Information's Effect on Voter Efforts to Hold Politicians to Account in Senegal

ABLE AND MOSTLY WILLING:AN EMPIRICAL ANATOMY OF INFORMATION’S EFFECT ON VOTER EFFORTS TO HOLD POLITICIANSTOACCOUNTIN SENEGAL ∗

ABHIT BHANDARI† HORACIO LARREGUY‡ JOHN MARSHALL§

SEPTEMBER 2018

Political accountability may be constrained by the reach and relevance of information campaigns in developing democracies and—upon receiving information— voters’ ability and will to hold politicians to account. To illuminate voter-level constraints without dissemination constraints, we conducted a ﬁeld experiment around Senegal’s 2017 parliamentary elections to examine the core theoretical steps linking personal delivery and explanation of different types of incumbent performance information to electoral and non-electoral accountability. Voters immediately processed information as Bayesians, found temporally benchmarked local performance most informative, and durably updated their beliefs for at least a month. However, information about incumbent duties had little independent or complementary effect. Learning that more projects than expected reached their department, voters durably increased non- electoral accountability—making costly requests of incumbents—but only increased incumbent vote choice among likely-voters and voters heavily weighting performance in their voting calculus. Voters are thus able and mostly willing to use relevant information to support political accountability.

∗We thank Fode´ Sarr and his team of enumerators for invaluable research assistance, and Elimane Kane, Thierno Niang, and LEGS-Africa for partnering with us to implement this project. We thank Antonella Bandiera, Nilesh Fernando, Matthew Gichohi, Lakshmi Iyer, Kate Orkin, Julia Payson, Amanda Robinson, Arturas Rozenas, Cyrus Samii, Moses Shayo, Jay Shon, Alberto Simpser, and participants at talks at APSA, MPSA, Notre Dame, NYU, NYU CESS Experimental Political Science Conference, and WGAPE-NYU Abu Dhabi for excellent comments. This project received ﬁnancial support from the Spencer Foundation, and was approved by the Columbia Institutional Review Board (IRB-AAAR3724) and the Harvard Committee on the Use of Human Subjects (IRB17-0880). Our pre- analysis plan was registered with the Social Science Registry, and is available at socialscienceregistry.org/trials/2324. †Department of Political Science, Columbia University. Email: [email protected]. ‡Department of Government, Harvard University. Email: [email protected]. §Department of Political Science, Columbia University. Email: [email protected].

1 1 Introduction

As a cornerstone of effective democracy, political accountability has received substantial attention from scholars and policy-makers (Ashworth 2012). Indeed, it is often argued that providing voters with information about their incumbent’s performance in office helps them to retain high-quality politicians (Fearon 1999) and engage in non-electoral accountability-seeking behavior(Aker, Col- lier and Vicente 2017; Gottlieb 2016). Particularly in contexts where governance is weak, the distribution of goods reflects patronage, and politician opportunities for rent-seeking abound, effective accountability mechanisms are needed most. In practice, recent studies identifying the effects of informational campaigns on electoral accountability and community action yield mixed findings.1 Given the complex chain of conditions linking the provision of information to better governance (Dunning et al. forthcoming; Lieberman, Posner and Tsai 2014), it is often hard to know where accountability breaks down. For example, the Metaketa initiative that coordinated similar interventions across six countries found limited electoral effects of providing information that benchmarked incumbent performance against other comparable incumbents in office at the same time (Dunning et al. forthcoming). However, as indicated by low levels of information retention, these limited effects may reflect the difficulties of disseminating relevant information, rather than voters’ low capacity or willingness to electorally reward (punish) better(worse)-performing incumbents. Furthermore, while the effects of community-driven development and civic education programs on informed local participation have received significant attention (Casey 2018), little is known about whether incumbent performance information also influences non-electoral forms of political accountability. This article dissects voters’ ability and will to use different types of incumbent performance information to hold politicians to account. By personally distributing and explaining information, 1See Banerjee et al.(2011), Casey(2018), Chong et al.(2015), Dunning et al.(forthcoming), Ferraz and Finan(2008), Humphreys and Weinstein(2012), Lieberman, Posner and Tsai(2014), and Olken(2007).

2 we abstract from dissemination challenges to focus on three links between providing information about the performance of parliamentary deputies and voter attempts to engage in electoral and non- electoral accountability-seeking political engagement. First, we examine the extent to which information provision causes voters to both immediately and durably update their beliefs in a Bayesian manner. Second, we vary the information’s content to understand what information voters regard as relevant. Specifically, we inform voters about the responsibilities of their deputies and help voters to abstract from district-specific factors influencing the performance of all incumbents by temporally benchmarking current incumbent performance metrics against previous incumbents within the same district. Third, we study whether such beliefs translate into electoral and non-electoral accountability-seeking behavior, the latter in the form of making requests from politicians, and the extent to which this behavior varies with the information’s relevance to individual voters. We designed a field experiment in rural Senegal around the 2017 parliamentary elections to examine these voter-level mechanisms underpinning political accountability. Across 450 villages from five of Senegal’s 45 departments (which serve as parliamentary districts), we trained enumerators to personally distribute and explain informational leaflets to nine young registered voters (aged 20-38) in treated villages in the month preceding the election. Our 2 × 3 factorial design varied whether respondents were informed about: (1) parliamentary deputies’ duties; (2) their current deputy’s activity in the legislature and the number and value of projects and transfers received by their department; and/or (3) how such incumbent performance metrics compared with their department’s previous deputy. In our departments, incumbent performance generally exceeded voter expectations and previous incumbents’ performance. To separate immediate effects of treatment from equilibrium responses that could reflect subsequent interactions with voters and political op- erators, our panel survey measured voters’ beliefs and (intended and actual) behavioral responses immediately before and after treatment, and again after the election. Our findings first demonstrate that rural Senegalese citizens process incumbent performance information in a sophisticated manner. Immediately after receiving the information, voters updated

3 their beliefs in line with the Bayesian tenets that the direction of updating depends on voters’ prior beliefs, that the extent of updating varies with the signal’s content, and that voters with imprecise prior beliefs update more. The results show that voters care principally about local projects and transfers, rather than legislative activities within the Assemblee´ Nationale. Moreover, while information about deputy responsibilities did not affect beliefs, temporally benchmarked information substantially influenced the extent of voter updating and increased the precision of posterior beliefs. Perhaps most remarkably, we find similar—albeit somewhat smaller—effects around a month after leaflets were delivered. Furthermore, voters seek to hold politicians to account on the basis of the information provided. Immediately after receiving the information, the average treated voter—who updated favorably about their incumbent—became three percentage points more likely to intend to vote for the incumbent. Heterogeneity in such rewards reflected the extent to which voters updated their beliefs and to which performance information is the most important factor determining vote choices. Alongside electoral intentions, non-electoral accountability-seeking behaviors also increased: treated voters were more likely to request an incumbent visit and an opportunity to express their views to those candidates directly. While voters durably updated their beliefs and demonstrated an initial willingness to hold politicians to account, lasting political accountability proves feasible but more complex. In the case of electoral accountability, performance information’s effects on incumbent vote choice were diluted by lower propensities to turn out among the younger voters that we disproportionately sampled, while responses to the local component of incumbent performance intensified over time. Specifically, our treatments did not affect self-reported vote choices, on average. However, electoral accountability behavior did materialize among respondents that cared most about incumbents lobbying for local development projects or that turned out at the last election.When treated, such voters were more likely to support incumbents overseeing more local projects and transfers, and even penalize incumbents for nationally-oriented parliamentary activity.Consistent with sub-

4 stantial within-village diffusion of our information—by voters and political parties—to older and more experienced voters, we also document greater incumbent vote shares at polling stations en- compassing treated villages that received information revealing higher rates of local projects and transfers. Together, these results suggest that information induced electoral accountability among likely-voters on issues relating to local performance. In contrast, non-electoral accountability-seeking behavior increased more uniformly, even a month after receiving treatment. The average respondent continued to make more requests of the winning candidates, especially those that performed best. The 0.1 standard deviation increase in requests reflects not only relatively costless requests for victors to call respondents or visit their village, but also reflects citizens making payments to send messages to election victors using a hotline we created. This effect was also most pronounced among those receiving benchmarked information. Our core finding that receiving incumbent performance information, especially when including temporal benchmarks, can facilitate electoral and non-electoral accountability makes several main contributions. First, we build on the frameworks of Dunning et al.(forthcoming) and Lieberman, Posner and Tsai(2014) by combining behavioral outcomes and detailed panel surveys to unpack the key links in the accountability chain once voters receive incumbent performance information. We demonstrate that accountability failures are unlikely to reflect cognitive constraints—whether voters’ inability to process information in a Bayesian manner (Gomez and Wilson 2006) or to durably retain their beliefs (Zaller 1992)—afflicting voters or an unwillingness to hold politicians to account. In fact, as in Italy (Kendall, Nannicini and Trebbi 2014) and Mexico (Arias et al. 2018b), voters—even in a hyper-presidential context (Thomas and Sissokho 2005) where almost half of respondents lacked any formal schooling and bloc voting is common (Gottlieb 2017; Koter 2013)—demonstratea remarkable capacity to incorporate relevant information. Rather, ineffec- tive information dissemination campaigns are more likely to reflect failures to reach voters, an inability to provide credible information relevant to voters’ evaluations, or competing community

5 or political influences that may be partly precipitated by the intervention. Second, we show that incumbent performance information also influences non-electoral forms of political accountability. Our behavioral measures of voter requests suggest that creating the sense that politicians will be responsive is crucial for voters to engage in costly accountability- seeking efforts beyond elections. Finding that expectations of responsiveness, rather than information about duties, drive request-making complements the few extant studies on this topic showing that civic education can stimulate demand for non-electoral accountability (Aker, Collier and Vi- cente 2017; Gottlieb 2016). Third, we further illuminate the types of information that facilitate political accountability. We provide the clearest evidence yet that benchmarking can support political accountability by helping voters to separate individual incumbent qualities from common factors affecting all incumbents in the same department. This finding may reflect our temporal benchmark being more informative than spatial benchmarks that have not induced effects beyond incumbent-only information (Arias et al. 2018a). Furthermore, we find that Senegalese voters prioritize politicians bringing projects and higher-value transfers to their department, a distributional preference consistent with Dixit and Londregan(1996). Conversely, greater involvement in parliamentary activities is—if anything—punished by voters (see also Adida et al. 2017). However, we find little evidence that information about incumbent responsibilities influences voter appraisals on its own, or that it substitutes or complements the provision of performance information. This suggests that any accountability-enhancing effects of civic education programs (e.g. Gottlieb 2016) may operate through components of the program beyond information about incumbent responsibilities.

6 2 Incumbent performance information and political account-

ability

The canonical selection model of electoral accountability reﬂects the interaction between agents (politicians) and their principals (voters), whereby incumbent performance information helps forward- looking voters to identify competent or policy-aligned politicians. In its simplest formulation, voters use performance information to update their beliefs about the extent to which the incumbent possesses such qualities, and thus identify and retain high-quality incumbents likely to perform well if re-elected (e.g. Fearon 1999). Beyond the ballot box, voters may similarly be more willing to make costly requests of incumbents that they believe will be responsive to their interests. By providing district-speciﬁc information—that incumbents could not have anticipated being released—just before elections in which incumbents sought re-election, our design sidesteps strategic policy and candidacy choices. Furthermore, by directly providing information to voters, we also abstract from the process through which information is supplied and consumed. This sharp- ens our theoretical focus on whether, upon receiving incumbent performance information, voters can hold incumbents seeking re-election to account electoral and non-electorally. We adopt a simple learning framework to dissect how voters engage in political accountability. With respect to electoral accountability, an expressive voter i implements the following decision rule:

 vote I if w f E [q ],E [q ] + (1 − w )V ≥ c  i i i I i C i i i   vi Ei[qI],Ei[qC],wi,Vi,ci = vote C if wi fi Ei[qI],Ei[qC] + (1 − wi)Vi ≤− ci (1)    abstain if wi fi Ei[qI],Ei[qC] + (1 − wi)Vi < ci where the function fi(·) increases with i’s expectation of incumbent I’s underlying “quality,”

Ei[qI], and decreases with i’s expectation of challenger C’s underlying quality, Ei[qC], and Vi

7 is the relative utility i receives from voting for I over C from all other factors. Voter i attaches weight wi ∈ [0,1] to relative expectations about quality, and weight (1 − wi) to other factors en- tering their voting calculus. If the magnitude of this weighted average of expressive benefits is positive (negative) and exceeds cost ci ≥ 0 of turning out, i will vote for I (C). This simple model implies that information that alters prior beliefs about incumbent quality—relative to challenger quality, and on issues that matter to voters—can alter vote choice and turnout. With respect to non-electoral accountability-seeking political engagement, we instead propose that voters make requests of an incumbent when the expected benefits of responsiveness exceed the cost ei ≥ 0 of making a request. Specifically, voter i implements the following rule:

 make request from I if g (E [q ]) ≥ e  i i I i ri Ei[qI],ei = (2)  no request if gi(Ei[qI]) < ei

where the beneﬁts function gi(·) increases with expected incumbent quality. Voters thus make requests of incumbents when they expect a high probability of action or a more effective action by the incumbent on the voter’s behalf. In contrast with voting, wi does not inﬂuence non-electoral accountability-seeking behavior. The extent to which these models capture electoral and non-electoral accountability-seeking behavior rests upon the validity of the model’s assumptions regarding voters’ capacity to process novel information (i.e. use a signal s to update E[qI|s]), the extent to which information is relevant to voters (i.e. how much s affects E[qI|s]), and ultimately voters’ willingness to act on their updated beliefs about the incumbent’s quality (i.e. the weight wi attached to posterior beliefs about quality, and the costs ci and ei). To structure our empirical design, the following subsections theorize the conditions under which each element of this anatomy of political accountability may hold.

8 2.1 Internalization of novel information

Upon receiving credible information (signal s in the framework above), political accountability relies on voters comprehending incumbent performance information, linking it to incumbent quality, and durably updating their beliefs about incumbent quality until the election. Since the political information that voters read, hear, and observe is often complex, voters may only superficially understand the information and therefore not meaningfully update their beliefs (Gomez and Wil- son 2006). Moreover, voters may reject novel information challenging their pre-existing beliefs (Zaller 1992). Durable belief updating about incumbent quality is also not trivial over longer periods; Zaller(1992) suggests that opinions reflect recently encountered propositions, rather than an internalized set of beliefs based on information encountered over sustained periods. Information that is credible and comprehensible to voters is most likely to alter behavior when it differs from voters’ prior beliefs about the incumbent’s relative suitability for office. Bayesian voters should alter the position and precision of their posterior beliefs about incumbent suitability most where a signal is precise, the signal deviates from voters’ priors beliefs, and voters’ prior beliefs are relatively imprecise (e.g. Kendall, Nannicini and Trebbi 2014).2 We thus expect that signals will increase (decrease) incumbent support and requests to the extent that the signal causes voters to favorably (unfavorably) update about incumbent quality relative to their prior beliefs, such that E[qI|s] > (<)E[qI].

2.2 Relevance of novel information

Even if novel incumbent performance information is credible and internalized, voters must per- ceive it as relevant to inﬂuence political accountability. By relevance, we refer to signals conveying information about incumbent quality. This study focuses on two aspects of relevance that could complement the provision of incumbent performance indicators: information about incumbent du- 2Voters may also update about challengers from incumbent performance, if such signals are correlated.

9 ties, and temporal performance benchmarks.

2.2.1 Information about incumbent duties

Information about incumbent duties has the potential to help voters hold incumbents to account upon receiving performance information in at least two ways. First, voters may only recognize performance information as relevant upon learning that politicians possess the capacity to influence such performance indicators. For example, Gottlieb(2016) finds that voters systematically underestimate government capacity in Mali, requiring civics training before identifying poorly- performing incumbents. Second, where multiple layers of government exist (e.g. a president and a legislature), providing information about the duties of specific types of politicians may enable voters to assign rewards and sanctions to their incumbent on the basis of performance relating to their duties(Powell and Whitten 1993). Both channels thus help voters to infer incumbent quality from performance signals and focus on outcomes that incumbents could feasibly influence, and could thus accentuate voter responses to performance indicators. While incumbent duties are often implicit when performance information is provided, or outlined alongside performance information (Gottlieb 2016), we explicitly separate between providing information about duties and performance to understand whether either individual piece or the combination of both is required to support political accountability.

2.2.2 Temporal performance benchmarks

A signal s that benchmarks incumbent performance can increase the accuracy of voters’ posterior beliefs, E[qI|s] (and potentially E[qC|s]), through two main channels. First, observing more than one performance signal helps voters to filter out common shocks influencing the performance of all agents in a given period or location (Aytaç 2018; Meyer and Vickers 1997). Without a benchmark, voters might imprecisely attribute strong/weak performance to incumbent quality. Second, a performance benchmark also enables voters to update about the absolute quality level of other

10 politicians that proxy for challenger politicians. Ultimately, both channels facilitate more accurate and precise beliefs about absolute and relative incumbent and challenger candidate quality—the key drivers of our simple models of political accountability. Whether cross-sectional or inter- temporal benchmarks are more useful in a particular context reflects the accuracy and uncertainty of voters’ prior beliefs about time- and unit-specific shocks and their magnitude. In practice, finding informative benchmarks has proved challenging. Most extant studies have focused on spatial benchmarks, which filter out common shocks affecting all incumbents equally (e.g. changing national budgets) by comparing incumbents representing different districts at the same time. However, cross-national, cross-legislator, or cross-municipality comparisons have generally failed to differentially alter voter beliefs and voting behavior beyond simply providing performance information about incumbents in developing contexts (Arias et al. 2018a; Campello and Zucco 2016). More positive cross-national findings regarding economic voting in developed democracies (Aytaç 2018; Kayser and Peress 2012) cannot pinpoint the effect of spatial benchmarks because they do not exogenously vary access to such information. Accordingly, we instead focus on temporal benchmarking, whereby the current incumbent’s performance is set alongside their immediate predecessor’s performance. This helps voters to filter out the effects of time-invariant features of districts—e.g. geographical constraints or demographic political importance—that affect all incumbents. We predict that abstracting from the influence of persistent district-specific characteristics will, relative to only providing incumbent performance information, increase the magnitude of favorable (unfavorable) voter updating when incumbent performance is above (below) voters’ prior belief (because benchmarked signals convey greater precision) and/or when the previous incumbent’s performance was below (above) voters’ prior belief (because benchmarked signals lower the common shock’s expected value). Appendix section A.1 formally derives these expectations.

11 2.3 Acting on internalized beliefs

Even if incumbent performance information causes voters to meaningfully update their beliefs, political accountability requires that voters ultimately act on such beliefs by engaging in electoral and non-electoral accountability-seeking behavior. This likely requires that several conditions hold. First, voters must connect their beliefs about the incumbent to their available actions, by applying decision rules akin to equations (1) and (2). However, Gomez and Wilson(2006) suggest that voters may lack the cognitive capacity to translate their beliefs into political action. In developing contexts, voters may lack the civic education to draw these connections (Gottlieb 2016). Second, in the case of voting, voters must be willing to vote on the basis of incumbent competence or policy alignment. However, beliefs about competence or policy alignment represent only one factor that could determine vote choices. Indeed, if the beliefs influenced by previous incumbent performance receive a low weight—wi—in voters’ calculus, even large changes in beliefs may not influence voting behavior. For example, voters may be unwilling to sanction good performers on other dimensions (Chauchard, Klasnjaˇ and Harish 2017), punish clientelistic candidates(Adida et al. 2017), or deviate from bloc voting agreements (Koter 2013). Third, the process of providing information could set in motion other forces that override the influence of voter beliefs on vote choice. Mass information dissemination campaigns could induce tacit or explicit voter coordination around particular candidates by providing public signals(Mor- ris and Shin 2002). Moreover, parties may seek to counteract, or adapt their electoral strategies, following performance revelations. Recent studies in developing contexts suggest that incumbents may respond to information campaigns by confiscating leaflets and threatening disseminators, increasing campaign spending, or altering vote buying strategies(Arias et al. 2018 b; Banerjee et al. 2011; Bidwell, Casey and Glennerster 2016). If such equilibrium responses affect vote choices, changes in the individual beliefs of voters may not ultimately translate into accountability-seeking behaviors.

12 3 Parliamentary accountability in Senegal

Senegal is one of Africa’s oldest and strongest democracies. It has generally experienced robust multi-party political competition—including peaceful transitions in 2000 and 2012, following fair democratic elections—since 1981, and is known for its vibrant civil society and freedom of press and expression. However, voters are often poorly informed about politics, partly due to the mass media’s limited reach. Senegal thus represents a developing context where voters might—given credible and relevant information—feasibly hold their government to account.

