ) [0,1] A l(θ) ) l(θ)γ γ }, (8) ( ) where recall that A l(θ) is defined in equation (1) and l(θ) γ /γ is the cost of producing local content. Equilibrium local content provision by any media station local to municipality q and covering our generic precinct p q is given implicitly by the following first-order condition: σt [1 + l (L)] σ 1 p q P q(θ) f(p q ) M L (p q )[1 + l (L)] σ + M N (p q ) dp q l (L) γ 1 = 0 (9) Intuitively, it is clear that the marginal benefit of increasing local content (the first term), in terms of greater audience, increases with the total density of voters in their catchment area for whom their content would be local (i.e. p q P q(θ) f(p q)dp q ). A relatively large local audience share thus increases the provision of local content. Putting everything together yields the following result: Proposition 1. (Equilibrium uniqueness) A unique SPNE l (θ), h (p q, l (θ(p q ))), v (i, p q ) exists. Proof : see Appendix for all proofs. To understand the model s key insights, consider how an additional local media station affects a voter s total local media consumption. In the context of our empirical analysis, this could entail a local media station s reach extending slightly further to cover a precinct that would not have received the signal otherwise. An additional local media station has the following effect: l(p q ) M L (p q ) = l (L) h ( ) p q, l (L) = l (L) h (p q, l (L) + M L (p q ) M L (p q ) ) (10) T [1 + l (L)] σ ] ( ) 2 [M L (p q )[1 + l (L)] σ (1 l (L)) + M N (p q ) 0. M L (p q )[1 + l (L)] σ + M N (p q ) 16 The continuum of stations ensures that θ s choice of l (θ) does not affect the choices of other stations, while the continuum of precincts ensures that changes in demand in any given precinct do not affect a station s optimization. 11

12 The first line highlights two competing effects. On one hand, the first term inside the bracket shows that the additional local media station increases local consumption, because voters preference for variety causes them to consume the new station. Conversely, the countervailing effect in the negative second term reduces average consumption of all other stations. This countervailing force does not fully offset the increase because increased consumption shifts entirely toward local media stations, while the compensating reduction among all other stations afflicts both local and non-local stations. 17 This follows from the second line because 1 l (L) 0. Based on the model s implication that an additional local media station increases total local news consumption, the following comparative statics characterize the equilibrium: Proposition 2. (Comparative statics) In the unique SPNE l (θ), h (p q, l (θ(p q ))), v (i, p q ), the following comparative statics hold: 1. The incumbent vote share V Iq (p q ) in a given electoral precinct is decreasing (increasing) in the mass of local media stations M L (p q ) covering that precinct when x Iq > (<)κ Iq. 2. Under fairly non-restrictive conditions on ψi 2 q and τi 2 q (see Appendix), the magnitude of the marginal effect of M L (p q ) on V Iq (p q ) is decreasing in M N (p q ) and increasing in the size of the local audience of local media stations f(p p q P q(θ) q)dp q. 18 We next explain the intuition underlying these results as we derive testable hypotheses. Intuitively, the conditions required for the second claim entails that prior beliefs and news signals are not so precise that an additional signal has little effect on voter beliefs. In contexts where voters are poorly informed about politics, this condition is likely to hold. 2.3 Empirical implications The comparative statics in Proposition 2 highlight three testable implications of our model in the context of audit report revelations about municipal incumbent parties. These hypotheses guide our identification strategy and empirical analysis. 17 The continuum of precincts in a given municipality implies that changes in the behavior of any given precinct do not affect the share of local content that a media station provides. However, our main results continue to hold provided that l (L)/ M L (p q ) < 0 is sufficiently small. Intuitively, this implies that increasing the number of local media outlets reaching a particular precinct does not substantially reduce the amount of local content that each individual local station provides. Both assumptions are plausible in our empirical context where a precinct constitutes a tiny fraction of a station s total audience. 18 The conditions on ψ 2 I q and τ 2 I q ensure that ω(l(p q )) is not too concave. 12

13 The first part of Proposition 2 shows that sanctioning increases with each additional local media station. Intuitively, because voters have a preference for variety in media content, each additional local station increases voters consumption of local media content. Although voters may not necessarily consume local content to improve their ability to sanction incumbent parties, the consumption of local content nevertheless increases the number of signals voters receive about their incumbent party s malfeasance. 19 We thus hypothesize that: H1. Following the release of an audit report before an election, the municipal incumbent party s vote share decreases with the number of local media stations covering the precinct when the report reveals high malfeasance, and increases with the number of local media stations covering the precinct when the report reveals low malfeasance. The second part of Proposition 2 shows that the presence of non-local media stations may crowd out the sanctioning effects of local media stations. Here, voter preference for media station variety means that the marginal effect of an additional local media station on total local content consumption is smaller because voters continue to diversify over many pre-existing non-local alternatives. We thus hypothesize that: H2. Following the release of an audit report before an election, the sanctioning effect of an additional local media station (described in H1) is decreasing in the number of non-local media stations covering the precinct. Furthermore, the extent to which local media stations facilitate electoral sanctioning varies with the incentives for local media stations to attract a greater local audience at the expense of non-local audiences. To see this, define α(θ) := p q P q(θ) f(p q)dp q + q q f(p p q P q(θ) q)dp q pq P f(p q(θ) q )dp (11) q as the share of media station θ s potential audience that is local. The second part of Proposition 2 indicates that, when a media station s local potential audience is large in comparison to its nonlocal audience (i.e. α(θ) is larger), the media station is more likely to provide local content and thus support sanctioning. We thus hypothesize that: 19 By the same logic, additional non-local media stations may reduce voters ability to sanction their incumbents by crowding out the consumption of local news, at least in our stylized model where non-local media provides no local content at all and voters face a media consumption time constraint. Even if they do provide some content, we nevertheless expect weaker sanctioning effects than for local media. 13

14 H3. Following the release of an audit report before an election, the sanctioning effect of an additional local media station (described in H1) is increasing in the share of its audience located within the precinct s municipality. 3 Municipal government and accountability in Mexico Mexico s federal system is divided into 31 states (excluding the Federal District of Mexico City) containing around 2,500 municipalities. Municipalities are Mexico s lowest administrative unit, and in 2010 varied in population from 102 to 1.8 million people with a mean of 46,134. Municipal governments the focus of this study exercise significant control over local policy. Following major fiscal decentralization reforms in the 1990s, the average municipality s annual budget reached $9 million (USD), representing 20% of total government spending. 20 Municipal governments are led by mayors responsible for delivering basic public services and managing local infrastructure. Mayors are typically elected to three-year non-renewable terms, and municipal elections in different states follow different electoral cycles Municipal audits An important component of a mayor s budget is the Municipal Fund for Social Infrastructure (FISM). Under the 1997 Fiscal Coordination Law, FISM funds are direct federal transfers to municipalities mandated exclusively for public infrastructure projects benefiting the socially disadvantaged population living in extreme poverty. Mayors are free to invest in projects in the following categories: potable water, sewage, drainage and latrines, municipal urbanization, electrification of rural and poor suburban areas, basic health infrastructure, basic education infrastructure, improvement of housing, rural roads, and rural productive infrastructure. FISM funds are allocated according to a formula accounting for the prevalence of extreme poverty and previous FISM transfers, and represent 24% of the average municipality s budget. 22 The use of FISM funds is subject to independent audits by Mexico s Federal Auditor s Office (ASF). The ASF was established in 1999 in response to widespread concerns regarding the mismanagement of public resources (Chong et al. 2015), and possesses constitutionally-enshrined autonomy to audit federal funds spent by federal, state, and municipal governments. The ASF is 20 Education and health were decentralized between 1992 and 1996 and the decentralization of infrastructure projects followed in 1997 (Wellenstein, Núñez and Andrés 2006). 21 During transitions to new electoral calendars, some states have occasionally elected mayors to four-year terms. Mayors will be able to run for re-election for the first time in 2018 in most states. 22 For more information on the formula, see this example from Oaxaca. 14

15 perceived as a neutral, autonomous, and professional agency by both national and international experts (De La O and Martel García 2015). 23 Between 2007 and 2012, 14% of municipalities were audited at least once, with around 150 municipalities across all states audited each year. The list of municipalities to be audited in a given year is announced after spending occurred. Although municipalities are not randomly selected for audit, the timing of an audit is essentially random among audited municipalities. The ASF specifies the following criteria for auditing a municipality 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 the municipality has recently been audited and whether other federal audits are occurring simultaneously; and whether a specific mandate exists to examine a particular municipality. 24 Communications 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. 25 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 reflect the electoral cycle or the government s partisanship. The balance tests in Table 2 that we discuss below show that, conditional on a municipality being audited, the timing of audit reports is generally uncorrelated with 53 audit outcomes and various political, demographic, and socioeconomic indicators. The ASF audits the spending and management of FISM resources from the prior fiscal year. Auditors check that municipal officials abided by procurement rules and accounting procedures, that the status of the funded projects is in accordance with the books, and that funds were used as intended. Reports categorize the use of FISM funds across multiple dimensions; an example is provided in Appendix Figure A1. Most importantly, reports document the percentage of FISM funds spent on unauthorized projects and the percentage of funds spent on projects not benefiting the poor. Unauthorized spending captures the diversion of resources toward non-social infrastructure projects (e.g. personal expenses and election campaigns) and funds that are unaccounted for. This can broadly be interpreted as corruption. 26 Spending that does not benefit the poor entails 23 Due to the ASF s limited resources, a smaller share of municipalities is audited by the Superior Audit Entities of States (De La O and Martel García 2015). 24 This 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. 25 Based on a personal interview with the Licentiate Jaime Alvarez Hernández, General Director of Research and Evaluation of the Special Audit of Federal Spending, in July 2012, and formal correspondence with the ASF. 26 This definition resembles Ferraz and Finan (2008), in that we focus on violations that include procurement fraud, 15

