How Equalizing Access to Political

Home , Electoral precinct

LEVELING THE PLAYING FIELD: HOWEQUALIZINGACCESSTOPOLITICAL ADVERTISING HELPS LOCALLY NON-DOMINANT PARTIES IN NON-CONSOLIDATEDDEMOCRACIES ∗

HORACIO A.LARREGUY† JOHN MARSHALL‡ JAMES M.SNYDER JR.§

JUNE 2015

To combat disparities in political advertising—a key challenge in many non-consolidated democracies—many countries have regulated access to political advertising. Since voters are often uncertain about challengers, equalizing advertising opportunities may especially benefit political parties that are not locally dominant. In 2007, Mexico implemented a reform allocating radio and television advertising slots according to the number of parties standing and their vote share at the previous state and national elections. Using fine-grained signal coverage data, we exploit within-neighbor variation induced by cross-state spillovers to identify the effects of exposure to political advertising in a major consolidating democracy. We find that political advertising— particularly AM radio—increases the vote shares of the PAN and PRD, but not the previously-hegemonic PRI. Consistent with our formal model, such advertising is most effective in poorly informed and politically uncompetitive electoral precincts, and among non-dominant parties facing a locally dominant party of intermediate strength.

∗We wish to thank Scott Ashworth, Andy Baker, Taylor Boas, Ernesto Dal Bo,´ Aditya Dasgupta, Jorge Dom´ınguez, Ruben Enikolopov, Leopoldo Fergusson, Andy Hall, Brian Knight, Chappell Lawson, Devra Moehler, Jonathan Phillips, Maxim Pinkovskiy, Gilles Serra, Edoardo Teso and participants at the LACEA Annual Meeting 2014, Second Annual Formal Theory and Comparative Politics Conference 2014, APSA Annual Meeting 2014 and Harvard political economy workshop for comments on earlier drafts. Data and replication code will be made available online upon publication. All errors are our own. †Department of Government, Harvard University, [email protected]. ‡Department of Government, Harvard University, [email protected]. §Department of Government, Harvard University, [email protected].

1 1 Introduction

The effect of political advertising has received considerable attention in consolidated democracies with competitive media markets, generally ﬁnding small effects on electoral outcomes (e.g. Gerber et al. 2011; Lenz 2009; Zhao and Chaffee 1995). Conversely, in authoritarian regimes, the media is often controlled or manipulated by the state and opposition parties possess few opportunities to express their political preferences and platforms (e.g. Djankov et al. 2003; King, Pan and Roberts 2013). However, little is known about the electoral effects of political advertising by political parties in non-consolidated democracies.1 This is a particularly prescient issue since many such democracies have recently introduced reforms guaranteeing all political parties relatively equitable access to political advertising. As indicated in Table1, advertising time across the world is typically allocated according to the number of parties and/or previous electoral performance. In this article, we identify the effects of political advertising in Mexico, a large consolidating democracy that in 2007 introduced regulations guaranteeing all political parties advertising time on every AM radio, FM radio and television station in the country. Political advertising in non-consolidated democracies—states holding relatively fair multi- party elections, but lacking in terms of the quality of political participation, representation and party competition (see Schedler 1998)—is distinctive in terms of media access, partisan ties and political competition. First, new advertising opportunities have the potential to inform voters about party platforms, because although the media is less free than in consolidated democracies it is less susceptible to capture than in authoritarian regimes (Lawson and McCann 2005). Second, since partisan identities and voter-party ties are typically weaker than in developed democracies, voters are potentially more susceptible to advertising (Baker, Ames and Renno 2006; Lawson and McCann 2005; McCann and Lawson 2003). Third, voters are better informed about locally domi- 1The few existing studies include Boas(2014), Boas and Hidalgo(2011), Da Silveira and De Mello(2011), Greene(2011), Lawson(2002) and Lawson and McCann(2005). We discuss their differences and limitations below.

2 Level of democracy Partial democracies Non-consolidated democracies Consolidated democracies (Polity score below 5) (Polity score between 5 and 8) (Polity score above 8) Equal share Ethiopia, Nigeria, Russia Ecuador, Ghana, Paraguay Denmark, Italy, Uruguay for all parties Previous vote Uganda, Venezuela South Korea Chile, Germany, Spain share Combination Argentina, Brazil, Mexico Finland, Netherlands, Peru of both

Table 1: Media access regulation in 2012 (source: Administration and Cost of Elections Project) Notes: The ACE Project provide details on the media access regulations in 220 countries around around the world. We present only a subset of countries here along the main divisions in the type of access provided; other countries employ either these rules, similar rules, no restrictions, or ad hoc allocations determined by special committees. We define the level of democracy by 2012 Polity scores. nant political parties—parties whose entrenched advantages allow them to dominate local electoral politics—relative to opposition parties. This important incumbency advantage contrasts with standard models of political competition, where the policy positions and competence of the major parties are assumed to be equally well-known (e.g. Downs 1957). We analyze the impact of changing a party’s share of political advertising on vote choice in such contexts using a stylized formal framework in the spirit of Zaller(1992). 2 In our model, a party is locally dominant in two key respects. First, there is an in-built “ideological” bias toward the dominant party among all voters that could originate from non-performance based factors such as clientelistic ties, risk aversion or voter loyalty. Second, the utility a voter receives if the dominant party wins office—reflecting the challenger’s policy position, policy emphasis or competence—is known with certainty, while risk-averse voters are not certain whether the challenger will increase their utility if elected (see also Shepsle 1972). Upon reaching voters, political advertising is informative about the utility level associated with the challenger winning office. Political advertising 2We differ from Zaller(1992) in arguing that uninformed voters in non-consolidated democracies are not uninformed because they are completely disinterested in politics. Furthermore, the pervasiveness of political advertising in the Mexican case that we analyze ensures that virtually all voters are highly exposed to advertising.

3 thus allows voters to learn about the relative benefits of each party, but also decreases the uncertainty surrounding the utility that they expect to receive if they elect the challenger. Our learning model predicts that political advertising’s effect on voting behavior is greatest among uninformed voters with weak prior beliefs about the consequences of electing the non- dominant party, in locations (or elections) where political competition—and thus other local political activity—is low, and where the ideological advantage of the dominant party is not insur- mountable. Political advertising is therefore most electorally beneficial for non-dominant parties where the locally dominant party is neither very strong nor facing severe competition. However, this non-linear relationship with party strength only applies to non-dominant parties, given voters are already well-informed about locally dominant ones. In sum, our model thus suggests that equalizing political advertising has the potential to support multi-party competition, and ultimately democratic consolidation, by empowering non-dominant political parties among less informed voters and in relatively uncompetitive areas. Mexico represents an important test of political advertising’s potential to break ingrained support for locally dominant parties. Until the late 1990s, the Institutional Revolutionary Party (PRI) had held power for seven decades on the basis of a broadly populist platform alongside substantial clientelism and electoral manipulation (e.g. Cornelius 2004; Greene 2007; Magaloni 2006). Despite losing the Presidency in 2000, the PRI remained dominant in poorer and more rural parts of Mexico (Langston 2003, 2006). However, Mexico’s other main political parties—the National Action Party (PAN), which won the Presidency in 2000 and 2006, and the Party of the Democratic Revolution (PRD)—have also developed local strongholds. These are generally located in more urban and developed areas, although the PRD also has significant rural presence in the states where it split from the PRI. A major factor in Mexico’s contested 2006 presidential election—where PAN candidate Felipe Calderon´ defeated PRD candidate Andres´ Manuel Lopez´ Obrador by only 0.56% of the vote—was the “dirty war” leveled by the PAN and big business against the PRD in the media. This highlighted

4 continuing concerns regarding inequalities in media access (see Greene 2011; Hughes and Lawson 2004; Lawson 2002, 2004a; Lawson and McCann 2005). In response, Mexico’s independent Federal Electoral Institute (IFE)3—which regulates federal elections—mandated that access to radio and television advertising for political parties be more equitably distributed (Serra 2012). Beginning in 2009, media access throughout the formal campaign period around five months prior to federal elections has been allocated exclusively by the IFE. On every radio and television station in the country, each party receives a fixed share of 82 30-second daily advertising slots. These shares are assigned according to a formula taking into account the total number of parties and their previous vote share at the state and national elections. The shares vary across states, but are fixed across all media stations broadcasting from within each state. In contrast to the highly targeted strategies used prior to the reform, the huge influx of advertising opportunities—which would reach all types of voters—induced political parties to adopt a more uniform approach to political advertising. Political parties generally used their spots to highlight salient policy issues such as the economy and domestic security, but also emphasized the principles, experience and skills of the party’s candidates. Leveraging Mexico’s 2006 political advertising reform, we investigate the conditions under which providing political parties with access to media resources penetrating almost all parts of the country affects their precinct-level electoral performance. Utilizing spatial data detailing the commercial quality coverage of each AM, FM and television station in the country, our empirical strategy exploits within-neighbor differences in average exposure to political advertising from different political parties. Such differences arise when a precinct is covered by media stations broadcasting from multiple states with different distributions of political advertising. Figure1 illustrates this identification strategy using an example of two electoral precincts in the state of Campeche. These precincts differ in that precinct 362 is covered by an AM station (XEQAA-AM) located in the neighboring state of Quintana Roo—where advertising is allocated 3IFE has since become the National Electoral Institute.

5 Figure 1: Neighboring electoral precincts that differ in their commercial quality radio signal coverage from out-of-state AM radio stations

to parties according to a different distribution than in Campache—while precinct 360 is not. Where multiple neighboring precincts with different commercial quality radio coverage exist, we use matching to enhance the efﬁciency of our estimation by focusing on the most similar comparison units. Since broadcast signals weaken gradually rather than abruptly at the coverage boundary, we identify the effect of an increase in the probability of exposure to political advertising.4 To validate our design, we demonstrate that the distribution of political advertising between neighboring electoral precincts is well balanced across 29 political, economic and demographic variables. Furthermore, using the pre-reform 2006 election as a placebo test, we demonstrate that party ad- 4Given differences in exposure across the boundary are probabilistic rather than deterministic, there exist “non-compliers” on either side of the signal coverage boundary. Our estimates thus capture the intent to treat effect of advertising exposure.

6 vertising shares are not proxying for cross-state differences in media content. We focus primarily on AM radio. Its extensive signal coverage extends beyond urban areas and often crosses state borders, and thus yields a large and disproportionately poor and rural sample. As noted above, our theory predicts that political advertising should be most effective among this population. Conversely, weaker FM radio and television signals principally serve more urban areas where elections are more competitive and voters are better informed. As a robustness check we show that our findings generalize to the smaller and more urban FM radio and television samples. Pooling the 2009 and 2012 Congressional elections, we first show that greater political advertising on AM radio increases the vote shares of the PAN and PRD. Supporting our theoretical model, we demonstrate that the electoral efficacy of such PAN and PRD political advertising varies considerably by precinct. First, in less economically developed precincts—where our survey evidence indicates that voters are less politically informed—advertising is more effective at winning votes. Second, advertising is less effective in more competitive precincts, as measured by the (pre- reform) effective number of political parties. Third, political advertising is less effective in races held concurrently with the intensely contested 2012 presidential election. Finally, and consistent with our theoretical predictions, the effects of political advertising for non-dominant parties are non-linear in the vote share of the dominant party, such that advertising is least effective in both the most competitive and most locally dominated precincts. Conversely, we find no evidence that PRI political advertising affected the PRI’s vote share in any type of precinct. Although the PRI is not dominant in all parts of the country, the ineffectiveness of PRI advertising suggests that an important legacy of its time in power is that voters retain precise beliefs about its suitability for office that are not susceptible to advertising.5 Together, these results demonstrate that equalizing access to political advertising can significantly increase support for locally non-dominant parties in a major non-consolidated democracy, 5In our interviews, one prominent political consultant argued that, while the PRI possesses a large base of loyal supporters, many swing voters are skeptical about the PRI and are instead deciding between other parties.

7 and particularly among the less politically engaged voters. This suggests that broad political advertising opportunities can help to foster multi-party competition and informed political participation. As many non-consolidated democracies have recently introduced media access reforms, these results contain great promise for supporters of democracy. Our findings also contribute to the nascent empirical literature identifying the causal effects of media outside consolidated democracies. While Da Silveira and De Mello(2011) exploit differences in television advertising allocations between the first and second round of Brazilian gubernatorial elections to show that such advertising substantially increases a candidate’s vote share, we examine political advertising over an entire election campaign in a context which is not confounded by strategic behavior between the first and second round of candidate-centric runoff elections. Our finding that advertising by non-dominant parties has large effects on vote choices accords with Conroy-Krutz and Moehler(2014), who find that exposure to rival arguments on the radio increases moderate political attitudes in Ghana. Unlike Boas(2014) and Boas and Hidalgo(2011), who respectively find that local politicians in Brazil often capture local broadcasters or acquire radio licenses to support their re-election campaigns, we focus on how more equitable media access can reverse rather than entrench media advantages. The extant literature has also found that media coverage increases voter punishment of incumbent indiscretions (Ferraz and Finan 2008; Larreguy, Marshall and Snyder Jr. 2015) and voting for non-authoritarian parties (Enikolopov, Petrova and Zhuravskaya 2011). Our novel empirical design extends survey analyses pointing to the importance of media in Mexico. For example, Lawson(2004 b), Lawson and McCann(2005) and Greene(2011) use data from the Mexico Panel Study to respectively show that television biases in favor of the PAN at the 2000 and 2006 elections significantly increased the likelihood that respondents voted for the PAN’s presidential candidate. Our results complement these studies by providing causal evidence of advertising’s effects on vote outcomes to support these survey correlations. However, our main contribution is to provide a theory explaining when particular parties benefit from political ad-

8 vertising in non-consolidated democracies, which also explains variation in political advertising’s effects across Mexico’s electoral precincts. The paper proceeds as follows. Section2 provides a brief overview of politics and media in Mexico, focusing on Mexico’s election advertising reform. Section3 develops a simple model to analyze the voting implications of political advertising in a non-consolidated democracy with locally dominant parties. Section4 details our data and identiﬁcation strategy. Section5 presents our main results and robustness checks. Section6 concludes.

2 Politics and media in Mexico

Mexico is divided into 31 states (and the federal district of Mexico City), and operates a presidential form of government. National legislative elections are held every three years, with members of the Chamber of Deputies (House) and Senate elected to single three- and six-year terms respectively.6 We focus on the Chamber of Deputies, which contains 300 members elected via plurality rule from single-member districts and 200 members elected via proportional representation. Between 1929 and 2000, widespread clientelistic practices and electoral manipulation ensured that the PRI maintained a stranglehold on the Presidency and almost always retained Congressional majorities. However, Mexico has undergone rapid democratic progress in the last two decades as the PRI’s grip on power has subsided. In 2009 and 2012, three main political parties competed for political control: the left-wing PRD, the populist PRI, and the right-wing PAN. In this section, we provide an overview of political competition, before describing political advertising in Mexico and the 2007 media reforms. 6A constitutional reform in 2014 permitted re-election up to three times for deputies and once for senators elected from the 2018 election onward.

9 2.1 Political competition

Following Mexico’s revolution in 1929, the PRI retained hegemonic status up until the 1990s. The masses were co-opted into the regime, campaigning relied heavily upon distributing public resources and mobilizing voter turnout, and dissension within the party was minimized by main- taining a high political cost of exit (e.g. Cornelius 1996, 2004; Fox 1994; Greene 2007; Magaloni 2006). Nevertheless, PRI politicians frustrated by the party’s hierarchy ultimately formed the left- wing offshoot National Democratic Front. This became the PRD in 1989, and has since built a strong base in Mexico City and among poor southern states. The PRI continued to govern in the 1990s, but was forced to engage in significant constitutional reforms in order to receive the Congressional support from the right-wing PAN required to pass pressing legislation to address the economic crisis. In the more competitive electoral environ- ment, the PRI first lost control of the House in 1997 before PAN candidate Vicente Fox won the Presidency in 2000 (Greene 2007; Magaloni 2006), based on strong support in Mexico’s business- friendly northern and western states. By 2006, the PAN and PRD became the largest parties in the legislature, and the PAN narrowly retained the Presidency. Although the PRI’s vertical hierarchy dispensing patronage was destroyed and the party became regionally fractionalized (Langston 2003), its powerful regional presence remained. In almost one-third of states the PRI never lost to another party, while the opportunity for local gov- ernments to influence local elections was left relatively untouched by the IFE. These advantages, combined with decentralized mechanisms selecting higher-quality candidates popular in their local area (Langston 2006) and continued vote- and turnout-buying (Larreguy 2013; Larreguy, Marshall and Querub´ın forthcoming; Nichter and Palmer-Rubin forthcoming), helped the PRI reclaim its majority status in 2009 in coalition with the party of the teachers’ union (PNA) and the green party (PVEM), and the Presidency in 2012.