3.1 The Assemblee´ Nationale’s role

The Assemblee´ Nationale (Parliament) continues to play a limited role in democratic representation in Senegal’s hyper-presidential context (Thomas and Sissokho 2005), which has contributed to preventing full democratic consolidation (Beck 2012). Our focus on parliamentary deputies may thus inform the incentives required for deputies to become more accountable to voters. Assemblee´ Nationale deputies are elected by a mixed system, where competing political coali- tions form a national list and lists for each of Senegal’s 45 departments. In each majoritarian department-level race, the coalition winning most votes receives all seats allotted to the department. In 2017, 105 deputies were elected from 12 single and 33 multi-member departments and the remaining 60 seats were allotted in proportion to a coalition’s national vote share. In the 2012 legislative elections, president Macky Sall’s coalition—Benno Bokk Yakkar (BBY)—won 87 of 90 majoritarian departmental seats and approximately half the proportionally-allocated seats. Our study examines deputies elected from departmental majoritarian lists, as their distinct constituen- cies create stronger local accountability linkages than deputies elected from national lists. The primary constitutional role of elected deputies is amending and voting on laws drafted by government ministries. Deputies can also initiate laws themselves, although this is rare in practice (Thomas and Sissokho 2005). Rather, Senegal’s hyper-presidentialism and consolidated executive

13 power show in parliament.Few laws are rejected by the Assembl ee´ Nationale, and its role as a check on executive power is often questioned by civil society. Parliamentary deputies themselves echoed these sentiments in our interviews. Nonetheless, deputies can—and do—affect legislative decisions through their parliamentary duties. First, deputies can serve on one or more of the 11 parliamentary committees. Leaders of committees are considered influential in the amending of laws, as these committees meet in closed sessions and make recommendations and amendments to a ministry’s bills before they are debated in open plenary sessions. The finance committee’s role in national budget deliberations is particularly influential (Thomas and Sissokho 2005). Second, deputies can submit written or oral questions to the government, which relevant ministers answer in open sessions. One deputy explained that participating in these debates and ordinary plenary sessions enables deputies to de- fend and publicize their constituents’ interests. Third, although deputies do not receive specific funds for local development projects, they are widely believed to influence the allocation of local projects and government transfers by lobbying ministers. Constituents often see local development projects as their primary responsibility, while one deputy described the biggest difference between good and bad deputies as their “capacity to lobby successfully.” Various deputies noted that effective lobbying revolved around access to government ministers allocating budgets.

3.2 Voter engagement with parliamentary elections and deputies

Voter turnout in Senegal is typically around 60% for presidential elections, and reached 54% in the 2017 parliamentary elections. Although Senegal’s rising participation rates fall below the mean within sub-Saharan Africa (see Appendix Figures A1 and A2), conﬁdence in the quality of democracy within Senegal is trending upwards. In Senegal’s 2016 Afrobarometer round, 87% of respondents viewed Senegal as a democracy, with 64% reporting that they are satisﬁed with the way democracy functions in the country. Furthermore, while direct citizen interaction with deputies is rare (only 9% of respondents in our sample had contacted a deputy within the last 12 months) and

14 Village or community of deputy 14

Ethnicity or religion of deputy 3

Education or profession of deputy 7

Party of deputy 4

Political experience of deputy 7

Ability of deputy to amend national laws and budgets 8

Ability of deputy to lobby for projects/transfers to department 46

Campaign promises of deputy 7

Electoral gifts distributed by deputy 2

0 10 20 30 40 50 Percentage of voters citing factor as the most important in determining vote choice

Figure 1: The most important factor driving an individuals’ vote choices

voters are pessimistic about whether deputies listen to voters and respond to requests, interactions with party ofﬁcials and brokers who report to deputies are relatively common. These patterns of beliefs and behaviors align with perceptions of low accountability within the Assemblee´ Nationale, but also suggest that voters retain faith that informed engagement could be effective. Indeed, our baseline survey data indicates that many voters seek to select deputies on the basis of expected performance on local budgetary issues. Figure1 indicates that 46% of voters claim that a deputy’s potential to lobby for projects and transfers beneﬁting their department is the most important factor driving their vote choice. Fewer voters regard national-level policy engagement as key. Moreover, when asked whether they would prefer a hypothetical deputy seeking to improve their constituents’ welfare by focusing on improving national policies and budgets or another hypothetical deputy solving everyday problems in the communities in their constituency, 71% of

15 respondents favored the locally-oriented politician. However, actually holding deputies to account has proved challenging for several reasons. First, voters often lack the information needed to select the best-performing deputies. Only 35% of voters in our sample could name at least one of their parliamentary representatives, and—despite the party-centric nature of Senegalese parliamentary politics—only 61% could correctly identify the(majoritarian) incumbent party in their department. Moreover, Figure5 below shows that voters’ prior beliefs are uncorrelated with the incumbent performance metrics that our treatment provides. This paucity of reliable information likely reflects a dependence on radio broadcasts (only 18% and 56% of voters in our sample ever obtain news via newspapers and national-level television, respectively), and the fact that political parties are the main source of information in rural communities. Second, attempts to hold deputies to account often compete against clientelistic incentives and coordinated group voting pushing vote choices in different directions. Clientelistic returns, which reflect the home village of the candidate as well as their ethnicity and religion, heavily influence rural voters. Furthermore, the Sufi networks—to which the majority of Senegalese citizens belong—intimately shape the political process across the country and moderate access to political resources (Villalon´ 1995). Political parties also attempt to influence rural vote choice via village chiefs as intermediaries (Gottlieb 2017; Koter 2013) and village-level meetings.

4 Research design

To examine whether and how Senegalese voters seek to hold elected deputies to account, we designed an information dissemination campaign to personally deliver and explain incumbent performance information to voters prior to the July 30, 2017 parliamentary elections. Our ﬁeld experiment randomized several key components of the information’s content across villages, and deployed a panel study tracking voter beliefs and (intended) actions before the treatments were

16 delivered, immediately after their delivery, and a month after the election. This design enables us to trace in unprecedented detail the chain of events between voters’ receipt of information about incumbent performance and political accountability.

4.1 Sample selection

We conducted our study in the ﬁve departments shown in Figure2: Fatick, Foundiougne, Kanel, Oussouye, and Ranerou´ Ferlo. These departments with one or two majoritarian deputies were selected because single incumbents were seeking re-election for the ﬁrst time, with the exception of Kanel where two incumbents both sought re-election.3 Within these departments, we selected 450 rural villages containing 200-4,000 people for our sample. Appendix Table A1 shows that our sample of rural villages is notably less educated and developed than the national average. Within each village, we aimed to survey nine registered voters aged 20-38 that had lived in the village prior to the age of primary school enrollment.4 The only logistical restriction was that respondents must have a cellphone number, which virtually all young Senegalese satisfy. Respon- dents were generally selected from the center of the village after receiving the chief’s approval.

4.2 Information treatments

Our treatments entailed distributing and explaining scorecards detailing combinations of legislator duties, current incumbent performance, and previous incumbent legislator performance in the month preceding the 2017 election. Regarding legislator duties, we highlighted that legislators can: (1) serve on one or several of the 11 parliamentary committees, where they can influence legisla- tion and budgets in addition to taking on influential leadership conditions; (2) represent constituent 3Appendix section A.4 provides further sampling information. 4This sampling strategy reflected our intention to identify differential treatment effects by educational attainment, leveraging cross-cohort variation in the effect of a 2002 secondary school construction program as an instrument. We do not examine such cross-cutting variation because access to new schools did not robustly increase schooling in our sample.

17 Figure 2: Distribution of departments across Senegal

interests in parliamentary debates; and (3) lobby government ministers to increase transfers to their departments for specific projects. Regarding incumbent legislator performance, we provided five nationally- and locally-oriented measures of performance in office over the full five-year electoral cycle that relate to deputies’ primary duties: (1) the committees to which they belonged; (2) positions of leadership within parliament; (3) the number of parliamentary debates in which they participated; (4) the number of local projects budgeted for their department in parliamentary documents; and (5) the number and (inflation-adjusted) per capita per year value of ministry transfers received by the department, decomposed by transfer category.5 The latter two locally-oriented measures are department-level 5Annual transfer data was available from 2010 to 2016. To ensure comparability between current and previous incumbents and across departments, we converted ministry transfers to a per year basis, adjusted for inflation, and normalized by 2013 population size. Transfers affecting multiple departments were distributed

18 Table 1: Factorial treatment conditions

Performance information provided: None Incumbent Benchmark Duties information None 75 villages [pure control] 75 villages 75 villages provided: Duties 75 villages 75 villages 75 villages

variables that do not vary across deputies in multi-member departments. All deputy performance data was obtained from either the Assemblee´ Nationale or a government department. The accuracy, relevance, and impartiality of the information on deputy duties and performance was validated by the head of legislative services at the Assemblee´ Nationale, the librarians and archivists at the Assemblee´ Nationale, and several active and former deputies. Based on the performance metrics just described, we used a 2 × 3 factorial design to randomly assign villages to one of the six experimental conditions in Table1. Treatment conditions vary along two dimensions of content, with the top-left element in the matrix constituting a pure control group where no intervention occurred. First, the “duties” dimension informed voters of the three main functions (enumerated above) that their parliamentary deputies can perform. Second, the “performance” dimension varies whether voters receive “incumbent” information relating to the/an incumbent representative’s performance on the ﬁve measures described above or “benchmark” information additionally providing the same information pertaining to the performance of the department’s previous incumbent representative. We maximized treatment homogeneity by showing only one current deputy and one benchmark deputy in multi-member departments. Figure3 reports the distribution of the main performance metrics provided, where each point represents a current incumbent-previous incumbent pairing.6 Points above the 45o line are cases where the current incumbent outperformed the previous incumbent.In general, the current incumbent outperformed preceding incumbents within the same department, especially with respect to in proportion to each department’s 2013 population. 6Although there are eight pairings across our ﬁve departments, some comparisons are exact duplicates.

19 Committee memberships Leaderships positions Debates participated in Kanel Kanel Oussouye Kanel Oussouye 3 1 10

Fatick Fatick 8 .8 Oussouye 2 Fatick FatickFoundiougne 6 .6 4 .4

1 Ranérou Ferlo Kanel Kanel 2 Current incumbent Current incumbent Current incumbent .2 Ranérou Ferlo Kanel Foundiougne Kanel Kanel

0 0 Ranérou Ferlo 0 FatickFoundiougne Foundiougne 0 1 2 3 0 .2 .4 .6 .8 1 0 2 4 6 8 10 Previous incumbent Previous incumbent Previous incumbent

Local projects Number of transfers Transfer value per capita-year 6

15 Fatick Fatick

Fatick 15000 4 10 10000

5 2 Kanel

5000 Oussouye Foundiougne Current incumbent Kanel Current incumbent Kanel Oussouye Current incumbent Ranérou Ferlo Oussouye Ranérou Ferlo Ranérou Ferlo

0 0 Foundiougne 0 Foundiougne 0 5 10 15 0 1 2 3 0 2000 4000 6000 Previous incumbent Previous incumbent Previous incumbent

Figure 3: Distribution of treatment information across departments (45o line in gray) Note: Cases within departments where previous incumbents performed identically are not duplicated. debates, projects, and transfers. On this basis, we expected that our information would increase voters’ favorability towards current incumbents, on average, across departments. Each information treatment was distributed to voters through leaflets like the one in Figure4. 7 The leaflets were created by a graphic artist and designed in partnership with our partnering civil association, LEGS-Africa. At the top of each leaflet variant is the LEGS-Africa logo alongside a statement that the organization is non-partisan, while a description of data sources and (redacted) contact information was provided at the bottom. The example in Figure4 from Oussouye depicts the duties and benchmark treatment variant; this contains the maximum amount of information that could be provided. The three paragraphs below the LEGS-Africa logo were provided to all participants receiving a “duties” variant. The current incumbent performance information on the left of the remainder of the leaflet was provided to participants receiving the “incumbent” variant, 7Appendix section A.3 shows all leaflet variants.

20 Figure 4: Example of “duties + benchmark” treatment in Oussouye

while the performance information on the left and right was provided to participants receiving the “benchmark” variant. We piloted the leaflet to ensure comprehensibility. The leaflet was delivered and explained in person to respondents as part of our baseline survey on behalf of LEGS-Africa. Respondents were informed that we were surveying several thousand voters across the country. Our enumerators gave each voter several minutes to read the leaflet and then walked them through the meaning of each component in the respondent’s local language, since only 61% of respondents could read French. On average, treatment delivery took around five minutes. Our training ensured that enumerators—who were generally recent university graduates— themselves understood and could clearly explain the leaflets’ content. After receiving treatment,

21 82% of treated respondents reported that the leaflet came from an NGO, while only 19% believed that it originated from the current deputy or national government. Our intervention is thus heavier-handed than most information dissemination campaigns. Pre- vious campaigns have typically posted fliers, sent SMS messages, or leveraged media opportunities (e.g. Banerjee et al. 2011; Chong et al. 2015; Dunning et al. forthcoming; Humphreys and We- instein 2012). In contrast, ensuring that voters received and understood the information allows us to focus on belief updating and voter behavior in a field experimental context that sidesteps the concern that dissemination technologies and voter consumption decisions have limited the reach of previous information campaigns.

4.3 Information provision randomization

Leaﬂet treatment conditions were randomly assigned at the village level to mitigate contamination arising from within-village spillovers and “John Henry” effects. Speciﬁcally, we constructed 75 blocks, each containing six similar villages from within the same department.8 Within each block, we used complete randomization to assign one village to each experimental condition. In multi-member departments, we used complete randomization to assign an incumbent-previous incumbent pair to each block. Consequently, all villages within a block were eligible to receive the same information and faced the same parliamentary election.

4.4 Data collection

To examine the effects of information treatments on political accountability outcomes, we designed a two-wave panel survey and collected polling station-level electoral returns. The baseline survey was conducted in person between July 4 and July 29, and our treatments were administered after 8After stratifying by department, village similarity was determined by closeness across 24 pre-treatment covariates. We used the R package “blockTools” to assign blocks within departments, based on a greedy Mahalanobis distance-based algorithm.

22 enumerators collected respondents’ baseline beliefs, characteristics, previous behaviors, and intentions. The shorter post-election survey was conducted by telephone between August 4 and August 26. We also mapped each village’s electoral returns to its associated polling station based on the electoral register.

4.4.1 Measurement of primary outcomes

Our primary outcomes focus on voter beliefs about the quality of current and previous incumbent performance in office, willingness to hold incumbents to account, and the extent to which such willingness translates into electoral action and non-electoral requests. Crucially, we measure self-reported attitudes and intended behaviors and behavioral outcomes both immediately after treatment was administered during the pre-election survey, as well as after the election. We focus on two classes of attitudinal and intentional outcomes in our panel surveys. First, we measured three main types of voter beliefs on five-point scales from “very bad” (1) to “very good” (5): how well the incumbent has done overall since they were elected in 2012; how such performance compares to the previous incumbent; and how respondents think that current incumbent seeking re-election would do if re-elected.9 For each variable, we also elicited the strength of voters’ assessment on a ten-point scale ranging from “not at all certain” (1) to “completely certain” (10).10 Each belief, and its associated certainty, was elicited in the baseline survey before and after information treatments were delivered for both treated and control respondents. The first two questions were repeated at endline. Second, we also elicited self-reported (intended) vote choices: vote intention before and after treatment in the baseline survey, and self-reported turnout and vote choice at endline. We use indicators for respondents stating that they would or did vote for the incumbent BBY coalition(to which all incumbents in our departments belonged). We also address self-reporting concerns by 9“Don’t know” responses are coded at the mid-level of the scale. 10“Don’t know” responses to scales eliciting voter beliefs are coded as the lowest level of certainty.

23 only counting votes as valid where the respondent correctly recalled both that the ballot contained a picture of a candidate and a logo and the color of the ballot corresponding to the party that they reported voting for. At baseline, we also elicited certainty about intended vote choice on a ten-point scale. While these outcomes permit a nuanced understanding of voter beliefs and actions, we use several behavioral outcomes that are not vulnerable to social desirability biases. First, we use ofﬁcial electoral returns to calculate incumbent party vote share—both as a share of turnout and registered voters—at the polling station corresponding to each village in our sample. Although we expect noisy estimates because fewer than 2% of voters were treated within polling stations, information could spread within our tight-knit set of rural villages. Second, we created several opportunities for voters to make costly requests of the incumbent. In the baseline survey, we offered respondents the opportunity to: request a poster from the party or candidate they intended to vote for, and/or other parties or candidates; request a visit from the party or candidate they intended to vote for, and/or other parties or candidates; and sign up to be contacted to express their views to the party or candidate they intended to vote for, and/or other parties or candidates.11 The ﬁrst measure captures another form of expressing support for the incumbent. The latter two measures capture non-electoral forms of accountability-seeking political engagement, akin to Aker, Collier and Vicente(2017) and Mvukiyehe and Samii(2015). In the post-election endline survey, we again offered respondents the opportunity to request a visit from and sign up to be contacted by the winning candidate, who was the incumbent in all instances. After observing high willingness to engage in the preceding actions at baseline, at endline we also created a hotline where respondents could send text messages (costing around US$0.01) or leave voicemails (costing around US$0.25) requesting to be contacted by the winning candidate. We measure this by linking telephone numbers to the respondent.12 11For each action, the voter’s name and village were shared with the party. 12Respondents were informed that all requests and messages sent to the hotline would be submitted to candidates on behalf of LEGS-Africa.

24 Given the large number of outcomes—which engender concerns about multiple comparisons and noise in speciﬁc variables—we also combine similar individual-level outcomes using indexes. Separately within baseline and endline panel waves, we created inverse-covariance weighted (ICW) indexes to summarize two groups of items (see Appendix A.6 for details): self-reported incumbent support, i.e. all attitudinal and (intended) vote choice outcomes; and behavioral indicators of support for and requests from incumbents. We selected the ICW approach to account for correlation among items, but report an average across standardized items in Appendix section A.8.4. By standardizing all indexes with respect to the control group, effect magnitudes represent standard deviation changes in control group outcomes.

4.4.2 Compliance

As Appendix A.5 describes in greater detail, we encountered two minor forms of data missingness. First, we could not access 7 of our 450 villages. However, since villages were identically surveyed by enumerators and not informed of treatment status in advance, our inability to start conducting surveys in these villages was unaffected by treatment assignment. Second, 4% of respondents attrited between baseline and endline surveys, but not differentially so across treatment conditions (see Appendix Table A4). Unsurprisingly, our baseline and endline balance tests in Tables A2 and A3 suggest that the randomization’s integrity was maintained.

4.5 Estimation

Following our pre-analysis plan, our fully-saturated speciﬁcation estimates the average treatment effect of different informational components of the leaﬂet using baseline regressions of the form:13

Yivb = β1dutiesv + β2incumbentv + β3benchmarkv + β4 (incumbentv × dutiesv)

baseline +β5 (benchmarkv × dutiesv) + αYivb + ηb + θe + εivb, (3) 13Appendix section A.7 justiﬁes minor deviations from the pre-analysis plan.

25 where Yivb is an outcome for individual i in village v in randomization block b, ηb are randomization

block fixed effects, and θe are survey enumerator fixed effects. Wherever possible, the outcome’s baseline pre-treatment baseline counterpart Yivb is included as a control to increase efficiency. For polling station-level outcomes, we replace the iv subscript with a p subscript. To recover the village-level average treatment effect, all survey-based regressions are weighted by the inverse of the number of respondents in the corresponding baseline or endline survey. Standards errors are clustered by village, asterisks denote significance tests based on one-sided t tests for pre-specified hypotheses, and daggers indicate significance tests based on two-sided t tests. Two-sided t tests are applied to hypotheses that were not pre-specified or pre-specified without a hypothesized direction, and to estimates in the opposite direction to our pre-specified hypothesis. To test additional hypotheses underpinning the accountability process, we also estimate heterogeneity in treatment effects by the content of the information provided and other predetermined covariates such as a voter’s prior beliefs or the importance of content for a voter’s decision-making. Taking the simpler case that pools across the duties treatment dimension, which we do frequently in the analysis, we estimate specifications of the form:

Yivb = β1incumbentv + β2benchmarkv + β3 (incumbentv × Xiv)

baseline +β4 (benchmarkv × Xiv) + γXiv + αYivb + ηb + θe + εivb, (4)

where Xiv is a predetermined covariate.