16 Density Density Proportion of spending unauthorized Proportion of spending not on the poor After election Before election After election Before election (a) Unauthorized spending (b) Spending not on the poor Figure 1: Distribution of audit report results, by release around election Notes: The audit outcome kernel density distributions are based on electoral precincts in municipalities that received an audit just before and just after an election. Precincts are weighted by the number of registered voters. We use an Epanechnikov kernel with a bandwidth of the allocation of FISM funds to social infrastructure projects that do not principally benefit voters in the (as of 2010) 79% of localities classified as high or very high on the federal government s marginalization scale. For example, this includes diverting funds to support agricultural production in areas without poverty or paving streets in richer urban areas. Audit reports are presented in Congress every February, the calendar year after the audit was conducted, and thus two calendar years after the spending occurred. Reports are then immediately posted on the ASF s website, Although the targeting of FISM funds toward their intended recipients has generally been more successful than previous social programs (Wellenstein, Núñez and Andrés 2006), funds are often misallocated. Between 2007 and 2012, ASF audits found that the average municipality allocated 5.8% of funds to unauthorized projects and 6.6% to infrastructure projects not benefiting the poor. However, as the spending distributions in Figure 1 show, malfeasant spending in our sample is concentrated among relatively few municipalities engaged in substantial malfeasance. The ASF has imposed various punishments on malfeasant public officials. Between December 2006 and July 2012, the Ministry of Public Function recovered $2 million (USD), sanctioned 9,000 public employees for serious misdemeanors, and incarcerated one hundred officials. 27 However, diversion, and over-invoicing. Unlike Ferraz and Finan (2008), who count the number of corruption violations, our measure quantifies monetarily the monetary value of malfeasance. 27 El Universal, A la cárcel, solamente 100 ex servidores, 29th May 2014, link. 16

17 the greatest punishment may be electoral. Since Mexican mayors could not stand for re-election, any electoral penalty hits the party of a malfeasant mayor. 28 There are good reasons to believe that a mayor s political party may be rewarded or punished by voters at the next election. First, although some voters are aware of particular candidates, Mexico s strong party system ensures that local party labels play a key role in voting decisions. Over the period of our sample, Mexico s three main parties the left-wing Party of the Democratic Revolution (PRD), the populist PRI, and the right-wing National Action Party (PAN) almost exclusively controlled municipal offices, in some cases in coalition with Mexico s many smaller parties. 29 Second, the top-down internal structure of Mexican parties at the state and local levels ensures that parties generally select similar candidates across elections (Langston 2003). Ultimately, studies consistently show that voters are willing to punish the parties of local mayors for their actions in office (e.g. Chong et al. 2015; Marshall 2017). 3.2 Voter information and local broadcast media Voters widely perceive Mexican politicians as malfeasant, but are not generally aware of audit reports. In the 2011 National Survey of Quality and Impact of Government, 84% of respondents believed that corrupt practices occurred frequently in their municipality, 30 while 50% cited corruption among the three most important issues in their state. Moreover, only 34% of voters believe that municipal governments are transparent about the use of public resources (Castañeda Sabido 2011). 31 However, voters are poorly informed about the resources available to mayors and their responsibility to provide basic public services (Chong et al. 2015). Furthermore, voters know little about the institutions responsible for auditing the use of public resources: only 25% of surveyed individuals can name a public auditing institution, and only 1.4% of those individuals identify the ASF as principally responsible for auditing the use of public funds (Castañeda Sabido 2011). Consequently, radio and television voters principal sources of news play an essential role in providing voters with information about incumbent malfeasance. The 2009 Latinobarometer in- 28 This feature further differentiates this article from preceding studies (e.g. Banerjee et al. 2011; Ferraz and Finan 2008; Humphreys and Weinstein 2012). 29 The 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). However, MORENA did not stand for elected office over the period of our sample. 30 Respondents perceive similar corruption levels at state and federal levels. 31 In principle, local governments are required to inform the public about the arrival of FISM funds. However, 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). 17

18 dicates that 83% of respondents gather political information from television, 41% from radio, 30% from newspapers, and 41% from family, friends, and colleagues (many of whom, of course, gather their information from television, radio, and newspapers). In 2008, around 10% listened to radio every day, while the average person watched four hours of television per day (Ibope/AGB México 2009). Furthermore, voters exhibit some preference for consuming various television channels: of the individuals who choose to purchase cable television, the 2016 National Survey on the Consumption of Audiovisual Content conducted by the Mexican Federal Telecommunications Institute shows that 72% indicate that they do so because they want more options of content and channels and because they transmit programs that they cannot see otherwise. Internet has only recently become widespread: according to the 2010 Census, only 21% of households had internet access, while 3G coverage only started to expand in While the majority of entertainment content is shared within Mexico s large radio and television networks, news is often provided locally. Content sharing is common within major radio groups such as ACIR, MVS Radio, Radio Fórmula, and Radiorama. These networks share branding and formats among their owned-and-operated and affiliated stations, but all individual stations provide localized content. 32 A survey of radio stations that we conducted in 2016 indicates that 83% of radios report news more than once a day, and principally report on municipal rather than national political, economic and security issues (Larreguy, Lucas and Marshall 2016). As a Federal Electoral Institute (IFE) study of radio news programming in 2009 indicates, hundreds of stations produce their own daily news programs. 33 Mexico s two main television networks, Televisa and TV Azteca, also share national content among their subsidiaries and affiliates. Televisa operates 190 stations solely retransmitting nationwide content from its main channels (Canal de las Estrellas and Canal 5), but also owns 33 regional stations and affiliates with 9 local stations across most states that broadcast local news. Similarly, TV Azteca owns 45 stations broadcasting only national content (Azteca 7 and Azteca Trece) and 134 stations complementing national content with state-specific programming particularly news serving the largest municipalities in a given state. Conditional on providing news, media stations provide around two hours of news coverage a day. 34 Commercial media outlets rely on advertising revenues, and thus face strong financial motives to target their programming at local audiences. The release of municipal audit report results each February is a major media event. As in other developing countries (see Pande 2011), revelations of political malfeasance are relatively 32 In 2009, all AM and FM radio stations were classified by the IFE as providing distinct content; see here. 33 See here. The IFE has since become the National Electoral Institute (INE). 34 These 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. 18

19 Standardized search activity (sigma) Standardized search activity (sigma) Feb 06 Feb 07 Feb 08 Feb 09 Feb 10 Feb 11 Feb 12 (a) Searches for ASF Feb 06 Feb 07 Feb 08 Feb 09 Feb 10 Feb 11 Feb 12 (b) Searches for FISM Figure 2: Google searches related to audit reports by month, Notes: Extracted using Google Correlate ( on 15th July The data cover Google searches in Mexico for the period used in our sample. frequently reported and can remain salient news for months when they are reported. According to our survey of radio stations, 76% believe that their audience care a lot about news relating to corruption. Moreover, 66% believe that it is very important to report on municipal-level corruption and 48% reported actually covering such news, while only 2% state that threats from politicians represent major costs to producing this type of news. News reports generally cover local mayors within the state, and typically focus on cases of unauthorized spending and mayors not spending FISM funds on projects targeting the poor. 35 Importantly, 83% of individuals report receiving information about malfeasance in the management of public resources through the media, and 61% regard such information as reliable (Castañeda Sabido 2011). Many media 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 examples of such audit report coverage. 36 In a particularly egregious case of unauthorized spending from 2013, BBM radio station reported that Oaxaca de Juarez s mayor had created a fake union to collect payments, presided over many public works contracts without offering open tender, diverted payments for advertising and consulting fees, and failed to provide 35 Little 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. 36 For example, see: BBM Noticias, ASF: desvió Ugartchechea 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. 19

20 details of considerable quantities of spending. 37 Such reports are not uncommon: the media often points to mayors diverting payments, using FISM funds for personal and family expenses and manipulating tender processes. Reports documenting failures to spend FISM funds on the poor are also common, including cases where public works projects were undertaken in urban and affluent parts of the city, or never materialized despite being paid for. To illustrate media coverage of the audit report releases more systematically, Figure 2 shows notable spikes in Google searches for the terms ASF and FISM around February and March each year. Unfortunately, we cannot systematically collect malfeasance reports in the media retrospectively. 4 Research Design This section first describes our main sources of data, before outlining our empirical strategy combining plausibly exogenous variation in both the timing of audit report releases and access to local broadcast media. 4.1 Data and main variables Municipal election outcomes Mexico s municipalities are divided into around 67,000 electoral precincts. Using data from the IFE and State Electoral Institutes, we collected municipal electoral returns for all available precincts between 2004 and We thus accumulated up to four election results per precinct, enabling us to identify the municipal incumbent party and their past vote share in all the elections in our period of analysis, President Felipe Calderón of the PAN held office throughout this period. Our main outcome is the change in the municipal incumbent party s vote share, as a share of turnout, at the precinct level. 38 To classify cases where the incumbent was a coalition of parties that split before the next election, we determined the party affiliation of the mayor by researching their identity and party ties. 39 Given that the candidates and coalitions of smaller parties can differ 37 BBM Noticias, ASF: desvió Ugartchechea mdp, October 21st 2013, here. 38 Appendix Table A6 shows that turnout is not systematically affected by audit reports, while normalizing incumbent vote share by the number of registered voters reinforces our main findings. 39 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 candidates stand for election: the PT, PVEM, and PC candidates stood against candidates from three coalitions, the PAN-ADC, PRI-PANAL, and PRD- PSD. Because the outgoing PRI-PVEM mayor came from the PRI, we use the PRI-PANAL vote as the incumbent s vote share. 20

21 Figure 3: Distribution of audit report outcomes by municipality Notes: Only the 311 unique municipalities in our final sample are included. Where more than one audit occurs, we take the average level of malfeasance. substantially, we focus only on the 97% of coalitions led by candidates from the three main parties. An advantage of mayors being unable to seek re-election is that, unlike individual candidates who may strategically choose not to stand for re-election in response to ASF revelations, incumbent parties always stand for election. The incumbent party received 43% of votes in the average electoral precinct in our sample, which represents a 5.1 percentage point decline in their vote share Audit reports We collected all available audit reports between 2007 and 2012, and extracted from each the proportion of funds spent in an unauthorized manner and the proportion of funds spent on projects not benefiting the poor. This yielded 742 municipal audits relatively evenly spread across years and covering 351 unique municipalities. Of these, 442 reports from 311 different municipalities form our sample of reports concerning incumbents from major parties released in an election year or the year after. We henceforth focus on this subsample, which is shaded by the extent of malfeasance in Figure 3. We operationalize malfeasance severity using indicators for mayors in the third and fourth quartiles of malfeasant FISM spending. Unauthorized spending in the third quartile ranges from 0.6% to 11.4% of available FISM funds with a mean of 5.4%, while unauthorized spending in the fourth quartile exceeds 11.4% with a mean of 30.1%. For spending not allocated to projects benefiting the poor, the third quartile contains any positive value up to 12.9% of available FISM funds with a mean of 5.8%, while the fourth quartile exceeds 12.9% with a mean of 36.9%. Since around 50% of precincts did not experience any unauthorized or spending on projects did not benefit the poor, the 25th percentile of each distribution is 0% for both malfeasance indicators; 21