10 2.2 Political advertising and the 2007 IFE reform

Disproportionate access to political advertising in the media became a political issue as Mexico transitioned toward competitive democracy in the 1990s, despite changes in journalistic norms and competition in media markets (Lawson 2002). Although a series of constitutional reforms were approved in 1989 and the operational establishment of the IFE in 1990, which became politically independent in 1996, contributed to substantially reducing vote fraud, the PRI enjoyed privileged access to public resources and lower commercial advertising costs, as well as signiﬁcantly greater coverage and positive appraisals across media formats (Hallin 2000; Lawson 2002; Lawson and McCann 2005). However, the IFE has progressively increased its regulation and monitoring of advertising spending by political parties, and become more willing to punish violations with ﬁnes. As the PRI’s dominance in the new democratic arena subsided, the PAN capitalized by dom- inating media coverage and strategically targeting marginal voters. Lawson(2004 b) and Lawson and McCann(2005) argues that more equal access to television time was essential to Vicente Fox’s victory in the 2000 presidential election. Similarly, Greene(2011) argues that differential media access—in particular controlling 66 percent of television advertising time—was the primary reason for Felipe Calderon’s´ narrow victory in 2006. The result was highly contentious, given the PAN’s vicious media attacks against Andres´ Manuel Lopez´ Obrador and the 240 cases of irregularities highlighted by the PRD. Despite upholding all such irregularities, the IFE nevertheless declared that they did not impact the electoral outcome, despite the fact that Calderon’s´ victory margin was only 0.56% of the vote. Ultimately, the IFE overhauled political advertising regulations in 2007, following the passage of major electoral reforms following the contentious 2006 elections (Serra 2012). The new regulations, in force in federal elections since 2009, specify that neither political parties nor independent groups can buy political advertising on radio and television stations. The IFE is instead responsi- ble for allocating all advertising slots to political parties during the pre-campaign and full electoral

11 campaign that span the five-to-six months prior to federal elections. Every media station in the country provides 41 minutes of 30-second political advertising slots throughout each day (until the final week of the campaign), with the timing of individual ad slots allocated randomly. Media stations are legally bound by the distribution applied in the state from which their signal is emitted. The IFE determines the number of slots available to each political party using a clearly-defined formula, detailed in the Online Appendix, that varies across states. In states not holding concurrent state-level elections, each party is allocated a minimum advertising share split equally between all parties (30% of total advertising time) and additional time according to their vote share at the previous national legislative election (70% of total advertising time). In states holding concurrent state-level elections, however, 15 of the 41 daily minutes available for advertising are apportioned according to the number of parties that stood and the vote share at the previous state legislative election. In 2009, 11 states simultaneously held state-level elections, while 15 states held concurrent elections in 2012.7 Our hand-coded transcription of the 683 unique federal ads broadcast on radio and television during the 2012 election campaign indicates that parties principally used relatively uniform positive messages to convey their policy positions, the salience of particular issues, and emphasize their candidate’s competence.8 Of the 69% of ads that mentioned policy issues, the vast majority fo- cused on valence issues like public security and employment and economic development, although education, health, corruption and rural development also received significant attention. While ads emphasized particular issues and in some cases detailed policies to address them, parties did not generally seek to distinguish their solutions from those of other parties. Explicitly negative ads were outlawed as part of the 2007 reforms (Serra 2012), and the vast majority of ads are positive. 7The 15 in 2012, shown in Figure3, were: Campeche, Chiapas, Colima, Distrito Federal, Guanajuato, Guerrero, Jalisco, Mexico,´ Morelos, Nueva Leon,´ Queretaro,´ San Luis Potos´ı, Sonora, Tabasco, and Yucatan. Chiapas, Guerrero, Tabasco, and Yucatan´ did not hold concurrent elections in 2009. 8These ads are publicly available at http://pautas.ife.org.mx/transparencia/camp. State-level ads were not systematically collected.

12 However, 8% of ads solely attacking opposition parties; for example, some PRD ads alluded to the PRI’s history of corruption during their 70 years in power, while some PAN ads attacked the past record of the PRI’s presidential candidate Enrique Pena˜ Nieto. While notably less frequent than policy issues, the competence of individual candidates— predominantly the principles, previous experience, and specific skills of federal candidates—was mentioned in 43% of ads. Consistent with a relatively uniform advertising strategy across the country, candidate mentions were heavily skewed toward presidential candidates: presidential candidate was mentioned in 54% of ads, while the many legislative candidates were mentioned in only 44% of ads. The emphasis on the party and its presidential candidates likely reflects low name recogni- tion for federal deputies. For example, the 2009 Comparative Study of Electoral Systems survey found that only 18% of voters knew even one federal legislative candidate in their district. These relatively nationally-uniform advertising strategies differ significantly from those used up until 2006. Before the reform, parties targeted clearly defined audiences, such as women watch- ing afternoon telenovelas, and bought the corresponding air time to reach such audiences. After the reform, as one political strategist explained, parties were forced to fill many more slots catering to more diverse audiences, and instead adopted a more homogeneous strategy that involved less advertising segmentation. Beyond the widespread belief that uneven political advertising opportunities have electoral implications, surveys also suggest that Mexican voters are responsive to electoral campaigns. For example, the 2012 National Survey of Political Culture and Civil Practices reveals that 58% of voters state that campaign proposals were the most important factor in deciding how they voted. A further 10% cited candidate image as the most important factor. Our interviews also indicated that prominent media experts, including a consultant that has led high-profile election communication strategies, believed that a significant proportion of the population were susceptible to political advertising. Greene’s (2011) estimates from the 2006 Mexican Panel Study indicate that more than one third of voters changed their vote intentions over the course of the campaign.

13 3 Political advertising and vote choice in non-consolidated democ-

racies

Theories of special interest politics have typically assumed that greater campaigning and political advertising exposure translate into votes (see Persson and Tabellini 2000). In these models, campaign contributions increase the probability that any voter supports the party being campaigned for in a homogeneous way. However, there now exists considerable evidence that providing fac- tual and partisan politically-relevant information affects voters very differently (e.g. Bullock 2011; Greene 2011; Lupu 2013). Particularly in developing contexts where electorates are generally poorly politically informed, and where voters can be beholden to parties through local ties, the effects of political advertising could differ substantially (e.g. Lawson and McCann 2005; Mc- Cann and Lawson 2003). We thus ask, when is political advertising effective at winning votes in non-consolidated democracies?

3.1 Theoretical model

To clarify the role of political advertising in non-consolidated democracies, we motivate our predictions in a simple decision-theoretic model that examines vote choice where one party is locally dominant. As in many non-consolidated democracies, Mexico’s main political parties are regionally concentrated: as noted above, the PRI remained dominant in many states despite losing its stranglehold on national ofﬁces, while the PRD inherited and retained strong support in southern areas dominated by the PRI before the PRD broke off, and the PAN has often controlled urban areas and some northern states. Furthermore, between 2000 and 2012, only 14% of electoral precincts had a third party with more than 20% of the vote. Accordingly, our model focuses on the widespread case where two political parties A and B compete for voters locally, and without loss of generality we let party B be locally dominant.

14 Parties. Party B is dominant in two respects. First, every voter i receives an ideological

bias v + bi > 0 inclining them to vote for B. This represents favorability toward entrenched local politicians, including factors such as loyalty biases, clientelistic beneﬁts, and candidate-speciﬁc

attributes. While v is ﬁxed across voters, bi allows for this bias toward B to vary across voters,

where bi is distributed according to cumulative distribution function F. To further capture B’s 00 dominance we assume F > 0, such that the mass of voters with a larger bi is greater than the mass

9 with a smaller bi. Second, B’s “policy” outcome xB—which we construe broadly to include B’s policy position, emphasis on particular programs, and valence factors such as expected competence in ofﬁce—is known with certainty by all voters.10 Conversely, the outcome associated with party

A is uncertain. The prior belief of all voters is normally distributed according to N(δ,τ2), where δ is the mean and τ2 > 0 is its variance. The model thus captures the idea that a locally dominant party has both an ideological and an informational advantage. For example, where the PRI is dominant, voters often receive material benefits from the PRI, which they expect to receive if they continue voting for the PRI. However, they are uncertain of the benefits of voting for the PAN or PRD. The asymmetric treatment of the parties is similar in spirit to previous models of incumbent politicians facing challengers (Shepsle 1972), but contrasts with many models of political competition where uncertainty is assumed to be symmetric across parties (e.g. Downs 1957). Voters. Voters, which differ only in their ideological bias toward party B, must decide whether to vote for party A or party B. Each voter maximizes their expected utility, where their utility from policy outcome x is given by u(x) = −exp(−x).11 We assume that i’s ideological bias 9With the exception of one result, all apply for any uniform distribution (in addition to cases where F00 < 0 is relatively small.) 10We obtain similar results when party B’s position is known with relatively greater certainty than party A’s. 11This constant absolute risk aversion utility function is chosen because of its convenient mathe- matical properties when taking expectations over normally distributed lotteries. For simplicity, we set the coefficient of risk aversion to unity.

15 substitutes for policy beneﬁts. Normalizing xB = 0, voter i therefore chooses party A over party B

if Eu(xA) ≥ u(v + bi). However, voters also learn about xA from political advertising. Advertising. Voters update their beliefs in response to political advertising according to their prior beliefs and the persuasiveness of the information they receive. Each voter receives n sig-

12 nals, each from an advertisement from party A. Each signal x j is independently drawn from the

2 distribution N(xA,σ ), where xA is party’s A true (but unknown) policy and its known variance is σ 2 > 0;13 when σ 2 is small, A’s signals are more informative. The model assumes that σ 2 > nτ2, which ensures that the signal does not overwhelm the prior.14

−1 n The mean signal observed by a voter is x¯ = n ∑ j=1 x j, and voters form expectations using these signals to update their posterior belief about the benefits of A winning office. Moreover, akin to Peterson’s (2009) finding that uncertainty about the policies and traits of candidates in the U.S. decreases over the course of an electoral campaign, advertising reduces voter uncertainty about the benefits of the non-dominant party. For the model to be interesting, we focus on the behavior of voters for whom A’s policy outcome is sufficiently beneficial relative to B’s, and thus on those voters for whom information could cause them to switch away from the dominant party.15

σ 2 Consequently, we assume that x¯A > 0 and x¯− δ > 2n . In words, A’s policy outcome is sufficiently better for voters than B’s, while voters’ prior beliefs are centered on an expectation below A’s true policy outcome. Applying Bayes’ rule, a voter’s posterior distribution over the policy outcome if party A wins 12Since B’s position is known with certainty, we ignore any signals sent by B. 13The posterior distribution depends upon the prior parameters δ and τ2 in similar ways in the more complex case where the variance is also unknown. However, this approach loses analytical tractability because the posterior distribution is no longer simply normally distributed. 14This condition prevents advertising from overwhelming the prior to a sufficient extent that the marginal effect of additional advertising is essentially uninformative. Furthermore, as the signal dominates the prior, party B no longer dominates in terms of information. 15Although some voters may have an overly-optimistic prior about non-dominant party A, these are likely to be sufficiently few in number that we can focus on voters positively updating about A in our empirical analysis. Voters that are already biased toward the dominant party will not change their voting behavior.

16 is:16

δ nx¯ −1! 2 + 2 1 n N τ σ , + . (1) 1 + n τ2 σ 2 τ2 σ 2

Consequently, a voter’s expected utility from party A winning ofﬁce is given by:

" δ + nx¯ −1!# τ2 σ 2 1 1 n Eu(xA) = −exp − − + 1 + n 2 τ2 σ 2 τ2 σ 2 δσ 2 + nx¯τ2 1 τ2σ 2 = −exp − − , (2) σ 2 + nτ2 2 σ 2 + nτ2

where the first term is the voter’s expectation of A’s policy outcome, and the negative second term reflects their disutility from risking the election of a candidate whose policy outcomes are uncertain. Morgenstern and Zechmeister(2001) have shown that risk-aversion was a significant factor in explaining continuing support for the PRI at the 2000 presidential elections. Defining

2 2 2 2 R ≡ δσ +nx¯τ − 1 τ σ , voter i thus chooses to vote for party A over party B only when R > v+b . σ 2+nτ2 2 σ 2+nτ2 i Equation (2) highlights several important implications of political advertising. First, voters are more likely to believe that party A’s policy outcome will beneﬁt them as the number of ads, n, increases. Second, as in Zaller(1992), voters with strong priors—smaller τ2—are less responsive to an additional ad from A. Third, the effect of political advertising on the belief that A will be beneﬁcial increases with the precision of the signal, or as σ 2 decreases. Combining voter beliefs with the decision to vote for party A or party B yields our main result determining when a voter supports a non-dominant party.

Proposition 1. The proportion of votes for party A, the non-dominant party, is VA ≡ F(R−v). The following comparative statics hold:

∂VA (a) The probability of voting for A, VA, is increasing in n (i.e. ∂n > 0). 16We obtain this posterior distribution by conjugating the prior with the n independent and iden- tically distributed signals.

17 2 ∂VA 2 2 ∂ VA (b) The effect of n on VA, ∂n , is decreasing in v and σ , and increasing in τ (i.e. ∂n∂v < 0, 2 2 ∂ VA < 0 and ∂ VA > 0). ∂n∂σ 2 ∂n∂τ2

Proof : see Online Appendix.

3.2 Empirical implications for Mexico

We now derive specific empirical predictions by aggregating voters at the electoral precinct level, which our empirical analysis will focus on. The most obvious prediction of the model—from Proposition1(a)—is that greater political advertising ( n) by a non-dominant party increases the probability that any individual votes for that party. Supporting the key assumption driving this prediction, the Online Appendix examines three surveys conducted after Mexico’s 2006, 2009 and 2012 federal elections and shows that voters are 5 percentage points more likely to know the PAN presidential candidate in precincts where the PAN is the largest party, and 2 percentage points more likely to know the PRD candidate in precincts where the PRD is the largest party.17 Although voters are more likely to know the PRI candidate when the PRI is the largest party, the lack of a significant difference could reflect decades of PRI rule. Importantly, since no party dominates all parts of the country, political advertising has the potential to help all political parties. We thus hypothesize that:

H1. An increase in a party’s political advertising increases its vote share.

The model also identiﬁes the types of precincts where a party’s advertising is most effective.

Proposition1(b) predicts that less well-informed voters—those with a weak prior, or large τ2—are the most responsive to persuasive new information provided by political parties. Greene(2011) and Lawson and McCann(2005) argue that a legacy of Mexico’s recent competitive authoritarian past is low levels of political knowledge. This is likely to be particularly true of poor and rural voters, which are easier to measure empirically. Conﬁrming this correlation, the Online Appendix 17McCann and Lawson(2006) ﬁnd similar results before 2006.

18 shows that our measure of local development (deﬁned below) is positively and signiﬁcantly correlated with the respective probabilities that respondents know of, and have an opinion on, the PAN, PRD and PRI presidential candidates, as well as an index of political knowledge probing a respondent’s knowledge of political institutions and the current political landscape. Consequently, we hypothesize that Mexico’s impoverished voters, who are the least well informed, are most likely to internalize political messages, and therefore most likely to change their vote when confronted with the overwhelming quantity of political advertising mandated by Mexico’s electoral reform.18

H2. Political advertising is most effective at winning votes in less developed parts of the country.

However, political advertising is only one tool deployed by political parties. All of the main political parties in Mexico also engage in signiﬁcant local-level clientelism (e.g. Cornelius 2004; Larreguy 2013; Magaloni 2006) and vote buying (e.g. Larreguy, Marshall and Querub´ın forthcoming; Nichter and Palmer-Rubin forthcoming), and these are especially concerted in swing districts (Diaz-Cayeros, Estevez´ and Magaloni forthcoming).19 The PRI has traditionally used these tactics to compensate unfavorable coverage in media (Lawson 2004a). In competitive localities where multiple political parties use a variety of tactics to win votes, the effect of political advertising— which is ﬁxed in quantity by law—may be crowded out by other activities.20 Consistent with this possibility, the Online Appendix shows that voters in more competitive districts are more knowledgeable about politics and their local candidates. In terms of the model, the presence of alternative

modes of persuasion may either increase the precision of voter’s prior (i.e. reduce τ2) or reduce the precision of advertising’s signal (i.e. increase σ 2). Proposition1(b) therefore implies that: 18Although such voters are least likely to acquire information (Zaller 1992), the reform we analyze ensured that political advertising hit all types of radio and television audiences. 19Swing voter models similarly predict that parties will allocate resources to competitive districts (Lindbeck and Weibull 1987). 20Theoretically, political advertising could complement other activities. However, it is not clear why complementarities with one party’s advertising should overcome both advertising and non- advertising counterveiling forces emanating from other political parties. Furthermore, strategies like vote buying are unlikely to serve as complements since they are designed to overcome political preferences. Ultimately, this is an empirical question.