5 Immediate effects of information provision

We start by examining immediate voter beliefs, intentions, and behavior in response to treatment in the baseline survey. This enables us to examine the mechanisms linking incumbent performance information and electoral and off-election accountability without contamination from subsequent

26 Table 2: Effects of information treatments on leaﬂet comprehension (baseline survey)

Respondent correctly states...... number of ...deputies ...number of ...number of parliamentary lack incumbent’s previous committees department local incumbent’s fund projects debates (1) (2) (3) (4) Duties 0.663*** 0.459*** (0.024) (0.027) Incumbent 0.729*** (0.027) Benchmark 0.717*** 0.461*** (0.027) (0.034)

Two-sided null: Incumbent = Benchmark (p value) 0.60 Observations 3,999 3,999 3,999 3,999 Outcome range {0,1}{0,1}{0,1}{0,1} Control outcome mean 0.05 0.14 0.08 0.07 Control outcome std. dev. 0.22 0.35 0.26 0.26

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects. All observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. interactions with other voters or political actors.

5.1 Voters understand the information contained in the leaﬂets

Although we maximized the probability that voters comprehended the information, our first step is to verify that voters indeed understood the information provided. We asked respondents four factual multiple-choice questions pertaining to different components of the information contained in our leaflets—the number of committees in the Assemblee´ Nationale (four choices), the lack of a department-specific development fund available to deputies (true or false), the number of projects in the department initiated over the current incumbent’s term (four options), and the number of debates that the previous incumbent participated over their term in office (four options). Our detailed visual and verbal explanations successfully ensured that most treated respondents

27 immediately recalled the information contained in the leaflet. Columns (1) and (2) of Table2 indicate that receiving any duties information increased the proportion of respondents correctly identifying the number of parliamentary committees from 5% to 71% and that deputies lack individual funds for department projects from 14% to 60%. Column (3) demonstrates that incumbent performance information, whether on its own or with a benchmark, increased the proportion of respondents correctly identifying the number of local projects received under the current incumbent from 8% to 79%. Finally, column (4) shows that the benchmark leaflet increased correct answers regarding the number of debates that the previous incumbent participated in from 7% to over 53%. At least when personally distributing and explaining informational leaflets to voters, voters’ inability to comprehend the information provided—even in a low-education environment—does not seem to be a bottleneck in the political accountability process.

5.2 Voters update their beliefs in a Bayesian manner

Given that information may not be perceived as credible or relevant, voter comprehension does not necessarily imply that our leaflets’ information would alter voter beliefs. Furthermore, regurgitat- ing information may not imply internalization. However, columns (1)-(3) in Table3 show that voters favorably updated their posterior beliefs on average immediately after receiving the leaflet. Panel A indicates that voters who received incumbent-only, and especially benchmarked, performance information experienced around a third of a standard deviation increase in favorability toward their incumbent deputy across each assessment of their suitability for office. Appendix Table A5 shows that treatment caused around 25% of voters to update favorably, relative to their pre-treatment prior beliefs; in contrast, treatment did not cause a significant number of voters to update unfavorably. Further demonstrating that voters’ beliefs were meaningfully affected, Appendix Table A7 shows that the leaflets increased treated voters’ reported certainty in their beliefs about the incumbent’s current and future performance by around half a point—or nearly 0.2 standard deviations—on our 10-point scale. These

28 Table 3: Average effects of information treatments on beliefs about incumbent performance, intention to vote for the incumbent, and requests from the incumbent (baseline survey)

Benchmark - Incumbent 0.076** 0.089** 0.066** 0.005 -0.002 0.054** -0.022 -0.008 -0.034 (0.040) (0.041) (0.035) (0.010) (0.013) (0.030) (0.013) (0.014) (0.029) Observations 3,942 3,932 3,928 3,999 3,891 3,998 3,999 3,998 3,997 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1} [-2.4,1.9] {0,1}{0,1} [-1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.59 0.67 0.02 0.70 0.70 0.01 Control outcome std. dev. 1.07 0.90 1.09 0.49 0.47 1.01 0.46 0.46 1.01

Notes: See Table2. All specifications include a lagged dependent variable as a control; in columns (5)-(9), pre-treatment incumbent vote is used as a proxy. significant effects suggest that voters regarded the information as credible and had systematically underestimated incumbents. The large magnitude of favorable updating is consistent with voters being poorly informed—Figure5 reports a somewhat negative correlation between performance and pre-treatment beliefs—and receiving performance indicators that generally exceeded the plausible anchor of the previous incumbents’ performance (see Figure3). Conversely, information about deputies’ duties did not systematically affect voter evaluations— whether on its own or in conjunction with performance indicators. This indicates that voters may not need additional information about incumbent duties to feel confident in using the incumbent performance indicators that voters regard as most relevant to evaluate incumbents. The limited effect of information about duties is a consistent pattern throughout this study, and may reflect the fact that the plurality of voters care most about deputies bringing projects back to their department

29 Very good

Good

Neither

Bad

Very bad

Village mean evaluation of incumbent overall performance -3 -2 -1 0 1 Overall performance (ICW)

Figure 5: Correlation between (village mean) pre-treatment incumbent overall performance assessments and incumbent deputy performance Notes: All points are jittered around the eight deputy performance points. Point size reflects the number of respondents in the village. The black line is a linear regression slope. and thus already believed that incumbents were capable of influencing this process. Henceforth, we focus primarily on the comparison between benchmarked and non-benchmarked incumbent performance information. Panel B—which pools the control and duties conditions—reports significant differential effects of within-department benchmarks. Consistent with voters using the previous incumbents’ comparatively poor performance to infer that the current incumbents are performing better than normal for their department, columns (1) and (3) demonstrate that the benchmark increases voters’ overall and prospective posterior assessments more than receiving incumbent-only information. Suggesting that benchmarked information increased the weight attached to the signal of incumbent performance, column (2) also reports that the benchmarked information had a larger effect on the relative comparison between current and previous incumbents than incumbent-only information.14 The pre-specified one-sided tests of the difference between incumbent-only and bench- 14Appendix Table A9 shows that previous incumbent performance benchmarks did not systematically influence prospective challenger performance evaluations. It is thus unlikely that previous incumbent per-

30 marked information at the foot of each column reject the null hypothesis that the benchmark is no more likely to increase incumbent appraisals than incumbent-only information treatment at the 5% level. Furthermore, Appendix Table A7 shows that the benchmark was particularly important in increasing belief precision, consistent with voters regarding it as a more precise signal for drawing both absolute and relative comparisons. In contrast with extant experimental findings that cross- sectional benchmarks do not provide important additional information, we thus find that temporal benchmarks can significantly influence beliefs by adding a yardstick for appraising incumbent performance. Finally, the heterogeneous effect results in Table4 show that the interactions between information provision, performance indicator content, and pre-treatment beliefs are consistent with three central predictions of Bayesian updating. Each interacting variable is standardized within the control group. First, panel A shows that voters updated significantly more favorably about the incumbent when the leaflet indicated higher performance—an ICW scale combining our six reported performance indicators. Panel B reports similar results when incumbent performance is instead weighted by whether voters stated that national or local factors were most important to them before receiving treatment.15 Panel C further shows that voters’ increased favorability toward the incumbent almost entirely reflects the local projects and transfers components of the performance index, suggesting that—consistent with their own claims—voters are mostly concerned with the resources that deputies bring to their departments. As in Adida et al.(2017), appearing to work on national or parliamentary issues, if anything, sends a negative signal.16 Second, panel D reports that the voters with least favorable prior beliefs updated most favorably about the incumbent. Third, panel E shows that favorable updating was less pronounced among respondents with more formance falling below prior expectations drives the differential effects of benchmarked information. 15The relevance-weighted performance index assigns a respondent the national, local, or both performance indicators corresponding to whether they listed these among the three most important factors in determining their vote choice. An indicator for respondents not listing national or local performance as important is also interacted with treatments. 16Unreported two-tailed F tests confirm that these heterogeneous effects are statistically distinguishable.

31 Table 4: Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (baseline survey)

Incumbent evaluation outcomes Accountability-seeking outcomes Incumbent Relative Prospective Incumbent Request Incumbent Request Request Accountability overall performance incumbent vote incumbent evaluation incumbent incumbent seeking performance (v. previous) performance intention poster index (ICW) visit conversation index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) (9) Panel A: Heterogeneity by (standardized) reported performance level Incumbent 0.367*** 0.297*** 0.269*** 0.031*** 0.033*** 0.266*** 0.041*** 0.030*** 0.079*** (0.044) (0.039) (0.038) (0.009) (0.014) (0.030) (0.013) (0.013) (0.028) Incumbent × Overall performance (ICW) 0.244*** 0.198*** 0.185*** 0.022*** 0.002 0.114*** 0.005 0.005 0.011 (0.042) (0.040) (0.038) (0.008) (0.012) (0.031) (0.012) (0.014) (0.027) Benchmark 0.440*** 0.381*** 0.332*** 0.036*** 0.030*** 0.319*** 0.018* 0.022** 0.045* (0.048) (0.052) (0.042) (0.011) (0.013) (0.038) (0.013) (0.013) (0.028) Benchmark × Overall performance (ICW) 0.220*** 0.149*** 0.154*** 0.034*** -0.002 0.116*** 0.005 0.002 0.008 (0.038) (0.044) (0.041) (0.009) (0.012) (0.032) (0.012) (0.012) (0.025) Panel B: Heterogeneity by (standardized) relevance-weighted reported performance level Incumbent 0.389*** 0.320*** 0.299*** 0.038*** 0.041*** 0.294*** 0.048*** 0.041*** 0.100*** (0.044) (0.041) (0.037) (0.011) (0.016) (0.033) (0.015) (0.015) (0.032) Incumbent × Relevant performance (ICW) 0.267*** 0.251*** 0.213*** 0.032*** 0.001 0.161*** -0.003 0.002 -0.001 (0.044) (0.042) (0.043) (0.013) (0.017) (0.036) (0.016) (0.016) (0.035) Benchmark 0.474*** 0.400*** 0.348*** 0.041*** 0.038*** 0.337*** 0.024* 0.029** 0.060** (0.052) (0.058) (0.045) (0.013) (0.015) (0.043) (0.015) (0.016) (0.033) Benchmark × Relevant performance (ICW) 0.194*** 0.145*** 0.131*** 0.031** 0.011 0.121*** -0.015 0.012 -0.004 32 (0.046) (0.044) (0.042) (0.016) (0.018) (0.039) (0.018) (0.020) (0.041) Panel C: Heterogeneity by (standardized) local and national reported performance level Incumbent 0.391*** 0.309*** 0.286*** 0.034*** 0.033*** 0.282*** 0.047*** 0.032*** 0.089*** (0.047) (0.041) (0.041) (0.010) (0.015) (0.031) (0.013) (0.013) (0.029) Incumbent × National performance (ICW) 0.001 0.050 -0.023 -0.001 0.008 -0.014 0.003 0.019 0.025 (0.059) (0.054) (0.047) (0.009) (0.018) (0.032) (0.018) (0.016) (0.037) Incumbent × Local performance (ICW) 0.295*** 0.220*** 0.246*** 0.023** -0.004 0.191*** 0.006 -0.006 0.000 (0.058) (0.051) (0.051) (0.012) (0.019) (0.038) (0.018) (0.016) (0.036) Benchmark 0.450*** 0.400*** 0.345*** 0.040*** 0.034*** 0.336*** 0.023** 0.028** 0.057** (0.049) (0.045) (0.042) (0.012) (0.014) (0.035) (0.014) (0.014) (0.030) Benchmark × National performance (ICW) -0.048 -0.157+ -0.101+ 0.004 -0.010 -0.104+ 0.002 -0.000 0.002 (0.068) (0.093) (0.053) (0.013) (0.018) (0.059) (0.018) (0.017) (0.038) Benchmark × Local performance (ICW) 0.206*** 0.256*** 0.214*** 0.031*** 0.017 0.208*** -0.002 0.017 0.016 (0.061) (0.073) (0.050) (0.016) (0.019) (0.052) (0.018) (0.019) (0.041) Observations 3,942 3,932 3,928 3,999 3,998 3,890 3,999 3,998 3,998 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1} [-2.4,1.9] {0,1}{0,1} [-1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.59 0.67 0.02 0.70 0.70 0.01 Control outcome std. dev. 1.07 0.90 1.09 0.49 0.47 1.01 0.46 0.46 1.01 Overall performance (ICW) range [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] Relevant performance (ICW) range [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] National performance (ICW) range [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] Local performance (ICW) range [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21]

Notes: See Table3. Lower-order interaction terms are included but not shown. Panel B also includes the interaction between incumbent and benchmark treatments and an indicator for respondents that did not regard local or national performance as one of the top three most important factors in determining their vote choice. (Continued...) Table4 (continued): Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (baseline survey)

Incumbent evaluation outcomes Accountability-seeking outcomes Incumbent Relative Prospective Incumbent Request Incumbent Request Request Accountability overall performance incumbent vote incumbent evaluation incumbent incumbent seeking performance (v. previous) performance intention poster index (ICW) visit conversation index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) (9) Panel D: Heterogeneity by (standardized) prior belief level Incumbent 0.355*** 0.292*** 0.264*** 0.032*** 0.032*** 0.260*** 0.041*** 0.031*** 0.081*** (0.047) (0.043) (0.041) (0.009) (0.014) (0.033) (0.013) (0.013) (0.028) Incumbent × Prior index (ICW) -0.132*** -0.086*** -0.097*** -0.041*** -0.031*** -0.113*** -0.029*** -0.022** -0.058*** (0.040) (0.035) (0.037) (0.011) (0.014) (0.033) (0.013) (0.013) (0.029) Benchmark 0.435*** 0.378*** 0.331*** 0.038*** 0.029*** 0.315*** 0.018* 0.024** 0.047* (0.050) (0.052) (0.043) (0.012) (0.014) (0.039) (0.014) (0.014) (0.029) Benchmark × Prior index (ICW) -0.163*** -0.083** -0.056* -0.029*** -0.023* -0.091*** -0.006 -0.008 -0.016 (0.043) (0.044) (0.041) (0.011) (0.015) (0.035) (0.015) (0.014) (0.032) Panel E: Heterogeneity by (standardized) prior belief precision Incumbent 0.362*** 0.297*** 0.280*** 0.039*** 0.026** 0.267*** 0.034*** 0.033*** 0.075*** (0.047) (0.045) (0.042) (0.010) (0.014) (0.034) (0.013) (0.013) (0.028) Incumbent × Prior precision index (ICW) -0.005 0.003 -0.041* -0.031*** -0.023* -0.048** -0.012 -0.009 -0.023 (0.038) (0.030) (0.030) (0.010) (0.016) (0.026) (0.013) (0.015) (0.030) Benchmark 0.440*** 0.388*** 0.352*** 0.044*** 0.028** 0.328*** 0.018* 0.029*** 0.052** (0.052) (0.054) (0.043) (0.012) (0.014) (0.040) (0.014) (0.014) (0.029) Benchmark × Prior precision index (ICW) -0.029 -0.055** -0.043 -0.025*** -0.030** -0.078*** -0.008 -0.017 -0.028 (0.038) (0.030) (0.038) (0.012) (0.016) (0.026) (0.017) (0.016) (0.035) 33 Panel F: Heterogeneity by importance of performance in determining vote choice Incumbent 0.355*** 0.284*** 0.261*** 0.031*** 0.032*** 0.259*** 0.039*** 0.029*** 0.077*** (0.048) (0.043) (0.040) (0.009) (0.014) (0.032) (0.013) (0.013) (0.028) Incumbent × Performance most important -0.007 -0.008 0.009 0.020*** 0.013 0.018 0.021** 0.019* 0.046** (0.033) (0.030) (0.030) (0.009) (0.013) (0.023) (0.012) (0.012) (0.026) Benchmark 0.431*** 0.373*** 0.328*** 0.036*** 0.030*** 0.313*** 0.017 0.022* 0.043* (0.051) (0.052) (0.043) (0.012) (0.013) (0.038) (0.013) (0.013) (0.029) Benchmark × Performance most important -0.001 -0.023 -0.010 0.016** 0.030*** 0.019 0.024** 0.019* 0.048** (0.031) (0.028) (0.029) (0.009) (0.014) (0.022) (0.012) (0.013) (0.027) Panel G: Heterogeneity by preference for locally-oriented deputies Incumbent 0.370*** 0.192*** 0.198*** -0.018 0.039* 0.210*** 0.041* 0.035 0.086* (0.079) (0.069) (0.065) (0.024) (0.029) (0.056) (0.027) (0.028) (0.059) Incumbent × Prefer locally-oriented deputies -0.019 0.130** 0.089* 0.068*** -0.011 0.068 -0.002 -0.009 -0.013 (0.075) (0.069) (0.066) (0.028) (0.034) (0.063) (0.031) (0.032) (0.068) Benchmark 0.502*** 0.286*** 0.302*** 0.014 0.040 0.298*** 0.028 0.028 0.063 (0.077) (0.088) (0.067) (0.023) (0.031) (0.067) (0.030) (0.031) (0.066) Benchmark × Prefer locally-oriented deputies -0.098 0.123* 0.036 0.029 -0.015 0.021 -0.016 -0.010 -0.029 (0.073) (0.080) (0.067) (0.024) (0.037) (0.063) (0.033) (0.033) (0.072) Observations 3,908 3,906 3,905 3,922 3,921 3,890 3,922 3,921 3,921 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1} [-2.4,1.9] {0,1}{0,1} [-1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.60 0.68 0.02 0.70 0.70 0.02 Control outcome std. dev. 1.08 0.90 1.10 0.49 0.47 1.01 0.46 0.46 1.00 Prior index (ICW) range [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] Prior precision (ICW) range [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] Performance most important range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Prefer locally-oriented deputies range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}

Notes: See Table3. Lower-order interaction terms are included but not shown. precise prior beliefs. Together, these ﬁndings demonstrate sophisticated voter learning about incumbent quality, suggesting that—at least in this context—neither cognitive capacity nor resistance to information impede political accountability.

5.3 Performance information alters vote intentions

We next examine whether these changes in beliefs translate into an intention to re-elect incumbent deputies. Column (4) of Table3 indicates that incumbent performance information increased the probability of intending to vote for the incumbent BBY candidate by 3 percentage points.17 This effect is again larger for benchmarked than incumbent-only information, although not significantly so. We also again find no evidence to suggest that incumbent duties information moderates the effects of performance information on vote intention. Furthermore, increased incumbent vote intentions vary according to key Bayesian predictions of the electoral accountability logic. First, the heterogeneous effects in column (4) of panels A-E of Table4 show that changes in vote intention mirror changes in voter beliefs, such that treatments revealing better performance are substantially more likely to increase incumbent voting. As with belief updating, a standard deviation change in overall reported performance produced an effect equivalent to the average treatment effect. Second, panel F demonstrates that the 54% of respondents who (before treatment) ranked the incumbent’s ability to amend laws and budgets or lobby for projects in the department as the most important determinant of their vote choice were significantly more likely to shift their vote intention in favor of the incumbent. Similar results hold in panel G among the 71% of voters expressing a preference for a locally-oriented, as opposed to nationally-oriented, politician our pre-treatment vignette.As our simple electoral accountability model predicts, increased intention to vote for the incumbent is thus driven by those induced to update most positively on issues they regard as important. These changes in self-reported vote intentions are generally reinforced by our behavioral indi- 17Voters’ already-high vote choice certainty was not significantly altered by performance information.

34 cator of incumbent support. Column (5) of Table3 reports a 3 percentage point increase in requests for a poster from the incumbent, while panel F of Table4 shows that this increase is concentrated among respondents for whom performance metrics are the primary determinant of vote choice. Interactions with the level of reported performance content and voters’ prior beliefs follow similar patterns, but are not always statistically signiﬁcant.

5.4 Performance information increases requests from incumbents

In addition to exhibiting greater willingness to engage in electoral accountability, our behavioral outcomes capturing forward-looking requests from the incumbent demonstrate that revelations of better-than-expected performance also encourage non-electoral accountability-seeking political engagement. Columns (7) and (8) in panel B of Table3 show that, on average, treated respondents also became signiﬁcantly more willing to request a visit from, or a conversation with, incumbents if those candidates were re-elected. The index outcome estimates in column (9) imply around a 0.05 standard deviation increase relative to the control group. For such behaviors, benchmarked information does not differentially increase requests. Broadly in line with vote intentions, Table 4 reports that larger treatment effects on incumbent requests are generally concentrated where incumbent performance levels were greatest, voter prior beliefs were lowest, and especially where voters cared most about performance. Given high baseline interest in contacting incumbents, these results indicate that voters with a low initial willingness to contact politicians became more likely to do so after learning that the incumbent was more responsive than expected. Non-electoral requests represent the one area where information about deputies’ duties complements performance indicators. Columns (7)-(9) of panel A in Table3 show that learning that the incumbent is generally performing better than expected primarily translates into accountability- seeking engagement when voters are aware of what incumbents can do. This suggests that voters must ﬁrst believe that politicians possess the capacity to respond effectively before engaging in costly requests, and that this may be complemented by the perception that there is also a will.