22 hence, there is no distinction between the first and second quartiles. These indicators are more flexible than a linear measure, and may better capture models suggesting that voter sanctioning involves cutoff rules (e.g. Barro 1973; Ferejohn 1986). Robustness checks below show similar results using a linear measure of malfeasance or cutoffs for different malfeasance thresholds Local and non-local broadcast media coverage We match our precinct-level electoral data with fine-grained 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. 40 For each media station we are thus able to identify (i) the municipality from which it broadcasts, and (ii) the commercial quality coverage range of its signal. 41 Inside a station s commercial quality coverage area, the signal is of high quality and voters have good access to the station s broadcasts. Figure 4 maps the location and commercial quality signal coverage associated with each of Mexico s 852 AM, 1,097 FM, and 1,255 television antennae. Broadcast media coverage is extensive, with most precincts receiving many more than one media signal and most municipalities containing at least one media station. However, there is considerable variation in the number of media stations covering each precinct that emit from within the precinct s own municipality. 42 The figure also clearly indicates that the commercial quality coverage range of AM radio is substantially greater than for FM and television. As explained below, we restrict attention to more urban precincts that vary primarily in FM and television coverage. We measure access to local media as the total number of local AM, FM, or television stations covering a given electoral precinct. A local station is defined by: (i) providing some local content, rather than (in the case of some televisions stations) retransmitting the nationwide signal; 43 and (ii) emitting from within the precinct s municipality. This definition thus captures the media stations with strongest market incentives to report malfeasance revelations within a precinct s municipality. To avoid counting signals that barely cover voters within a precinct, we use (urban) 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 inside the commercial quality 40 For 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. 41 The 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). 42 The number of radio and television stations has remained fairly constant over our period of study, so we cannot exploit temporal variation in media coverage. 43 We thus do not count television stations solely broadcasting national Televisa or TV Azteca content as local. 22

23 (a) AM radio stations (b) FM radio stations (c) Television stations Figure 4: Media station commercial quality signal coverage areas (source: IFE) coverage boundary. 44 The average precinct is covered by 4.5, 5.9, and 1.5 local AM, FM, and television stations respectively. The total number of local media stations covering a precinct ranges from 0 to 40. To compare the effects of local media and non-local media, we similarly count the number of media stations providing any mix of local or national content that cover a precinct but transmit from outside their municipality. Due to the power of their antennae, the average precinct receives many more non-local than local AM signals. To a lesser extent, this is also true of FM and television signals. The likelihood that a voter follows the news is associated with the availability of local media. Using data on media consumption from the 2009 CIDE-CSES Survey, Table 1 shows a strong positive correlation between access to local media stations and consumption in a predominantly urban sample. Specifically, each additional local radio and television station respectively is associated with a 0.4 and 2.4 greater probability that an individual regularly consumes a news program, as well as a significantly higher total number of news programs regularly consumed. 45 In total, 80% of respondents listed at least one television program, and 30% listed one radio program. Conversely, the evidence that the number of non-local media stations are associated with news consumption is significantly weaker. Given that non-local media stations also report news just not relevant local news this suggests that voters are somewhat more likely to consume local than non-local media stations. 44 Table 5 shows that the results are robust to using alternative population thresholds. We selected 20% for our baseline specification because information could spill over even within precincts when only a small fraction of voters can access the local media station. 45 In the absence of detailed media consumption data and a comparable sample, these figures cannot be combined with our main estimates to compute a persuasion rate. 23

24 Table 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 0.024*** 0.045** (0.009) (0.018) Local media 0.004*** 0.015*** (0.001) (0.004) Observations 1,055 1,055 2,110 2,110 2,110 1,055 Outcome mean Outcome standard deviation Media mean Media standard deviation 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) Non-local television stations 0.015* (0.008) (0.013) Non-local media (0.001) (0.003) Observations 1,055 1,055 2,110 2,110 2,110 1,055 Outcome mean Outcome standard deviation Media mean Media standard deviation Notes: All respondents were asked the identities of any news programs that they regularly watch on television (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. 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 <

25 4.2 Identification strategy We seek to estimate the effect of an additional local media station on the incumbent party s electoral performance in municipalities where audit reports reveal malfeasance just before an election, as well as how this effect varies with the local media market structure. Identification thus requires plausibly exogenous variation in the release of audit reports before mayoral elections and access to local media Variation in the release of audit reports before municipal elections To identify the effect of audit reports revealing mayoral malfeasance on incumbent party electoral performance, as in Ferraz and Finan (2008), we exploit the timing of audit report releases with respect to municipal elections. Specifically, we compare municipalities where an audit report was released just before an election to a control group of municipalities where the audit was released just after a municipal election. We define a pre-election audit report release as an audit released in February of an election year, which is typically four or five months before the election. 46 Our control group contains municipalities where the audit was released in February the year after the election. It is important to emphasize that reports released in both the election year and the year after refer to spending by the same mayor, respectively in her first and second year in office. Restricting attention to municipalities that have been audited at least once yields a maximum sample of 46,419 precinct-election observations. The assumption required to identify the effects of audits released just before an election is that the timing of report releases are effectively random. The exogeneity of the timing of report releases is subject to two main types of concern. First, despite the ASF s claim to the contrary, selection for audit may not be independent of the electoral calendar. For example, political motivations could lead the ASF to target certain incumbent parties or particularly clean malfeasant mayors before elections. Second, audit findings could differ across reports released before and after elections. This could reflect auditors being more or less lenient or meticulous in the knowledge that a report will be released in an election year (see Ferraz and Finan 2008), mayors anticipating the possibility that their first year in office could be subjected to an audit released before an election (Bobonis, Fuertes and Schwabe 2016), or the relative inexperience of first-year mayors about whom audit reports are released before an election inducing them to allocate their funds differently from second-year mayors. However, we find no empirical evidence to support these concerns. Table 2 confirms that differences between electoral precincts in municipalities where an audit was released in the year before 46 Although states differ in the month in which they hold elections, only Baja California Sur holds elections before mid-february. We adjust for Baja California Sur accordingly. 25

26 Table 2: Summary statistics by pre-election audit status (full sample) Difference between post- and pre-election audit Coef. Std. err. Control (post-election audit) mean Municipal-level variables Unauthorized spending (0.029) Unauthorized spending Q (0.063) Unauthorized spending Q (0.081) Spending not on the poor (0.033) Not poor Q (0.076) Not poor Q (0.081) PAN incumbent (0.087) PRI incumbent (0.089) PRD incumbent (0.059) PAN state Governor (0.088) PRI state Governor (0.051) PRD state Governor (0.086) Mayor aligned with Governor (0.086) Total coalition partners (0.231) Municipal incumbent victory margin (lag) (0.019) Municipal effective number of political parties (lag) (0.079) Precinct-level variables Incumbent party vote share (lag) (0.012) Effective number of political parties (lag) (0.063) Registered voters (log) (0.062) Turnout (lag) (0.019) Distance to the municipal head from precinct border (log) (0.068) Distance to the municipal head from precinct centroid (log) (0.056) Area (log) (0.116) Population (log) (0.057) Population density (log) (0.212) Local media * (2.112) Non-local media (4.582) Average children per woman (0.034) Share households with male head (0.005) Share indigenous speakers (0.008) Average years of schooling (0.156) Share illiterate (0.005) Share no schooling (0.004) Share incomplete primary schooling (0.004) Share complete primary schooling (0.010) Share incomplete secondary schooling (0.013) Share complete secondary schooling (0.013) Share higher education (0.015) Share economically active (0.006) Share without health care (0.015) Share state workers health care (0.005) Average occupants per dwelling (0.038) Average occupants per room (0.024) Share non-dirt floor (0.005) Share toilet at home (0.005) Share running water (0.016) Share drainage (0.011) Share electricity (0.002) Share washing machine (0.018) Share fridge (0.013) Share cell phone (0.016) Share car (0.023) Share computer (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 observations are weighted by electorate size. Each regression contains 46,419 observations; this weighting ensures that municipal-level variables are equivalent to municipal-level regressions. Standard errors clustered by municipalityelection are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p <

27 an election and precincts where an audit was released the following year are well-balanced across 53 municipal and precinct-level variables. We find only one statistically significant difference at the 10% level. 47 Of particular note, pre-election audit releases are not significantly correlated with either unauthorized spending or spending on projects not benefiting the poor, the incumbent s partisanship (and thus alignment with the President, in the case of the PAN) or alignment with the state Governor, previous electoral performance, or proxies for socioeconomic development. Appendix Tables A1 and A2 also show that the interaction of pre-election audit reports with incumbent party and measures of political competition does not systematically predict audit outcomes. Furthermore, Figure 1 above compares the distributions of audit report outcomes across precincts from municipalities where audit reports were released before and after the election. 48 The distributions of unauthorized spending and spending not on the poor are very similar; in neither case does a Kolmogorov-Smirnov test reject equality of distribution. 49 This is consistent with neither mayors nor auditors acting differently when an audit report will be released before an election. Together, these checks support our identifying assumption that the municipalities where audits were released in election years, and the content of those reports, are comparable to municipalities where audit reports were released after the election Variation in access to local media To identify the effect of access to local media stations, we exploit differences between neighboring electoral precincts from the same municipality in the number of local media stations that they are covered by. Our design is most closely related to the approaches of Enikolopov, Petrova and Zhuravskaya (2011) and Spenkuch and Toniatti (2016). Focusing on neighboring precincts near local media coverage boundaries allows us to exploit differences in access arising from 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). 50 Furthermore, leveraging only within-municipality neighbors removes municipal political differences and ensures that neighboring precincts vote on the same incumbent party. Broadcast signals decay continuously and whether any given household receives a signal may depend upon the quality of their receiver. Consequently, discrete differences in commercial quality 47 This difference in the number of local media is controlled for in all main specifications. 48 Further illustrating the independence of the selection procedure from electoral considerations, this graph is very similar if we compare audits from election years to any non-election year. 49 Collapsing to the municipality-audit level (with 442 observations) to avoid duplication across precincts, the p values of the unauthorized spending and not spending on poor distributions are respectively 0.63 and 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. 27