19 H3. Political advertising is most effective at winning votes in less politically competitive parts of the country.

Similarly, while local political competition may differentially crowd out the effects of political advertising across electoral precincts, some elections are more salient than others.21 In Mexico, the president is elected to a six year term simultaneous to every other federal legislative election. As in many other developing democracies, presidential elections in Mexico are particularly hard fought, and political parties dedicate many more resources to their various mobilization and persuasion strategies. However, under the IFE’s rules, the quantity of political advertising is constant across national elections. We thus also hypothesize that τ2 is larger and σ 2 is smaller in mid-term elections, and thus:

H4. Political advertising is most effective at mid-term elections.

Finally, and bringing together the key insights of our theoretical model, the relationship between advertising and local dominance is not necessarily linear. When there is little bias toward the locally dominant party there are fewer votes for the challenger to win and the election is likely to be more competitive (decreasing τ2 and increasing σ 2). At interim levels of local dominance, voters are more susceptible to political advertising because they possess weaker priors (larger τ2) and advertising is not crowded out as much by political competition (smaller σ 2). However, propo- sition1(b) shows that advertising ultimately becomes less effective once the ideological bias ( v) toward the dominant party becomes sufficiently large that no amount of advertising can convince voters to abandon the locally dominant party. Together, these insights imply that the effects of a non-dominant party’s advertising are non-linear in the level of local dominance: where a dominant party is relatively strong, but not completely commanding, we expect advertising to be most effective. Importantly, since the model assumes that the policies of locally dominant parties are well known, we do not expect to find any effect of advertising locally dominant parties. 21In the U.S., Anzia(2013) finds that special interest groups are particularly important in mid- term elections where turnout is relatively low.

20 H5. Political advertising by non-locally dominant parties is most effective at intermediate levels of local dominance, while political advertising by locally dominant parties is ineffective.

4 Research design

To identify the causal effects of political advertising on party vote share, we compare neighboring electoral precincts which receive differential exposure to political advertising by virtue of lying at the limit of the commercial quality signal coverage provided by an out-of-state media station. Although we ﬁnd similar results for FM radio and television, this paper principally focuses on AM radio advertising. AM radio typically reaches more rural and less well informed voters than FM radio and television, and thus has the greatest potential for diminishing the strength of locally dominant parties.

4.1 Data

We collected data from various sources to produce a dataset combining political advertising shares for each political party, local economic and demographic characteristics, and national election vote share outcomes for each of Mexico’s c.67,000 electoral precincts. Electoral precincts make up the legislative districts (within states) which elect national representatives, the smallest area for which media coverage and electoral data could be matched. Given that political advertising and signal coverage data at the media outlet-level were ﬁrst collected after Mexico’s media reforms, we examine the 2009 and 2012 elections. In 2012, a presidential election was held concurrently. We now describe our main variables. More detailed deﬁnitions, sources and summary statistics are provided in the Online Appendix.

21 4.1.1 Dependent variable: vote share

Our outcome of interest is a party’s vote share. The IFE provides polling station-level election outcomes for the 2000-2012 federal legislative elections. For the 2009 and 2012 legislative elections following Mexico’s political advertising reform, we aggregate polling station party vote shares up to the electoral precinct level to match the level of our other data.22

4.1.2 Independent variable: party political advertising share

In their new regulatory role, the IFE collected data from every media station in the country after the 2007 media reforms.23 The data includes the location of the signal’s antennae, which allows us to identify the advertising distribution mandated in the associated state, and the coverage area for each station. AM radio coverage was typically calculated using the Kirke (or equivalent distance) method, which adjusts for local terrain disrupting ground conductivity.24 The IFE deﬁnes the boundary of the coverage area using a 60 dBµ threshold for signal strength.25 This is the threshold commonly used to determine a radio station’s audience and sell advertising space commercially.26 Inside a station’s coverage area the signal is of high quality, ensuring that interior precincts have good access to the station’s broadcasts. Precincts outside the coverage area experience sharply decreasing coverage quality as the distance from the boundary increases. We exclude the Federal District given the small size of its electoral precincts reduces the validity of this comparison, while our identiﬁcation strategy (below) ensures that our sample is disproportionately rural. The number 22The correlation between PAN, PRI and PRD legislative and presidential vote shares always exceeds 0.91. Polling stations within a precinct are often located in the same or an adjacent building (Larreguy, Marshall and Querub´ın forthcoming). 23This data was obtained from IFE using a freedom of information request. 24Stromberg¨ (2004) shows that ground conductivity is a good predictor for the number of households with radios in the U.S. in the 1930s. 25The coverage for FM radio and television stations was calculated similarly. 26In the U.S. it “is recognized as the area in which a reliable signal can be received using an ordinary radio receiver and antenna” (NTIA link).

22 Figure 2: Commercial quality signal coverage of all AM radio stations (source: IFE)

of media stations has not recently changed.27 Figure2 maps the signal coverage of all AM radio stations. There are no instances of a radio station emitting the same signal from different states, while 87% of electoral precincts are covered by at least one AM radio station. Since the uncovered precincts may be systematically different, our empirical analysis focuses on comparing differences in party political advertising among precincts receiving AM coverage from at least one radio station.28 The Online Appendix shows that FM radio and television stations are more numerous, but emit weaker signals that cover signiﬁcantly fewer and more urban precincts than AM stations. 27Although we were unable to obtain data for 2012, the number of radio and television stations did not change in any year between 2003 and 2010. 28Balance across covariates declines when comparing precincts with and without AM coverage.

23 Our principal independent variable is the share of AM political advertising received from a given party in a given electoral precinct. Speciﬁcally, we calculate the average share of political advertising for party i across all AM media stations g covering precinct j at election t:

1 advertisingsharei jt = ∑ mediashareigt, (3) |G j| g∈G j where G j ≡ {g : g covers j} is the set of AM stations covering precinct j. Since some precincts may be covered by AM stations from multiple states, this average advertising distribution varies with the share of slots assigned to each party in the state from which each stations emits. Across the country, the range of the PAN, PRD and PRI advertising shares are respectively 21-35%, 9-20% and 19-35%. Since we cannot accurately measure media station audiences, and the decision to listen to political advertisements is likely to be correlated with other relevant political variables , we rely on a measure of exposure rather than consumption (see also Huber and Arceneaux 2007).29 Nev- ertheless, previous studies strongly suggest that the volume and breadth of media access translates into the consumption of political information (Barabas and Jerit 2009; Prior 2007). Furthermore, Larreguy, Marshall and Snyder Jr.(2015) show that radio coverage increases the likelihood that voters own a radio.

4.1.3 Precinct-level variables

We also collected precinct-level data to test the heterogeneous effects predicted by our model. To test H2, we measure local socioeconomic development, as a proxy for voter knowledge of politics, using ﬁve variables: 2006 electorate density; the proportion of the precinct population that has 29Ideally, we could also identify the electoral effect of receiving or consuming an additional media station using instrumental variable techniques. However, in the absence of detailed individual- level variables measuring which radio or television stations voters have access to or actually con- sume, we cannot estimate an appropriate ﬁrst stage. Furthermore, the exclusion restriction is unlikely to hold since voters are likely to discuss the news that they receive with their friends and family.

24 non-dirt floors, running electricity, running water, a toilet, and drainage; the employment rate; the literate proportion of the population aged above 15; and the share of the population aged above 15 that completed primary school. Given the strong correlation between these theoretically-related variables, we combine them by taking the first factor from a principal components factor analysis.30 We refer to this standardized variable as “basic development.” To test H3, we measure political competition by the effective number of political parties (ENPV) at the precinct level. One effective party represents complete local dominance by a single party, while larger values represent greater political competition.31 To ensure that competition is not affected by political advertising during or following the 2009 or 2012 elections, we calculated ENPV using the vote share of every party that stood in each precinct in the 2006 legislative election.32 To test H5, we define the locally dominant party as the party that received the most votes in the precinct in the 2006 election. We use linear and quadratic terms in the locally dominant party’s vote share to capture the non-linearity in H5, and thus allow the effect of political advertising to differ depending upon whether a party is itself locally dominant and the extent of its dominance. Additional electoral data and the 2010 Census facilitate the matching component of our identification strategy. The Census provides precinct-level information about education, economic activity and poverty indicators; descriptions of these variables are provided in the Online Appendix. We use 29 variables for matching (see Table3). 30In our main sample (see below), the first factor has an eigenvalue of 1.78, while the second factor’s eigenvalue is only 0.53. This indicates that a single factor captures most of the variation. 31Although most elections are two-party races, smaller parties remain sufficiently large that they should not be ignored. 32In our main sample, the correlation between 2006 ENPV and (endogenous) contemporaneous ENPV is 0.50.

25 4.2 Identiﬁcation strategy

To address the important concern that electoral precincts receiving different political advertising distributions also systematically differ in other electorally-relevant respects (e.g. arising from geographic proximity to economically and politically different states), our identification strategy exploits within-neighbor variation in political advertising shares. In particular, we compare neighboring electoral precincts that receive a different distribution of political advertising because they receive a different mix of radio signals from stations based inside and outside the state. Our design thus relies on differences in advertising shares that arise from cross-state “spillovers” in AM radio coverage.33 More specifically, we focus on “treated” precincts that differ from at least one neighboring “control” precinct in terms of the distribution of political advertising that they receive from AM radio stations. To increase the efficiency of our estimation, we compare the most similar neighboring precincts by selecting as control units the (up to) three neighboring precincts that are most similar to each treated unit in terms of Mahalanobis distance over 29 demographic, economic and political matching variables. We use both the best match and all possible matches as robustness checks below. We also restrict attention to matches located within 1 kilometer (km) of a coverage boundary to further increase the comparability of media access. Since broadcast signal strength decays gradually with distance, the commercial coverage boundary is not a sharp difference between receiving or not receiving a station’s signal. Rather, some households beyond the boundary can nonetheless receive signals from the media outlet (perhaps not regularly, or depending on time of day), while signal quality may be erratic for some households 33See also U.S. studies exploiting differences in media market boundaries (e.g. Ansolabehere, Snowberg and Snyder 2006; Huber and Arceneaux 2007; Snyder and Stromberg¨ 2010). Note, that our design differs from geographic regression discontinuity designs in two key respects. First, differences in the number of commercial quality local media signals between neighbors are non- binary because neighbors can differ by more than one media station. Second, the multidimension- ality of these differences determining the distribution of political advertising do not clearly define a continuous running variable.

26 Figure 3: AM radio neighboring precinct sample used in our analysis

inside the boundary. As noted above, we are unable to measure media consumption. Consequently, by identifying the effect of an increase in the probability of exposure to AM radio signals, we thus estimate the “intent to treat” effect of political advertising. Pooling across the 2009 and 2012 elections, our design yielded a total of 31,969 matched groups containing a single “treated” unit and up to three neighboring “control” units. This produced 108,871 observations in total, while Figure3 shades in grey the 16,236 unique electoral precincts included in our sample. Unsurprisingly, our sample is clustered around the borders of states holding concurrent state-level elections. Furthermore, the summary statistics in Table2 show that the electoral precincts in our sample are more rural and less economically developed than the national average. Given that parts of Mexico—particularly urban areas—had already been exposed

27 Table 2: Comparison of AM radio neighboring precinct sample and population summary statistics

AM radio neighboring precinct sample National population (2009 and 2012) Obs. Mean Std. dev. Min. Max. Obs. Mean Std. dev. Min. Max. Dependent variables PAN vote share 108,871 0.271 0.156 0 0.959 131,346 0.264 0.144 0 0.972 PRD vote share 108,871 0.162 0.142 0 0.953 131,346 0.164 0.147 0 1 PRI vote share 108,871 0.365 0.137 0 0.967 131,346 0.362 0.142 0 1

Treatment variables PAN advertising share 108,871 0.277 0.024 0.211 0.347 131,369 0.239 0.096 0 0.347 PRD advertising share 108,871 0.148 0.033 0.093 0.204 131,369 0.127 0.059 0 0.204 PRI advertising share 108,871 0.262 0.059 0.189 0.354 131,369 0.228 0.107 0 0.354

28 Covariates Largest vote share (2006) 108,871 0.472 0.106 0.126 0.994 128,406 0.482 0.103 0.002 0.995 PAN largest (2006) 108,871 0.421 0.494 0 1 131,369 0.396 0.489 0 1 PRD largest (2006) 108,871 0.275 0.446 0 1 131,369 0.310 0.463 0 1 PRI largest (2006) 108,871 0.305 0.460 0 1 131,369 0.294 0.456 0 1 Electorate density (2006, log) 108,871 4.600 2.291 0.026 10.704 126,452 6.480 2.840 0.007 11.792 Share basic necessities 108,871 0.675 0.295 0 1 120,142 0.747 0.302 0 1 Share illiterate above 15 108,871 0.111 0.082 0 0.894 120,136 0.086 0.090 0 1 Share employed 108,871 0.947 0.054 0.017 1 120,134 0.954 0.045 0.017 1 Share primary complete 108,871 0.472 0.093 0.046 0.917 120,136 0.432 0.114 0 1 Basic development (factor) 108,871 0.000 1.000 -4.680 2.101 ENPV (2006) 108,871 2.741 0.523 1.007 4.600 128,406 2.729 0.493 1 4.600 2012 presidential election 108,871 0.531 0.499 0 1 131,369 0.506 0.500 0 1

Notes: Summary statistics are for the full AM radio neighboring precinct sample (allowing for up to three matches within 1 km of a coverage boundary) and full 2009 and 2012 national population of electoral precincts. to the PAN and PRD (as well as the PRI), we expect to find larger effects for political advertising in the AM matched sample. The key identifying assumption is that neighboring precincts differ only in their AM radio political advertising shares. There are good reasons to believe this assumption. First, by restricting attention to within-neighbor comparisons, variation in access to radio signals is in large part determined by fixed signal impediments such as terrain and salt water that inhibit or enhance ground-level electrical conductivity (see Stromberg¨ 2004). Second, given that out-of-state AM radio stations are unlikely to specifically target audiences at the extremities of their coverage area, both because such audiences represent a small share of their potential listenership and because they lack the technology to precisely differentiate precincts,34 the direction and reach of cross- state spillovers are unlikely to be correlated with precinct characteristics. Third, if voters choose where to live according to media availability, they would likely choose a location much closer to the antennae, rather than near the commercial quality coverage boundary where high-quality signal coverage cannot be guaranteed. Empirically, we demonstrate the plausibility of our identification strategy below by showing that advertising shares are balanced across a variety of political, demographic and economic variables.

4.2.1 Estimation

Provided that differences in political advertising arising from cross-state spillovers in AM signals occur effectively randomly, we can estimate the average effect of exposure to political advertising from each political party using the following OLS regression equation:

votesharei jt = β advertisingsharei jt + µmt + εi jt, (4) 34The power output in watts for the AM radio stations in our sample are almost exclusively round thousands and divisible by 5.

29 where votesharei jt is the vote share of party i ∈ {PAN,PRI,PRD} in precinct j (of district d) at election t ∈ {2009,2012}. We include matched group-year fixed effects, µmt, to ensure that our estimates are only identified out of differences within matched neighboring precincts at a given election.35 In all specifications, we weight by the inverse of the number of precincts per matched group to ensure that each matched group is weighted equally.36 Standard errors are clustered by state throughout.

To estimate the heterogeneous effects of media conditional on Xi jt, we add interaction terms and estimate:

votesharei jt = β advertisingsharei jt + Xi jt γ + (advertisingsharei jt × Xi jt )δ + µmt + εi jt. (5)

We test H2 by interacting advertising share with basic development, H3 by interacting the advertising share with the ENPV at the 2006 legislative election, H4 using an interaction for the 2012 election, H5 by interacting a party’s political advertising share with quadratic terms for the largest local party’s vote share and an indicator for whether party i is locally dominant.