35 However, this ﬁnding is tentative because it does not persist at endline.

6 Longer-term effects of information provision

Our analysis of voters’ immediate responses indicates that personally delivering and explaining credible performance information, at least on issues that matter to voters, can enhance electoral and non-electoral accountability. However, it is not obvious whether immediate changes in voter actions translate into action at the ballot box and persisting efforts to make costly requests of incumbents. To understand electoral behavior and requests of newly re-elected incumbents, we turn to our endline survey and polling station electoral returns.

6.1 Voters correctly recall leaﬂet content type after the election

While our respondents clearly immediately retained the information provided, such changes could quickly dissipate. The endline survey examines this possibility by asking respondents if they recalled receiving our leaflet, and (if so) what type of information the leaflet contained. Importantly, no enumerator was permitted to re-interview a respondent that they interviewed at baseline and enumerators were not informed of endline respondents’ treatment status. The results in Table5 report high levels of recall around a month after leaflet dissemination. Column (1) shows that virtually all treated respondents correctly recalled receiving the LEGS- Africa leaflet, while only 7% of control respondents incorrectly recalled receiving the leaflet. Columns (2)-(4) further demonstrate that almost as many respondents correctly remember the dif- ferentiating features of the leaflet’s content. These high recall rates far exceed those documented in other field experiments using less direct forms of information delivery (e.g. Dunning et al. forthcoming). This suggests that interventions generating five minutes of intensive reading and explanation of leaflets—at least in contexts with limited competing information sources—can ensure that limited recall is unlikely to represent a significant impediment to political accountability.

36 Table 5: Effects of information treatments on leaﬂet content type recall (endline survey)

Received Received Received Received leaﬂet duties incumbent previous information information incumbent information (1) (2) (3) (4) Any treatment 0.921*** (0.011) Duties 0.881*** (0.011) Incumbent 0.920*** (0.009) Benchmark 0.937*** 0.924*** (0.009) (0.008)

Two-sided null: Incumbent = Benchmark (p value) 0.04 Observations 3,875 3,875 3,875 3,875 Outcome range {0,1}{0,1}{0,1}{0,1} Control outcome mean 0.07 0.01 0.01 0.00 Control outcome std. dev. 0.25 0.09 0.09 0.06

Notes: See Table2.

6.2 Beliefs about incumbent performance persist after the election

Young rural voters’ overall beliefs about the incumbent also persisted to a large degree a month after treatment.Columns (1) and (2) of Table6 show that voters who received performance information continue to register significantly higher ratings of the incumbent, and believe that the incumbent performed better than previous incumbents. Appendix Table A6 similarly shows that performance information treatments caused 5-10% of treated voters to evaluate the incumbent more positively than at baseline, relative to the control group, while Table A8 reports that treated voters continue to express greater certainty about their beliefs. Treatment effects decay by about half the immediate effect of providing incumbent performance information, although this in part reflects average evaluations in the control group increasing by around a quarter of a standard deviation. Given the lack of evidence that information spilled across villages (see Table A18), these estimates suggest that the smaller treatment effects at endline may reflect both decay in information’s effects

37 Table 6: Average effects of information treatments on beliefs about incumbent performance, reported vote for the incumbent, and requests from the incumbent (endline survey)

Benchmark - Incumbent 0.085*** 0.102*** 0.026* 0.024* 0.106*** 0.009** 0.080 0.000 0.010 0.060** (0.032) (0.034) (0.02) (0.018) (0.038) (0.004) (0.066) (0.007) (0.015) (0.036) Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5}{1,...,5}{0,1}{0,1} [-2.4,1.9] {0,1}{0,1}{0,1}{0,1} [-1.6,0.7] Control outcome mean 3.08 3.46 0.64 0.53 -0.00 0.98 0.98 0.95 0.11 -0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 0.99 0.14 0.14 0.21 0.32 1.01

Notes: See Table2. All speciﬁcations include a lagged dependent variable as a control; in columns (5)-(10), pre-treatment incumbent vote is used as a proxy. and differentially effective campaign efforts to increase incumbent support in the control group after treatment delivery. The persistent effects on beliefs remain generally consistent with Bayesian updating. First, the one-sided hypothesis tests at the foot of panel B in Table6 again indicate that—consistent with previous incumbent performance falling below expectations—benchmarked information induced more favorable updating than receiving only incumbent performance indicators. This difference is also more pronounced than immediately after treatment delivery. Second, the heterogeneous treatment effects with respect to reported incumbent performance in Table7 also point in the same direction as immediately after treatment delivery. Importantly, voters also appear to focus on certain types of information over time: a comparison of panel A with panels B and C suggests that voters increasingly prized higher levels of local performance, and also became more likely to view national-oriented legislative activity negatively. The interactions with prior beliefs are no longer

38 statistically significant, suggesting that the election campaign could have significantly reshaped posterior beliefs. The results thus suggest that our leaflets had enduring effects on voter beliefs, but also that the passing of time also altered how performance information was viewed: voters upweighting local performance indicators, and attaching relative greater weight to benchmarked information.

6.3 Performance information inﬂuences the vote choices of likely-voters

We next examine whether the vote intentions registered at baseline, and the beliefs that persisted through endline, were ultimately acted upon. We test this crucial step in electoral accountability by first examining self-reported vote choices, before analyzing polling station-level returns. The self-reported survey data provides mixed evidence that incumbent performance information enhances electoral accountability. First, panel A of Table6 offers little evidence of a significant increase in incumbent voting on average, even after column (4) applies our vote validation criteria.18 The pooled estimates in panel B are also indistinguishable from zero. Nevertheless, consistent with voters’ more favorable updating from benchmarked information, panel B indicates that the benchmark significantly increased BBY voting by 2 percentage points more than the incumbent-only performance information. Second, the heterogeneous effects in Table7 also yield mixed evidence: while the leaflet’s content and voters’ prior beliefs did not influence the self-reported voting of the average respondent, panels F and G show that both performance information treatments did increase incumbent support among respondents that (at baseline) regarded performance information as the most important factor in determining their vote choice or preferred locally-oriented deputies. Appendix Table A16 shows that the information did not increase the importance attached to national and local incumbent performance in office in determining vote choices, or the most important factor in determining vote choice. These findings suggest that only a small share of the younger rural voters 18Appendix Table A22 shows that treatment did not significantly affect self-reported turnout.

39 Table 7: Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (endline survey)

Incumbent evaluation outcomes Accountability-seeking outcomes Incumbent Relative Incumbent Incumbent Incumbent Request Request Request Called Accountability overall performance vote vote evaluation incumbent incumbent hotline hotline seeking performance (v. previous) (validated) index (ICW) visit conversation number index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Heterogeneity by (standardized) reported performance level Incumbent 0.165*** 0.147*** -0.022 -0.016 0.116*** 0.000 -0.017 0.007 0.011 0.031 (0.034) (0.033) (0.020) (0.019) (0.035) (0.004) (0.028) (0.007) (0.014) (0.034) Incumbent × Overall performance (ICW) -0.022 -0.016 0.001 0.002 -0.027 -0.004 -0.007 0.015* 0.021 0.042 (0.030) (0.034) (0.021) (0.024) (0.037) (0.004) (0.009) (0.009) (0.019) (0.045) Benchmark 0.249*** 0.248*** 0.005 0.009 0.223*** 0.009*** 0.063 0.006 0.021* 0.086*** (0.033) (0.035) (0.021) (0.020) (0.038) (0.004) (0.054) (0.007) (0.015) (0.036) Benchmark × Overall performance (ICW) -0.021 -0.021 0.019 0.015 -0.016 -0.007++ -0.019 0.001 0.012 -0.021 (0.030) (0.036) (0.019) (0.021) (0.037) (0.004) (0.015) (0.009) (0.022) (0.048) Panel B: Heterogeneity by (standardized) relevance-weighted reported performance level Incumbent 0.184*** 0.139*** -0.027 -0.008 0.110*** -0.000 -0.013 0.006 0.020* 0.038 (0.039) (0.037) (0.023) (0.022) (0.042) (0.005) (0.023) (0.008) (0.015) (0.039) Incumbent × Relevant performance (ICW) 0.072* 0.068* 0.004 0.006 0.043 -0.002 0.008 0.021*** 0.029** 0.076** (0.045) (0.045) (0.026) (0.026) (0.052) (0.004) (0.022) (0.009) (0.016) (0.042) Benchmark 0.250*** 0.241*** 0.017 0.025 0.229*** 0.005 0.069 0.005 0.033*** 0.082** (0.038) (0.040) (0.024) (0.023) (0.044) (0.004) (0.062) (0.008) (0.016) (0.041) Benchmark × Relevant performance (ICW) 0.022 0.026 0.006 0.004 -0.004 -0.006 -0.032 -0.003 0.031* -0.001 40 (0.040) (0.040) (0.026) (0.025) (0.048) (0.006) (0.025) (0.008) (0.020) (0.047) Panel C: Heterogeneity by (standardized) local and national reported performance level Incumbent 0.160*** 0.159*** -0.021 -0.018 0.121*** 0.002 -0.012 0.005 0.014 0.039 (0.034) (0.033) (0.021) (0.021) (0.036) (0.004) (0.025) (0.007) (0.014) (0.034) Incumbent × National performance (ICW) -0.069 -0.065 -0.001 -0.023 -0.060 -0.013 -0.010 0.000 0.029** -0.031 (0.045) (0.040) (0.020) (0.022) (0.046) (0.009) (0.025) (0.009) (0.016) (0.056) Incumbent × Local performance (ICW) 0.107*** 0.140*** 0.006 0.024 0.105*** 0.015*** 0.020 0.014** 0.004 0.114*** (0.043) (0.039) (0.022) (0.022) (0.045) (0.007) (0.024) (0.008) (0.015) (0.045) Benchmark 0.235*** 0.255*** 0.005 0.007 0.222*** 0.010*** 0.063 0.007 0.025* 0.102*** (0.033) (0.032) (0.022) (0.021) (0.038) (0.004) (0.053) (0.007) (0.016) (0.036) Benchmark × National performance (ICW) -0.070++ -0.119++ 0.016 0.001 -0.090+ -0.020+ -0.080 -0.027++ 0.015 -0.151++ (0.035) (0.054) (0.025) (0.023) (0.049) (0.008) (0.059) (0.010) (0.018) (0.065) Benchmark × Local performance (ICW) 0.052* 0.117*** 0.006 0.024 0.085** 0.014*** -0.017 0.027*** 0.005 0.143*** (0.036) (0.045) (0.027) (0.025) (0.048) (0.007) (0.027) (0.009) (0.016) (0.052) Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5}{1,...,5}{0,1}{0,1} [-2.8,1.9] {0,1}{0,1}{0,1}{0,1} [-7.8,1.6] Control outcome mean 3.08 3.46 0.64 0.53 -0.00 0.98 0.98 0.95 0.11 -0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 0.99 0.14 0.14 0.21 0.32 1.01 Overall performance (ICW) range [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] [-2.93,1.27] Relevant performance (ICW) range [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] [-2.94,2.03] National performance (ICW) range [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] [-1.63,2.03] Local performance (ICW) range [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21] [-1.10,1.21]

Notes: See Table6. Lower-order interaction terms are included but not shown. Panel B also includes the interaction between incumbent and benchmark treatments and an indicator for respondents that did not regard local or national performance as one of the top three most important factors in determining their vote choice. (Continued...) Table7 (continued): Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (endline survey)

Incumbent evaluation outcomes Accountability-seeking outcomes Incumbent Relative Incumbent Incumbent Incumbent Request Request Request Called Accountability overall performance vote vote evaluation incumbent incumbent hotline hotline seeking performance (v. previous) (validated) index (ICW) visit conversation number index (ICW) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel D: Heterogeneity by (standardized) prior belief level Incumbent 0.165*** 0.152*** -0.027 -0.020 0.112*** 0.001 -0.019 0.006 0.013 0.039 (0.033) (0.033) (0.021) (0.020) (0.036) (0.004) (0.029) (0.008) (0.015) (0.034) Incumbent × Prior index (ICW) -0.024 -0.011 0.010 0.026 -0.005 0.009 -0.022 0.009 -0.007 0.058 (0.039) (0.037) (0.020) (0.019) (0.039) (0.006) (0.026) (0.007) (0.013) (0.047) Benchmark 0.241*** 0.250*** -0.003 0.001 0.219*** 0.009*** 0.059 0.006 0.022* 0.089*** (0.032) (0.035) (0.021) (0.020) (0.038) (0.004) (0.050) (0.007) (0.016) (0.037) Benchmark × Prior index (ICW) -0.041 -0.025 -0.004 0.023 -0.032 0.004 -0.005 0.010 0.000 0.043 (0.038) (0.038) (0.019) (0.021) (0.041) (0.004) (0.018) (0.007) (0.015) (0.039) Panel E: Heterogeneity by (standardized) prior belief precision Incumbent 0.163*** 0.157*** -0.027 -0.018 0.116*** 0.002 -0.003 0.007 0.009 0.039 (0.034) (0.031) (0.022) (0.021) (0.035) (0.004) (0.005) (0.008) (0.015) (0.036) Incumbent × Prior precision index (ICW) 0.058 0.029 -0.002 -0.001 0.032 -0.002 -0.006 -0.002 -0.002 -0.017 (0.037) (0.037) (0.021) (0.021) (0.040) (0.006) (0.006) (0.008) (0.015) (0.042) Benchmark 0.244*** 0.270*** 0.002 0.007 0.230*** 0.009*** 0.003 0.008 0.019 0.091*** (0.033) (0.033) (0.021) (0.020) (0.038) (0.004) (0.004) (0.007) (0.016) (0.038) Benchmark × Prior precision index (ICW) -0.005 0.034 0.005 0.010 0.016 0.005 -0.002 0.011 -0.028* 0.014 (0.041) (0.039) (0.020) (0.021) (0.042) (0.005) (0.005) (0.009) (0.017) (0.047) 41 Panel F: Heterogeneity by importance of performance in determining vote choice Incumbent 0.161*** 0.144*** -0.024 -0.018 0.112*** -0.000 -0.016 0.006 0.011 0.027 (0.034) (0.034) (0.021) (0.020) (0.036) (0.004) (0.028) (0.008) (0.015) (0.034) Incumbent × Performance most important 0.021 -0.007 0.001 0.011 0.015 -0.006 0.007 -0.001 0.003 -0.029 (0.031) (0.032) (0.019) (0.019) (0.034) (0.004) (0.013) (0.008) (0.013) (0.038) Benchmark 0.246*** 0.246*** 0.002 0.006 0.218*** 0.009*** 0.063 0.006 0.021* 0.086*** (0.033) (0.035) (0.021) (0.020) (0.038) (0.004) (0.054) (0.007) (0.016) (0.036) Benchmark × Performance most important -0.018 0.002 0.025* 0.031** 0.031 -0.005 0.060 -0.003 0.028*** 0.005 (0.032) (0.031) (0.019) (0.018) (0.033) (0.004) (0.052) (0.007) (0.013) (0.036) Panel G: Heterogeneity by preference for locally-oriented deputies Incumbent 0.104** 0.103** -0.095+++ -0.073++ -0.003 -0.018++ -0.004 0.006 0.018 -0.057 (0.062) (0.056) (0.033) (0.035) (0.061) (0.008) (0.038) (0.015) (0.026) (0.072) Incumbent × Prefer locally-oriented deputies 0.079 0.057 0.099*** 0.077** 0.160*** 0.025*** -0.015 -0.001 -0.009 0.118* (0.075) (0.063) (0.039) (0.043) (0.071) (0.011) (0.046) (0.017) (0.028) (0.090) Benchmark 0.191*** 0.227*** -0.048 -0.045 0.123** -0.003 0.283 0.002 0.050** 0.054 (0.057) (0.056) (0.036) (0.035) (0.065) (0.005) (0.282) (0.014) (0.029) (0.055) Benchmark × Prefer locally-oriented deputies 0.077 0.025 0.070* 0.072* 0.131** 0.016*** -0.306 0.005 -0.041 0.045 (0.067) (0.066) (0.043) (0.045) (0.078) (0.007) (0.324) (0.015) (0.032) (0.064) Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5}{1,...,5}{0,1}{0,1} [-2.8,1.9] {0,1}{0,1}{0,1}{0,1} [-7.8,1.6] Control outcome mean 3.08 3.46 0.64 0.53 -0.00 0.98 0.98 0.95 0.11 -0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 0.99 0.14 0.14 0.21 0.32 1.01 Prior index (ICW) range [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] [-2.40,2.12] Prior precision (ICW) range [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] [-3.44,1.23] Performance most important range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Prefer locally-oriented deputies range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}

Notes: See Table6. Lower-order interaction terms are included but not shown. in our sample ultimately acted on their updated beliefs. However, our sample of voters may respond differently to incumbent performance information than the broader more politically-experienced electorate. First, younger voters appear to face greater costs of turning out: a 20-year-old is more than 20 percentage point less likely to turn out than a 33-year-old in our sample. Second, younger voters and those that have not previously voted are significantly less likely to value performance indicators and locally-oriented politicians. Third, younger voters may be malleable to competing party or village campaign influences when voting for the first time. If younger voters do not turn out or vote on the basis of other factors, even when they durably update about the incumbent’s performance, then our decision-theoretic framework suggests that electoral accountability may rely on more seasoned likely-voters receiving the information and updating similarly. To better approximate electorate-level electoral outcomes, Table8 first restricts our survey estimates to the 38% of voters that reported turning out in the 2012 parliamentary election. Such voters were 14 percentage points more likely to report voting in 2017. The point estimates in columns (1)-(6) indicate that previous voters immediately updated their vote intentions and durably updated their posterior beliefs similarly to our full sample of young registered voters, suggesting that any differences in behavior are unlikely to reflect differential priors or differential updating from our treatments. However, validated vote choice depicts a starker contrast. Unlike the full sample of voters, columns (4) and (7) show that previous voters remained 2-3 percentage points more likely to vote for the incumbent across baseline and endline surveys, although this is imprecisely estimated in this subsample. Furthermore, in contrast with the full sample, column (9) shows that the persisting belief that incumbents with higher performance scores on relevant/local dimensions— the primary drivers of differences in election-time beliefs—are better overall actually translates into a significantly higher probability of treatment increasing self-reported incumbent vote choice. These findings suggest that treatment may have influenced vote choices, with substantial variation across reported performance, but principally among experienced voters. Moreover, the increased

42 Table 8: Effects of information treatments on posterior beliefs and reported vote for the incumbent, among those that turned out in 2012 (baseline and endline survey)

Incumbent overall Incumbent vote Incumbent vote performance (endline) intention (validated) (1) (2) (3) (4) (5) (6) (7) (8) (9) Incumbent 0.134** 0.139** 0.131** 0.015 0.018 0.016 0.034 0.034 0.025 (0.059) (0.058) (0.056) (0.014) (0.014) (0.016) (0.031) (0.031) (0.033) Benchmark 0.235*** 0.229*** 0.215*** 0.022 0.023 0.023 0.034 0.037 0.037 (0.057) (0.057) (0.056) (0.015) (0.015) (0.016) (0.028) (0.028) (0.027) Incumbent × Relevant performance (ICW) 0.092 0.040** 0.008 (0.081) (0.019) (0.044) Benchmark × Relevant performance (ICW) 0.023 0.020 0.030 (0.071) (0.021) (0.040) Incumbent × National performance (ICW) -0.087 -0.030** -0.104*** (0.054) (0.013) (0.030) 43 Incumbent × Local performance (ICW) 0.117* 0.033* 0.095*** (0.061) (0.019) (0.036) Benchmark × National performance (ICW) -0.095 -0.017 -0.062* (0.082) (0.019) (0.035) Benchmark × Local performance (ICW) 0.043 0.031 0.116*** (0.075) (0.021) (0.034)

Observations 1,469 1,469 1,469 1,528 1,528 1,528 1,435 1,435 1,435 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Control outcome mean 3.10 3.10 3.10 0.63 0.63 0.63 0.59 0.59 0.59 Control outcome std. dev. 0.93 0.93 0.93 0.48 0.48 0.48 0.49 0.49 0.49 Interaction mean -0.01 -0.00 0.00 -0.00 -0.01 -0.04 Interaction std. dev. 0.86 1.05 0.86 1.04 0.87 1.04 Second interaction mean 0.01 0.01 -0.00 Second interaction std. dev. 1.00 1.00 1.01