28 Figure 5: 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. signal coverage do not necessarily imply that neighboring precincts differ strictly between receiving or not receiving a station s signal. Furthermore, we examine equilibrium outcomes that may underestimate the direct effect of media access if there are cross-neighbor information spillovers or (in the unlikely event that) campaigns are aware of differences in media access and reallocate resources to counteract the information. The design thus captures an intent to treat effect of differences in access to commercial quality local media. 51 Our estimates are thus likely to understate the effect of gaining access to an additional local media station, while any spillovers (e.g. arising from driving to work across radio coverage boundaries) should further attenuate our estimates toward zero. Figure 5 illustrates our identification 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 and broadcasting some local content. 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 from within 51 Ideally, we could also identify the electoral effect of consuming to an additional media station by using our measure of access to instrument for the media actually consumed. Unfortunately, we have not been able to obtain the detailed individual-level data required to estimate such a first stage. 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. Furthermore, voters may discuss the news that they receive with their friends and family, which would violate the exclusion restriction. 28

29 the same municipality that receive a different number of local media stations. 52 We include fixed effects for each such neighbor grouping to ensure that our estimates exploit only variation between neighbors (and thus within-municipality) variation in access to local media stations. Our identifying assumption is that neighboring precincts differ only in their local media coverage. Although within-neighbor variation undoubtedly reflects exogenous topological characteristics, it could also reflect sorting. Our estimates would be biased if certain types of voters live in areas with better local media coverage, or media stations strategically choose the strength or location of their emitter to target certain audiences. 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 do not drop off discontinuously, the antennae strengths that media stations purchase are highly discrete. 53 Nevertheless, to maximize the plausibility of our identification strategy, we restrict our sample to precincts with at most an area of 10km 2. Between small neighboring precincts, sorting is particularly unlikely because media stations cannot plausibly choose transmission technologies to separate between markets at such a fine-grained level. Given the substantial reach of AM stations, a second key advantage of this design is that more urban areas (containing smaller precincts) almost exclusively differ with respect to FM radio and television coverage. Removing the most volatile media signal increases the accuracy of our coverage measures. Moreover, our area restriction prevents rural-urban comparisons at the edges of towns and cities, which could be problematic because urban areas correlate with other politically-relevant characteristics. This yielded a sample containing 14,828 observations, where neighbors typically differ by one or two local media stations. To evaluate the plausibility of our design, we again test whether local media predict background covariates. Table 3 shows that only 4 of 39 precinct-level variables are significantly correlated with local media at the 10% significance level. 54 A particularly important concern is that precincts receiving more local media are closer to the municipal center, where antennas are typically located, and may differ in politically salient ways. However, the balance tests show that distance to the centroid of the municipal head (from both a precinct s border and its centroid), precinct area, total population, number of registered voters, and population density (all as natural logarithms) are all 52 We lack the sufficient power to compare no media to some media. 53 The power output in watts for the AM, FM, and television stations in our sample are almost exclusively round thousands and divisible by 5, Relative to Table 2, we removed municipality-level variables. 29

30 Table 3: Imbalance over local media (local media neighbor sample) Incumbent Effective Registered Turnout Distance to Distance to Area Population Population Non-local party vote number of voters (log) (lag) municipal municipal (log) (log) density media share (lag) political head from head from (log) parties (lag) border (log) centroid (log) Local media (0.0004) (0.0019) (0.0066) (0.0005) (0.0076) (0.0043) (0.0067) (0.0052) (0.0066) (0.2231) AM FM Television Average Share Share Average Share Share no Share antennae antennae antennae children households indigenous years of illiterate schooling incomplete in in in per woman with male speakers schooling primary precinct precinct precinct head schooling Local media ** *** * (0.0023) (0.0024) (0.0003) (0.0024) (0.0002) (0.0003) (0.0281) (0.0001) (0.0001) (0.0001) Share Share Share Share Share Share Share Average Average Share complete incomplete complete higher economically without state occupants occupants non-dirt primary secondary secondary education active health workers per per floor schooling schooling schooling care health care dwelling room Local media (0.0005) (0.0009) (0.001) (0.0015) (0.0002) (0.0005) (0.0002) (0.0034) (0.0016) (0.0002) 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 * (0.0003) (0.0006) (0.0004) (0.0001) (0.0006) (0.0005) (0.0007) (0.0012) (0.0013) Notes: Each coefficient is from a separate OLS regression including neighbor fixed effects, is estimated using OLS, and observations are weighted by electorate size in treated precincts and electorate size divided by the number of control precincts per neighbor set for control precincts (see equation (12) in the main text). Each regression contains 14,828 observations. Standard errors are clustered by municipality-election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p <

31 consistent with local media being exogenously assigned across neighboring precincts. Moreover, and crucially for distinguishing the effects of local and non-local media, there is no relationship between the number of local and non-local media stations. More generally, we find no evidence that the number of local media stations covering a precinct predicts proxies for socioeconomic development, such as education, basic housing needs, or owning more luxurious items like computers, washing machines, and cell phones. Similarly, we find no correlation between the presence of local media and prior political behavior: an additional media station is not associated with incumbent electoral performance, local political competition (proxied by the effective number of political parties), and electoral turnout. These balance tests add credibility to our identification strategy Estimation To estimate how local media moderate the effect of pre-election audit report revelations, and thus test hypothesis H1, we combine our variation in the timing of audit report releases and the number of local media stations covering an electoral precinct. Specifically, we use OLS to estimate the following baseline specification in our sample of neighboring precincts: Y pkmt = β 1 pre-election audit mt + β 2 outcome Q3 mt + β 3 outcome Q4 mt + β 4 local media ( ) ( pm ) +β 5 pre-election audit mt outcome Q3 mt + β 6 pre-election audit mt outcome Q4 mt ( ) +β 7 pre-election audit mt local media pm ( ) +β 8 pre-election audit mt outcome Q3 mt local media pm ( ) +β 9 pre-election audit mt outcome Q4 mt local media pm + ξ km + ε pkmt, (12) where Y pkmt is the incumbent party s vote share in precinct p in neighbor group k nested within municipality m in year t, pre-election audit mt is an indicator for an audit report being released before the election, outcome Q3 mt and outcome Q4 mt are indicators for municipalities in the third and fourth quartiles of either the unauthorized spending distribution or not spending on the poor distribution (whether the audit was released before or after the election), and local media pm is the total number of local media stations. 55 Given that we only consider neighboring precincts within 55 The demanding structure of our identification strategy means that we lack the power to non-parametrically estimate media s effect. The ideal non-parametric approach would allow media s effect to vary for each number of media stations. However, by requiring 4 coefficients for each of the 41 levels in our data, we quickly lose power and rely on cells containing few observations. A second best approach could use 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 31

32 the same muncipality, neighbor-group fixed effects ξ km ensure that our estimates are identified from within-neighbor variation in media coverage within municipalities. To both weigh each voter and each neighbor comparison equally, each observation is weighted by the number of registered voters. For control precincts, this weight is further divided by the number of control units per neighbor group. 56 Throughout, standard errors are clustered by municipality-election. Our main coefficients of interest β 7, β 8, and β 9 respectively identify the marginal effect of an additional local media station on electoral accountability 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 (Q3), and to be in the fourth quartile of malfeasance (Q4). Because, as in our model, individual precincts represent only a tiny fraction of a media station s potential audience, our estimates should not reflect the decision of media stations to target content at specific precincts. Based on our model, we expect rewards for incumbent parties correctly allocating federal transfers (β 7 > 0) and punishment for clearly highly malfeasant parties (β 9 < 0). The sign of β 8, however, is not obvious given that voters prior beliefs about incumbent party malfeasance are unlikely to be zero (e.g. Banerjee et al. 2011). The coefficients β 1, β 5, and β 6 similarly estimate the effects of audit report revelations in the absence of local media. Finding that β 1 > 0 and β 6 < 0 would indicate that voters sanction incumbent parties even in the absence of local media How the effect of local media varies with the media market structure A second empirical objective is to examine how the local media market structure moderates local media s role in supporting electoral accountability, and thus test hypotheses H2 and H3. Accordingly, we examine effect heterogeneity by media market characteristics using specifications of the 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. 56 Appendix Table A8 shows that unweighted results yield similar estimates. 32

33 form: Y pkmt = β 1 pre-election audit mt + β 2 outcome Q3 mt + β 3 outcome Q4 mt + β 4 local media pm ( ) +β 5 market structure pm + β 6 pre-election audit mt outcome Q3 mt ( ) ( ) +β 7 pre-election audit mt outcome Q4 mt + β 8 pre-election audit mt local media pm ( ) +β 9 pre-election audit mt outcome Q3 mt local media pm ( ) +β 10 pre-election audit mt outcome Q4 mt local media pm ( ) +β 11 pre-election audit mt market structure pm ( ) +β 12 pre-election audit mt outcome Q3 mt market structure pm ( ) +β 13 pre-election audit mt outcome Q4 mt market structure pm ( ) +β 14 pre-election audit mt local media pm market structure pm ( ) +β 15 pre-election audit mt outcome Q3 mt local media pm market structure pm (13) ( ) +β 16 pre-election audit mt outcome Q4 mt local media pm market structure pm + ξ km + ε pkmt, where market structure pm is one of our two measures of the media market structure. Following our theoretical argument, and in order to test H2, we first explore potential crowdout by examining how the effect of an additional local media station depends on the presence of non-local media stations. In this case, market structure pm in equation (13) represents the number of non-local media stations covering a precinct. If non-local media stations which face weaker incentives to report on incumbent malfeasance in a voter s municipality attract consumers from local media stations, we expect to find large effects for local media (β 8 > 0 and β 10 < 0), but also crowd-out effects of non-local media such that β 14 < 0 and β 16 > 0. In order to test H3, our second test examines how the composition of a local media station s audience influences its effect on electoral accountability. Media stations whose audience is principally located within the municipality have a stronger incentive to cover audit report outcomes in depth, given that this is likely to interest their audience. To test this argument, we employ a similar approach to the congruence measure introduced by Snyder and Strömberg (2010). Specifically, for each local media station, we calculate the proportion of the population that receives a commercial quality signal which resides in the municipality containing the antennae. Our local market congruence measure is the average of these proportions across all AM radio, FM radio, and television stations covering a given electoral precinct. If local media outlets indeed disproportionately cover local malfeasance when their audience is principally located within their municipality, we expect to find sanctioning where local media is highly congruent, i.e. β 14 > 0 and β 16 < 0. 33