4.2.2 Balance on political, demographic and economic covariates

The key concern for designs exploiting differences between neighboring locations is sorting. Al- though the discussion above argues that neither strategic sorting (on the part of either voters or radio station owners) nor incidental sorting are plausible in this case, Table3 assesses balance across the PAN, PRD and PRI political advertising shares by separately estimating equation (4) with each of our 29 matching variables as outcomes. The demographic and political variables in the ﬁrst seven columns were measured in 2006 (to avoid post-treatment bias), while the remaining 35Since differences in political advertising shares differ by matched groups, match-year ﬁxed effects ensure that our results are not being driven by neighbor differences correlated with differences in political advertising shares. 36The results are robust to further weighting by the number of registered voters per precinct (see Online Appendix).

30 Census variables—which represent long-run development outcomes—were measured in 2010. The results show that a party’s political advertising share on AM radio is rarely significantly correlated with potentially confounding demographic, political or economic factors. The few statistically significant correlations are consistent with chance: only 4 of 87 regressions yielded coefficients significant at the 10% level. Of particular importance, we find no significant correlation between advertising shares and precinct vote shares before the 2007 reform. The results therefore support the validity of our identification strategy claiming to generate exogenous variation in political advertising shares.37 Nevertheless, we ensure below that our results are robust to controlling for our these variables, as well as prior vote trends, and their interaction with political advertising.

5 Results

We now test the implications of our theoretical model. The results demonstrate that political advertising is effective at winning votes for the PAN and PRD. Consistent with the model, advertising’s impact is greatest in less developed and less competitive precincts, and, for locally non-dominant parties, it increases non-linearly with the vote share of the locally dominant party. Conversely, we ﬁnd no evidence that PRI advertising is effective.

5.1 Average effects of AM radio political advertising on party vote share

Table4 presents the average and heterogeneous effects of political advertising on AM radio. Re- spectively, the dependent variable in panels A, B and C are the precinct-level vote shares of the PAN, PRD and PRI. As noted above, all estimates of equations (4) and (5) include up to three comparable neighbors all within 1 km of a AM coverage boundary. To save space, lower-order interactions terms are omitted from the tables. 37Advertising shares are also well-balanced when we examine all possible neighboring precincts (and thus omit the matching component of our design).

31 Table 3: Balance checks—partial correlation between AM radio political advertising and 29 covariates

Registered Population Turnout PAN PRD PRI ENPV Share Share Share electorate density 2006 vote vote vote 2006 economically employed medical 2006 (log) 2006 2006 2006 2006 active insurance PAN advertising share 1007.225 -5.713 -0.065 0.593 -0.379 -0.229 -0.646 -0.026 0.065 0.731 (2111.497) (6.662) (0.248) (0.414) (0.395) (0.364) (1.300) (0.206) (0.151) (0.570) PRD advertising share 1277.448 -2.332 -0.004 -0.497 0.425 0.068 -0.050 -0.002 0.012 -0.425 (2133.898) (6.135) (0.177) (0.324) (0.392) (0.372) (1.371) (0.163) (0.111) (0.577) PRI advertising share -787.190 4.048 0.180 0.294 -0.211 -0.100 0.975 0.027 -0.178 0.037 (2591.932) (6.043) (0.197) (0.270) (0.348) (0.412) (1.504) (0.173) (0.139) (0.326) Primary Middle Secondary Share Share Share Share Share Share Share school school school 6-17 in 18-24 in illiterate no school primary primary secondary attendance attendance attendance school school above 15 incomplete complete incomplete complete PAN advertising share -0.062 -0.355 0.230 -0.069 0.060 -0.001 -0.028 -0.095 -0.166 -0.044 (0.183) (0.363) (0.484) (0.291) (0.217) (0.173) (0.199) (0.200) (0.153) (0.138) 32 PRD advertising share 0.139 0.262 -0.558 0.027 -0.284 0.033 0.077 0.273 0.254* 0.071 (0.172) (0.318) (0.412) (0.244) (0.224) (0.143) (0.183) (0.201) (0.149) (0.158) PRI advertising share -0.023 -0.005 0.834** 0.200 0.336 -0.059 -0.171 -0.236 -0.229 0.053 (0.199) (0.329) (0.331) (0.193) (0.234) (0.139) (0.184) (0.257) (0.222) (0.159) Share Share Share Share Share Share Share Basic Share secondary with non-dirt electricity piped with with basic with incomplete house ﬂoor water toilet drainage necessities internet PAN advertising share -0.039 -0.127 -0.020 -0.278 -0.001 -0.197 0.004 -0.161 0.177 (0.119) (0.381) (0.371) (0.577) (0.891) (0.403) (0.541) (0.767) (0.175) PRD advertising share 0.077 -0.100 -0.069 0.152 -0.732 -0.276 -0.377 -0.715 -0.388** (0.131) (0.259) (0.452) (0.584) (0.891) (0.390) (0.591) (0.761) (0.154) PRI advertising share -0.009 0.006 0.037 0.180 0.789 0.542 0.691 1.022 0.375** (0.138) (0.301) (0.479) (0.656) (0.736) (0.533) (0.684) (0.674) (0.158)

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 1.024*** 0.872** 4.368*** 1.564*** -1.170 3.782** (0.364) (0.417) (0.930) (0.473) (0.884) (1.649) × Basic development (factor) -0.237** -0.156 (0.108) (0.092) × ENPV (2006) -1.370*** -0.926*** (0.370) (0.311) × 2012 presidential election -0.949** -0.853** (0.456) (0.376) × Largest vote share 8.953** 3.507 (3.309) (3.411) × Largest vote share (squared) -9.220*** -7.379** (3.066) (2.929) × Largest vote share × PAN largest -9.772* -8.802* (5.095) (5.094) × Largest vote share (squared) × PAN largest 10.660** 9.574* (4.833) (4.827) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.486 0.382 1.347** 1.364*** -0.778 3.246*** (0.460) (0.506) (0.576) (0.417) (0.462) (0.873) × Basic development (factor) -0.144** -0.108** (0.059) (0.052) × ENPV (2006) -0.354** -0.676*** (0.129) (0.190) × 2012 presidential election -1.302* -1.178** (0.671) (0.542) × Largest vote share 4.597*** 0.810 (1.070) (1.060) × Largest vote share (squared) -4.205*** -3.160*** (0.987) (0.879) × Largest vote share × PRD largest -2.327 -2.265 (4.397) (4.401) × Largest vote share (squared) × PRD largest 2.175 2.153 (4.585) (4.608) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.088 -0.071 -0.068 -0.177 -0.121 -0.187 (0.336) (0.338) (0.336) (0.672) (0.381) (0.714) × Basic development (factor) -0.023 -0.050 (0.040) (0.038) × ENPV (2006) -0.007 0.007 (0.050) (0.067) × 2012 presidential election 0.163 0.105 (0.723) (0.727) × Largest vote share 0.115 -0.021 (1.069) (1.099) × Largest vote share (squared) 0.326 0.493 (1.099) (1.115) × Largest vote share × PRI largest 0.594 0.779 (1.833) (1.826) × Largest vote share (squared) × PRI largest -1.466 -1.663 (1.908) (1.893)

Notes: All specifications include match-year fixed effects, up to three matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. Lower order interaction terms are omitted. All specifications include 108,871 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. 33 Column (1) reports the average effect of political advertising, showing significant variation by political party across panels. In panel A, we find that the share of PAN political advertising substantially increases the PAN’s vote share. A percentage point increase in their advertising share increases their vote share by a percentage point. A standard deviation increase in political advertising thus corresponds to a 2.5 percentage point increase in the PAN’s vote share, or a 10% increase in their precinct vote share. For the PAN, we therefore find significant support for H1— that political advertising is effective on average. In panel B, PRD political advertising relatively substantially increases the party’s vote share, but is not precisely estimated. The positive coefficient indicates that a percentage point increase in advertising translates into a half percentage point increase in vote share, while a standard deviation increase in advertising also corresponds to a 10% increase in their vote share. The relative impre- cision reflects the ineffectiveness of PRD ads in 2012: column (4) demonstrates that the effect of PRD ads in 2009 was statistically significant and similar in magnitude to the average effect of PAN ads. There is no evidence in panel C, however, that PRI political advertising affects their vote share. This finding is consistent with the likelihood that voters have relatively strong priors about the PRI after seven decades in power, and are thus relatively unaffected by PRI advertising. Our interviews with political strategists also suggested that voter opinions of the PRI are highly polarized. During Chile’s 1988 plebiscite, Boas(forthcoming) similarly finds that opposition advertising was effective while pro-Pinochet advertising was not.

5.2 Heterogeneous effects of AM radio political advertising on party vote

We now turn to our interactive speciﬁcations in columns (2)-(6) to examine hypotheses H2-H5. We ﬁnd clear evidence that the characteristics of an electoral precinct and the type of election

34 affect the vote-winning efficacy of political advertising. Column (6) includes all heterogeneous effects simultaneously to demonstrate that the individual interaction estimates are not driven by correlations among our interaction variables. Column (2) shows that, consistent with H2, PAN and PRD political advertising is significantly more effective at winning votes in the less developed electoral precincts, which our survey data confirmed to be populated by the less informed voters.38 Our estimates suggest that a standard deviation increase in the development factor variable reduces the increase in vote share due to every percentage point increase in political advertising by 0.24 percentage points for the PAN and 0.14 percentage points for the PRD. In the least developed precincts (with a standardized development score of -4.7), the effects of political advertising are substantial, increasing the PAN and PRD vote share by 2.0 and 1.1 percentage points respectively for each additional percentage point of advertising share. The PRI’s political advertising is equally ineffective across more and less developed electoral precincts. Supporting H3, the results in columns (3) and (4) show that political advertising’s weakest effects are in competitive areas and elections, where parties are engaging in other political activities. First, the large and statistically significant interaction with the ENPV shows that PAN and PRD political advertising is most effective in precincts where a small number of parties garnered most of the votes in 2006. The differential is particularly large for PAN advertising, where a percentage point increase in their advertising share increases their vote share by 3 percentage points in the least competitive precinct in our sample and only reaches zero in the 20% of precincts with at least 3.2 effective parties. The effect of PRD advertising on the PRD’s vote share, which is 0.2 percentage points lower after a standard deviation increase in political competition, declines four times slower with ENPV but similarly hits zero in the 2% of precincts with at least 3.7 effective parties. These effects are robust to the simultaneous inclusion of our other interactions with political advertising 38These results are consistent with Greene’s (2011) survey results from the pre-reform 2006 election, and Da Silveira and De Mello(2011) in Brazil.

35 in column (6), where the PAN and PRD coefficients converge to similar magnitudes. Again, we find no difference in the effectiveness of PRI advertising in panel C. Second, providing support for H4, column (4) shows that AM radio advertising was less effective during the 2012 presidential election than the 2009 legislative election. Consistent with our crowding out argument, the impact of PAN advertising was significantly lower in 2012, although it continued to significantly increase their vote share on average.39 PRD ads had a large positive effect in 2009, on a par with PAN advertising. However, the interaction between political advertising and the election year indicates that PRD advertising, on average, was ineffective in 2012. The estimates in panel C show that in neither election was the effect of PRI advertising positive. Although the 2009 and 2012 elections of course differed in other important respects, the difference across elections provides further suggestive evidence consistent with our theory. Finally, the estimates in column (5) show that political advertising is most effective for non- dominant parties and where the locally dominant party has intermediate strength. For both the PAN and PRD, the coefficients in the second and third rows show that the marginal effect of political advertising is initially increasing in the vote share of the locally dominant party, but starts to decrease once that dominant party’s vote share reaches around 50% of the vote. The final two coefficients in these specifications show that the marginal effect, for any level of the locally dominant party’s vote share, is essentially zero when either party is themselves dominant. Figure4 illustrates these non-linear marginal effects graphically, providing clear support for H5 by demonstrating that PAN and PRD advertising are more effective in precincts dominated by other political parties until the locally dominant party becomes too strong. Again, PRI advertising is equally ineffective across all types of precinct. 39The effect of PAN advertising share in 2012 is significant just outside the 5% significance level.

36 (a) PAN political advertising 1 .5 0 -.5 -1 -1.5 Marginal effect of PAN political advertising -2 .2 .4 .6 .8 Vote share of largest party (2006)

Non-dominant party Dominant party

(b) PRD political advertising 1 .5 0 -.5 -1 -1.5 Marginal effect of PAN political advertising -2 .2 .4 .6 .8 Vote share of largest party (2006)

Non-dominant party Dominant party

Figure 4: Effects of political advertising by vote share of the largest party and party dominance Notes: The figures plot the estimated marginal effect of AM political advertising, based on the estimates in Table4. The figures show that political advertising is only effective for non-dominant parties, and particularly so when facing a locally dominant party of intermediate strength. The density of the data is shown in grey along the x axis; less than 1% of our sample lies outside the range depicted on the x axis. The insignificant relationships for the PRI are omitted. 37 5.3 Robustness checks

Given that our identification strategy leverages cross-state media spillovers and only exploits variation between comparable neighboring precincts, there are good reasons to be confident in the effects we identify. However, we conduct a variety of checks to ensure that our estimates are robust to potential violations of our identification assumptions and generalize to FM and television advertising. The regression estimates are presented in the Online Appendix. Measurement error in AM radio coverage is a potentially important concern. Such error occurs where changes in the probability of coverage around the boundary are smaller than the IFE maps suggest. If anything, however, this suggests that we are underestimating the effects of political advertising. Nevertheless, to check that our results are not driven by such measurement error, we restrict attention to boundaries arising from lower-powered AM radio signals—for whom coverage is less variable and more accurately measured—by excluding antennae with wattages above 10,000, and Table A2 (in the Online Appendix) shows very similar results.40 As an alternative check, Table A3 shows that controlling for the interaction between political advertising and precinct area—in order to partial out differences in our heterogeneous effects that could simply arise from differential coverage measurement error—does not affect our results. Another potential concern is that our findings are driven by coverage boundaries that corre- spond with political boundaries. Despite being neighbors, a precinct in a different legislative district might experience differential campaign activity. Although it is not obvious that this would be correlated with media coverage, we address this potential concern by also including district fixed effects. The results in Table A4 are very similar in magnitude. Similarly, to ensure that our results are not driven by precincts covered by different numbers of media stations, Table A5 shows that the results are robust to the inclusion of fixed effects for the total number of AM radio stations 40Stations with high wattage (high power) have larger total coverage areas and tend to have wider zones where signal strength is between 50 and 60 dBµ, in which coverage may be spotty or poor but often not zero.

38 covering an electoral precinct. More generally, we examined the sensitivity of our results to the inclusion of additional con- trols. Unsurprisingly, given the balance shown in Table3, our average effects are very similar when controlling for our matching variables. Furthermore, we control for the interaction between political advertising and each matching variable in separate regressions, as well as the interaction between political advertising and a party’s 2000, 2003 and 2006 precinct vote share to address the concern that differential trends across our interaction variables are driving the results. The results, available in our replication code to save space, show that our main estimates are not substantially affected. We also examined the sensitivity of our estimates to the choice of maximum distance from the coverage boundary and the number of matches. Tables A6 and A7 show that restricting attention to precincts within 0.5km or 5km of the nearest coverage boundary produced essentially identical results. Furthermore, we obtain similar coefficients when using the best available match, allowing for up to five matches, or using all possible matches (Tables A8-A10), although using only the best available match reduces the power of our tests. An important consideration regarding our interpretation of the results is the possibility that political advertising reflects other differences in media content across states, rather than solely the effects of advertising. For example, AM stations in states with larger distributions of PAN advertising, and thus higher PAN vote shares, may also more favorably or more frequently cover the PAN in the news. To address this concern, we examine the 2006 election as a placebo. Using the allocation formula specified by the 2007 reform, we calculated the advertising share that each party would have received in 2006 had the reform already been passed. Supporting our claim that it is political advertising—rather than biases in media content—that affect vote choice, Table A11 shows that the predicted 2006 advertising share does not increase the vote share of any party. Furthermore, there is no systematic evidence of the heterogeneous effects detailed above. A further issue with interpreting our findings is that our results also capture the response of

39 political parties to media coverage. However, conversations with a prominent political consultant indicate that parties are either unaware of the cross-state signal spillovers that we exploit, or do not take these spillovers into account when designing their political advertising strategies. As highlighted in Figure2, spillovers in AM radio signals across states are also not straightforward to detect and are likely to be second order in determining party strategies. Furthermore, we also regard the overall effect of access to advertising—which combines the equilibrium behavior of both parties and voters—as the primary estimate of policy interest. Finally, our results also generalize to other media formats. Although the smaller FM and television samples differ markedly from our main AM sample, the heterogeneous effects—which are similar to the AM results and generally remain statistically signiﬁcant—in Tables A16 and A17 further indicate that political advertising is most effective where voters are less informed, political competition is low, and a party is not locally dominant. Consistent with our theory, changes in sample composition ensure that the average effects of political advertising are lower in the more developed and competitive precincts that constitute these samples. Moreover, we again ﬁnd that neither FM nor television political advertising wins votes for the PRI.