Notes: See Table3. Lower-order interaction terms are included but not shown. Given that these hypotheses were not pre-speciﬁed, * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. Table 9: Effects of information treatments on polling station-level incumbent vote share, by leaﬂet content (polling station data)

Incumbent vote share Incumbent vote share (proportion of turnout) (proportion of registered voters) (1) (2) (3) (4) (5) (6) Incumbent 0.001 -0.004 -0.014 0.006 0.001 -0.010 (0.026) (0.026) (0.025) (0.019) (0.018) (0.017) Benchmark -0.003 -0.011 -0.022 -0.003 -0.011 -0.020 (0.024) (0.025) (0.025) (0.016) (0.016) (0.016) Incumbent × Relevant performance (ICW) 0.039* 0.046*** (0.028) (0.019) Benchmark × Relevant performance (ICW) 0.043* 0.048*** (0.030) (0.019) Incumbent × National performance (ICW) -0.025 -0.008 (0.041) (0.022) Incumbent × Local performance (ICW) 0.061** 0.045*** (0.033) (0.019) Benchmark × National performance (ICW) 0.010 -0.002 (0.038) (0.024) Benchmark × Local performance (ICW) 0.028 0.033** (0.031) (0.019)

Observations 284 284 284 284 284 284 Outcome range [.06,.99] [.06,.99] [.06,.99] [.02,.73] [.02,.73] [.02,.73] Control outcome mean 0.71 0.71 0.71 0.41 0.41 0.41 Control outcome std. dev. 0.17 0.17 0.17 0.13 0.13 0.13

Notes: See Table3. Lower-order interaction terms are included but not shown. Observations are not weighted, and polling stations where the village in our sample comprises less than 50% of registered voters at the polling station are excluded. Robust standard errors are in parentheses. sensitivity to performance suggests that, by endline, voters had become more willing to sanction low-performing locally-oriented incumbents and high-performing nationally-oriented incumbents. Due to high levels of within-village information diffusion, such responses could translate into precinct-level voting outcomes. Appendix Table A17 indicates that almost 40% of our nine treated voters per village, and at least one within each village, discussed the leaflet with others. Further- more, Table A19 suggests that the leaflets became a focal point of localized incumbent campaign- ing where incumbents performed well. To examine official electoral returns that are not susceptible to reporting biases, we restrict our analysis to polling stations containing the 284 villages in our

44 sample that comprise at least 50% of registered voters at their polling station.19 The results in Table9 largely mirror the self-reported behavior of the survey respondents that voted in 2012. We do not observe a notable average treatment effect of incumbent performance information on incumbent vote share in columns (1) and (4). However, columns (2), (3), (5), and (6), provide clear evidence that the effects of treatment increase with performance on the metrics that voters value most. The estimates imply that a standard deviation increase in performance increases the incumbent’s vote share by around four percentage points. Consistent with a significant fraction of voters not receiving the information, heterogeneous effect magnitudes are around half the size reported in Table8. Given that rewards vary substantially with revealed performance around the average treatment effect, this may also explain why the 2 percentage point increase in self-reported incumbent voting on average does not show up at the precinct level. In sum, these findings suggest that likely-voters throughout treated villages indeed learned about the level of local performance and engaged in significant electoral accountability. In contrast, younger and first-time voters facing higher costs did not.

6.4 Performance information increases requests from incumbents

While electoral accountability weakens between the point of information provision (and intended vote choice) and the election itself (actual vote choice), requests from incumbents represent a different form of accountability-seeking behavior.The requests that we study likely reﬂect different costs and depend less on the relative importance of performance metrics in vote choice and possible campaign-based interactions between survey waves. Indeed, we ﬁnd that voters receiving performance information continued to engage in elevated accountability-seeking political engagement with incumbent winners a month after information was delivered. As at baseline, columns (6)-(10) of Table6 indicate that the benchmarked informa- 19Appendix Table A21 reports similar results weighting the full set of villages by the share of registered voters from the experimental village.

45 tion significantly increased requests from the government. Given that low-cost requests for visits, conversations, and a hotline number to contact incumbents through were almost universally sought, the 20% increase in the costly act of actually texting or calling the hotline in column (9) provides the most compelling evidence. Aggregated as an index, columns (10) of panel B reports a 0.09 standard deviation increase across such behaviors. Like electoral accountability, the heterogeneous effects in Table7 further demonstrate that increased hotline usage was greatest in departments where the incumbent-only and benchmark treatments reported greatest incumbent performance. In line with self-reported behaviors among likely-voters, this exclusively reflected local performance. Unlike electoral accountability, voters did not engage significantly less with worst-performing incumbents. Consistent with the baseline results, requests are not greater among respondents that value performance information more in making vote choices.These results indicate that treatment caused voters to durably engage in costly efforts to demand accountability from better-performing politicians, consistent with voters expecting greater responsiveness.

7 Conclusion

Given the mixed evidence that information campaigns can support political accountability, this article examined the extent to which accountability failures reﬂect voter-level constraints. By abstracting from issues of information dissemination and consumption, we dissect the step-by-step process linking the personal delivery and explanation of incumbent performance information to electoral and non-electoral accountability. Our ﬁndings demonstrate that rural Senegalese voters are capable of sophisticated information processing, retain their updated beliefs, and regard local performance and temporal benchmarks as informative about incumbent quality. In contrast, information about national performance and incumbent duties has little effect in a context where delivering “pork” is widely-perceived as a legislator’s primary function. Intentions to hold politicians

46 to account translate into rewarding highly-performing incumbents at the ballot box for more experienced voters, but belief updating does not appear to translate into electoral accountability among voters that are harder to mobilize or that prioritize other issues. With respect to making costly requests of incumbents, voters that expect greater responsiveness engage in greater accountability- seeking behavior. These findings thus illustrate that, upon receiving information they deem credible and relevant, voters are able and mostly willing to hold politicians to account. This highlights the importance of understanding how other factors sustain low-accountability equilibria in developing contexts. First, following Dunning et al.(forthcoming), future research is needed to establish the most effective and scale-able means through which information can be communicated. Second, although we showed that voters will respond to information that they receive, we still know little about the factors driving voters’ desire to consume such information in the first place. Third, we also abstracted from the dynamics of incumbent action and candidate (de)selection that underpinned the performance information and vote choices the electorate faces. How candidate selection relates to performance in office represents a particularly important area for future research. Finally, our partial equilibrium focus only briefly addressed campaign responses to information dissemination. While significant responses are often documented (Banerjee et al. 2011; Bidwell, Casey and Glennerster 2016), little is known about when campaigns are effective at complementing or refut- ing information dissemination campaigns. This moderating role of competing influences merits a more detailed examination than this paper has scope to conduct. Although rural areas may be exposed to fewer competing influences, there are good reasons to believe that our anatomy of political accountability extends beyond rural Senegal. First, sophisticated responses among voters with low educations levels and limited previous exposure suggest that voters across the world may be capable of drawing similar inferences. Second, parliamentary elections in Senegal share many features with elections in other developing democracies, such as non-trivial levels of local clientelism and the ability of local agents to intervene in response

47 to campaigns like ours. Third, because our leaﬂets are similar in design to previous studies (e.g. Chong et al. 2015; Dunning et al. forthcoming; Gottlieb 2016; Humphreys and Weinstein 2012), our ﬁndings may help to direct researchers in other contexts toward the types of impediments to electoral accountability that we highlight here.

48 References

Adida, Claire, Jessica Gottlieb, Gwyneth McClendon and Eric Kramon. 2017. “Breaking the Clientelistic Voting Equilibrium: The Joint Importance of Salience and Coordination.” Working Paper.

Aker, Jenny C., Paul Collier and Pedro C. Vicente. 2017. “Is information power? Using mobile phones and free newspapers during an election in Mozambique.” Review of Economics and Statistics 99(2):185–200.

Arias, Eric, Horacio A. Larreguy, John Marshall and Pablo Querub´ın. 2018a. “Does the Content and Mode of Delivery of Information Matter for Electoral Accountability? Evidence from a Field Experiment in Mexico.” Working paper.

Arias, Eric, Horacio A. Larreguy, John Marshall and Pablo Querub´ın. 2018b. “Priors rule: When do malfeasance revelations help and hurt incumbent parties?” Working paper.

Ashworth, Scott. 2012. “Electoral accountability: recent theoretical and empirical work.” Annual Review of Political Science 15:183–201.

Aytac¸, Selim Erdem. 2018. “Relative Economic Performance and the Incumbent Vote: A Refer- ence Point Theory.” Journal of Politics 80(1):16–29.

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

Beck, Linda J. 2012. ‘Senegal’. In Countries at the Crossroads 2011: An Analysis of Democratic Governance, ed. Jake Dizard, Christopher Walker and Vanessa Tucker. Rowman & Littleﬁeld.

Bidwell, Kelly, Katherine Casey and Rachel Glennerster. 2016. “Debates: Voting and Expenditure Responses to Political Communication.” Working Paper.

49 Campello, Daniela and Cesar Zucco, Jr. 2016. “Presidential success and the world economy.” Journal of Politics 78(2):589–602.

Casey, Katherine. 2018. “Radical Decentralization: Does community driven development work?” Annual Review of Economics 10:139–165.

Chauchard, Simon, Marko Klasnjaˇ and S.P. Harish. 2017. “The Limited Impact of Information on Political Selection: An Experiment on Asset Disclosures in India.” Working paper.

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

Dixit, Avinash and John Londregan. 1996. “The determinants of success of special interests in redistributive politics.” Journal of Politics 58(04):1132–1155.

Dunning, Thad, Guy Grossman, Macartan Humphreys, Susan Hyde, Craig McIntosh and Gareth Nellis. forthcoming. Information, Accountability, and Cumulative Learning: Lessons from Metaketa I. Cambridge University Press.

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

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.

Gomez, Brad T. and J. Matthew Wilson. 2006. “Cognitive heterogeneity and economic voting: A comparative analysis of four democratic electorates.” American Journal of Political Science 50(1):127–145.

50 Gottlieb, Jessica. 2016. “Greater Expectations: A Field Experiment to Improve Accountability in Mali.” American Journal of Political Science 60(1):143–157.

Gottlieb, Jessica. 2017. “Explaining variation in broker strategies: A lab-in-the-ﬁeld experiment in Senegal.” Comparative Political Studies 50(11):1556–1592.

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

Kayser, Mark Andreas and Michael Peress. 2012. “Benchmarking across borders: electoral accountability and the necessity of comparison.” American Political Science Review 106(3):661– 684.

Kendall, Chad, Tommaso Nannicini and Francesco Trebbi. 2014. “How do voters respond to information? Evidence from a randomized campaign.” American Economic Review 105(1):322– 353.

Koter, Dominika. 2013. “King makers: Local leaders and ethnic politics in Africa.” World Politics 65(2):187–232.

Lieberman, Evan S., Daniel N. Posner and Lily L. Tsai. 2014. “Does information lead to more active citizenship? Evidence from an education intervention in rural Kenya.” World Development 60:69–83.

Meyer, Margaret A. and John Vickers. 1997. “Performance comparisons and dynamic incentives.” Journal of Political Economy 105(3):547–581.

Morris, Stephen and Hyun Song Shin. 2002. “Social value of public information.” American Economic Review 92(5):1521–1534.

Mvukiyehe, Eric and Cyrus Samii. 2015. “Promoting Democracy in Fragile States: Insights from a Field Experiment in Liberia.” World Development 95(1):254–267.

51 Olken, Benjamin A. 2007. “Monitoring corruption: evidence from a ﬁeld experiment in Indonesia.” Journal of Political Economy 115(2):200–249.

Powell, Jr., G. Bingham and Guy D. Whitten. 1993. “A Cross-National Analysis of Economic Vot- ing: Taking Account of the Political Context.” American Journal of Political Science 37(2):391– 414.

Thomas, Melissa A. and Oumar Sissokho. 2005. “Liaison Legislature: The Role of the National Assembly in Senegal.” Journal of Modern African Studies 43(1):97–117.

Villalon,´ Leonardo A. 1995. Islamic Society and State Power in Senegal: Disciples and Citizens in Fatick. Cambridge University Press.

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

52 A Online appendix

Contents

A.1 Formal derivation of the effect of providing incumbent-only and benchmarked information...... A3 A.1.1 Incumbent-only performance information...... A3 A.1.2 Benchmarked malfeasance information...... A5 A.1.3 Empirical implications...... A8 A.2 Senegalese parliamentary electoral participation in comparative context...... A9 A.3 All leaﬂet conﬁgurations...... A12 A.4 Additional details about sample selection...... A12 A.4.1 Selection of departments...... A12 A.4.2 Selection of villages...... A12 A.4.3 Selection of young voters as respondents...... A17 A.4.4 Sample characteristics relative to national averages...... A18 A.5 Compliance and experimental validation checks...... A18 A.6 ICW index construction...... A20 A.7 Deviations from pre-analysis plan...... A20 A.8 Additional results...... A24 A.8.1 Fraction of voters that updated favorably about the incumbent...... A24 A.8.2 Effects on the precision of voters beliefs...... A24 A.8.3 Effects on evaluations of challenger parties...... A27 A.8.4 Alternative scales for computing outcome and interaction indexes..... A29 A.8.5 The importance voters attach to incumbent legislative performance does not change...... A29

A1 A.8.6 Within-village information diffusion...... A29 A.8.7 Cross-village informational spillovers...... A38 A.8.8 Party responses to information dissemination...... A42 A.8.9 Weighted polling station level estimates...... A46 A.8.10 Effects on electoral turnout...... A47

A2 A.1 Formal derivation of the effect of providing incumbent-only and bench-

marked information

We formally derive the effects of providing incumbent-only and temporally- benchmarked information, relative to each other and to receiving no information at all, on Bayesian voters’ behavior. We adopt a simple Normal learning framework where a given voter learns about the current incumbent t’s unobserved underlying quality, the previous incumbent t − 1’s unobserved underlying quality, and an unobserved time-invariant district/department-specific component of performance that affects all incumbents within the district equally. Specifically, denote the voter’s prior belief about the current incumbent’s quality as qt, the voter’s prior belief about the previous incumbent’s quality as qt−1, and the voter’s prior belief about the district-specific component as qc. We assume that our voter’s prior beliefs over these quantities are independently given by N(θt,1/pt ),

N(θt−1,1/pt−1), and N(θC,1/pc) respectively. Our goal is to examine the differential effect of different types of incumbent performance signal on voters’ absolute and relative posterior beliefs. These are two key outcomes in our empirical analysis that directly inﬂuence vote choice and, in the case of the level, request-making in our simple decision-theoretic model—equations (1) and (2) in the body of the article.

A.1.1 Incumbent-only performance information

We start with the baseline case where the voter receives a given realization of the incumbent-only

performance signal, sˆt, drawn from signal distribution N(qt + qc,1/ρt ), where the signal’s preci-

sion ρt is known to the voter but its expectation qt + qc is not. The performance signal thus reﬂects both the unobserved quality of the current incumbent and unobserved time-invariant characteristics of the district/department. The following proposition establishes voters’ posterior inferences about incumbent’s quality qt and the district-speciﬁc shock qc:

Proposition 1. (Incumbent-only performance information) Upon receiving realized signal sˆt, a

A3 voter’s posterior expectation of current incumbent t’s quality is wt (sˆt −θc) + (1−wt )θt and of the

district-specific shock is wc(sˆt −θt ) + (1−wc)θt, where wt and wc are weights (defined within the proof) that both increase with ρt and respectively increase in pc and pt.   1/p 0 0 0 −1  t  −1 Proof : We first define q = [qt,qc] , µ = [θt,θc] , Λ =  , A = [1,1], and L = 0 1/pc

[1/ρt ]. Applying a standard multivariate updating result (e.g. ?:93) implies that posterior beliefs are distributed according to:

0 −1 0 0 −1 p(q|sˆt ) ∼ N (Λ + A LA) (A Lsˆt + Λµ),(Λ + A LA) , (A1)

where the application of matrix operations to the model in hand implies:

 −1 p + ρ ρ 0 −1  t t t  (Λ + A LA) =   ρt pc + ρt   1 pc + ρt −ρt =   := Σ, (A2) pt pc + ptρt + pcρt   −ρt pt + ρt   ρ sˆ + p θ 0  t t t t  (A Lsˆt + Λµ) =  . (A3) ρtsˆt + pcθc

Combining these results yields probability distribution:

   w (sˆ − θ ) + (1 − w )θ  t t c t t   p(q|sˆt ) ∼ N  ,Σ, (A4) wc(sˆt − θt ) + (1 − wc)θc

where w = pcρt and w = pt ρt . t : pt pc+pt ρt +pcρt c : pt pc+pt ρt +pcρt This first proposition demonstrates that current incumbent-only performance information influences voter beliefs about current incumbent quality to the extent that the district-specific shock-

A4 adjusted signal (sˆt − θc) differs from the voter’s prior belief θt. Since the district-speciﬁc shock is also unobserved from the perspective of the voter, the voter has limited capacity to update about

the value of this shock, and thus relies on their prior belief θc. Indeed, relative to receiving no

information about incumbent performance, and thus retaining the prior belief θt, a voter upwardly

(downwardly) updates their expectation of t’s quality when θt < (>)sˆt −θc. This intuitively shows that, after netting out prior expectations of the district-speciﬁc shock, voters update favorably about the incumbent party when the signal exceeds the voter’s prior expectation. The same expression pertains to evaluating the posterior belief regarding the expected difference in t’s quality relative to previous incumbent t − 1’s quality, the latter of which may serve as a proxy for challengers and thus drive vote choices.

A.1.2 Benchmarked malfeasance information

Turning to our main result, we now consider the benchmark scenario case where the voter receives performance signal sˆt−1 pertaining to the previous incumbent, as well as the current incumbent performance signal sˆt. We similarly assume that sˆt−1 is drawn from signal distribution

N(qt−1 + qc,1/ρt−1), where the signal’s precision ρt−1 is again known to the voter. This second signal enables the voter to draw more precise inferences by ﬁltering out their more accurate updated beliefs about the district-speciﬁc component of performance, as well as learn more about the performance of previous incumbents that may be informative about current challengers. The following proposition now establishes voters’ posterior beliefs following the provision of such benchmarked performance information:

Proposition 2. (Benchmarked performance information) Upon receiving realized signals sˆt and

sˆt−1, a voter’s posterior expectation of current incumbent t’s quality is wt,ssˆt −wt,cθc −wt,∆(sˆt−1 −

θt−1) + wt,tθt, of the previous incumbent’s t − 1’s quality is wt−1,ssˆt−1 − wt−1,cθc − wt−1,∆(sˆt −

θt ) +wt−1,t−1θt−1, and of the district-speciﬁc shock is wc,t (sˆt −θt ) +wc,t−1(sˆt−1 −θt−1) +wc,cθc, where the weights are deﬁned within the proof.