34 5 Results We now present our findings. Our first main contribution demonstrates that voter sanctioning of incumbent party malfeasance increases with each additional local media station. Consistent with our model, we then show that such effects are indeed moderated by the media market structure: local media induces greatest sanctioning of incumbent party malfeasance in the absence of competing non-local media stations, and where local media stations audiences are most local. 5.1 Audits, local media, and electoral accountability Table 4 presents our results examining whether voters sanction parties for mayoral malfeasance. Before turning to the role of local media, columns (1) and (2) first report the average effect of audit reports released before an election on the precinct-level change in incumbent party vote share, regardless of the local media environment. This entails estimating equation (12) without measures of media in the full sample of audits. 57 Appendix Table A4 shows similar results in the subsample we focus on for estimating the effect of media. The results tentatively suggest that, on average across precincts, pre-election audit report revelations could have electoral consequences. Column (1) estimates that a pre-election audit revealing the incumbent party s mayor to be in the top quartile of unauthorized spending reduces that party s vote share by 4.5 percentage points. Column (2) similarly estimates that revealing the incumbent party s mayor to have spent FISM funds on projects not benefiting the poor reduces the incumbent party s vote share by 2.5 percentage points for mayors in the third quartile and a further 3.8 percentage points for mayors in the fourth quartile. In both cases, the pre-election 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 likely enhanced by the audits may slightly increase their vote share. However, it is important to emphasize that although these point estimates are relatively large, they are imprecise and statistically insignificant by conventional standards. 58 One potential explanation for this lack of precision is the heterogeneity in the effect arising from audit reports only influencing voter behavior in the presence of local media stations. Turning to columns (3) and (4) our main estimates of the sanctioning effect of access to an additional local media station, in the sample of neighboring precincts that differ in access to local media we indeed find that each additional local media station substantially increases electoral sanctioning. Column (3) shows that, on average, mayors in the third and fourth quartiles of the 57 Appendix Table A5 shows similar results examining the municipal-level probability of re-election. 58 These estimates are similar in magnitudes to those in Ferraz and Finan s (2008) larger Brazilian sample. 34

35 Table 4: Effects of local media and pre-election audit reports publicizing malfeasance before an election (full sample and local media neighbor sample) Change in incumbent party vote share (1) (2) (3) (4) Constant *** *** * *** (0.015) (0.015) (0.043) (0.034) Pre-election audit *** (0.019) (0.021) (0.057) (0.036) Unauthorized Q (0.026) (0.057) Pre-election audit Unauthorized Q (0.037) (0.075) Unauthorized Q (0.038) (0.063) Pre-election audit Unauthorized Q * (0.047) (0.087) Not poor Q ** (0.023) (0.065) Pre-election audit Not poor Q (0.036) (0.102) Not poor Q *** (0.043) (0.051) Pre-election audit Not poor Q (0.048) (0.071) Local media (0.002) (0.002) Pre-election audit Local media 0.006* (0.003) (0.003) Unauthorized Q3 Local media (0.004) Pre-election audit Unauthorized Q3 Local media (0.005) Unauthorized Q4 Local media (0.004) Pre-election audit Unauthorized Q4 Local media * (0.005) Not poor Q3 Local media (0.004) Pre-election audit Not poor Q3 Local media (0.005) Not poor Q4 Local media (0.003) Pre-election audit Not poor Q4 Local media ** (0.005) Observations 46,419 46,419 14,828 14,828 Outcome mean Outcome standard deviation Pre-election audit mean Malfeasance Q3 mean Malfeasance Q4 mean Local media mean Local media standard deviation Notes: Specifications (1) and (2) use the full sample of all possible precincts, and weight precinct by electorate size. Specifications (3) and (4) restrict the sample to neighboring groups that differ in the number of local media they are covered by; these specifications include neighbor group fixed effects, and observations are weighted by electorate size in treated precincts and electorate size divided by the number of control precincts per neighbor set for control precincts. All specifications are estimated using OLS. The omitted category for unauthorized spending and not spending on the poor is Q1 and Q2. Standard errors are clustered by municipality-election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p <

36 unauthorized spending distribution respectively experience a 0.6 and 0.9 percentage point loss in their precinct-level vote share for each additional local media station. A standard deviation increase in the number of local media stations, which entails 10 more such stations, thus respectively reduces the malfeasant incumbent party s vote share by one or two thirds of the average municipal victory margin in our sample of 15 percentage points. Moreover, the positive and statistically significant coefficient on the interaction between being audited before an election and the number of local media stations shows that each additional local media station increases the vote share of clean incumbent parties by 0.6 percentage points. In contrast with these findings, the positive and statistically significant coefficient for the interaction between a pre-election audit and a mayor s unauthorized spending quartile shows that a mayor revealed to be malfeasant in this respect is not punished in precincts covered by no (or few) local media stations. We find similar local media-induced sanctioning of revelations that funds were not spent on projects benefiting the poor. Column (4) shows that an additional local media station reduces a neglectful mayor s vote share by 0.5 percentage points for mayors in the third quartile and 1.2 percentage points for mayors in the most malfeasant quartile. A standard deviation increase in the number of local media stations thus entails a 12 percentage point decrease in the vote share a quarter reduction in their precinct-level total vote share of the most neglectful mayors if their malfeasance is revealed before an election. Furthermore, the positive but statistically insignificant interaction between revealing an audit and the total number of local media stations again suggests that parties revealed to have correctly spent all FISM funds on projects benefiting the poor receive a boost at the polls, although this is smaller than the degree to which malfeasant mayors are punished. The insignificant interaction between the pre-election audit release and not spending on the poor again indicates that in locations with very few local media stations the incumbent party is not meaningfully electorally sanctioned. The interactions between local media and audit report outcomes released after the election also provide a placebo test. Specifically, local media should not amplify sanctioning of revelations that do not become public until after the election. The results in both columns (3) and (4) show insignificant interactions, and thus provide further evidence that our findings are not driven by other characteristics of municipalities that correlate with malfeasance and also reduce incumbent support. A second placebo test similarly examines whether future pre-election audit reports affect voting behavior. Appendix Table A7 shows that malfeasant audit outcomes do not decrease the incumbent party s vote share at the election before they were released, suggesting that our findings do not reflect differential trends between precincts that vary in their municipal malfeasance and local media access. 36

37 5.1.1 Robustness checks We now demonstrate the robustness of our main finding that local media facilitates electoral sanctioning. In addition to the balance and placebo tests noted above, Table 5 presents the results of various sensitivity checks and alternative definitions of malfeasance. We focus on the triple interactions estimating the effects of local media publicizing mayoral malfeasance. The sensitivity checks in Panel A show that the results are robust to adding potential confounds, restricting the sample to comparisons equidistant from media antennae, and varying our area restriction and media coverage definition. First, we include pre-election audit mt outcome Q3 mt non-local media pm and pre-election audit mt outcome Q4 mt non-local media pm (and lowerorder terms) to interactively control for the possibility that the sanctioning effect of an additional local media station reflects differential access to non-local media stations. Consistent with the balance shown in Table 3, columns (1) and (2) indicate that the effects of local media are unaffected. Columns (3) and (4) more generally show that the results are robust to simultaneously further including flexible (quadratic) interactive controls for the four variables showing statistically significant imbalances across local media, as well as four other potentially important confounds. 59 Second, to address the potential concern that neighbors closer to antennae which are more likely to receive local media systematically differ, we restrict the sample to neighboring precincts that differ by less than 50m on average from their nearest media stations. 60 Such comparisons thus principally exploit topological variation arising from impediments that limit the receipt of radio and television signals between precincts effectively adjacent to the signal of the average media station (Olken 2009). At the cost of dropping more than 85% of the sample, and thus substantially reducing statistical power, columns (5) and (6) confirm that this restriction produces similar results. Third, columns (7) and (8) redefine local media as the natural logarithm of the number of local media stations. The results show that our findings are not driven by precincts containing many local media stations, and imply that a 10% increase in local media decrease the vote share of malfeasant incumbent parties by around percentage points. Fourth, we show that our results are not driven by restricting attention to relatively small neighboring precincts or precincts around the 20% cutoff for defining a precinct as covered by a local media station. Columns (9) and (10) show larger estimates when the restriction to precincts of 10km 2 or less is removed, while columns (11)-(14) show that our findings are robust to both relaxing and strengthening our definition of precinct coverage respectively to require that 10% or 50% 59 These are: non-local media, (log) distance to the municipal head from the precinct centroid, average years of schooling, and an index of five household assets. 60 We compute the average distance to the 36 (the sample average number of media stations covering a precinct) nearest stations. 37

38 of the population receive a commercial quality coverage signal. 61 Finally, we drop the post-election comparison to address the possibility that media and election dynamics in municipalities receiving audit reports after the election are different. Columns (15) and (16) show that this approach, exploiting only variation in pre-election audit reports publicized by local media before the election (e.g. Banerjee et al. 2011; Chong et al. 2015), yields more precise estimates at the expense of plausibly exogenous variation in audits. The alternative definitions of malfeasance in Panel B reinforce our findings, and suggest that voters primarily sanction high levels of malfeasance. Columns (1) and (2) first report linear measures of unauthorized spending and not spending on the poor. 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 4.1 percentage points in the case of unauthorized spending and 5.5 percentage points in the case of not spending on the poor. While substantial, only the estimate for not spending on the poor is precisely estimated. Together with our baseline estimates, these estimates suggest that unauthorized spending s effects may be non-linear. To examine when the effect of malfeasance kicks in, columns (3)-(8) sequentially operationalize malfeasant behavior using an indicator for unauthorized spending 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 unauthorized spending, the evidence suggests that only the highest levels greater than 20% are severely punished. This suggests that voters may possess negative prior beliefs about incumbent unauthorized spending. Finally, at the cost of sacrificing our plausibly exogenous variation in local media, we estimate equation (12) in the full sample of audited municipalities to check the external validity of the neighbor sample estimates. The results, provided in Appendix Table A10, are broadly similar. For both unauthorized spending and not spending 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, while malfeasance in the third quartile is not significantly sanctioned. Although the samples differ in terms of both composition and quality of identification, the similarity of the results suggests that our findings may generalize beyond the urban sample we focus on No sanctioning role for non-local media Although non-local media stations are not driving the sanctioning effects of local media, they could nevertheless also support electoral accountability. We examine this possibility using the same identification strategy, but instead exploiting variation in non-local media coverage across 61 Appendix Table A9 shows similar estimates when our 10km 2 restriction is relaxed to 25km 2 and 50km 2. 38