6 Conclusion

Little is known about how political advertising affects electoral outcomes in non-consolidated democracies, where the media is partially free but local politics is often dominated by a single party. Given that informational advantages are a key facet of local dominance, we theorize that access to political advertising is particularly effective for non-locally dominant parties that would otherwise struggle to reach or decide not to target uninformed voters in less competitive areas. We use an innovative empirical design that exploits within-neighbor differences in cross-state media coverage spillovers to test the implications of our theoretical argument in the aftermath of a major media regulation reform in Mexico. Our ﬁndings show that political advertising principally

40 benefits locally non-dominant parties by winning the votes of less politically informed voters in contexts with intermediate levels of political competition. A key implication of our findings is that equalizing political advertising opportunities across political parties can support democratic consolidation in two ways. First, greater equality in political advertising opportunities increases informed political participation, which presumably increases political representation by better matching electoral outcomes with voter preferences. In turn, this may increase support for democracy (e.g. Mattes and Bratton 2007; Mishler and Rose 1997). Second, by increasing the vote share of non-dominant parties in less competitive precincts, greater equality in political advertising opportunities promotes multi-party competition and may force incumbents to address the concerns of a wider range of citizens. Conversely, as Boas and Hidalgo(2011) show, when increased media access is concentrated among incumbents, cycles of political dominance can instead be perpetuated. Our results therefore suggest that recent reforms providing equitable access to election advertising could cultivate a process of democratic deepen- ing in many parts of the world. Nevertheless, further work is required to understand exactly how political advertising wins votes. Given the difficulty of distinguishing mobilization from persuasion empirically, our research does not separate these explanations for our results. Particularly outside consolidated democracies, it is not obvious which mechanisms apply where voters are less knowledgeable about politics and political parties employ a wide variety of strategies to win votes. A related question is how political parties allocate their advertising in non-consolidated democracies. Should parties focus on positive or negative advertising? Should parties use messages targeted to a broad audience or a narrow one? The answer to these important questions, which affect the effect of equalization of media access on democratic consolidation, should be the subjects of future research. Finally, our findings also complement recent evidence from the U.S., and suggest that our theoretical insights apply both within non-consolidated democracies but also for some types of races in consolidated democracies. Our results support Lenz’s (2009) and Peterson’s (2009) findings that

41 a key function of electoral campaigns—via political advertising in our case—is to reduce voter uncertainty about the policy positions and characteristics of different candidates. However, since the positions of major U.S. parties are relatively well-known, it is unsurprising that political advertising’s effects are small and transient (Gerber et al. 2011). Nevertheless, where incumbents are well-established—an informational incumbency advantage (e.g. Shepsle 1972) creating dynamics akin to Mexico’s local dominance—increasing awareness of elections (Panagopoulos and Green 2008) and challenger campaign spending (Goldstein and Ridout 2004) has signiﬁcantly reduced incumbent victory margins.

42 References

Ansolabehere, Stephen, Erik C. Snowberg and James M. Snyder. 2006. “Television and the incumbency advantage in US elections.” Legislative Studies Quarterly 31(4):469–490.

Anzia, Sarah F. 2013. Timing and Turnout: How Off-cycle Elections Favor Organized Groups. University of Chicago Press.

Baker, Andy, Barry Ames and Lucio R. Renno. 2006. “Social context and campaign volatility in new democracies: networks and neighborhoods in Brazil’s 2002 Elections.” American Journal of Political Science 50(2):382–399.

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

Boas, Taylor. forthcoming. “Voting for Democracy: Campaign Effects in Chile’s Democratic Transition.” Latin American Politics and Society .

Boas, Taylor C. 2014. “Media Barons on the Ballot: Politically-Controlled Broadcasting in Brazil.” Presented at the 2014 Midwest Political Science Association meeting.

Boas, Taylor C. and F. Daniel Hidalgo. 2011. “Controlling the airwaves: Incumbency advantage and community radio in Brazil.” American Journal of Political Science 55(4):869–885.

Bullock, John G. 2011. “Elite inﬂuence on public opinion in an informed electorate.” American Political Science Review 105(3):496–515.

Conroy-Krutz, Jeffrey and Devra C. Moehler. 2014. “Moderation from Bias: A Field Experiment on Partisan Media in a New Democracy.” Working paper.

Cornelius, Wayne A. 1996. Mexican politics in transition: The breakdown of a one-party-dominant regime. La Jolla: Center for US-Mexican Studies, University of California, San Diego.

43 Cornelius, Wayne A. 2004. Mobilized voting in the 2000 elections: The changing efﬁcacy of vote buying and coercion in Mexican electoral politics. In Mexico’s Pivotal Democratic Elec- tions: Candidates, Voters, and the Presidential Campaign of 2000, ed. Jorge I. Dom´ınguez and Chappell Lawson. Stanford University Press pp. 47–65.

Da Silveira, Bernardo S. and Joao˜ M.P. De Mello. 2011. “Campaign advertising and election outcomes: Quasi-natural experiment evidence from gubernatorial elections in Brazil.” Review of Economic Studies 78(2):590–612.

Diaz-Cayeros, Alberto, Federico Estevez´ and Beatrix Magaloni. forthcoming. Strategies of Vote Buying: Democracy, Clientelism, and Poverty Relief in Mexico. New York: Cambridge Univer- sity Press.

Djankov, Simeon, Caralee McLiesh, Tatiana Nenova and Andrei Shleifer. 2003. “Who Owns the Media?” Journal of Law and Economics 46(2):341–382.

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

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

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

Fox, Jonathan. 1994. “The difﬁcult transition from clientelism to citizenship: Lessons from Mex- ico.” World Politics 46(2):151–184.

Gerber, Alan S., James G. Gimpel, Donald P. Green and Daron R. Shaw. 2011. “How large and long-lasting are the persuasive effects of televised campaign ads? Results from a randomized ﬁeld experiment.” American Political Science Review 105(1):135–150.

44 Goldstein, Kenneth and Travis N. Ridout. 2004. “Measuring the effects of televised political advertising in the United States.” Annual Review of Political Science 7:205–226.

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

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

Hallin, Daniel. 2000. Media, political power, and democratization in Mexico. In De-Westernizing Media Studies, ed. James Currant and Myung-Jin Park. London: Routledge pp. 97–110.

Huber, Gregory A. and Kevin Arceneaux. 2007. “Identifying the Persuasive Effects of Presidential Advertising.” American Journal of Political Science 51(4):957–977.

Hughes, Sallie and Chappell H. Lawson. 2004. “Propaganda and crony capitalism: Partisan bias in Mexican television news.” Latin American Research Review 39(3):81–105.

King, Gary, Jennifer Pan and Margaret E. Roberts. 2013. “How censorship in China allows government criticism but silences collective expression.” American Political Science Review 107(2):326–343.

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

Langston, Joy. 2006. “The Changing Party of the Institutional Revolution Electoral Competition and Decentralized Candidate Selection.” Party Politics 12(3):395–413.

Larreguy, Horacio. 2013. “Monitoring Political Brokers: Evidence from Clientelistic Networks in Mexico.” Working paper.

45 Larreguy, Horacio, John Marshall and James M. Snyder Jr. 2015. “Publicizing malfeasance: When media facilitates electoral accountability in Mexico.” Working paper.

Larreguy, Horacio, John Marshall and Pablo Querub´ın. forthcoming. “Parties, Brokers and Voter Mobilization: How Turnout Buying Depends Upon the Party’s Capacity to Monitor Brokers.” American Political Science Review .

Lawson, Chappell H. 2002. Building the Fourth Estate: Democratization and the Rise of a Free Press in Mexico. Berkeley: University of California Press.

Lawson, Chappell H. 2004a. Mexico’s Great Debates: The Televised Candidate Encounters of 2000 and Their Electoral Consequences. In Mexico’s Pivotal Democratic Election: Candidates, Voters, and the Presidential Campaign of 2000, ed. Jorge I. Dom´ınguez and Chappell H. Law- son. Stanford, CA: Stanford University Press pp. 210–241.

Lawson, Chappell H. 2004b. Television Coverage, Vote Choice, and the 2000 Campaign. In Mex- ico’s Pivotal Democratic Election: Candidates, Voters, and the Presidential Campaign of 2000, ed. Jorge I. Dom´ınguez and Chappell H. Lawson. Stanford, CA: Stanford University Press pp. 187–211.

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

Lenz, Gabriel S. 2009. “Learning and opinion change, not priming: Reconsidering the priming hypothesis.” American Journal of Political Science 53(4):821–837.

Lindbeck, Assar and Jorgen¨ W. Weibull. 1987. “Balanced-budget redistribution as the outcome of political competition.” Public choice 52(3):273–297.

Lupu, Noam. 2013. “Party brands and partisanship: Theory with evidence from a survey experiment in Argentina.” American Journal of Political Science 57(1):49–64.

46 Magaloni, Beatriz. 2006. Voting for Autocracy: Hegemonic Party Survival and its Demise in Mexico. New York: Cambridge University Press.

Mattes, Robert and Michael Bratton. 2007. “Learning about democracy in Africa: Awareness, performance, and experience.” American Journal of Political Science 51(1):192–217.

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

McCann, James A. and Chappell H. Lawson. 2006. “Presidential campaigns and the knowledge gap in three transitional democracies.” Political Research Quarterly 59(1):13–22.

Mishler, William and Richard Rose. 1997. “Trust, distrust and skepticism: Popular evaluations of civil and political institutions in post-communist societies.” Journal of Politics 59(2):418–451.

Morgenstern, Scott and Elizabeth Zechmeister. 2001. “Better the devil you know than the saint you don’t? Risk propensity and vote choice in Mexico.” Journal of Politics 63(1):93–119.

Nichter, Simeon and Brian Palmer-Rubin. forthcoming. Clientelism, Declared Support and Mex- ico’s 2012 Campaign. In Mexico’s Evolving Democracy: A Comparative Study of the 2012 Elec- tions, ed. Jorge I. Dom´ınguez, Kenneth G. Greene, Chappell Lawson and Alejandro Moreno. John Hopkins University Press.

Panagopoulos, Costas and Donald P. Green. 2008. “Field experiments testing the impact of radio advertisements on electoral competition.” American Journal of Political Science 52(1):156–168.

Persson, Torsten and Guido Tabellini. 2000. Political Economics: Explaining Economic Policy. The MIT Press.

Peterson, David A.M. 2009. “Campaign learning and vote determinants.” American Journal of Political Science 53(2):445–460.

47 Prior, Markus. 2007. Post-broadcast democracy: How media choice increases inequality in political involvement and polarizes elections. Cambridge University Press.

Schedler, Andreas. 1998. “What is democratic consolidation?” Journal of Democracy 9(2):91– 107.

Serra, Gilles. 2012. “The Risk of Partyarchy and Democratic Backsliding Mexico.” Taiwan Journal of Democracy 8(1):93–118.

Shepsle, Kenneth A. 1972. “The strategy of ambiguity: Uncertainty and electoral competition.” American Political Science Review 66(2):555–568.

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

Stromberg,¨ David. 2004. “Radio’s impact on public spending.” Quarterly Journal of Economics 119(1):189–221.

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

Zhao, Xinshu and Steven H. Chaffee. 1995. “Campaign advertisements versus television news as sources of political issue information.” Public Opinion Quarterly 59(1):41–65.

48 Online Appendix

Proof of Proposition1

Voter i votes for party A only when R ≥ v + bi, given −exp(−R) ≥ −exp(−(v + bi)). Given bi is randomly distributed according to F, the proportion of votes for party A is given by VA = F(R−v). We henceforth denote that cut point (for voting for A) by b ≡ R − v.

We now prove the comparative static predictions by differentiating VA. First, VA is increasing in n given:

∂V ∂R A = F0(b) > 0, ∂n ∂n which is positive because F0 is a probability density function with everywhere positive support, and 2 2 2 the second term (the effect of n on the expected utility of A’s policy outcome, ∂R = σ τ (x¯−δ +τ /2) ) ∂n (σ 2+nτ2)2 σ 2 2 is positive because x¯− δ > 0 (given we assume that x¯− δ > 2n , σ > 0, and n > 0). Second, we identify the following cross-partial effects:

∂ 2V ∂R A = −F00(b) < 0, ∂n∂v ∂n 2 ∂ 2V nσ 2(x¯− δ − σ ) ∂R σ 2[(x¯− δ )(σ 2 − nτ2) + σ 2τ2] A = F00(b) 2n + F0(b¯) > 0, ∂n∂τ2 (σ 2 + nτ2)2 ∂n (σ 2 + nτ2)3 ∂ 2V nτ2(x¯− δ + τ2/2) ∂R τ2(x¯− δ + τ2/2)(σ 2 − nτ2) A = −F00(b) − F0(b) < 0. ∂q∂σ 2 (σ 2 + nτ2)2 ∂n (σ 2 + nτ2)3

The ﬁrst inequality holds because F00 > 0 and ∂R/∂n > 0. The second inequality holds given the

2 2 σ 2 assumptions σ > nτ and x¯ > δ + 2n ensure that each term is positive. The third inequality holds because both fractions are positive, where the ﬁrst follows from x¯ − δ > 0 and τ2 > 0, and the

2 2 second from σ > nτ .

49 Media allocation formula

Here, we provide exact media allocation formulas. IFE speciﬁes that the 30 second advertising slots available to party i be allocated according to the following formula in states s without concurrent state-wide elections:

3 1 1 7 nationalshareit = + voteshareit−1, (6) 10 |Cit| |Ct| 10

where voteshareit−1 is i’s national vote share in the previous election, |Cit| is the number of par-

ties in i’s federal coalition, and |Ct| is the total number of federal coalitions. This formula says that 30% of time is distributed evenly between electoral coalitions (and then between parties in a given coalition), while 70% of time is allocated to parties based on their vote share at the last election. Because the rule is based on the national-level vote share, there is no variation in political advertising time across states without local elections. Crucially for our empirical strategy, media slots are shared with state-level elections when state elections are held simultaneously. Of the 41 minutes allotted to political advertising, 15 minutes are allocated according to the analogous state-level formula:

3 1 1 7 stateshareist = + voteshareist−1, (7) 10 |Cist| |Cst| 10

where the subscript st − 1 denotes that these variables are calculated using the previous state legislative election in state s. Combined, the media share of party i in state s at election t is:

  nationalsharei if s has no concurrent election mediashareist = (8)  26 15  41 nationalsharei + 41 stateshareis if s has concurrent election.

50 Technical details of empirical strategy

Speciﬁcally, our empirical strategy entailed conducting the following procedure:

1. Identify neighboring potential matches. For each precinct j, we restrict the set of possible matches to the set of neighboring precincts k that have different political advertising shares, and for whom some part of the precinct is within b kilometers (kms) of the media signal boundary/boundaries m( j,k) (which induces the difference in political advertising shares).

0 The set of possible matches is denoted M j(b) ≡ {k : d(m( j,k),k) ≤ b}, where d(a,a ) is the minimum Euclidean distance in kilometers between a and a0.

2. Calculate distance between potential matches. Calculate the Mahalanobis distance D(Xj,Xk) ≡

p 0 −1 (Xj − Xk) C (Xj − Xk) between precinct j and each possible precinct match k ∈ M j(b)

using our vector of 29 covariates (Xj and Xk) and the covariance matrix C for the set of all

possible matches, before rank-ordering the quality of matches according to D(Xj,Xk).

3. Choose control units. Take the n nearest matches within b km of the signal boundary m( j,k),

which deﬁnes the set N j(n,b) ≡ {k ∈ M j(b) : rank(D(Xj,Xk)) ≤ n}.

In our main speciﬁcations, we set b = 1 and n = 3.