A5   1/pt 0 0   0 0 0 −1   Proof : We ﬁrst deﬁne sˆ = [sˆt,sˆt−1] , q = [qt,qt−1,qc] , µ = [θt,θt−1,θc] , Λ =  0 1/p 0 ,  t−1    0 0 1/pc     1 0 1 1/ρ 0   −1  t  A =  , and L =  . We then apply the same theorem as in the previous 0 1 1 0 1/ρt−1 proof, where the application of matrix operations to the model in hand implies:

 −1 pt + ρt 0 ρt   0 −1   (Λ + A LA) =  0 p + ρ ρ   t−1 t−1 t−1    ρt ρt−1 pc + ρt + ρt−1 1 = ptρI(pt−1 + ρt−1) + pt−1ρt−1(pt + ρt ) + pc(pt + ρt )(pt−1 + ρt−1)   (pt−1 + ρt−1)(pc + ρt ) + pt−1ρt−1 ρt ρt−1 −ρt (pt−1 + ρt−1)   × ρt ρt−1 (pt + ρt )(pc + ρt−1) + pt ρt −ρt−1(pt + ρt )    −ρt (pt−1 + ρt−1) −ρt−1(pt + ρt )(pt + ρt )(pt−1 + ρt−1)

:= ΣB, (A5)   ρtsˆt + ptθt   0   (A Lsˆ + Λµ) =  ρ sˆ + p θ . (A6)  t−1 t−1 t−1 t−1    ρtsˆt + ρt−1sˆt−1 + pcθc

Combining these results yields the probability distribution:

   wt,ssˆt − wt,cθc − wt, (sˆt−1 − θt−1) + wt,tθt  ∆      p(q|sˆt,sˆt−1) ∼ N w sˆ − w θ − w (sˆ − θ ) + w θ ,ΣB, (A7)  t−1,s t−1 t−1,c c t−1,∆ t t t−1,t−1 t−1     wc,t (sˆt − θt ) + wc,t−1(sˆt−1 − θt−1) + wc,cθc

ρt (pt−1 pc+pcρt−1+pt−1ρt−1) pcρt (pt−1+ρt−1) where the weights are given by wt,s := D , wt,c := D , wt,∆ :=

pt−1ρt ρt−1 pt (pt−1 pc+pcρt−1+pt−1ρt−1+pt−1ρt +ρt ρt−1) ρt−1(pt pc+pcρt +pt ρt ) D , wt,t := D , wt−1,s := D , wt−1,c :=

A6 pcρt−1(pt +ρt ) pt ρt ρt−1 pt−1(pt pc+pcρt +pt ρt +pt ρt−1+ρt ρt−1) pt ρt (pt−1+ρt−1) D , wt−1,∆ := D , wt−1,t−1 := D , wc,t := D ,

pt−1ρt−1(pt +ρt ) pc(pt +ρt )(pt−1+ρt−1) wc,t−1 := D , and wc,c := D , where D := [ptρt (pt−1 +ρt−1)+ pt−1ρt−1(pt + −1 ρt ) + pc(pt + ρt )(pt−1 + ρt−1)] and all weights are positive. This proposition demonstrates that voter posterior beliefs about the level of current incumbent quality increase with the extent to which indicators of performance exceed expectations that now explicitly adjust for updated beliefs about the district-speciﬁc shock. In contrast with incumbent- only information, a Bayesian voter now also uses the benchmarked signal to better account for the possibility that high incumbent malfeasance could reﬂect a high realization of the common shock,

i.e. sˆt−1 − θt−1. Consequently, relative to receiving no information, benchmarked performance

information will induce upward (downward) updating when: θt < (>)wt,ssˆt −wt,cθc −wt,∆(sˆt−1 −

θt−1) + wt,tθt. This will hold when performance indicators, adjusted for updated expectations of the district-speciﬁc shock, exceed prior expectations of quality. The same logic applies to evaluations of the previous incumbent. The voter’s posterior belief about the common shock itself, wc,t (sˆt − θt ) + wc,t−1(sˆt−1 − θt−1) + wc,cθc, is intuitively increasing in the extent to which performance signals exceed prior expectations. Combining our two propositions, benchmarked information induces a more favorable (unfavorable) posterior expectation of incumbent quality than an incumbent-only signal when:

wt (sˆt − θc) + (1 − wt )θt < (>)wt,ssˆt − wt,cθc − wt,∆(sˆt−1 − θt−1) + wt,tθt. (A8)

There are thus two primary forces pushing voters to update favorably about the current incumbent: (1) the increased weight attached to the current incumbent’s performance signal (it is easy to show

that wt < wt,s), when sˆt > θt; and (2) when the weights attached to the signal and prior beliefs do

not drastically differ (and thus sˆt, θc, and θt cancel out), voters will generally update favorably

when sˆt−1 < θt−1, i.e. when the previous incumbent performed worse than expected. Force (1) reﬂects the sharper inferences that can be drawn from a given signal when it is benchmarked, while

A7 force (2) reflects the second signal inducing the voter to infer that there was a larger-than-expected district-specific shock. Turning to relative evaluations and vote choice, voter behavior is instead likely to reflect a relative comparison between posterior belief about current and previous incumbent quality. In this case, benchmarked information induces a larger difference in expected quality between the current incumbent and the previous incumbent (which may proxy for challengers, when it comes to vote choice) relative to incumbent-only information when:

∗ ∗ wt,ssˆt + wt,tθt − wt−1,ssˆt−1 − wt−1,t−1θt−1 > wt (sˆt − θc − θt−1) + (1 − wt )(θt − θt−1), (A9) where the district-specific shock is identically accounted for when comparing posterior beliefs about the current and previous incumbents (but adjusts the weighting coefficients to account for extracting the district-specific shock). As with absolute beliefs, a voter will become relatively more

∗ favorable toward the incumbent when: (1) wt < wt,s, where sˆt > θt; and (2) sˆt−1 − (θt−1 + θc) < 0, where the weights on comparable terms are similar in magnitude.

A.1.3 Empirical implications

With respect to absolute posterior beliefs, we expect to observe the following predictions and comparative statics:

• The effect of incumbent-only information v. no information on overall beliefs about current

incumbent quality is positive when sˆt > θt + θc, and is increasing in (sˆt − θt ).

• The effect of benchmarked information v. no information on overall beliefs about current

incumbent quality is positive when, approximately, sˆt > θt + E[qc|sˆt,sˆt−1], and is increasing

in (sˆt − θt ) and decreasing in (sˆt−1 − θt−1).

• The effect of benchmarked information v. incumbent-only information on beliefs about

A8 current incumbent quality is positive when, approximately, θt−1 > sˆt−1, and is increasing in

(sˆt − θt ) and decreasing in (sˆt−1 − θt−1).

Relationships are deﬁned as “approximate” where we assume that the weights do not meaningfully differ. With respect to relative comparisons between current and previous incumbents—a plausible proxy for challengers, and thus vote choices—incumbent-only and benchmarked information provision implies the following comparative statics:

• The effect of incumbent-only information v. control on incumbent vote share is positive

when sˆt > θt + θc, and is increasing in (sˆt − θt ).

• The effect of benchmark information v. control on incumbent vote share is positive when,

approximately, sˆt > sˆt−1, and is decreasing in (sˆt − sˆt−1).

• The effect of benchmark v. incumbent-only information on incumbent vote share is negative

when, approximately, θt−1 + θc > sˆt−1, and is increasing and (sˆt − θt ) decreasing in (sˆt−1 −

θt−1).

A.2 Senegalese parliamentary electoral participation in comparative con-

text

Figure A1 reports national turnout rates across sub-Saharan Africa in the most recent parliamentary elections, while Figure A2 shows the percentage of Afrobarometer (round 6) respondents that have contacted a member of parliament within the last year. Consistent with the limited role of the Assemblee´ Nationale, turnout rates are relatively low in Senegalese parliamentary elections.

A9 Somalia Rwanda Ethiopia Guinea-Bissau Sao Tome and Principe Seychelles Botswana Sierra Leone Equatorial Guinea Kenya Cameroon Angola Mauritius Burundi Mauritania South Africa Liberia Namibia Comoros Malawi Uganda Ghana Niger Togo Cape Verde Benin Guinea Tanzania Djibouti Burkina Faso Democratic Republic of Congo Chad Zambia Senegal Madagascar Mozambique Sudan Lesotho Central African Republic Nigeria Congo Gambia Libya Zimbabwe Mali Gabon Cote d'Ivoire

0 20 40 60 80 100 Most recent parliamentary turnout rate (%)

Figure A1: Parliamentary turnout rates in the most recent election across sub-Saharan Africa

Note: All data was downloaded from International Institute for Democracy and Electoral Assistance.

A10 Liberia Swaziland Sierra Leone Botswana Tanzania São Tomé and Príncipe Malawi Zambia Zimbabwe Morocco Benin Uganda Ghana Nigeria Kenya Cameroon Sudan Gabon Mauritius Togo Cote d'Ivoire Cape Verde Lesotho Senegal Mali Guinea Mozambique Algeria Niger Namibia Burkina Faso South Africa Burundi Madagascar Tunisia 0 10 20 30 40 Percentage of voters who contacted MP within the past year

Figure A2: Parliamentary deputy contact rates across sub-Saharan Africa

Note: All data is from Afrobarometer round 6.

A11 A.3 All leaﬂet conﬁgurations

In addition to the “duties + benchmark” leaflet shown in Figure4, Figures A3-A6 show our other leaflet configurations.

A.4 Additional details about sample selection

A.4.1 Selection of departments

We conducted our study in the ﬁve departments: Fatick, Foundiougne, Kanel, Oussouye, and Ranerou´ Ferlo. These departments were selected because they satisﬁed four criteria that prior theoretical arguments suggest would increase the likelihood of performance information helping voters hold incumbents to account: (1) only a single incumbent was seeking re-election through the majoritarian vote (with the exception of Kanel where two were standing); (2) there were no incumbents from the proportional list attached to the department (with the exception of Kanel); (3) the incumbent’s performance could be compared with the previous incumbent(s), because no incumbent was seeking re-election for a second time and the department was not a newly-created administrative unit; and (4) given the preceding criteria, the selected departments have the lowest number of deputies representing the department. Oussouye and Ranerou´ Ferlo had only one incumbent deputy, although Oussouye had two in the previous legislature. Fatick and Foundiougne had two majoritarian deputies and no deputies from the proportional list. However, both had deputies from the proportional list in the previous legislature. Only Kanel, which has two majoritarian deputies, also has one deputy elected from the proportional list; in Kanel, one majoritarian and the one proportional list incumbent were standing for re-election on the majoritarian list.

A.4.2 Selection of villages

Within the ﬁve departments chosen for our study, we selected 450 rural villages for our sample. Starting from the 859 possible villages in these departments, we excluded all villages with fewer

A12 Figure A3: Example of “duties” treatment

A13 Figure A4: Example of “incumbent” treatment in Oussouye

A14 Figure A5: Example of “duties + incumbent” treatment in Oussouye

A15 Figure A6: Example of “benchmark” treatment in Oussouye

A16 than 200 people and all villages with more than 4,000 people. Logistical concerns and access to newly-constructed schools further restricted the set of potential villages. Of logistical concerns, 19 villages were dropped because they were too expensive to reach, e.g. because they are located on islands. In the hope of leveraging cross-cohort variation in access to schooling following a 2002 school construction program(to instrument for educational attainment and identify heterogeneous effects by educational attainment, see more below), we excluded villages where the ﬁrst post-2002 school was built between 2006 and 2010. We ignore this cross-cutting variation because access to new schools did not robustly increase educational attainment among our survey respondents. By virtue of our randomization, access to schooling is orthogonal to our informational treatments.

A.4.3 Selection of young voters as respondents

Our sampling strategy stratified the sample into three age groupings of roughly equal size within each village (i.e. 3 respondents per village from each category): 20-26, 27-31, and 32-38. As briefly noted in the main text, this decision reflected our pre-registered intention to examine the effects of educational attainment. In particular, we aimed to leverage a difference-in-differences (or regression discontinuity) design exploiting cross-cohort variation in access to schools constructed as part of a 2002 secondary schooling expansion program as an instrument for educational attainment (?). The 20-26 age grouping contained cohorts that were counted as fully treated if a school been constructed within 6km of their village, while the 27-31 age grouping comprised partially treated respondents (who were already in secondary school at the time of the reform) if a school has been constructed within 6km of their village, and the 32-38 category is a control group. Un- fortunately, we were unable to obtain a first stage showing that the instrument robustly increased educational attainment using either the difference-in-differences or regression discontinuity ap- proaches.

A17 A.4.4 Sample characteristics relative to national averages

Table A1 compares 2013 Census in our sample of 450 villages with the Senegalese national averages.

A.5 Compliance and experimental validation checks

We were unable to conduct surveys in 7 of our 450 villages. In three cases we were refused entry, while the remaining cases reflected a lack of identity cards among villagers, inability to locate the village, heavy rain, and a village falling under judicial control. Given that we conduced surveys in all villages and did not allocate different treatment assignments to villages according to their characteristics, our inability to conduct surveys in these villages was unaffected by treatment assignment. We estimate equation (3) to demonstrate that treatment assignments for the 443 villages where surveys were conducted are indeed orthogonal to predetermined covariates, and thus that the randomization’s integrity was maintained after dropping the seven villages that we could not access. Appendix Table A2 shows that 90 predetermined individual- and village-level covariates are well- balanced across treatment conditions at baseline: the two-sided joint F test of the restriction that each treatment group is indistinguishable from the others was only rejected at the 10% level in 12 cases. Analysis of endline data is generally more complex, since estimates using endline data could be confounded by selective attrition in response to treatment. As noted in the main text, we successfully re-interviewed 96% of the baseline sample. This remarkably high recontact rate for a telephone followup survey may have reflected the low frequency with which rural Senegalese voters have opportunities to express their views to survey teams, especially those that offered to pass on requests to politicians at baseline. Unsurprisingly, given this low rate of attrition, Table A4 reports no significant difference in attrition rates across treatment groups. Moreover, Table A3 shows

A18 Table A1: Baseline sample balance tests

Data from our 2017 sample Data from Senegal 2013 census Weighted by... 20-38 age group 20-38 age group, Variable Unweighted Population Baseline resps Entire census From our villages all villages from our villages Average age 28.50 28.50 28.43 22.08 21.40 27.47 27.65 % female 36.33 36.30 36.09 50.83 50.89 52.35 54.24 % with some primary education 52.20 52.05 48.39 34.01 25.88 34.57 20.99 % with some secondary education 35.04 34.96 32.96 14.66 5.68 24.36 10.53 % read/write French 37.70 28.86 35.40 21.37 A19 % read/write Wolof 1.53 0.57 2.20 1.05 % read/write Pular 1.23 1.33 1.92 2.21 % read/write Serer 0.33 1.69 0.47 2.72 % read/write Mandingue 0.15 0.10 0.22 0.20 % read/write Diola 0.09 0.06 0.13 0.10 % read/write Soninke 0.03 0.04 0.05 0.05 % Muslim 89.26 89.26 89.86 95.66 89.84 95.42 90.96 % Christian 9.09 9.08 8.25 3.91 7.86 4.19 7.32 % with piped water 57.81 57.77 51.93 53.37 27.31 57.57 26.39 % with electricity 28.96 28.91 23.19 38.90 3.45 44.70 4.09 Average number of bedrooms 7.34 7.34 7.27 4.95 4.86 4.91 4.91 % from rural villages 100 100 100 59.27 100.00 53.43 100.00 that the balance we observed at baseline continues to hold within the endline response sample: for 15 of 102 predetermined (baseline and endline) variables do we observe signiﬁcant differences across treatment conditions.

A.6 ICW index construction

Following ?, a given inverse-covariance weighted index for individual i is deﬁned by:

0 −1 −1 0 −1 (1 Σ 1) (1 Σ x˜i), (A10)

where Σ is the covariance matrix between items xi1,...,xiK, x˜i is the vector of standardized items, and 1 is a vector of 1s.

A.7 Deviations from pre-analysis plan

All reported analyses follow our pre-analysis plan, with the following minor exceptions:

1. Although we pre-speciﬁed that standard error would be clustered by randomization block, we instead cluster standard errors by village. This change does not substantively affect our conclusions, but was nevertheless implemented to reﬂect best practice (e.g. ?).

2. Although we pre-specified that requesting an incumbent poster (in the baseline survey) would be included in the ICW index of behavioral outcomes, we instead include it in the ICW scale of incumbent evaluations. We made this change to reflect its better conceptual fit alongside incumbent evaluations. As the point estimates for individual items suggest, this switch does not alter the substantive conclusions drawn from either the baseline incumbent evaluation index or the baseline accountability-seeking behavior index.

3. Although we did not pre-specify that we would restrict our polling station-level analysis to polling stations that contain experimental villages that comprise at least 50% of the polling