39 Panel A: Sensitivity tests Table 5: Robustness checks (local media neighbor sample) Change in incumbent party vote share (1) (2) (3) (4) (5) (6) (7) (8) Unauthorized Q3 Local media ** (0.003) (0.002) (0.009) (0.039) Pre-election audit Unauthorized Q3 Local media * * (0.003) (0.003) (0.010) (0.045) Unauthorized Q4 Local media (0.003) (0.002) (0.010) (0.036) Pre-election audit Unauthorized Q4 Local media ** *** *** (0.004) (0.003) (0.011) (0.048) Not poor Q3 Local media (0.006) (0.004) (0.009) (0.047) Pre-election audit Not poor Q3 Local media (0.006) (0.005) (0.010) (0.062) Not poor Q4 Local media (0.004) (0.003) (0.010) (0.037) Pre-election audit Not poor Q4 Local media * ** ** *** (0.006) (0.005) (0.010) (0.056) Observations 14,828 14,828 14,828 14,828 5,640 5,640 14,556 14,556 Interactive controls Non-local media 7 quadratic covariates None None None None Local media transformation None None None None None None Natural log Panel A (cont.) Change in incumbent party vote share (9) (10) (11) (12) (13) (14) (15) (16) Unauthorized Q3 Local media 0.010*** 0.008* * (0.004) (0.004) (0.004) (0.001) Pre-election audit Unauthorized Q3 Local media *** ** (0.004) (0.005) (0.005) Unauthorized Q4 Local media *** (0.004) (0.004) (0.004) (0.002) Pre-election audit Unauthorized Q4 Local media *** (0.005) (0.005) (0.005) Not poor Q3 Local media (0.004) (0.005) (0.005) (0.003) Pre-election audit Not poor Q3 Local media (0.005) (0.006) (0.006) Not poor Q4 Local media 0.010*** * ** (0.004) (0.003) (0.004) (0.006) Pre-election audit Not poor Q4 Local media *** *** *** (0.005) (0.004) (0.005) Observations 33,659 33,659 14,014 14,014 13,391 13,391 9,148 9,148 Sample restriction <10km 2 <10km 2 <10km 2 <10km 2 <10km 2 <10km 2 <10km 2 & Pre-election Cutoff for media coverage None None 10% 10% 50% 50% 20% 20% Panel B: Malfeasance definitions Change in incumbent party vote share (1) (2) (3) (4) (5) (6) (7) (8) Pre-election audit Local media Unauthorized (0.028) Pre-election audit Local media Not poor *** (0.020) Pre-election audit Local media Unauthorized 5% (0.006) Pre-election audit Local media Not poor 5% ** (0.004) Pre-election audit Local media Unauthorized 10% ** (0.004) Pre-election audit Local media Not poor 10% ** (0.005) Pre-election audit Local media Unauthorized 20% *** (0.013) Pre-election audit Local media Not poor 20% *** (0.005) Observations 14,828 14,828 14,828 14,828 14,828 14,828 14,828 14,828 Notes: All specifications include neighbor fixed effects, are estimated using OLS, and observations are weighted by electorate size in treated precincts and electorate size divided by the number of control precincts per neighbor set for control precincts. All other interactions are omitted. In specifications (3) and (4) of Panel A, the controls are an indicator for precincts containing an AM antennae, the share of households with a male head, the illiterate share of the population, the share of households with a washing machine, the number of non-local media, distance to the municipal head from the precinct centroid (log), average years of schooling, and an index of five household assets. Columns (7) and (8) of Panel A exclude precincts with no local media stations. Standard errors are clustered by municipality-election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p <

40 neighboring precincts that differ in the number of non-local media stations that they are covered by. The large number of non-local media stations increases our sample to 40,208 observations. The results presented in Appendix Table A11 show that access to non-local media does not systematically influence voter sanctioning of incumbent malfeasance revelations. 62 This indicates that local media is the key driver of electoral accountability, and thus access to any type of media is not sufficient to punish the parties of malfeasant mayors Rich-poor distributional conflict over FISM spending An important difference between unauthorized spending and not spending on the poor is its distributional implications. While unauthorized spending still reaches the intended recipients or reaches no recipients at all, not spending on the poor redistributes FISM funds toward richer voters. Consequently, we may expect that richer (poorer) voters are less (more) likely to punish an incumbent party for not spending funds designed for the poor actually on those in extreme poverty. To examine this possibility, we further interact the effect of local media with two proxies for the extent of voter poverty: a standardized scale for precinct-level socioeconomic under-development that combines 25 measures of impoverishment from the Census; and an indicator for precincts containing voters that FISM funds should directly target. The results in Table 6 suggest that the extent of media-induced sanctioning indeed reflects distributional incentives. Columns (2) and (4) show that voters in poorer precincts, by both definitions of poverty are more likely to sanction incumbent parties revealed to have spent funds on projects not benefiting the poor. In each case, the coefficient on under-development or being a FISM target is negative, and statistically significant or borderline statistically significant in three of four instances. Conversely, columns (1) and (3) show that the sanctioning of unauthorized spending does not vary by precinct socioeconomic development. Although the poor also suffer relatively more from unauthorized spending than the rich, these results indicate that the distributional conflict is weaker and thus further suggest that voters are responding to the specific nature of malfeasance revelations covered by local media. 5.2 The moderating role of media market structure To better understand the conditions under which local media stations facilitate electoral accountability, we further investigate how the media market structure crowd-out by non-local media and the geographic composition of a station s potential audience moderates local media s effects. 62 The significant negative coefficient for Q3 unauthorized spending is not consistent with the smaller effects for Q4 and positive coefficients for not spending on the poor. 40

41 Table 6: Effects of local media and audit reports publicizing malfeasance before an election, by probability of benefiting from FISM funds (local media neighbor sample) Change in incumbent party vote share (1) (2) (3) (4) Pre-election audit Local media 0.006** (0.003) (0.003) (0.003) (0.003) Pre-election audit Local media Under development ** (0.003) (0.003) Pre-election audit Unauthorized Q3 Local media (0.005) (0.005) Pre-election audit Unauthorized Q3 Local media Under development (0.005) Pre-election audit Unauthorized Q4 Local media (0.005) (0.005) Pre-election audit Unauthorized Q4 Local media Under development (0.005) Pre-election audit Not poor Q3 Local media (0.005) (0.005) Pre-election audit Not poor Q3 Local media Under development *** (0.003) Pre-election audit Not poor Q4 Local media *** ** (0.004) (0.004) Pre-election audit Not poor Q4 Local media Under development ** (0.005) Pre-election audit Local media FISM target (0.003) (0.003) Pre-election audit Unauthorized Q3 Local media FISM target (0.006) Pre-election audit Unauthorized Q4 Local media FISM target (0.005) Pre-election audit Not poor Q3 Local media FISM target (0.008) Pre-election audit Not poor Q4 Local media FISM target *** (0.005) Observations 14,828 14,828 14,828 14,828 FISM beneficiary proxy mean FISM beneficiary proxy standard deviation Notes: All specifications include neighbor fixed effects, are estimated using OLS, and observations are weighted by electorate size in treated precincts and electorate size divided by the number of control precincts per neighbor set for control precincts. The omitted category for unauthorized spending and not spending on the poor is Q1 and Q2. Lower order terms are omitted. Standard errors are clustered by municipality-election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p <

42 5.2.1 Non-local media crowd out local media Since non-local media are less likely to cover municipal mayors beyond the outlet s own municipality, which is supported by the preceding analysis demonstrating 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 also rests upon the model s assumption that individuals care about variety, and thus that a new media station attracts listeners or viewers away from local media stations. The results in columns (1) and especially (2) of Table 7 suggest that non-local media indeed crowd out the sanctioning effects of local media. For both high and low malfeasance revelations, the positive quadruple interaction coefficients suggest that access to more non-local media stations weakens the effect of an additional local media station. Particularly for not spending on the poor, the significant interaction terms show that each additional non-local media station reduces the effect of an additional local media station by percentage points, 2.5-5% of an additional local media station s effect in the absence of non-local media. In the average precinct, covered by 17 non-local media stations, the effect of an additional local media station falls by around 2 percentage points. Although they also reduce the effect of local media, the interactions suggest that an additional non-local media station only reduces the effect of an additional local media station by 0.05 percentage points in the case of unauthorized spending. 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 spending that does not benefit the poor. Furthermore, the fact that an additional non-local media station reduces the effect of local media suggests both that voters indeed receive municipal news from local media otherwise, an additional media station should not affect voter behavior and that voters have a preference for variety Audience composition influences sanctioning To further explore the mechanism driving local media s effect, we next consider the composition of the media market. In particular, as highlighted by our theoretical model, local media outlets predominantly serving an audience within their municipality face stronger incentives to publicize audit report revelations. In line with H3, we thus expect to observe greater sanctioning in precincts covered by local media primarily serving their own municipality. Columns (3) and (4) of Table 7 support this hypothesis, providing evidence consistent with a media station s market influencing electoral accountability. Specifically, we find that a larger local market share increases the reward and punishment of incumbent parties. Again, this is par- 42

43 Table 7: Effects of local media and audit reports publicizing malfeasance before an election, conditional on the media market structure (local media neighbor sample) Change in incumbent party vote share (1) (2) (3) (4) Pre-election audit Local media ** ** ** (0.0055) (0.0065) (0.0048) (0.0027) Pre-election audit Non-local media Local media *** (0.0002) (0.0001) Pre-election audit Unauthorized Q3 Local media * (0.0106) (0.0074) Pre-election audit Unauthorized Q3 Non-local media Local media (0.0003) Pre-election audit Unauthorized Q4 Local media * (0.0109) (0.0072) Pre-election audit Unauthorized Q4 Non-local media Local media * (0.0003) Pre-election audit Not poor Q3 Local media *** (0.0097) (0.0090) Pre-election audit Not poor Q3 Non-local media Local media *** (0.0005) Pre-election audit Not poor Q4 Local media *** (0.0094) (0.0052) Pre-election audit Not poor Q4 Non-local media Local media *** (0.0002) Pre-election audit Local media Local market * (0.0076) (0.0121) Pre-election audit Unauthorized Q3 Local media Local market (0.0213) Pre-election audit Unauthorized Q4 Local media Local market (0.0161) Pre-election audit Not poor Q3 Local media Local market (0.0152) Pre-election audit Not poor Q4 Local media Local market *** (0.0157) Observations 14,828 14,828 14,828 14,828 Media market structure interaction mean Media market structure interaction standard deviation Notes: All specifications include neighbor fixed effects, are estimated using OLS, and observations are weighted by electorate size in treated precincts and electorate size divided by the number of control precincts per neighbor set for control precincts. In columns (3) and (4), the local market variable is the average share of a media outlet s potential audience that resides within the municipality of its antennae, averaging across all AM, FM, and television stations covering a precinct. The omitted category for unauthorized spending and not spending on the poor is Q1 and Q2. Media, unauthorized spending, and non poor spending lower order terms are omitted. Standard errors are clustered by municipality-election. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p <