Variable deﬁnitions

PAN/PRD/PRI vote share. Party legislative vote share in a given electoral precinct. One compli- cation that arises in measuring the vote share of an individual party is the existence of cross-party federal coalitions between larger and smaller parties in certain parts of the country. Voters may cast a vote for either an individual parties or a coalition. In 2009, the two coalition groups—PRI and PVEM, and PC and PT—received only 0.3% and 0.2% of the national vote share. Coalition voting was more prevalent in 2012 when the PRI-PVEM and PRD-PT-MC coalitions respectively received 3.6% and 3.3% of the national vote, with the three PRD-PT-MC sub-coalitions further

51 receiving a 1.3% vote share. We distribute the coalition vote share among the constituent parties according to their relative vote share in the precinct. Since coalition voting is rare and the large parties have dominated these coalitions, this re-allocation method does not affect our results. Source: IFE. PAN/PRD/PRI advertising share. Explained in main text. Source: constructed using data from IFE. ENPV (2006). Effective number of political parties, as deﬁned by the vote shares from the 2006 election according to the following formula:

1 ENPV (2006, unstandardized)i jt = 2 , ∑i∈I jt−1 (votesharei jt−1) where I jt−1 is the set of parties standing in precinct j at time t − 1. We then standardized this variable in each estimation sample. Precinct area (log). Natural logarithm of precinct area in square kilometers. Source: Own compu- tations in ArcGIS. Population density (log). Natural logarithm of the number of registered electors divided by precinct area. Source: IFE. Basic amenities. Percentage of households with electricity, piped water, toilet and drainage. Source: Mexican 2010 Census. Share employed. Percentage of the precinct population employed in 2010. Source: Mexican 2010 Census. Share illiterate. Percentage of the precinct population aged above 15 that is illiterate in 2010. Source: Mexican 2010 Census. Share primary complete. Percentage of the precinct population aged above 15 that completed primary education in 2010. Source: Mexican 2010 Census. Basic development (factor). The ﬁrst (standardized) factor from an iterated principal factors factor

52 analysis including population density (log), basic amenities, share employed, share illiterate and share primary complete. The factor was computed separately for each sample (to ensure that a unit increase is always a standard deviation change in that sample). The ﬁrst factor has an eigenvalue of 1.78, while the second factor has an eigenvalue of only 0.53. This indicates that our variables form a single coherent dimension. The Cronbach’s alpha for standardized versions of these variables is 0.58. Matching variables. Our matching variables are listed in Table3. They are drawn from the 2010 Census, with the exception of 2006 party vote shares, ENPV (2006), registered voters (2006) and (log) population density (2006).

Correlation between local development and political knowledge

As noted in the main text, we find a strong positive correlation between our precinct-level variables and political knowledge—both in general and of specific party candidates. Pooling the 2006, 2009 and 2012 Comparative Study of Electoral Systems (CSES) surveys, we define political knowledge using an index combining indicators for whether respondents can correctly identify the Congres- sional chambers, legislator term length, their Governor’s name, the Minister of Economics (2012 only), their legislative candidate (2009 only), the unemployment rate (2012 only), and the second group in the legislature (2012 only). Specifically, the index is the proportion of questions correctly answered by the respondent.41 Furthermore, separately for the PAN, PRD and PRI, we code indicators for respondents that both know a given party’s candidate and has an opinion about that candidate. Supporting the claims underlying our hypotheses, Table A1 shows that local development, ENPV and the locally dominant party are strong predictors of political knowledge, even after the inclusion of state fixed effects, survey fixed effects and individual level covariates. A standard 41In 2009, only four questions were asked, while six questions were asked in 2012. We use the proportion to ensure comparability between the samples.

53 Table A1: Correlation between basic development, political competition and local party dominance and political knowledge

Political Knows Knows Knows knowledge PAN PRD PRI candidate candidate candidate Panel A: Basic development (1) (2) (3) (4) Basic development (factor) 0.077*** 0.026*** 0.020*** 0.015** (0.009) (0.010) (0.006) (0.007)

Observations 10,934 10,934 10,934 10,934 Outcome mean 0.58 0.90 0.92 0.90 Outcome standard deviation 0.41 0.29 0.28 0.31 Panel B: Political competition (1) (2) (3) (4) ENPV (2006) 0.048*** 0.030*** 0.024*** 0.032*** (0.015) (0.011) (0.007) (0.010)

Observations 12,332 12,332 12,332 12,332 Outcome mean 0.59 0.91 0.92 0.90 Outcome standard deviation 0.41 0.29 0.27 0.30 Panel C: Local party dominance (1) (2) (3) PAN largest 0.052*** (0.008) PRD largest 0.021** (0.010) PRI largest 0.007 (0.011)

Observations 12,332 12,332 12,332 Outcome mean 0.91 0.92 0.90 Outcome standard deviation 0.29 0.27 0.30

Notes: All speciﬁcations are bivariate OLS regressions. Standard errors are clustered by district. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

54 deviation increase in development is associated with an eight percentage point increase in the proportion of correct answers, while an additional effective party is correlated with a ﬁve percentage point increase in correct answers. Furthermore, both development and ENPV are strongly positively correlated with knowledge of each major political party. Finally, the results for locally dominant parties indicate that PAN and PRD dominance are associated with greater knowledge of local candidates. Consistent with our ﬁndings that political advertising does not help the PRI, PRI local dominance is not correlated with greater knowledge of PRI candidates.

Robustness check results

We first check the sensitivity of our estimates to factors that affect the distinctiveness of our coverage boundary. Table A2 shows the results when the sample is restricted to lower-powered AM radio stations with signal strengths of 10,000 watts or fewer. The results are very similar in this sample. Table A3 includes an interaction between the linear political advertising term and the (natural logarithm of one plus the) area in kilometers of a given precinct in the heterogeneous effect specifications to show that our results (especially for development where population density is central) are not simply a function of differences in the coverage boundary’s reach. We obtain very similar results without using the logarithmic transformation. Table A4 shows our estimates following the inclusion of district fixed effects. Although the average effects decline slightly in magnitude, the heterogeneous effects are perhaps even stronger. The coefficients for the interactions are similar in magnitude, but are more precisely estimated. The results now indicate a statistically significant difference in the effects of PAN political advertising between mid-term and presidential elections. Table A5 includes total number of AM station fixed effects to allow for a separate intercept for precincts with different numbers of AM radio stations, and returns very similar estimates to our main specification. Tables A6 and A7 respectively show the results when using 0.5 and 5 km bandwidths to identify valid neighboring matches. In both cases, the estimates are very similar to those presented in the

55 Table A2: Effect of AM radio political advertising on PAN, PRD and PRI vote share, antennae with less than 10,000 watts only

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 0.922** 0.772* 4.609*** 1.403*** -1.391 4.175*** (0.373) (0.433) (0.965) (0.489) (0.981) (1.449) × Basic development (factor) -0.227* -0.144 (0.123) (0.108) × ENPV (2006) -1.494*** -1.054*** (0.377) (0.291) × 2012 presidential election -0.853 -0.792* (0.512) (0.407) × Largest vote share 9.272** 2.997 (3.569) (3.246) × Largest vote share (squared) -9.399*** -7.232** (3.344) (3.101) × Largest vote share × PAN largest -10.313* -8.918 (5.478) (5.439) × Largest vote share (squared) × PAN largest 11.404** 9.919* (5.469) (5.393) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.694* 0.579 1.497*** 1.762*** -0.467 3.575*** (0.402) (0.452) (0.517) (0.549) (0.405) (0.900) × Basic development (factor) -0.147** -0.117* (0.069) (0.061) × ENPV (2006) -0.334** -0.652*** (0.144) (0.158) × 2012 presidential election -1.583* -1.361* (0.872) (0.692) × Largest vote share 4.248*** 0.507 (1.276) (1.371) × Largest vote share (squared) -3.929*** -2.865** (1.138) (1.066) × Largest vote share × PRD largest -1.641 -1.654 (4.508) (4.451) × Largest vote share (squared) × PRD largest 1.218 1.249 (4.665) (4.670) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share 0.048 0.055 0.128 -0.035 -0.058 0.127 (0.388) (0.387) (0.385) (0.801) (0.405) (0.779) × Basic development (factor) -0.014 -0.037 (0.044) (0.041) × ENPV (2006) -0.029 -0.029 (0.049) (0.074) × 2012 presidential election 0.156 -0.028 (0.887) (0.874) × Largest vote share 0.246 -0.080 (0.955) (0.980) × Largest vote share (squared) 0.150 0.374 (0.985) (1.007) × Largest vote share × PRI largest 0.890 1.025 (2.006) (1.988) × Largest vote share (squared) × PRI largest -1.633 -1.770 (2.063) (2.032)

Panel A: PAN vote share (1) (2) (3) (4) (5) PAN advertising share 0.935** 4.231*** 1.344** -1.355 3.850** (0.389) (0.867) (0.526) (0.923) (1.644) × Basic development (factor) -0.241* -0.166 (0.135) (0.099) × ENPV (2006) -1.367*** -0.931*** (0.367) (0.313) × 2012 presidential election -0.964** -0.850** (0.425) (0.379) × Largest vote share 8.893** 3.451 (3.292) (3.405) × Largest vote share (squared) -9.169*** -7.350** (3.054) (2.924) × Largest vote share × PAN largest -9.718* -8.761* (5.084) (5.088) × Largest vote share (squared) × PAN largest 10.619** 9.555* (4.824) (4.822) × Precinct area (log) -0.024 0.033 0.060 0.053 -0.015 (0.036) (0.034) (0.041) (0.041) (0.024) Panel B: PRD vote share (1) (2) (3) (4) (5) PRD advertising share 0.510 1.326** 1.306*** -0.867* 3.268*** (0.472) (0.584) (0.396) (0.482) (0.878) × Basic development (factor) -0.196** -0.138** (0.073) (0.059) × ENPV (2006) -0.353** -0.666*** (0.131) (0.194) × 2012 presidential election -1.291* -1.198** (0.673) (0.536) × Largest vote share 4.547*** 0.892 (1.082) (1.060) × Largest vote share (squared) -4.171*** -3.208*** (0.999) (0.861) × Largest vote share × PRD largest -2.300 -2.292 (4.404) (4.384) × Largest vote share (squared) × PRD largest 2.160 2.167 (4.578) (4.593) × Precinct area (log) -0.046* 0.006 0.015 0.029 -0.023 (0.026) (0.020) (0.020) (0.021) (0.023) Panel C: PRI vote share (1) (2) (3) (4) (5) PRI advertising share -0.053 -0.082 -0.199 -0.145 -0.167 (0.345) (0.344) (0.674) (0.381) (0.717) × Basic development (factor) -0.031 -0.061 (0.042) (0.040) × ENPV (2006) -0.006 0.011 (0.050) (0.067) × 2012 presidential election 0.180 0.082 (0.725) (0.728) × Largest vote share 0.099 0.010 (1.067) (1.104) × Largest vote share (squared) 0.336 0.474 (1.098) (1.117) × Largest vote share × PRI largest 0.605 0.761 (1.832) (1.818) × Largest vote share (squared) × PRI largest -1.473 -1.653 (1.910) (1.882) × Precinct area (log) -0.005 0.004 0.004 0.009 -0.009 (0.016) (0.017) (0.017) (0.017) (0.015)

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 0.423 0.287 3.635*** 1.068 -1.616* 2.871* (0.320) (0.372) (0.867) (.) (0.830) (1.436) × Basic development (factor) -0.220* -0.150 (0.108) (0.089) × ENPV (2006) -1.313*** -0.807*** (0.378) (0.281) × 2012 presidential election -1.143 -1.095*** (.) (0.360) × Largest vote share 8.860*** 4.209 (3.036) (3.105) × Largest vote share (squared) -9.006*** -7.490*** (2.839) (2.688) × Largest vote share × PAN largest -10.090** -9.352* (4.858) (4.939) × Largest vote share (squared) × PAN largest 11.053** 10.201** (4.606) (4.687) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share -0.022 -0.141 0.765* 1.010** -1.380** 2.237*** (.) (0.391) (0.434) (0.385) (0.510) (0.789) × Basic development (factor) -0.160*** -0.133*** (0.054) (0.047) × ENPV (2006) -0.320** -0.559*** (0.122) (0.184) × 2012 presidential election -1.535** -1.426*** (0.588) (0.494) × Largest vote share 5.081*** 1.910* (1.164) (1.002) × Largest vote share (squared) -4.659*** -3.785*** (1.070) (0.919) × Largest vote share × PRD largest -2.166 -2.061 (4.077) (4.074) × Largest vote share (squared) × PRD largest 1.995 1.925 (4.244) (4.260) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share 0.010 0.023 0.023 0.211 -0.096 -0.070 (.) (0.343) (0.344) (0.654) (0.374) (0.682) × Basic development (factor) -0.026 -0.054 (0.039) (0.037) × ENPV (2006) -0.005 0.023 (0.050) (0.065) × 2012 presidential election -0.365 -0.201 (0.701) (0.668) × Largest vote share 0.207 0.148 (1.115) (1.147) × Largest vote share (squared) 0.224 0.373 (1.131) (1.145) × Largest vote share × PRI largest 0.744 0.939 (1.827) (1.816) × Largest vote share (squared) × PRI largest -1.602 -1.807 (1.897) (1.879)

Notes: All specifications include match-year and district fixed effects, up to three matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. The 2012 year dummy is subsumed by the match-year fixed effects. All specifications include 108,871 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, ***58 denotes p < 0.01. Table A5: Effect of AM radio political advertising on PAN, PRD and PRI vote share, controlling for number of AM radio stations fixed effects

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 1.011** 0.865** 4.348*** 1.554*** -1.203 3.677** (0.370) (0.421) (0.935) (0.482) (0.905) (1.626) × Basic development (factor) -0.230** -0.153 (0.109) (0.091) × ENPV (2006) -1.368*** -0.915*** (0.373) (0.309) × 2012 presidential election -0.949** -0.827** (0.455) (0.379) × Largest vote share 9.013** 3.664 (3.370) (3.420) × Largest vote share (squared) -9.270*** -7.477** (3.102) (2.948) × Largest vote share × PAN largest -9.941* -9.005* (5.144) (5.159) × Largest vote share (squared) × PAN largest 10.840** 9.786* (4.839) (4.851) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.493 0.388 1.344** 1.385*** -0.761 3.262*** (0.470) (0.511) (0.579) (0.429) (0.477) (0.872) × Basic development (factor) -0.140** -0.106** (0.058) (0.052) × ENPV (2006) -0.350** -0.674*** (0.130) (0.190) × 2012 presidential election -1.323* -1.206** (0.676) (0.537) × Largest vote share 4.542*** 0.779 (1.054) (1.037) × Largest vote share (squared) -4.161*** -3.125*** (0.963) (0.851) × Largest vote share × PRD largest -2.345 -2.282 (4.398) (4.406) × Largest vote share (squared) × PRD largest 2.199 2.179 (4.573) (4.601) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.104 -0.109 -0.087 -0.188 -0.136 -0.196 (0.329) (0.341) (0.334) (0.654) (0.369) (0.710) × Basic development (factor) -0.023 -0.050 (0.040) (0.038) × ENPV (2006) -0.005 0.008 (0.050) (0.068) × 2012 presidential election 0.154 0.084 (0.720) (0.737) × Largest vote share 0.128 -0.001 (1.068) (1.098) × Largest vote share (squared) 0.310 0.476 (1.098) (1.112) × Largest vote share × PRI largest 0.568 0.751 (1.828) (1.820) × Largest vote share (squared) × PRI largest -1.444 -1.640 (1.902) (1.888)

Notes: All specifications include match-year and number of AM radio station fixed effects, up to three matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. The 2012 year dummy is subsumed by the match-year fixed effects. All specifications include 108,871 observations. Standard errors clustered by state. * denotes p < 0.1, **59 denotes p < 0.05, *** denotes p < 0.01. Table A6: Effect of AM radio political advertising on PAN, PRD and PRI vote share, 0.5 km bandwidth