A20 Table A2: Baseline sample balance tests

Control Control Incumbent Benchmark F test (two- Outcome Observations mean std. dev. Duties Incumbent and Duties Benchmark and duties sided p value) Individual-level baseline survey variables Female 3,999 0.37 0.48 0.006 (0.031) -0.027 (0.036) 0.006 (0.036) -0.060 (0.041) -0.020 (0.034) 0.30 Married 3,999 0.66 0.48 -0.009 (0.033) -0.056* (0.029) -0.034 (0.032) -0.045 (0.033) -0.012 (0.033) 0.30 Age 3,999 28.34 5.78 0.444 (0.295) 0.173 (0.253) 0.133 (0.285) 0.204 (0.270) 0.300 (0.253) 0.54 Years of education 3,998 4.82 5.46 0.049 (0.412) 0.528 (0.359) 0.229 (0.422) 0.112 (0.365) -0.191 (0.382) 0.40 Diola ethnicity 3,999 0.07 0.25 0.006* (0.003) -0.007 (0.014) 0.000 (0.004) 0.000 (0.009) 0.004 (0.003) 0.58 Pulaar ethnicity 3,999 0.18 0.39 -0.021 (0.025) 0.002 (0.023) -0.018 (0.020) -0.022 (0.022) -0.005 (0.024) 0.66 Peul ethnicity 3,999 0.16 0.37 0.002 (0.019) 0.018 (0.027) 0.008 (0.028) -0.020 (0.030) 0.003 (0.022) 0.77 Serer ethnicity 3,999 0.41 0.49 -0.004 (0.026) 0.017 (0.028) 0.009 (0.024) -0.016 (0.032) -0.014 (0.030) 0.78 Toucouleur ethnicity 3,999 0.01 0.08 -0.005 (0.004) 0.013 (0.011) 0.007 (0.007) 0.015* (0.009) -0.008 (0.006) 0.07* Wolof ethnicity 3,999 0.16 0.37 0.023 (0.024) -0.027 (0.026) 0.000 (0.023) -0.046* (0.024) 0.029 (0.025) 0.10 Christian 3,997 0.10 0.30 -0.002 (0.025) -0.014 (0.028) -0.012 (0.022) -0.011 (0.025) -0.011 (0.027) 0.99 Muslim 3,997 0.88 0.32 -0.001 (0.025) 0.027 (0.031) 0.023 (0.023) 0.000 (0.025) 0.010 (0.027) 0.71 Household has electicity 3,999 0.32 0.47 0.007 (0.053) -0.057 (0.053) -0.041 (0.047) -0.012 (0.051) -0.053 (0.060) 0.61 Household has water 3,999 0.61 0.49 0.014 (0.046) -0.001 (0.046) -0.066* (0.039) -0.016 (0.050) -0.048 (0.051) 0.27 Number of bedrooms 3,889 7.08 5.22 0.638 (0.457) 0.406 (0.340) 0.300 (0.352) 0.206 (0.322) 0.420 (0.295) 0.57 Income scale 3,454 1.72 1.86 -0.067 (0.123) -0.265** (0.118) -0.208 (0.125) -0.165 (0.138) -0.207 (0.127) 0.23 Frequency discuss politics 3,996 2.06 0.80 -0.035 (0.051) -0.014 (0.053) -0.065 (0.053) 0.022 (0.052) 0.025 (0.047) 0.45 Interest in public affairs 3,999 1.97 1.01 0.049 (0.054) 0.082 (0.057) -0.013 (0.064) 0.040 (0.059) 0.083 (0.051) 0.33 Radio news frequency 3,999 4.01 2.23 -0.006 (0.174) 0.251** (0.120) -0.065 (0.161) 0.269* (0.137) 0.059 (0.161) 0.01** Television news frequency 3,999 2.42 2.49 0.139 (0.225) 0.165 (0.193) -0.042 (0.221) 0.045 (0.226) -0.087 (0.229) 0.60 Newspaper news frequency 3,999 0.67 1.62 -0.003 (0.097) -0.069 (0.110) -0.024 (0.119) -0.038 (0.099) -0.047 (0.099) 0.98 Satisﬁed with National Assembly 3,999 2.01 0.99 0.028 (0.062) 0.036 (0.068) 0.027 (0.066) -0.047 (0.060) 0.080 (0.060) 0.27 Believe deputies listen to voters 3,999 0.58 0.73 0.033 (0.049) 0.030 (0.050) 0.032 (0.050) -0.017 (0.044) 0.050 (0.048) 0.71 Believe deputies respond to requests 3,999 1.99 0.89 0.033 (0.055) 0.020 (0.065) -0.067 (0.060) -0.040 (0.062) -0.036 (0.056) 0.36 Frequency of contacting deputy 3,999 0.13 0.48 -0.009 (0.025) -0.013 (0.024) -0.029 (0.025) 0.011 (0.033) -0.016 (0.030) 0.73 Turnout in 2012 3,999 0.42 0.49 -0.030 (0.032) -0.043 (0.032) -0.044 (0.034) -0.049* (0.027) -0.031 (0.033) 0.61 Incumbent vote in 2012 3,999 0.32 0.47 -0.049 (0.030) -0.066** (0.030) -0.062** (0.031) -0.089*** (0.025) -0.064** (0.030) 0.02** Believe deputy is from own commune 3,997 0.28 0.45 -0.024 (0.037) -0.058 (0.039) -0.038 (0.042) -0.070 (0.044) -0.034 (0.042) 0.42 Believe deputy is from own village 3,997 0.04 0.20 -0.006 (0.014) 0.005 (0.013) 0.006 (0.013) 0.036 (0.025) 0.011 (0.019) 0.67 Believe deputy is of same ethnicity 3,990 0.57 0.50 -0.017 (0.034) -0.039 (0.029) -0.007 (0.028) -0.067** (0.033) -0.054 (0.035) 0.16 Know incumbent party 3,999 0.64 0.48 -0.056* (0.030) 0.020 (0.025) -0.005 (0.033) -0.025 (0.030) -0.029 (0.033) 0.14 Know incumbent name 3,999 0.35 0.48 -0.021 (0.033) -0.004 (0.029) 0.014 (0.030) 0.014 (0.031) 0.008 (0.032) 0.84 Know incumbent commune 3,999 0.66 0.47 -0.014 (0.038) 0.049 (0.038) 0.044 (0.043) 0.098** (0.039) -0.021 (0.045) 0.01*** Know incumbent village 3,999 0.91 0.29 -0.003 (0.022) 0.004 (0.021) 0.010 (0.021) 0.014 (0.022) -0.021 (0.031) 0.81 Know incumbent ethnicity 3,999 0.54 0.50 -0.040 (0.034) -0.036 (0.034) -0.041 (0.035) 0.000 (0.035) -0.016 (0.034) 0.58 Know deputies make laws 3,999 0.46 0.50 -0.027 (0.028) 0.023 (0.032) 0.025 (0.033) 0.040 (0.036) 0.004 (0.036) 0.43 Know deputies approve budget 3,999 0.54 0.50 0.020 (0.030) 0.018 (0.038) 0.000 (0.034) 0.011 (0.037) -0.014 (0.035) 0.92 Know deputies do not select local projects 3,999 0.15 0.35 0.021 (0.023) 0.057** (0.027) 0.045* (0.025) 0.024 (0.024) 0.030 (0.028) 0.26 Believe proposing laws is a main role 3,999 0.13 0.33 0.005 (0.020) 0.012 (0.024) 0.009 (0.024) -0.009 (0.023) 0.028 (0.022) 0.65 Believe passing laws is a main role 3,999 0.22 0.41 -0.008 (0.029) -0.012 (0.029) 0.009 (0.034) -0.015 (0.029) 0.009 (0.029) 0.84 Believe committees are a main role 3,999 0.05 0.22 -0.004 (0.014) 0.014 (0.017) 0.014 (0.018) -0.023* (0.013) 0.019 (0.016) 0.01*** Believe budgeting is a main role 3,999 0.07 0.26 -0.006 (0.016) 0.002 (0.017) 0.003 (0.016) -0.024 (0.016) 0.009 (0.018) 0.13 Believe constituency petitions are a main role 3,999 0.20 0.40 0.002 (0.022) -0.019 (0.019) 0.002 (0.022) -0.021 (0.024) -0.015 (0.022) 0.52 Believe local transfer lobbying is a main role 3,999 0.16 0.37 0.016 (0.024) 0.035 (0.027) -0.004 (0.020) 0.007 (0.026) 0.002 (0.022) 0.31 Believe local project lobbying is a main role 3,999 0.38 0.49 0.018 (0.034) 0.051 (0.038) 0.041 (0.032) 0.029 (0.034) 0.034 (0.031) 0.72 Believe local project implementation is a main role 3,999 0.23 0.42 -0.023 (0.022) 0.025 (0.024) -0.006 (0.027) -0.014 (0.021) -0.004 (0.023) 0.27 Passing laws is a main role 3,999 0.08 0.27 -0.020 (0.017) 0.021 (0.019) 0.003 (0.017) -0.036** (0.017) -0.008 (0.016) 0.06* Passing laws is a main role 3,999 0.21 0.40 -0.004 (0.022) 0.029 (0.020) 0.018 (0.024) -0.014 (0.028) 0.022 (0.023) 0.14 Prefer nationally-oriented deputies 3,999 0.24 0.42 -0.005 (0.026) -0.016 (0.029) 0.022 (0.026) 0.025 (0.026) -0.026 (0.025) 0.17 Prefer locally-oriented deputies 3,999 0.73 0.45 -0.016 (0.033) 0.022 (0.031) -0.043 (0.026) -0.035 (0.029) 0.010 (0.027) 0.04** Deputy’s village or community is among three most important voting factors 3,999 0.33 0.47 0.017 (0.026) -0.016 (0.027) 0.047 (0.029) 0.019 (0.025) 0.022 (0.030) 0.15 Deputy’s ethnicity or religion is among three most important voting factors 3,999 0.14 0.35 0.032 (0.023) 0.005 (0.021) 0.023 (0.022) -0.012 (0.024) 0.016 (0.021) 0.15 Deputy’s education or profession is among three most important voting factors 3,999 0.28 0.45 0.001 (0.029) -0.025 (0.026) -0.003 (0.028) -0.023 (0.028) 0.003 (0.026) 0.68 Deputy’s party is among three most important voting factors 3,999 0.21 0.41 -0.047* (0.028) -0.013 (0.024) -0.040 (0.024) -0.065*** (0.021) -0.027 (0.027) 0.01*** Deputy’s political experience is among three most important voting factors 3,999 0.36 0.48 0.046* (0.026) 0.030 (0.029) -0.005 (0.029) 0.046 (0.031) 0.024 (0.029) 0.26 Deputy’s amending/approving of laws is among three most important voting factors 3,999 0.33 0.47 -0.025 (0.023) -0.007 (0.028) -0.030 (0.030) 0.011 (0.027) 0.005 (0.021) 0.62 Deputy’s parliamentary lobbying is among three most important voting factors 3,999 0.74 0.44 0.032 (0.027) 0.036 (0.027) 0.024 (0.028) 0.037 (0.029) 0.009 (0.028) 0.60 Deputy’s campaign promises is among three most important voting factors 3,999 0.20 0.40 -0.012 (0.021) 0.022 (0.025) 0.017 (0.022) -0.007 (0.025) -0.002 (0.027) 0.38 Deputy’s election gifts is among three most important voting factors 3,999 0.08 0.27 -0.008 (0.013) 0.003 (0.016) 0.009 (0.015) 0.009 (0.015) -0.006 (0.017) 0.60 No listed factor is among most important voting factor 3,999 0.21 0.41 -0.010 (0.025) -0.020 (0.027) -0.032 (0.022) -0.004 (0.025) -0.014 (0.026) 0.68

Village-level variables Turnout (2012) 3,999 0.59 0.10 0.008 (0.011) 0.003 (0.012) 0.015 (0.013) 0.010 (0.011) 0.023** (0.012) 0.18 Incumbent vote share (2012) 3,999 0.68 0.17 0.000 (0.021) 0.035* (0.019) -0.026 (0.022) 0.020 (0.020) 0.023 (0.022) 0.01** Village x coordinate 3,999 440370.08 147848.55 222.972 (2621.374) -748.300 (3235.710) 1,990.090 (3442.184) 735.476 (3010.644) -1,755.535 (2689.351) 0.78 Village y coordinate 3,999 1583885.99 81743.16 -1,551.907 (2415.548) -4416.385* (2429.440) -2,476.073 (2289.450) -3647.528* (2099.871) -4983.795** (2132.448) 0.33 Village population 3,999 863.05 686.09 -138.344 (93.605) -36.580 (99.191) -130.982* (69.817) 0.282 (110.842) -63.088 (85.988) 0.25 Share of village with any middle school 3,999 0.86 0.19 -0.011 (0.011) 0.005 (0.015) 0.009 (0.012) -0.003 (0.012) -0.003 (0.009) 0.45 Distance to nearest school 2,792 4.52 2.68 0.152 (0.508) 0.072 (0.450) -0.364 (0.503) -0.505 (0.557) 0.482 (0.495) 0.20 Share of village completing middle school 3,999 0.04 0.06 0.003 (0.006) 0.005 (0.005) -0.002 (0.005) 0.002 (0.007) -0.005 (0.007) 0.19 Share of village households with a good toilet 3,999 0.06 0.11 0.016 (0.016) -0.010 (0.012) -0.008 (0.015) -0.008 (0.011) -0.003 (0.010) 0.43 Share of village households with piped toilet 3,999 0.08 0.11 -0.008 (0.016) 0.004 (0.015) 0.011 (0.015) -0.008 (0.016) 0.004 (0.015) 0.68 Share of village households with electricity 3,999 0.01 0.04 -0.006 (0.005) -0.002 (0.009) -0.007 (0.005) 0.010 (0.012) 0.010 (0.010) 0.24 Share of village households with good walls 3,999 0.75 0.32 0.008 (0.028) -0.012 (0.033) -0.025 (0.031) 0.020 (0.033) 0.017 (0.035) 0.37 Share of village households with a good roof 3,999 0.03 0.07 0.001 (0.009) 0.002 (0.010) 0.015 (0.010) -0.001 (0.010) 0.007 (0.009) 0.37 Share of village households with good ﬂoors 3,999 0.22 0.20 0.030 (0.026) 0.013 (0.028) 0.027 (0.028) 0.025 (0.024) 0.068** (0.028) 0.08* Share of village households with a radio 3,999 0.73 0.18 -0.035 (0.022) -0.053** (0.021) -0.041** (0.020) -0.045** (0.022) -0.041** (0.017) 0.08* Share of village households with a good television 3,999 0.03 0.04 0.001 (0.006) 0.003 (0.007) 0.007 (0.006) 0.007 (0.007) 0.003 (0.007) 0.84 Share of village households with a car 3,999 0.02 0.06 0.001 (0.008) -0.004 (0.008) -0.004 (0.007) -0.005 (0.006) 0.002 (0.008) 0.87 Bambara share of village 3,999 0.03 0.09 -0.005 (0.015) -0.016 (0.013) -0.008 (0.014) -0.007 (0.017) -0.018 (0.012) 0.52 Diola share of village 3,999 0.07 0.25 0.001 (0.001) -0.012 (0.011) -0.006 (0.004) -0.011 (0.009) -0.002 (0.002) 0.32 Lebou share of village 3,999 0.00 0.00 0.000 (0.000) 0.000 (0.000) 0.001 (0.000) 0.000 (0.000) 0.001 (0.000) 0.36 Manding share of village 3,999 0.03 0.11 -0.021 (0.013) -0.022* (0.011) -0.006 (0.017) -0.006 (0.009) -0.016 (0.010) 0.15 Manjag share of village 3,999 0.00 0.00 0.000 (0.000) 0.000 (0.000) 0.001 (0.001) 0.000 (0.000) 0.000 (0.000) 0.90 Maure share of village 3,999 0.00 0.01 -0.001 (0.002) 0.000 (0.002) -0.001 (0.001) -0.001 (0.001) -0.001 (0.002) 0.94 Peul share of village 3,999 0.21 0.38 0.018 (0.015) 0.030 (0.021) 0.019 (0.017) 0.010 (0.019) 0.019 (0.026) 0.66 Pulaar share of village 3,999 0.06 0.17 -0.018 (0.011) 0.009 (0.020) 0.003 (0.025) 0.014 (0.025) 0.014 (0.023) 0.37 Serer share of village 3,999 0.39 0.43 0.002 (0.020) 0.028 (0.019) 0.003 (0.018) 0.014 (0.026) -0.010 (0.019) 0.49 Soce share of village 3,999 0.01 0.04 0.001 (0.003) -0.002 (0.003) 0.003 (0.005) 0.004 (0.004) 0.001 (0.004) 0.74 Soninke share of village 3,999 0.01 0.07 0.001 (0.012) -0.007 (0.009) -0.003 (0.003) 0.012 (0.012) -0.007 (0.009) 0.30 Toucouleur share of village 3,999 0.04 0.13 0.002 (0.019) -0.010 (0.014) -0.002 (0.023) 0.006 (0.022) -0.018 (0.016) 0.76 Wolof share of village 3,999 0.15 0.30 0.021 (0.021) -0.001 (0.017) -0.001 (0.022) -0.038* (0.023) 0.037** (0.018) 0.02**

Notes: Each row represents a single regression, and all speciﬁcations include block and enumerator ﬁxed effects and are weighted by the inverse of the number of respondents completing the survey in each village. Standard errors clustered by village are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01.

A21 Table A3: Endline sample balance tests

Control Control Incumbent Benchmark F test (two- Outcome Observations mean std. dev. Duties Incumbent and Duties Benchmark and duties sided p value) Individual-level endline variables Read French 3,876 0.61 0.49 0.004 (0.035) 0.004 (0.028) -0.007 (0.030) 0.004 (0.031) -0.006 (0.032) 1.00 Speak French 3,876 0.37 0.48 0.041 (0.034) 0.022 (0.029) 0.039 (0.031) 0.019 (0.032) 0.015 (0.031) 0.68 Speak Wolof 3,876 0.78 0.42 0.012 (0.023) 0.013 (0.027) 0.029 (0.022) -0.002 (0.020) 0.022 (0.023) 0.54 Speak Serer 3,876 0.41 0.49 0.013 (0.026) 0.021 (0.026) 0.032 (0.026) 0.009 (0.025) -0.007 (0.023) 0.51 Speak Pulaar 3,876 0.31 0.46 -0.013 (0.013) 0.027 (0.018) 0.009 (0.021) 0.025 (0.015) 0.008 (0.015) 0.06* Speak Diola 3,876 0.07 0.26 0.002 (0.003) 0.000 (0.003) 0.001 (0.003) -0.005 (0.004) 0.003 (0.003) 0.72 Speak Soninke 3,876 0.02 0.14 -0.005 (0.017) -0.017 (0.015) -0.009* (0.005) 0.012 (0.018) -0.018 (0.015) 0.24 Speak Mandingue 3,876 0.05 0.22 -0.021 (0.015) -0.020 (0.013) -0.008 (0.015) 0.005 (0.018) -0.022 (0.016) 0.33 Speak Mandjak 3,876 0.00 0.04 0.000 (0.000) -0.002 (0.002) -0.001 (0.001) 0.000 (0.003) -0.002 (0.001) 0.37 Speak other language 3,876 0.05 0.23 -0.015 (0.011) -0.014 (0.012) -0.007 (0.015) 0.020 (0.013) -0.018 (0.014) 0.17 Attempted phone calls required for endline 3,876 2.01 2.28 0.051 (0.148) -0.124 (0.129) -0.030 (0.129) -0.177 (0.137) -0.009 (0.137) 0.26 Number of days endline was completed after election 3,876 16.93 10.05 0.142 (0.652) -0.727 (0.587) 0.062 (0.592) -0.590 (0.608) 0.194 (0.604) 0.33

Individual-level baseline variables Female 3,876 0.37 0.48 0.006 (0.032) -0.025 (0.036) 0.005 (0.036) -0.054 (0.042) -0.021 (0.034) 0.39 Married 3,876 0.66 0.48 -0.012 (0.033) -0.056* (0.030) -0.042 (0.032) -0.042 (0.034) -0.015 (0.034) 0.30 Age 3,876 28.37 5.77 0.374 (0.316) 0.238 (0.268) 0.116 (0.280) 0.256 (0.276) 0.289 (0.265) 0.75 Years of education 3,875 4.93 5.50 -0.049 (0.411) 0.431 (0.356) 0.155 (0.419) 0.008 (0.372) -0.304 (0.385) 0.36 Diola ethnicity 3,876 0.07 0.25 0.006* (0.004) -0.007 (0.015) -0.001 (0.004) 0.000 (0.010) 0.005* (0.003) 0.48 Pulaar ethnicity 3,876 0.18 0.39 -0.021 (0.024) 0.005 (0.023) -0.016 (0.020) -0.024 (0.022) -0.006 (0.023) 0.57 Peul ethnicity 3,876 0.16 0.37 0.000 (0.019) 0.016 (0.026) 0.007 (0.028) -0.019 (0.031) 0.004 (0.022) 0.80 Serer ethnicity 3,876 0.41 0.49 -0.001 (0.026) 0.019 (0.027) 0.014 (0.024) -0.013 (0.032) -0.010 (0.030) 0.73 Toucouleur ethnicity 3,876 0.01 0.08 -0.006 (0.004) 0.013 (0.011) 0.007 (0.007) 0.016 (0.010) -0.009 (0.006) 0.04** Wolof ethnicity 3,876 0.16 0.37 0.019 (0.024) -0.031 (0.027) -0.001 (0.024) -0.051** (0.024) 0.028 (0.025) 0.08* Christian 3,874 0.10 0.30 -0.002 (0.024) -0.012 (0.028) -0.009 (0.023) -0.008 (0.025) -0.011 (0.027) 1.00 Muslim 3,874 0.88 0.33 -0.002 (0.025) 0.025 (0.031) 0.021 (0.024) -0.002 (0.025) 0.011 (0.027) 0.77 Household has electicity 3,876 0.32 0.47 0.003 (0.052) -0.057 (0.053) -0.042 (0.047) -0.015 (0.050) -0.054 (0.060) 0.65 Household has water 3,876 0.62 0.49 0.016 (0.046) -0.004 (0.045) -0.065 (0.040) -0.017 (0.050) -0.056 (0.051) 0.20 Number of bedrooms 3,769 7.11 5.25 0.622 (0.466) 0.393 (0.353) 0.308 (0.349) 0.258 (0.322) 0.414 (0.030) 0.64 Income scale 3,346 1.74 1.87 -0.080 (0.123) -0.294** (0.121) -0.225* (0.128) -0.190 (0.139) -0.232* (0.132) 0.17 Frequency discuss politics 3,873 2.06 0.80 -0.040 (0.050) -0.013 (0.051) -0.069 (0.053) 0.022 (0.050) 0.019 (0.047) 0.44 Interest in public affairs 3,876 1.97 1.01 0.046 (0.054) 0.078 (0.057) -0.015 (0.066) 0.054 (0.058) 0.076 (0.051) 0.42 Radio news frequency 3,876 4.02 2.22 0.030 (0.175) 0.254** (0.120) -0.047 (0.159) 0.241* (0.136) 0.034 (0.163) 0.03* Television news frequency 3,876 2.43 2.49 0.142 (0.222) 0.160 (0.202) -0.040 (0.222) 0.058 (0.226) -0.097 (0.234) 0.59 Newspaper news frequency 3,876 0.68 1.63 0.019 (0.099) -0.074 (0.115) -0.024 (0.121) -0.037 (0.102) -0.067 (0.103) 0.91 Satisﬁed with National Assembly 3,876 2.01 0.99 0.024 (0.062) 0.037 (0.068) 0.041 (0.067) -0.034 (0.059) 0.080 (0.062) 0.42 Believe deputies listen to voters 3,876 0.57 0.72 0.047 (0.050) 0.043 (0.052) 0.048 (0.050) -0.005 (0.045) 0.057 (0.050) 0.61 Believe deputies respond to requests 3,876 2.00 0.89 0.041 (0.058) 0.018 (0.065) -0.059 (0.061) -0.041 (0.064) -0.053 (0.058) 0.30 Frequency of contacting deputy 3,876 0.13 0.48 -0.003 (0.025) -0.011 (0.025) -0.027 (0.026) 0.018 (0.033) -0.010 (0.030) 0.74 Turnout in 2012 3,876 0.42 0.49 -0.038 (0.033) -0.037 (0.034) -0.038 (0.035) -0.043 (0.026) -0.033 (0.034) 0.76 Incumbent vote in 2012 3,876 0.31 0.46 -0.054* (0.031) -0.061* (0.031) -0.056* (0.031) -0.085*** (0.025) -0.058* (0.031) 0.06* Believe deputy is from own commune 3,874 0.28 0.45 -0.020 (0.037) -0.054 (0.040) -0.029 (0.042) -0.063 (0.045) -0.029 (0.043) 0.51 Believe deputy is from own village 3,875 0.04 0.20 -0.002 (0.015) 0.003 (0.013) 0.009 (0.013) 0.037 (0.025) 0.010 (0.019) 0.78 Believe deputy is of same ethnicity 3,867 0.57 0.50 -0.017 (0.033) -0.040 (0.029) -0.003 (0.028) -0.065* (0.033) -0.053 (0.035) 0.16 Know incumbent party 3,876 0.64 0.48 -0.061* (0.031) 0.022 (0.026) -0.004 (0.034) -0.026 (0.031) -0.025 (0.033) 0.09* Know incumbent name 3,876 0.35 0.48 -0.021 (0.032) 0.000 (0.030) 0.017 (0.030) 0.017 (0.031) 0.014 (0.032) 0.74 Know incumbent commune 3,876 0.67 0.47 -0.012 (0.039) 0.053 (0.039) 0.045 (0.043) 0.099** (0.040) -0.021 (0.046) 0.00*** Know incumbent village 3,876 0.91 0.29 -0.001 (0.022) 0.011 (0.022) 0.014 (0.022) 0.016 (0.023) -0.018 (0.031) 0.74 Know incumbent ethnicity 3,876 0.54 0.50 -0.042 (0.034) -0.038 (0.034) -0.038 (0.035) 0.007 (0.036) -0.018 (0.034) 0.57 Know deputies make laws 3,876 0.46 0.50 -0.023 (0.030) 0.024 (0.032) 0.028 (0.033) 0.043 (0.037) 0.010 (0.037) 0.43 Know deputies approve budget 3,876 0.54 0.50 0.021 (0.030) 0.021 (0.038) 0.004 (0.035) 0.012 (0.037) -0.010 (0.035) 0.92 Know deputies do not select local projects 3,876 0.15 0.36 0.014 (0.023) 0.059** (0.028) 0.039 (0.025) 0.020 (0.024) 0.030 (0.029) 0.24 Believe proposing laws is a main role 3,876 0.13 0.34 -0.001 (0.020) 0.011 (0.024) 0.004 (0.023) -0.010 (0.023) 0.026 (0.022) 0.74 Believe passing laws is a main role 3,876 0.22 0.42 -0.010 (0.030) -0.012 (0.029) 0.003 (0.033) -0.011 (0.029) 0.010 (0.028) 0.88 Believe committees are a main role 3,876 0.05 0.22 -0.008 (0.015) 0.014 (0.016) 0.012 (0.017) -0.026* (0.014) 0.019 (0.016) 0.00*** Believe budgeting is a main role 3,876 0.07 0.26 -0.009 (0.016) 0.003 (0.017) 0.003 (0.015) -0.024 (0.016) 0.011 (0.019) 0.14 Believe consituency petitions are a main role 3,876 0.20 0.40 0.005 (0.021) -0.021 (0.019) 0.004 (0.022) -0.016 (0.022) -0.010 (0.021) 0.39 Believe local transfer lobbying is a main role 3,876 0.16 0.37 0.012 (0.023) 0.030 (0.026) -0.004 (0.020) 0.004 (0.025) 0.001 (0.022) 0.51 Believe local project lobbying is a main role 3,876 0.38 0.49 0.017 (0.034) 0.053 (0.038) 0.042 (0.032) 0.031 (0.034) 0.043 (0.031) 0.61 Believe local project implementation is a main role 3,876 0.22 0.42 -0.020 (0.022) 0.022 (0.025) -0.009 (0.027) -0.018 (0.020) -0.001 (0.023) 0.36 Passing laws is a main role 3,876 0.08 0.27 -0.020 (0.017) 0.021 (0.019) 0.001 (0.018) -0.037** (0.017) -0.005 (0.017) 0.06* Passing laws is a main role 3,876 0.20 0.40 0.002 (0.022) 0.035* (0.019) 0.022 (0.024) -0.013 (0.028) 0.027 (0.023) 0.14 Prefer nationally-oriented deputies 3,876 0.24 0.43 -0.012 (0.025) -0.016 (0.029) 0.019 (0.027) 0.025 (0.027) -0.028 (0.025) 0.18 Prefer locally-oriented deputies 3,876 0.72 0.45 -0.006 (0.031) 0.026 (0.031) -0.040 (0.027) -0.030 (0.030) 0.016 (0.027) 0.04** Deputy’s village or community is among three most important voting factors 3,876 0.33 0.47 0.008 (0.028) -0.017 (0.028) 0.041 (0.030) 0.020 (0.026) 0.019 (0.032) 0.27 Deputy’s ethnicity or religion is among three most important voting factors 3,876 0.14 0.35 0.032 (0.024) 0.004 (0.021) 0.021 (0.022) -0.012 (0.023) 0.018 (0.022) 0.18 Deputy’s education or profession is among three most important voting factors 3,876 0.28 0.45 0.002 (0.028) -0.028 (0.027) -0.001 (0.028) -0.031 (0.028) -0.001 (0.027) 0.56 Deputy’s party is among three most important voting factors 3,876 0.21 0.41 -0.044 (0.028) -0.010 (0.024) -0.045* (0.024) -0.066*** (0.021) -0.034 (0.027) 0.01*** Deputy’s political experience is among three most important voting factors 3,876 0.36 0.48 0.048* (0.027) 0.032 (0.030) 0.003 (0.030) 0.047 (0.031) 0.028 (0.029) 0.39 Deputy’s amending/approving of laws is among three most important voting factors 3,876 0.32 0.47 -0.020 (0.025) -0.006 (0.027) -0.030 (0.029) 0.015 (0.027) 0.012 (0.023) 0.57 Deputy’s parliamentary lobbying is among three most important voting factors 3,876 0.74 0.44 0.028 (0.029) 0.031 (0.027) 0.021 (0.028) 0.034 (0.029) 0.013 (0.029) 0.77 Deputy’s campaign promises is among three most important voting factors 3,876 0.21 0.40 -0.015 (0.022) 0.017 (0.025) 0.010 (0.023) -0.011 (0.027) -0.002 (0.027) 0.46 Deputy’s election gifts is among three most important voting factors 3,876 0.07 0.26 -0.008 (0.014) 0.004 (0.017) 0.009 (0.016) 0.012 (0.016) -0.007 (0.017) 0.55 No listed factor is among most important voting factor 3,876 0.20 0.40 -0.009 (0.025) -0.019 (0.027) -0.026 (0.023) -0.001 (0.026) -0.015 (0.027) 0.81