44 ticularly 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 0.7 percentage points for mayors in the third quartile and 4.0 percentage points for mayors in the fourth quartile. 63 Similarly, mayors that correctly spend FISM funds on projects benefiting the poor gain 1.7 percentage points for an additional media station exclusively serving their market. Although the analogous coefficients for unauthorized spending are again not quite statistically significant, they paint a similar picture where a mayor from the third and fourth quartile is respectively punished by 2.6 and 2.1 percentage points. The party of a mayor that does not engage in unauthorized spending receives a significant boost of 1.5 percentage points for each additional local media station serving only the local municipal market. Notably, we observe no reward or punishment of incumbent parties when a media station s main market is outside the municipality where it broadcasts from, as indicated by the insignificant baseline triple interaction coefficients. 6 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 can play important roles in supporting local electoral accountability. While the influence of national media has received considerable attention, we demonstrate the importance of local media and the media market structure for electoral accountability in a major developing context where local governments play an important role in service delivery. In particular, we find that local media increase voter sanctioning of Mexico s malfeasant incumbent parties, particularly in conducive media environments containing relatively few non-local media stations and where local media primarily serve a segmented local audience. Variation in the presence of media stations with incentives to provide local content has the potential to explain the substantially varying levels of electoral accountability observed across the world. However, the implications for democracies across the world are somewhat mixed. On the one hand, increasing access to media in developing contexts offers unprecedented opportunities for supporting democratic accountability. 64 Our estimates imply that extensive nationwide access to local media stations could remove malfeasant incumbent parties from office and increase the probability of re-electing high-performing parties. By reporting on incumbent performance in office, local media also encourages better future performance. For example, Avis, Ferraz and Fi- 63 These estimates respectively reflect β 9 + β 15 and β 10 + β E.g. State of the Media, Why U.S. Newspapers Suffer More than Others, stateofthemedia.org/2011/mobilesurvey/international-newspaper-economics 44

45 nan (forthcoming) show that previously-audited Brazilian municipalities exhibit lower corruption, while Casey (2015) finds that the media can encourage parties to shift away from ethnic appeals towards more programmatic spending in Sierra Leone. On the other, local news is increasingly being replaced by national and international content. 65 To the extent that this crowds out and replaces local content, this is likely to adversely affect local accountability as voters consume less content relevant to appraising local politicians. Moreover, a lack of regulation in media markets, combined with an understanding of media s potential impact, can create opportunities for media capture by politicians and other interests (Besley and Prat 2006). For example, recent research in Latin America shows that incumbents strategically purchase their own radio stations (Boas and Hidalgo 2011) and introduce regulation outlawing defamation (Stanig 2015). Petrova (2011) similarly shows that falling advertising revenues reduced media independence from political parties in the U.S. in the nineteenth century. More generally, Reporters Without Borders report that global press freedoms have decreased 14% since In the U.S., see e.g. Nieman Lab, As giant platforms rise, local news is getting crushed, September 1st 2015, niemanlab.org/2015/09/as-giant-platforms-rise-local-news-is-getting-crushed, and Washington Post, Charting the years-long decline of local news reporting, March 26th 2014, washingtonpost.com/lifestyle/style/charting-the-yearslong-decline-of-local-news-reporting/2014/03/26/977bf088-b457-11e3-b de76985 story.html. In China, see e.g. State of the Media, Why U.S. Newspapers Suffer More than Others, stateofthemedia.org/2011/mobilesurvey/international-newspaper-economics. 66 Reporters Without Borders, A deep and disturbing decline in media freedom, rsf.org/en/deep-and-disturbingdecline-media-freedom. 45

46 References Ansolabehere, Stephen, Erik C. Snowberg and James M. Snyder Television and the incumbency advantage in US elections. Legislative Studies Quarterly 31(4): Ashworth, Scott Electoral accountability: recent theoretical and empirical work. Annual Review of Political Science 15: Avis, Eric, Claudio Ferraz and Frederico Finan. forthcoming. Do Government Audits Reduce Corruption? Estimating the Impacts of Exposing Corrupt Politicians. Journal of Political Economy. Banerjee, Abhijit V., Selvan Kumar, Rohini Pande and Felix Su Do Informed Voters Make Better Choices? Experimental Evidence from Urban India. Working paper. Barro, Robert J The control of politicians: An economic model. Public Choice 14(1): Baum, Matthew A Sex, lies, and war: How soft news brings foreign policy to the inattentive public. American Political Science Review 96(1): Besley, Timothy and Andrea Prat Handcuffs for the grabbing hand? Media capture and government accountability. American Economic Review 96(3): Besley, Timothy and Robin Burgess The political economy of government responsiveness: Theory and evidence from India. Quarterly Journal of Economics 117(4): Boas, Taylor C. and F. Daniel Hidalgo Controlling the airwaves: Incumbency advantage and community radio in Brazil. American Journal of Political Science 55(4): Bobonis, Gustavo J., Luis R. Cámara Fuertes and Rainer Schwabe Monitoring Corruptible Politicians. American Economic Review 106(8): Casey, Katherine Crossing Party Lines: The Effects of Information on Redistributive Politics. American Economic Review 105(8): Castañeda Sabido, Fernando 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. 46

47 Chang, Eric C.C., Miriam A. Golden and Seth J. Hill Legislative malfeasance and political accountability. World Politics 62(2): Chong, Alberto, Ana De La O, Dean Karlan and Leonard Wantchekon Does Corruption 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): Costas, Elena, Albert Solé-Ollé and Pilar Sorribas-Navarro Do voters really tolerate corruption? Evidence from Spanish Mayors. Working paper. de Figueiredo, Miguel F.P., F. Daniel Hidalgo and Yuri Kasahara When Do Voters Punish Corrupt Politicians? Experimental Evidence from Brazil. Working paper. De La O, Ana L. and Fernando Martel García Can Intrastate Accountability Reduce Local Capture? Results from a Field Experiment in Mexico. Working paper. DellaVigna, Stefano and Ethan Kaplan The Fox News effect: Media bias and voting. Quarterly Journal of Economics 122(3): Delli Carpini, Michael X. and Scott Keeter What Americans Know about Politics and Why It Matters. New Haven: Yale University Press. Durante, Ruben and Brian Knight Partisan control, media bias, and viewer responses: Evidence from Berlusconi s Italy. Journal of the European Economic Association 10(3): Durante, Ruben and Ekaterina Zhuravskaya. forthcoming. Attack When the World Is Not Watching? U.S. News and the Israeli-Palestinian Conflict. Journal of Political Economy. Durante, Ruben, Paolo Pinotti and Andrea Tesei The Political Legacy of Entertainment TV. Working paper. Eggers, Andrew Partisanship and Electoral Accountability: Evidence from the UK Expenses Scandal. Quarterly Journal of Political Science 9(4): Eisensee, Thomas and David Strömberg News droughts, news floods, and US disaster relief. Quarterly Journal of Economics pp Enikolopov, Ruben, Maria Petrova and Ekaterina Zhuravskaya Media and political persuasion: Evidence from Russia. American Economic Review 101(7):

48 Ferejohn, John Incumbent performance and electoral control. Public Choice 50(1):5 25. Ferraz, Claudio and Frederico Finan Exposing Corrupt Politicians: The Effects of Brazil s Publicly Released Audits on Electoral Outcomes. Quarterly Journal of Economics 123(2): Gentzkow, Matthew and Jesse M. Shapiro Media bias and reputation. Journal of Political Economy 114(2): Hamilton, James All the news that s fit to sell: How the market transforms information into news. Princeton: Princeton University Press. Humphreys, Macartan and Jeremy Weinstein Policing Politicians: Citizen Empowerment and Political Accountability in Uganda Preliminary Analysis. Working paper. Ibope/AGB México Audiencias y Medios en México.. Keefer, Phillip 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 Klašnja, Marko. forthcoming. Uninformed Voters and Corrupt Politicians. American Politics Research. Langston, Joy Rising from the ashes? Reorganizing and unifying the PRI s state party organizations after electoral defeat. Comparative Political Studies 36(3): Larreguy, Horacio A., Christopher Lucas and John Marshall When do media stations support political accountability? A field experiment in Mexico. AEA RCT Registry AEARCTR Larreguy, Horacio A., John Marshall and James M. Snyder, Jr. forthcoming. Leveling the playing field: How campaign advertising can help non-dominant parties. Journal of the European Economic Association. Lavielle, Briseida, Kristina Pirker and Alberto Serdán Monitoreo de Programas Sociales en Contextos Electorales. Fundar Working paper. Lawson, Chappell H Building the fourth estate: Democratization and the rise of a free press in Mexico. Berkeley: University of California Press. 48

49 Marshall, John Political information cycles: When do voters sanction incumbent parties for high homicide rates? Working paper. Mauro, Paolo Corruption and Growth. Quarterly Journal of Economics 110(3): Melitz, Marc J The impact of trade on intra-industry reallocations and aggregate industry productivity. Econometrica 71(6): Mullainathan, Sendhil and Andrei Shleifer The Market for News. American Economic Review 95(4): Olken, Benjamin A Monitoring Corruption: Evidence from a Field Experiment in Indonesia. Journal of Political Economy 115(2): Olken, Benjamin A Do TV and Radio Destroy Social Capital? Evidence from Indonesian Villages. American Economic Journal: Applied Economics 1(4):1 33. Pande, Rohini Can informed voters enforce better governance? Experiments in lowincome democracies. Annual Review of Economics 3(1): Petrova, Maria Newspapers and parties: How advertising revenues created an independent press. American Political Science Review 105(4): Prat, Andrea and David Stromberg The Political Economy of Mass Media. In Advances in Economics and Econometrics: Volume 2, Applied Economics: Tenth World Congress. Vol. 50 Cambridge University Press p Prior, Markus 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 Newspaper coverage of political scandals. Journal of Politics 73(3): Serra, Gilles The Risk of Partyarchy and Democratic Backsliding Mexico. Taiwan Journal of Democracy 8(1): Snyder, James M. and David Strömberg Press Coverage and Political Accountability. Journal of Political Economy 118(2): Spenkuch, Jorg L. and David Toniatti Political Advertising and Election Outcomes. Working Paper. 49