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 1.112*** 0.975** 4.506*** 1.654*** -1.268 3.962** (0.401) (0.455) (0.937) (0.520) (0.887) (1.699) × Basic development (factor) -0.211* -0.136 (0.115) (0.101) × ENPV (2006) -1.395*** -0.985*** (0.382) (0.307) × 2012 presidential election -0.949* -0.820* (0.474) (0.403) × Largest vote share 9.576*** 3.741 (3.229) (3.359) × Largest vote share (squared) -9.790*** -7.775** (2.985) (2.837) × Largest vote share × PAN largest -10.893** -9.801* (5.127) (5.149) × Largest vote share (squared) × PAN largest 11.774** 10.565** (4.772) (4.805) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.571 0.463 1.431** 1.411*** -0.556 3.508*** (0.415) (0.460) (0.536) (0.443) (0.446) (0.911) × Basic development (factor) -0.144** -0.108* (0.062) (0.054) × ENPV (2006) -0.355** -0.701*** (0.133) (0.191) × 2012 presidential election -1.255* -1.062* (0.627) (0.525) × Largest vote share 4.136*** 0.180 (1.073) (0.997) × Largest vote share (squared) -3.709*** -2.600*** (0.972) (0.797) × Largest vote share × PRD largest -0.970 -0.927 (4.688) (4.697) × Largest vote share (squared) × PRD largest 0.618 0.610 (4.837) (4.858) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.123 -0.109 -0.109 -0.237 -0.085 -0.257 (0.348) (0.351) (0.347) (0.724) (0.413) (0.768) × Basic development (factor) -0.028 -0.054 (0.040) (0.038) × ENPV (2006) -0.003 0.025 (0.048) (0.069) × 2012 presidential election 0.209 0.141 (0.813) (0.811) × Largest vote share -0.169 -0.191 (1.134) (1.169) × Largest vote share (squared) 0.574 0.696 (1.192) (1.207) × Largest vote share × PRI largest 0.561 0.735 (1.889) (1.895) × Largest vote share (squared) × PRI largest -1.396 -1.584 (1.965) (1.961)

Notes: All specifications include match-year fixed effects, up to three matches within 0.5 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. The 2012 year dummy is subsumed by the match-year fixed effects. All specifications include 104,521 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p60< 0.01. Table A7: Effect of AM radio political advertising on PAN, PRD and PRI vote share, 5 km bandwidth

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 0.798* 0.660 4.170*** 1.356*** -0.953 4.058** (0.404) (0.448) (1.029) (0.488) (0.899) (1.730) × Basic development (factor) -0.222* -0.153* (0.110) (0.083) × ENPV (2006) -1.375*** -0.936*** (0.379) (0.314) × 2012 presidential election -0.952* -0.898** (0.533) (0.380) × Largest vote share 7.601** 2.219 (3.209) (3.455) × Largest vote share (squared) -8.018*** -6.240** (2.871) (2.823) × Largest vote share × PAN largest -8.777* -7.725 (4.839) (4.905) × Largest vote share (squared) × PAN largest 9.785** 8.593* (4.555) (4.623) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.331 0.238 1.177** 1.277*** -0.867* 3.232*** (0.480) (0.532) (0.541) (0.392) (0.488) (0.792) × Basic development (factor) -0.132** -0.099** (0.055) (0.048) × ENPV (2006) -0.349*** -0.676*** (0.123) (0.184) × 2012 presidential election -1.349** -1.235** (0.602) (0.514) × Largest vote share 4.611*** 0.841 (1.004) (0.984) × Largest vote share (squared) -4.297*** -3.254*** (0.947) (0.853) × Largest vote share × PRD largest -2.357 -2.405 (4.214) (4.246) × Largest vote share (squared) × PRD largest 1.995 2.099 (4.457) (4.508) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.068 -0.052 -0.106 -0.008 -0.067 -0.091 (0.310) (0.311) (0.327) (0.583) (0.352) (0.623) × Basic development (factor) -0.022 -0.049 (0.039) (0.036) × ENPV (2006) 0.015 0.022 (0.051) (0.072) × 2012 presidential election -0.108 -0.113 (0.628) (0.655) × Largest vote share -0.124 -0.186 (1.133) (1.096) × Largest vote share (squared) 0.523 0.682 (1.149) (1.133) × Largest vote share × PRI largest 0.890 1.105 (1.729) (1.713) × Largest vote share (squared) × PRI largest -1.870 -2.099 (1.805) (1.788)

Notes: All specifications include match-year fixed effects, up to three matches within 5 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. The 2012 year dummy is subsumed by the match-year fixed effects. All specifications include 115,816 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p61< 0.01. main paper. The results in Tables A8-A10 show the results in the 1 km neighbors sample for different matching specifications. Table A8 allows for only the best match, Table A9 allows for up to five matches per group, and Table A10 includes all possible matches per group. The results are very similar to the case of three matches reported in the main paper.42 The balance, however, is slightly worse when using all possible matches; this is why we use just the three best matches in the main paper. To conduct placebo tests, we computed the predicted advertising share for each state using prior national and state legislative election results in accordance with the formula used for the 2009 and 2012 elections. This test addresses the concern that our advertising results are reflecting underlying differences in media coverage across states, rather than differences in political advertising shares: finding the same results in 2006 for the placebo advertising share as for 2009 and 2012 would suggest that our results are driven by differences in AM radio across states. Using the same identification strategy, we use a sample of neighboring precincts that differ in their advertising distribution. Since we find a large imbalance for the 2003 PAN vote share, we linearly control for this imbalance.43 The results in Table A11 support the effects attributable to political advertising. With the exception of the interactions with local dominance, there is no evidence that the predicted political advertising share increases support for any party on average, or produces heterogeneous effects akin to those found in the main paper. The significant interactions with ENPV for the PAN and basic development for the PRD are in the opposite direction to the effect found in the post- reform elections. In the case of local dominance, which shows similar interactions to the main results, the effect of political advertising is barely ever positive for any largest vote share. Finally, we also weighted our results by the number of registered voters (in addition to weight- 42This may not be hugely surprising given that we also weight by the number of precincts per matched group. 43Since registered voters was not reported together with the electoral data in 2003, we exclude electorate density from the factor variable.

62 Table A8: Effect of AM radio political advertising on PAN, PRD and PRI vote share, best match only

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 0.819** 0.708* 3.568*** 1.392*** -0.652 3.733** (0.316) (0.349) (0.911) (0.351) (0.985) (1.774) × Basic development (factor) -0.189* -0.132 (0.095) (0.081) × ENPV (2006) -1.125*** -0.826** (0.293) (0.309) × 2012 presidential election -0.964** -0.981** (0.379) (0.403) × Largest vote share 6.336 2.177 (4.006) (4.325) × Largest vote share (squared) -6.824* -5.725 (3.656) (3.634) × Largest vote share × PAN largest -8.851 -8.469 (5.601) (5.638) × Largest vote share (squared) × PAN largest 9.693* 9.119* (5.201) (5.251) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.441 0.368 1.202 1.678** -0.706 3.841*** (0.740) (0.776) (0.794) (0.710) (0.634) (1.298) × Basic development (factor) -0.110 -0.070 (0.069) (0.068) × ENPV (2006) -0.310** -0.731*** (0.130) (0.262) × 2012 presidential election -1.815* -1.633* (0.898) (0.876) × Largest vote share 4.508*** 0.677 (1.270) (1.709) × Largest vote share (squared) -4.512*** -3.537*** (1.162) (1.225) × Largest vote share × PRD largest -2.491 -2.552 (5.274) (5.413) × Largest vote share (squared) × PRD largest 2.411 2.583 (5.675) (5.848) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share 0.326 0.345 0.321 0.742 -0.219 0.251 (0.592) (0.594) (0.652) (1.050) (0.669) (1.058) × Basic development (factor) 0.006 -0.018 (0.041) (0.037) × ENPV (2006) 0.002 -0.036 (0.092) (0.086) × 2012 presidential election -0.829 -0.575 (1.208) (1.127) × Largest vote share 1.329 0.945 (1.451) (1.405) × Largest vote share (squared) -0.808 -0.528 (1.316) (1.300) × Largest vote share × PRI largest 0.345 0.561 (3.126) (3.107) × Largest vote share (squared) × PRI largest -1.460 -1.686 (3.118) (3.099)

Notes: All specifications include match-year fixed effects, the best match within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. The 2012 year dummy is subsumed by the match-year fixed effects. All specifications include 63,938 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p63< 0.01. Table A9: Effect of AM radio political advertising on PAN, PRD and PRI vote share, up to 5 matches

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 1.191*** 1.049** 4.870*** 1.628*** -1.325 4.334*** (0.345) (0.398) (0.943) (0.500) (0.793) (1.508) × Basic development (factor) -0.234** -0.141 (0.112) (0.097) × ENPV (2006) -1.504*** -1.066*** (0.387) (0.342) × 2012 presidential election -0.785 -0.770* (0.502) (0.403) × Largest vote share 9.862*** 3.338 (3.003) (2.889) × Largest vote share (squared) -9.913*** -7.569*** (2.786) (2.510) × Largest vote share × PAN largest -12.705** -11.319** (4.716) (4.698) × Largest vote share (squared) × PAN largest 13.394*** 11.888** (4.528) (4.459) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.645 0.546 1.544*** 1.337*** -0.688 3.423*** (0.422) (0.462) (0.547) (0.401) (0.430) (0.821) × Basic development (factor) -0.138** -0.099** (0.054) (0.047) × ENPV (2006) -0.372** -0.725*** (0.139) (0.179) × 2012 presidential election -1.029* -0.918** (0.579) (0.448) × Largest vote share 4.780*** 0.677 (1.032) (0.966) × Largest vote share (squared) -4.300*** -3.161*** (0.905) (0.778) × Largest vote share × PRD largest -0.739 -0.714 (4.452) (4.444) × Largest vote share (squared) × PRD largest 0.397 0.395 (4.667) (4.675) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.243 -0.228 -0.261 -0.419 -0.101 -0.424 (0.284) (0.286) (0.302) (0.638) (0.354) (0.722) × Basic development (factor) -0.025 -0.055 (0.040) (0.037) × ENPV (2006) 0.008 0.035 (0.051) (0.066) × 2012 presidential election 0.313 0.289 (0.700) (0.749) × Largest vote share -0.093 -0.039 (1.039) (1.041) × Largest vote share (squared) 0.426 0.512 (1.059) (1.065) × Largest vote share × PRI largest 1.040 1.207 (1.664) (1.659) × Largest vote share (squared) × PRI largest -1.811 -1.991 (1.740) (1.725)

Notes: All specifications include match-year fixed effects, up to 5 matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. The 2012 year dummy is subsumed by the match-year fixed effects. All specifications include 131,984 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p64< 0.01. Table A10: Effect of AM radio political advertising on PAN, PRD and PRI vote share, all possible matches

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 1.224*** 1.076** 5.006*** 1.542*** -1.314 4.654*** (0.346) (0.400) (0.944) (0.496) (0.791) (1.451) × Basic development (factor) -0.249** -0.136 (0.116) (0.098) × ENPV (2006) -1.548*** -1.145*** (0.396) (0.347) × 2012 presidential election -0.581 -0.614 (0.513) (0.416) × Largest vote share 9.852*** 2.842 (3.061) (2.855) × Largest vote share (squared) -9.963*** -7.456*** (2.872) (2.565) × Largest vote share × PAN largest -13.153*** -11.714** (4.654) (4.654) × Largest vote share (squared) × PAN largest 13.967*** 12.382*** (4.507) (4.458) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.702 0.603 1.592*** 1.266*** -0.689 3.444*** (0.424) (0.462) (0.561) (0.362) (0.452) (0.848) × Basic development (factor) -0.139** -0.099** (0.053) (0.047) × ENPV (2006) -0.369** -0.748*** (0.144) (0.187) × 2012 presidential election -0.845 -0.757* (0.560) (0.439) × Largest vote share 5.030*** 0.729 (1.130) (1.041) × Largest vote share (squared) -4.492*** -3.269*** (0.973) (0.824) × Largest vote share × PRD largest 0.191 0.291 (4.484) (4.460) × Largest vote share (squared) × PRD largest -0.718 -0.795 (4.698) (4.689) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.259 -0.249 -0.294 -0.549 -0.042 -0.468 (0.296) (0.298) (0.319) (0.651) (0.357) (0.716) × Basic development (factor) -0.031 -0.064 (0.042) (0.039) × ENPV (2006) 0.015 0.044 (0.055) (0.070) × 2012 presidential election 0.513 0.389 (0.698) (0.729) × Largest vote share -0.191 -0.089 (1.059) (1.044) × Largest vote share (squared) 0.467 0.533 (1.061) (1.066) × Largest vote share × PRI largest 1.048 1.208 (1.569) (1.576) × Largest vote share (squared) × PRI largest -1.750 -1.924 (1.628) (1.625)

Notes: All specifications include match-year fixed effects, all possible matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. The 2012 year dummy is subsumed by the match-year fixed effects. All specifications include 146,140 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p65< 0.01. Table A11: Effect of the AM radio 2006 placebo on PAN, PRD and PRI vote share

Panel A: PAN vote share (1) (2) (3) (4) (5) PAN advertising share (placebo) 0.042 0.038 -0.461 -5.226*** -5.591*** (0.204) (0.201) (0.320) (1.784) (1.723) × Basic development (factor) 0.048 0.048 (0.079) (0.049) × ENPV (2003) 0.209** 0.107* (0.084) (0.061) × Largest vote share 21.488*** 21.739*** (7.124) (7.103) × Largest vote share (squared) -20.387*** -20.461*** (6.720) (6.763) × Largest vote share × PAN largest -18.496*** -18.912*** (6.686) (6.640) × Largest vote share (squared) × PAN largest 17.845*** 18.174*** (6.455) (6.452) Panel B: PRD vote share (1) (2) (3) (4) (5) PRD advertising share (placebo) -0.158 -0.259 -0.257 -0.686 -0.526 (0.177) (0.179) (0.380) (1.065) (1.159) × Basic development (factor) 0.163** 0.041 (0.068) (0.046) × ENPV (2003) 0.049 -0.064 (0.152) (0.080) × Largest vote share 3.304 3.308 (4.488) (4.520) × Largest vote share (squared) -3.880 -3.995 (4.446) (4.439) × Largest vote share × PRD largest 1.029 1.083 (5.235) (5.342) × Largest vote share (squared) × PRD largest -0.097 -0.252 (5.134) (5.169) Panel C: PRI vote share (1) (2) (3) (4) (5) PRI advertising share (placebo) 0.117 0.126 0.100 -0.291 -0.227 (0.175) (0.175) (0.214) (0.648) (0.678) × Basic development (factor) -0.084 -0.041 (0.054) (0.029) × ENPV (2003) 0.007 -0.037 (0.076) (0.050) × Largest vote share 1.517 1.627 (2.576) (2.477) × Largest vote share (squared) -1.351 -1.440 (2.404) (2.285) × Largest vote share × PRI largest -1.179 -1.494 (3.284) (3.271) × Largest vote share (squared) × PRI largest 1.219 1.475 (3.142) (3.126)

Notes: All specifications include match-year fixed effects, all possible matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. The basic development variable has mean zero and a standard deviation of one, while ENPV66 ranges from 1 to 4.6. All specifications include 49,749 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. Table A12: Effect of AM radio political advertising on PAN, PRD and PRI vote share, weighting by the number of registered voters

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 0.820** 0.756** 4.324*** 1.179*** -1.711 3.523* (0.299) (0.340) (0.685) (0.362) (1.057) (1.839) × Basic development (factor) -0.261** -0.219* (0.121) (0.118) × ENPV (2006) -1.383*** -0.998*** (0.273) (0.360) × 2012 presidential election -0.678* -0.564* (0.352) (0.311) × Largest vote share 10.721** 4.862 (3.967) (3.845) × Largest vote share (squared) -11.011*** -9.162** (3.703) (3.445) × Largest vote share × PAN largest -9.218 - 7.891 (7.513) (7.378) × Largest vote share (squared) × PAN largest 8.822 7.567 Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.341 0.262 1.318** 0.944** -1.039*** 2.856*** (0.381) (0.407) (0.495) (0.350) (0.328) (0.661) × Basic development (factor) -0.176*** -0.133*** (0.048) (0.045) × ENPV (2006) -0.389*** -0.702*** (0.118) (0.136) × 2012 presidential election -0.844 -0.895* (0.512) (0.501) × Largest vote share 5.245*** 1.944 (1.027) (1.201) × Largest vote share (squared) -5.056*** -4.603*** (0.979) (1.036) × Largest vote share × PRD largest -2.817 - 3.202 (4.244) (4.194) × Largest vote share (squared) × PRD largest 2.652 3.052 Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.304 -0.262 -0.169 -0.454 -0.461 -0.416 (0.315) (0.316) (0.329) (0.697) (0.383) (0.804) × Basic development (factor) -0.018 -0.042 (0.043) (0.041) × ENPV (2006) -0.049 -0.033 (0.046) (0.075) × 2012 presidential election 0.248 0.306 (0.865) (0.752) × Largest vote share 1.003 0.559 (1.234) (1.354) × Largest vote share (squared) -0.452 -0.113 (1.334) (1.407) × Largest vote share × PRI largest 1.704 2.011 (2.333) (2.320) × Largest vote share (squared) × PRI largest67 -2.796 -3.113 Notes: All specifications include match-year fixed effects, up to three matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping multiplied by the number of registered voters in the precinct. The basic development variable has mean zero and a standard deviation of one, while ENPV ranges from 1 to 4.6. Lower order interaction terms are omitted. All specifications include 108,871 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. ing by the inverse of the number of neighbor comparisons). This could provide estimates that more accurately reflect population averages. The results in Table A12 indicate that the results are effectively unchanged by this weighting scheme.

FM radio and television signals

In the body of the paper, we focus on AM radio stations. This is mainly because of the larger and far more rural sample that they allow. Additionally, the AM radio sample provides greater variation in types of electoral precinct, and thus allows for a better test of our heterogeneous effects. However, our findings generalize to other media formats. Here, we confirm that the FM radio and television media samples are relatively different to the AM media one and more internally homogeneous, before demonstrating that we find very similar heterogeneous effects across all samples. Examining the full sample radio adverts placed by political parties during the 2012 federal elections, we find that advertising does not substantively differ across AM and FM radio stations. Advertisements were fairly evenly distributed across AM and FM frequencies: of the 330 radio ads, only 15 (5%) and 11 (4%) were respectively broadcast disproportionately on AM stations and FM stations.44 Therefore, there does not appear to be any meaningful difference between the types of ads broadcast over the different wavelengths. Comparable television advert data were not available, but ads were very general and targeted the same kinds of national political issues noted in the body of the paper. Furthermore, by identifying off cross-state radio signal spillovers, the locations our effects are identified for are very unlikely to be the targets of locally-specific ads targeted at different states. Figures5 and6 map the coverage areas of all FM radio and television stations, and show that the level of coverage associated with any given media outlet is far lower than for AM radio (in Figure2 in the main paper). 45 Due to the relative limited reach of FM radio and television signals, 44We tested for whether the proportion on AM and FM radio stations differ, retaining all those that differ at the 10% level for more detailed analysis. 45Note that there are some television channels which emit from multiple antennae across the

68 Figure 5: FM radio signal coverage (source: IFE)

Figure 6: TV signal coverage (source: IFE)

69 Figure 7: FM neighbor-matched 1 km sample of electoral precincts

Figure 8: TV neighbor-matched 1 km sample of electoral precincts

70 Table A13: Summary statistics for FM radio neighbor-matched sample

Obs. Mean Std. dev. Min. Max. Dependent variables PAN vote share 44,358 0.258 0.149 0 0.959 PRD vote share 44,358 0.162 0.124 0 0.839 PRI vote share 44,358 0.369 0.127 0 0.966

Treatment variables PAN advertising share 44,358 0.274 0.024 0.211 0.347 PRD advertising share 44,358 0.154 0.034 0.094 0.209 PRI advertising share 44,358 0.253 0.058 0.186 0.343

Covariates Electorate density (2006, log) 44,358 5.253 2.328 0.070 10.518 Share basic necessities 44,358 0.750 0.256 0 1 Share illiterate above 15 44,358 0.091 0.068 0 0.584 Share employed 44,358 0.948 0.043 0.243 1 Share primary complete 44,358 0.483 0.091 0.054 0.778 Basic development (factor) 44,358 0.000 1.000 -3.757 1.993 ENPV (2006) 44,358 2.886 0.532 1.193 4.600 2012 presidential election 44,358 0.503 0.500 0 1

in combination with the fact that the antennae are predominately located in and around towns and cities, precincts at the boundary between receiving and not receiving a signal from a neighboring state are far more urban and closer to the state boundaries. This is conﬁrmed in Tables A13 and A14, which provide the summary statistics for these matched samples. Furthermore, Tables A15 and A16 (below) show that the FM and especially television samples are imperfectly balanced across political advertising distributions, and therefore provide somewhat less reliable estimates. Column (1) of Tables A17 and A18 shows that the average effects of FM and television political advertising decline substantially in ways consistent with the differences in the AM, FM and television samples and our heterogeneous effects. This is consistent with our theory, which ex- country. Our variable deﬁnitions adjust for any double-counting such that the same channel reaches a given precinct via multiple antennae.

71 Table A14: Summary statistics for television neighbor-matched sample

Obs. Mean Std. dev. Min. Max. Dependent variables PAN vote share 42,623 0.263 0.148 0 0.867 PRD vote share 42,623 0.169 0.123 0 0.852 PRI vote share 42,623 0.350 0.126 0 0.966

Treatment variables PAN advertising share 42,623 0.274 0.025 0.211 0.347 PRD advertising share 42,623 0.153 0.032 0.094 0.209 PRI advertising share 42,623 0.252 0.056 0.186 0.343

Covariates Electorate density (2006, log) 42,623 5.698 2.324 0.070 10.846 Share basic necessities 42,623 0.774 0.235 0 1 Share illiterate above 15 42,623 0.085 0.066 0 0.568 Share employed 42,623 0.945 0.043 0.243 1 Share primary complete 42,623 0.482 0.096 0.050 0.818 Basic development (factor) 42,623 0.000 1.000 -3.366 1.993 ENPV (2006) 42,623 2.891 0.520 1.199 4.600 2012 presidential election 42,623 0.502 0.500 0 1

72 Table A15: Balance checks—partial correlation between FM radio political advertising and 29 matching covariates

Registered Population Turnout PAN PRD PRI ENPV Share Share Share electorate density 2006 vote vote vote 2006 economically employed medical 2006 (log) 2006 2006 2006 2006 active insurance PAN advertising share -162.804 8.965 0.314 0.350 -0.253 -0.187 -0.230 -0.080 0.181 0.356 (1047.833) (7.151) (0.355) (0.226) (0.190) (0.266) (0.588) (0.188) (0.140) (0.458) PRD advertising share -1170.859 -16.995* -0.470 -0.545* 0.442 0.277 -0.708 0.218 -0.141 -0.538 (1559.843) (8.493) (0.387) (0.266) (0.257) (0.332) (1.044) (0.155) (0.166) (0.473) PRI advertising share 871.519 21.429** 0.142 0.497 -0.500** -0.080 0.779 -0.086 -0.041 0.210 (1803.384) (8.712) (0.372) (0.298) (0.232) (0.380) (1.366) (0.190) (0.203) (0.363) Primary Middle Secondary Share Share Share Share Share Share Share school school school 6-17 in 18-24 in illiterate no school primary primary secondary attendance attendance attendance school school above 15 incomplete complete incomplete complete PAN advertising share -0.017 0.306 1.004** 0.233 0.420 0.178 0.167 -0.759*** -0.641*** -0.224* (0.173) (0.436) (0.430) (0.288) (0.266) (0.152) (0.224) (0.250) (0.130) (0.120) 73 PRD advertising share 0.024 -0.479 -1.294** -0.394 -0.577** -0.110 -0.120 0.704** 0.621** 0.184 (0.159) (0.443) (0.605) (0.321) (0.261) (0.170) (0.184) (0.328) (0.230) (0.196) PRI advertising share -0.084 0.138 0.878 0.225 0.267 0.026 0.042 -0.280 -0.339 -0.202 (0.139) (0.456) (0.744) (0.351) (0.330) (0.124) (0.179) (0.473) (0.339) (0.163) Share Share Share Share Share Share Share Basic Share secondary with non-dirt electricity piped with with basic with incomplete house ﬂoor water toilet drainage necessities internet PAN advertising share -0.171 0.067 -0.146 0.205 0.437 0.134 0.641 0.386 0.453 (0.102) (0.177) (0.320) (0.130) (0.685) (0.370) (0.589) (0.752) (0.304) PRD advertising share 0.143 0.110 0.137 -0.261 -0.424 -0.089 -0.613 -0.668 -0.411 (0.185) (0.214) (0.472) (0.177) (0.657) (0.390) (0.679) (0.709) (0.276) PRI advertising share -0.185 -0.175 0.119 0.355 0.597 0.027 0.787 1.234 0.224 (0.155) (0.238) (0.435) (0.224) (0.742) (0.499) (0.788) (0.834) (0.386)

Registered Population Turnout PAN PRD PRI ENPV Share Share Share electorate density 2006 vote vote vote 2006 economically employed medical 2006 (log) 2006 2006 2006 2006 active insurance PAN advertising share 2938.201 9.567 0.124 0.339* -0.423* 0.126 0.439 0.242* 0.251*** -0.441** (2079.540) (12.902) (0.228) (0.165) (0.208) (0.117) (0.752) (0.127) (0.065) (0.167) PRD advertising share -7300.702*** -12.948 0.036 -0.179 0.164 0.024 -0.512 -0.173 -0.045 0.478 (2574.938) (8.968) (0.259) (0.319) (0.405) (0.216) (1.337) (0.171) (0.169) (0.403) PRI advertising share 7783.211** 7.416 0.005 0.029 0.047 -0.128 0.219 0.067 -0.235* -0.075 (3062.922) (12.639) (0.183) (0.378) (0.259) (0.240) (1.273) (0.227) (0.123) (0.500) Primary Middle Secondary Share Share Share Share Share Share Share school school school 6-17 in 18-24 in illiterate no school primary primary secondary attendance attendance attendance school school above 15 incomplete complete incomplete complete PAN advertising share -0.078 0.078 0.979*** 0.236 0.814** -0.374* -0.322 -0.882*** -0.429** -0.170* (0.113) (0.245) (0.276) (0.176) (0.331) (0.215) (0.221) (0.262) (0.200) (0.086) 74 PRD advertising share 0.262** 0.022 -1.133** -0.181 -1.036*** 0.289* 0.218 1.291*** 0.709*** 0.311** (0.094) (0.197) (0.405) (0.172) (0.259) (0.154) (0.129) (0.248) (0.151) (0.144) PRI advertising share -0.155 0.086 0.704 0.201 0.557 0.012 0.076 -0.889* -0.484* -0.173 (0.093) (0.325) (0.675) (0.258) (0.553) (0.312) (0.276) (0.467) (0.270) (0.174) Share Share Share Share Share Share Share Basic Share secondary with non-dirt electricity piped with with basic with incomplete house ﬂoor water toilet drainage necessities internet PAN advertising share -0.129 -0.151 0.377** -0.002 1.147* 0.853 1.622*** 2.126** 0.669 (0.094) (0.333) (0.173) (0.205) (0.665) (0.620) (0.531) (0.825) (0.402) PRD advertising share 0.255** 0.155 -0.374 0.199 -1.816** -0.515* -1.322** -2.293*** -0.889*** (0.122) (0.363) (0.279) (0.218) (0.794) (0.251) (0.519) (0.803) (0.255) PRI advertising share -0.128 0.037 -0.015 -0.292 0.265 -0.220 0.121 0.101 0.532 (0.125) (0.324) (0.331) (0.215) (0.820) (0.726) (1.053) (1.301) (0.441)

Notes: Each coefficient is estimated separately from a regression of the outcome on a party’s advertising share and match-year fixed effects. All Census variables, in columns (8)-(29) are from 2010. All specifications include up to three matches within 1 km of a coverage boundary, and weight by the inverse of the number of precincts per match-year grouping. All specifications include 42,623 observations. Standard errors clustered by state. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01. plains that more urban, developed and politically competitive precincts should experience smaller average effects. Combined with our estimates showing that political advertising via all media formats is less effective in more developed and politically competitive precincts, the change in sample composition is expected to reduce the average effects of ads on FM radio and television. The heterogeneous effects further support our theory. While standard errors inevitably increase as the sample size more than halves, columns (2)-(5) show that our heterogeneous effects are generally similar to the AM results and remain statistically significant in many cases, especially for television stations. Only in the case of differences between mid-term and presidential elections in the FM sample do our results slightly differ. Moreover, we again find that political advertising never wins votes for the PRI. These results further highlight that political advertising is most effective in the areas least exposed to democratic political competition and most vulnerable to clientelistic practices.

75 Table A17: Effect of FM radio political advertising on PAN, PRD and PRI vote share

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 0.416 0.218 3.206* 0.419 -1.957** 0.725 (0.268) (0.362) (1.564) (0.302) (0.855) (3.483) × Basic development (factor) -0.240 -0.204 (0.245) (0.135) × ENPV (2006) -1.144* -0.545 (0.602) (0.640) × 2012 presidential election -0.010 -0.029 (0.371) (0.347) × Largest vote share 9.612*** 5.941 (3.318) (5.334) × Largest vote share (squared) -9.259** -7.998** (3.421) (3.538) × Largest vote share × PAN largest -7.430* -5.634 (3.886) (4.158) × Largest vote share (squared) × PAN largest 6.921 5.543 (4.665) (4.555) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share -0.056 -0.240 0.689 -0.112 -0.980** 1.387 (0.244) (0.219) (0.489) (0.289) (0.466) (0.830) × Basic development (factor) -0.206*** -0.167*** (0.057) (0.049) × ENPV (2006) -0.293 -0.537*** (0.172) (0.159) × 2012 presidential election 0.096 0.122 (0.298) (0.289) × Largest vote share 4.053** 1.279 (1.451) (1.672) × Largest vote share (squared) -4.059*** -3.437** (1.432) (1.447) × Largest vote share × PRD largest -0.085 -0.172 (5.664) (5.936) × Largest vote share (squared) × PRD largest -0.368 -0.459 (5.246) (5.613) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.084 0.012 0.292 -0.932 -0.567 -0.071 (0.449) (0.427) (0.463) (0.976) (0.592) (0.895) × Basic development (factor) -0.045 -0.036 (0.044) (0.046) × ENPV (2006) -0.144** -0.182* (0.065) (0.105) × 2012 presidential election 1.351 0.756 (1.083) (0.738) × Largest vote share 2.216 1.173 (1.952) (1.857) × Largest vote share (squared) -1.595 -1.259 (1.950) (1.992) × Largest vote share × PRI largest -3.024 -3.234 (3.142) (3.189) × Largest vote share (squared) × PRI largest 2.139 2.315 (3.208) (3.235)

Panel A: PAN vote share (1) (2) (3) (4) (5) (6) PAN advertising share 0.375* 0.138 3.971** 0.484* -0.639 2.374 (0.196) (0.199) (1.479) (0.273) (1.344) (1.698) × Basic development (factor) -0.350** -0.283** (0.140) (0.104) × ENPV (2006) -1.379** -0.614* (0.531) (0.320) × 2012 presidential election -0.332 -0.296 (0.372) (0.357) × Largest vote share 3.783 0.207 (6.526) (6.248) × Largest vote share (squared) -3.110 -2.035 (6.908) (6.757) × Largest vote share × PAN largest -3.260 -2.819 (9.154) (9.106) × Largest vote share (squared) × PAN largest 0.828 0.643 (9.831) (9.728) Panel B: PRD vote share (1) (2) (3) (4) (5) (6) PRD advertising share 0.152 -0.057 0.810 0.416 -1.046 0.565 (0.402) (0.398) (0.546) (0.457) (0.820) (1.617) × Basic development (factor) -0.257*** -0.258*** (0.051) (0.051) × ENPV (2006) -0.236* -0.309 (0.133) (0.245) × 2012 presidential election -0.459 -0.428 (0.334) (0.267) × Largest vote share 4.664** 3.004 (1.964) (2.446) × Largest vote share (squared) -4.458** -4.134** (1.604) (1.593) × Largest vote share × PRD largest -3.091 -2.506 (4.517) (5.181) × Largest vote share (squared) × PRD largest 2.416 1.681 (5.039) (5.874) Panel C: PRI vote share (1) (2) (3) (4) (5) (6) PRI advertising share -0.481* -0.471* -0.417* -1.456 -0.472 -0.560 (0.276) (0.240) (0.241) (1.099) (0.310) (0.784) × Basic development (factor) -0.031 -0.041 (0.050) (0.044) × ENPV (2006) -0.023 -0.126 (0.052) (0.100) × 2012 presidential election 1.191 0.890 (1.091) (0.690) × Largest vote share 0.030 -0.619 (0.995) (0.857) × Largest vote share (squared) -0.140 -0.006 (0.979) (0.911) × Largest vote share × PRI largest -3.472 -3.417 (2.669) (2.678) × Largest vote share (squared) × PRI largest 3.918 3.831 (2.705) (2.683)