50 Stanig, Piero Regulation of Speech and Media Coverage of Corruption: An Empirical Analysis of the Mexican Press. American Journal of Political Science 59(1): Strömberg, David Mass media competition, political competition, and public policy. Review of Economic Studies 71(1): Treisman, Daniel What have we learned about the causes of corruption from ten years of cross-national empirical research? Annual Review of Political Science 10: 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 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

51 A Appendix [for online publication] Contents A.1 Proofs A1 A.2 Variable definitions A4 A.3 Examples of audit reports A7 A.4 Additional audit balance tests A10 A.5 Lack of balance across media stations in the population A10 A.6 Additional results A10 A.7 Effects of an additional non-local media station A18 A.1 Proofs Proof of Proposition 1. We solve for the subgame perfect equilibrium by working backwards. Starting with sanctioning, each voter i in precinct p q chooses sanctioning optimal rule v (i, p q ). This follows straightforwardly from the[ decision to vote for I q when the ] expected utility from voting for the incumbent, i.e. δ(i, p q ) ω(l(p q ))x Iq + [1 ω(l(p q ))κ Iq ], exceeds the expected utility from voting for C q, i.e. κ Cq. Integrating over δ(i, p q ) then yields V Iq (p q ), as given in equation (6). Continuing to work backwards, we turn to voter media consumption, holding l(θ(p q )) fixed. Voters solve the following Lagrangian: L(p q ) = U ρ (p q ) + λ [ T ML (p q)+m N (p q) 0 ( ) ] h p q, θ(p q ) dθ(p q ), (A1) where the strictly increasing transformation of U to U ρ allows us to simplify the ( derivation ) without altering the solutions. We now have a continuum of first-order conditions in h p q, θ(p q ) : L(p q ) ρ 1 ( ) = 0 = ρ[1 + l(θ(p q ))]h (p q, θ(p q )) = λ, (A2) h p q, θ(p q ) which is a maximum because ρ < 1 implies that ( 2 L(p q) h p q,θ(p q) ) 2 = ρ(ρ 1)[1+l(θ(p q ))]h (p q, θ(p q )) ρ 2 < A1

52 ( ) 0. Solving for h p q, θ(p q ) yields: ) ( ) σ h (p λ q, θ(p q ) = (A3) ρ[1 + l(θ(p q ))] where recall that σ = 1. Remove λ by dividing through by the analogous first-order condition 1 ρ for θ (p q ) θ(p q ), and then rearrange to give: ) ( ) h (p ( ) q, θ(p q ) = h p q, θ 1 + l(θ σ (p q )) (p q ). (A4) 1 + l(θ(p q )) Given that the budget constraint binds, integrating over θ on both sides yields: ML (p q)+m N (p q) 0 h ( ML (p q)+m N (p q) p q, θ(p q ) )dθ = T = h(θ (p q ))[1 + l(θ (p q ))] σ [1 + l(θ(p q ))] σ dθ(p q )(A5) which rearranges to give the equilibrium consumption function for θ(p q ): 0 h (p q, θ(p q )) = = = T [1 + l(θ(p q ))] σ ML (p q)+m N (p q) [1 + l(θ(p 0 q ))] σ dθ(p q ) T [1 + l(θ(p q ))] σ ML (p q) [1 + l(θ(p 0 q ))] σ dθ(p q ) + M L (p q)+m N (p q) M L (p q) 1 σ dθ(p q ) T [1 + l(θ(p q ))] σ M L (p q )[1 + l (L)] σ + M N (p q ). (A6) where the separation in the denominator of the second line reflects the fact that l(n) = 0 for all non-local stations (which can be ordered in the index without loss of generality). The third line follows from noting that the first term of the denominator in the second line reflects the solution to an identical optimization problem for a continuum of identical local stations covering p q. Finally, we turn to the choice of l(l) for a media station local to municipality q. As noted above, each local station faces the maximization problem in equation (8). Given the continuum of local stations, the denominator of h (p q, θ(p q )) a local-content index of media stations is taken as fixed. The first-order condition in equation (9) for station θ then implicitly characterizes the optimal l. The second-order condition for a maximum requires that: σ(σ 1)T [1 + l (L)] σ 2 p q P q(θ) f(p q ) M L (p q )[1 + l (L)] σ + M N (p q ) dp q < (γ 1)l (L) γ 2. (A7) Given that σ > 1 and γ > 1, an interior solution requires requires that γ is sufficiently large, such A2

53 that the cost is sufficiently convex. If this condition does not hold, then we have a corner solution where l (L) = 1. In sum, the unique SPNE is thus characterized by: l (θ) γ 1 = σt [1 + l (θ)] σ 1 f(p q ) p q P q(θ) M L (p q )[1 + l (θ)] σ + M N (p q ) dp q, where θ is local to muncipality q ) h (p q, l (θ(p q )) = v (i, p q ) = T [1+l (L)] σ M L (p q)[1+l (L)] σ +M N (p q) T M L (p q)[1+l (L)] σ +M N (p q) I q C q (A8) if θ(p q ) = L (A9) if θ(p q ) = N [ ] if δ(i, p q ) ω(l(p q ))x Iq + [1 ω(l(p q ))κ Iq ] κ Cq (A10) otherwise Proof of Proposition 2. The first part follows from differentiating V Iq (p q ) with respect to M L (p q ): V Iq (p q ) M L (p q ) = l(p q ) φω (l(p q ))[x Iq κ Iq ] M L (p q ) < (>)0, (A11) where the negative (positive) sign follows from φ > 0, and ω (l(p q )) > 0 and equation (10) l(p proving that q) > 0, when x M L (p q) I q > (<)κ Iq. The second part of the proposition follows from further differentiating by M N (p q ) and f(p q ): 2 V Iq (p q) M L (p q) M N (p q) 2 V Iq (p q) M L (p q) f(p q) [ ] = φ[x Iq κ Iq ] ω 2 l(p q) (l(p q)) M L (p q) M N (p + l(p q) l(p q) q) ω (l(p q)) M N (p q) M L (p q) > (<)0, = φω (l(p q))[x Iq κ Iq ] 2 l(p q) M L (p q) M N (p + 2 l(p q) l(p q) q) ψi 2 q + l(p q)τi 2 > (<)0,(A12) M N (p q q) M L (p q) [ ] = φ[x Iq κ Iq ] ω 2 l(p q) (l(p q)) M L (p + q) f(p q) ω (l(p l(pq) l(p q) q)) f(p q) M L (p q) < (>)0, = φω (l(p q))[x Iq κ Iq ] 2 l(p q) M L (p + 2 l(p q) l(p q) q) f(p q) ψi 2 q + l(p q)τi 2 > (<)0. f(p q q) M L (p q) (A13) where, given x Iq > (<)κ Iq and ω (l(p q )) > 0, both conditions hold if the first term in brackets dominates the second. For both comparative statics, it is easier to see that 2 l(p q) M L (p q) M N (p q) < 0 A3

54 and 2 l(p q) l(p M L (p q) f(p q) > 0, while q) M N (p q) < 0 and l(pq) f(p q) > 0. The first term dominates when the second-order effect through ω (l(p q )) is small in magnitude relative to the first-order effect through ω (l(p q )). This occurs when the variances of the signal and prior, ψ 2 I q and τ 2 I q, are sufficiently large; intuitively, this ensures that ω(l(p q )) is not too concave in equilibrium. Due to the continuum of precincts, an increase in f(p q ) in a single precinct does not affect local news provision. However, given the discrete number of municipalities, an increase in f(p q ) across all p q P q i.e. p q P q f(p q )dp q entails 2 l(p q) M L (p q) f(p q) > 0. A.2 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 that changes in the following election, we count the vote share of the coalition of the party affiliation of the mayor at the next election. We hand-coded the mayor s affiliation for around a quarter of municipalities in the sample where the mayor s affiliation could not be immediately established. Source: IFE and State Electoral Institutes. Incumbent party 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. Change in turnout. The change in turnout in a given electoral precinct. Source: IFE and State Electoral Institutes. Unauthorized spending. Percentage of FISM funds spent in an unauthorized manner. See text for further discussion. The variables Unauthorized Q3 and Unauthorized Q4 are defined by the third and fourth quartiles of our full 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 defined by the third and fourth quartiles of our full sample. Source: ASF audit reports. PAN/PRI/PRD incumbent. Indicator coded one for the incumbent mayor represents a coalition 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 largest 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 A4

55 in the municipal, defined by the following formula: 1 P, p=1 v2 p where v p 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 the 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 the head from precinct centroid (log). The (logged) distance in meters from the municipal head to the centroid of the electoral precinct. Source: computed from INEGI data. Area (km 2 ). 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 television stations with local content emitting from within an electoral precinct s municipality and covering at least 20% of voters in a given precinct. Source: computed from IFE data. Local AM/FM/television. The total number of AM/FM/television stations with local content emitting from within an electoral precinct s municipality and covering at least 20% of voters in a given precinct. Source: computed from IFE data. AM/FM/television antennae in precinct. Indicator coded 1 for precincts containing one or more AM/FM/television antennae. Non-local media. The total number of AM, FM and television stations emitting from outside an electoral precinct s municipality or broadcasting only national content. Source: computed from A5

56 IFE data. Under-development. Standardized scale combining the following precinct-level measures of socioeconomic development: average number of children per woman, share indigenous speakers, average years of schooling, share illiterate, share no schooling, share incomplete primary schooling, share complete primary schooling, share incomplete secondary schooling, share complete secondary schooling, share higher education, share economically active, share without health insurance, share state workers without health care, average occupants per dwelling, average occupants per room, share non-dirt floor, share toilet at home, share running water, share drainage, share electricity, share fridge, share washing machine, share cell phone, share car and share computer. The scale was constructed so that higher values reflect lower levels of development. FISM target. Indicator coded 1 for precincts containing at least one locality with high or very high levels of marginalization. Local market. The share of covered voters located within the municipality from which the media station emits, averaged across all local media stations. 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 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 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. A6

57 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 refrigerator. 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. Share car. Share of the electoral precinct population in a private dwelling with a car or truck. Source: Mexican 2010 Census. Share computer. Share of the electoral precinct population in a private dwelling with a computer. Source: Mexican 2010 Census. Household assets index. Standardized scale combining the following precinct-level measures of household assets: share fridge, share washing machine, share cell phone, and share computer. These variables are defined above, and higher values of the index indicate a higher proportion of voters within precinct possessing these goods. A.3 Examples of audit reports Figures A1 and A2 provide an example of an audit report from 2008 for the municipality of Ajalpan in the state of Puebla. A7

58 Figure A1: 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 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. A8

59 Figure A2: Sample ASF audit report (page 2) Note: See Figure A1. A9