Village-level variables Turnout (2012) 3,876 0.59 0.09 0.007 (0.011) 0.002 (0.011) 0.014 (0.012) 0.010 (0.011) 0.022** (0.011) 0.16 Incumbent vote share (2012) 3,876 0.68 0.17 0.001 (0.021) 0.035* (0.019) -0.023 (0.022) 0.019 (0.019) 0.023 (0.022) 0.01** Village x coordinate 3,876 440940.150 148667.070 45.916 (2559.880) -1084.797 (3232.221) 1656.006 (3406.412) 318.279 (2937.082) -1724.714 (2657.693) 0.78 Village y coordinate 3,876 1584208.080 81966.080 -1510.946 (2293.035) -4588.076* (2400.736) -2440.301 (2205.886) -3599.132* (2017.477) -4838.398** (2099.340) 0.33 Village population 3,876 867.25 682.24 -144.884 (94.458) -42.725 (99.207) -137.380* (70.087) -6.493 (111.676) -72.630 (85.373) 0.25 Share of village with any middle school 3,876 0.86 0.19 -0.012 (0.012) 0.006 (0.014) 0.009 (0.012) -0.003 (0.012) -0.003 (0.009) 0.45 Distance to nearest school 2,713 4.55 2.70 0.170 (0.501) 0.014 (0.436) -0.390 (0.05) -0.515 (0.544) 0.391 (0.485) 0.18 Share of village completing middle school 3,876 0.04 0.06 0.003 (0.006) 0.005 (0.005) -0.002 (0.005) 0.002 (0.007) -0.005 (0.007) 0.20 Share of village households with a good toilet 3,876 0.06 0.11 0.015 (0.016) -0.009 (0.011) -0.009 (0.014) -0.009 (0.011) -0.003 (0.010) 0.42 Share of village households with piped toilet 3,876 0.08 0.11 -0.011 (0.016) 0.002 (0.015) 0.010 (0.015) -0.011 (0.016) 0.003 (0.015) 0.70 Share of village households with electricity 3,876 0.01 0.04 -0.006 (0.005) -0.003 (0.009) -0.007 (0.005) 0.010 (0.012) 0.010 (0.010) 0.25 Share of village households with good walls 3,876 0.75 0.32 0.010 (0.029) -0.012 (0.033) -0.026 (0.032) 0.021 (0.034) 0.017 (0.035) 0.36 Share of village households with a good roof 3,876 0.03 0.08 0.001 (0.009) 0.000 (0.009) 0.014 (0.010) -0.002 (0.010) 0.007 (0.009) 0.37 Share of village households with good ﬂoors 3,876 0.22 0.20 0.028 (0.027) 0.009 (0.027) 0.024 (0.028) 0.021 (0.024) 0.069** (0.028) 0.09* Share of village households with a radio 3,876 0.73 0.18 -0.037* (0.022) -0.054** (0.021) -0.043** (0.020) -0.046** (0.023) -0.040** (0.017) 0.06* Share of village households with a good television 3,876 0.03 0.04 0.001 (0.006) 0.001 (0.006) 0.006 (0.006) 0.006 (0.007) 0.003 (0.007) 0.86 Share of village households with a car 3,876 0.02 0.06 0.000 (0.008) -0.004 (0.008) -0.004 (0.007) -0.005 (0.006) 0.002 (0.008) 0.89 Bambara share of village 3,876 0.03 0.09 -0.006 (0.015) -0.016 (0.013) -0.009 (0.014) -0.008 (0.017) -0.020* (0.012) 0.49 Diola share of village 3,876 0.07 0.25 0.001 (0.001) -0.013 (0.011) -0.006 (0.004) -0.011 (0.010) -0.001 (0.002) 0.33 Lebou share of village 3,876 0.00 0.00 0.000 (0.000) 0.000 (0.000) 0.001 (0.000) 0.000 (0.000) 0.001 (0.000) 0.36 Manding share of village 3,876 0.03 0.11 -0.020 (0.013) -0.021* (0.011) -0.006 (0.016) -0.006 (0.009) -0.016* (0.009) 0.15 Manjag share of village 3,876 0.00 0.00 0.000 (0.000) 0.000 (0.000) 0.001 (0.001) 0.000 (0.000) 0.000 (0.000) 0.91 Maure share of village 3,876 0.00 0.01 -0.001 (0.001) 0.000 (0.002) -0.001 (0.001) -0.001 (0.001) -0.001 (0.002) 0.95 Peul share of village 3,876 0.21 0.38 0.019 (0.016) 0.031 (0.022) 0.020 (0.017) 0.011 (0.019) 0.020 (0.027) 0.65 Pulaar share of village 3,876 0.06 0.18 -0.020* (0.012) 0.009 (0.021) 0.002 (0.026) 0.012 (0.026) 0.015 (0.024) 0.36 Serer share of village 3,876 0.40 0.43 0.001 (0.021) 0.026 (0.019) 0.003 (0.018) 0.015 (0.026) -0.010 (0.019) 0.46 Soce share of village 3,876 0.01 0.04 0.000 (0.004) -0.002 (0.004) 0.002 (0.004) 0.004 (0.004) 0.001 (0.004) 0.77 Soninke share of village 3,876 0.01 0.07 0.001 (0.011) -0.009 (0.009) -0.004 (0.004) 0.012 (0.012) -0.008 (0.009) 0.21 Toucouleur share of village 3,876 0.04 0.13 0.003 (0.020) -0.009 (0.013) 0.000 (0.024) 0.008 (0.023) -0.018 (0.016) 0.74 Wolof share of village 3,876 0.14 0.29 0.021 (0.021) -0.001 (0.016) 0.000 (0.022) -0.039* (0.023) 0.038** (0.017) 0.02**

Attrited (1) Duties 0.001 (0.009) Incumbent -0.009 (0.009) Incumbent and duties -0.014 (0.008) Benchmark 0.005 (0.010) Benchmark and duties 0.004 (0.011)

Observations 3,999 Outcome range {0,1} Control outcome mean 0.03 Control outcome std. dev. 0.17 F test: all coefﬁcients = 0 (two-sided p value) 0.21

A23 station’s registered voters, we believe that this is a natural restriction to minimize estimation imprecision arising from polling station containing a small number of villagers that could have received the treatment information via within-village information spillovers. As a robustness check, Table A21 shows similar results when using all polling stations, but weighting by the share of registered voters contributed by the village in our experimental sample.

4. Our pre-analysis plan proposed both first-differencing and a controlling for a lagged dependent variable. We ultimately chose only the latter due to its efficiency benefits (?) and because only baseline proxies are available for some outcomes.

A.8 Additional results

A.8.1 Fraction of voters that updated favorably about the incumbent

Tables A5 and A6 show the fraction of voters that updated their beliefs favorably (panel A) and unfavorably (panel B) relative to the prior beliefs that they expressed before treatment at baseline. The results in Table A5 indicate that around a quarter of the sample updated favorable immediately after receiving treatment, while treatment did not systematically induce any unfavorable updating; these people appear to then intend to vote for the incumbent. The results in Table A6 show that favorable updating generally persisted, while treatment also increased resistance to unfavorable updating.

A.8.2 Effects on the precision of voters beliefs

Tables A7 and A8 respectively show that our information treatments increased the precision of respondent beliefs about the incumbent at both baseline and endline.

A24 Table A5: Effects of information treatments favorable and unfavorable updating (baseline survey)

Indicator of updating posterior relative to prior on... Incumbent Relative Prospective Incumbent overall performance incumbent vote performance (v. previous) performance (1) (2) (3) (4) Panel A: Updated favorably Incumbent 0.209*** 0.188*** 0.159*** 0.035*** (0.023) (0.019) (0.021) (0.009) Benchmark 0.261*** 0.260*** 0.197*** 0.035*** (0.023) (0.022) (0.020) (0.010)

Two-sided null: Incumbent = Benchmark (p value) 0.01 0.00 0.03 1.00 Observations 3,942 3,932 3,928 3,999 Outcome range {0,1}{0,1}{0,1}{0,1} Control outcome mean 0.13 0.11 0.12 0.02 Control outcome std. dev. 0.34 0.31 0.33 0.15 Panel B: Updated unfavorably Incumbent -0.001 -0.017 -0.023* 0.002 (0.012) (0.015) (0.013) (0.005) Benchmark -0.018 -0.020 -0.040*** -0.002 (0.011) (0.018) (0.014) (0.005)

Two-sided null: Incumbent = Benchmark (p value) 0.10 0.84 0.20 0.45 Observations 3,942 3,932 3,928 3,999 Outcome range {0,1}{0,1}{0,1}{0,1} Control outcome mean 0.08 0.11 0.15 0.02 Control outcome std. dev. 0.28 0.31 0.35 0.14

A25 Table A6: Effects of information treatments favorable and unfavorable updating (endline survey)

Indicator of updating posterior relative to prior on... Incumbent Relative Incumbent Incumbent overall performance vote vote performance (v. previous) (validated) (1) (2) (3) (4) Panel A: Updated favorably Incumbent 0.069*** 0.035* -0.011 -0.008 (0.022) (0.021) (0.018) (0.017) Benchmark 0.083*** 0.095*** 0.001 -0.002 (0.022) (0.021) (0.017) (0.015)

Two-sided null: Incumbent = Benchmark (p value) 0.50 0.01 0.51 0.73 Observations 3,834 3,825 3,781 3,781 Outcome range {0,1}{0,1}{0,1}{0,1} Control outcome mean 0.43 0.42 0.24 0.20 Control outcome std. dev. 0.49 0.49 0.43 0.40 Panel B: Updated unfavorably Incumbent -0.026 -0.029* 0.012 0.008 (0.017) (0.016) (0.017) (0.019) Benchmark -0.054*** -0.055*** 0.005 -0.003 (0.016) (0.018) (0.017) (0.019)

Two-sided null: Incumbent = Benchmark (p value) 0.07 0.12 0.61 0.50 Observations 3,834 3,825 3,781 3,781 Outcome range {0,1}{0,1}{0,1}{0,1} Control outcome mean 0.21 0.23 0.19 0.26 Control outcome std. dev. 0.41 0.42 0.40 0.44

A26 Table A7: Effects of information treatments on posterior belief precision (baseline survey)

Incumbent Relative Prospective Incumbent overall performance incumbent vote performance (v. previous) performance precision precision precision precision (1) (2) (3) (4) Incumbent 0.351*** 0.291*** 0.305*** -0.032 (0.107) (0.105) (0.105) (0.041) Benchmark 0.517*** 0.650*** 0.486*** 0.024 (0.098) (0.101) (0.097) (0.046)

One-sided null: Incumbent ≥ Benchmark (p value) 0.07 0.00 0.04 0.12 Observations 3,963 3,942 3,945 3,615 Outcome range {1,...,10}{1,...,10}{1,...,10}{1,...,10} Control outcome mean 6.75 6.74 6.87 8.75 Control outcome std. dev. 2.79 2.60 2.63 1.87

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects and a lagged dependent variable. All observations are inversely weighted by the baseline number of respondents surveyed in the village. Standard errors are clustered by village. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis.

A.8.3 Effects on evaluations of challenger parties

Table A9 examines how our information treatments affected voter beliefs about prospective challenger performance in ofﬁce (if elected). Since the direction of the effect did not have a clear theoretical expectation, we use two-sided tests. Columns (1) and (2) indicate that voters receiving the benchmark also updated positively about challengers, albeit far less positively than about incumbents. In the context of the model in Appendix section A.1, this suggests that—to the extent that challengers and previous incumbents are believed to be correlated—-previous incumbent performance information exceeded expectations. However, the heterogeneous effects in columns (3)-(5) do not support this interpretation, given that treatment effects on prospective challenger evaluations are not increasing in previous incumbent performance. These results thus suggest that benchmarked information did not substantially affect perceptions of challengers. Rather, the results may thus instead reﬂect a perceived positive correlation across all types of current politician

A27 Table A8: Effects of information treatments on posterior belief precision (endline survey)

Incumbent Relative overall performance performance (v. previous) precision precision (1) (2) Incumbent 0.302*** 0.156** (0.068) (0.087) Benchmark 0.448*** 0.536*** (0.072) (0.090)

Null: Incumbent≥Benchmark (p value) 0.02 0.00 Observations 3,852 3,844 Outcome range {1,...,10}{1,...,10} Control outcome mean 5.85 6.10 Control outcome std. dev. 2.65 2.83

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects and a control for the corresponding pre-treatment outcome. All observations are inversely weighted by the baseline number of respondents surveyed in the village. Standard errors are clustered by village. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis.

A28 (Kendall, Nannicini and Trebbi 2014). Importantly, this ﬁnding suggests that the differential effects of benchmarked information on relative evaluations, including vote choice, likely reﬂect the increased weight attached to incumbent performance signals when an accompanying benchmark increases the precision of the signal, rather than the difference between posterior and prior expectations of challenger performance.

A.8.4 Alternative scales for computing outcome and interaction indexes

Tables A10-A15 report the results from all main tables using the ? approach to creating a summa- tive rating scale, instead of our preferred ICW index. This choice does not meaningfully affect the results.

A.8.5 The importance voters attach to incumbent legislative performance does not change

While voters’ evaluations of incumbents were durably affected, the provision of incumbent performance information could also influence the relative weight attached to incumbent legislative performance in making voting decisions. Any changes in voting behavior might then reflect changes in salience, rather than changes in beliefs. To examine such salience effects, we asked voters what the three most important factors in determining their vote choice in the 2017 election were. Table A16 shows that treatments do not affect the likelihood of reporting that national or local legislative performance is one of the three most important, or the most important factor in determining vote choice. This suggests that effects on vote choice are unlikely to reflect voters placing greater weight on the considerations that the information provided relates to.

A.8.6 Within-village information diffusion

The endline survey provides clear evidence of substantial voter engagement with the leaflets within their village. Table A17 shows that nearly 40% of treated respondents discussed the leaflet with others. This suggests that significant information diffusion occurred, which may account for the

A29 Table A9: Effects of information treatments on prospective challenger performance evaluations (baseline survey)

Prospective challenger performance (1) (2) (3) (4) (5) Duties -0.006 (0.047) Incumbent 0.044 0.051 0.051 0.051 0.050 (0.041) (0.032) (0.032) (0.032) (0.032) Incumbent × Duties 0.015 (0.063) Benchmark 0.074** 0.112*** 0.114*** 0.112*** 0.114*** (0.037) (0.027) (0.027) (0.027) (0.026) Benchmark × Duties 0.073 (0.059) Incumbent × Overall previous performance (ICW) 0.027 (0.032) Benchmark × Overall previous performance (ICW) -0.035 (0.026) Incumbent × Relevant previous performance (ICW) -0.013 (0.038) Benchmark × Relevant previous performance (ICW) -0.049 (0.031) Incumbent × National previous performance (ICW) 0.056* (0.030) Benchmark × National previous performance (ICW) 0.003 (0.021) Incumbent × Local previous performance (ICW) -0.019 (0.031) Benchmark × Local previous performance (ICW) -0.042 (0.025)

Observations 3,888 3,888 3,888 3,888 3,888 Control outcome mean 3.42 3.42 3.42 3.42 3.42 Control outcome std. dev. 0.88 0.88 0.88 0.88 0.88 Interaction mean -0.01 -0.02 -0.01 Interaction std. dev. 1.69 1.14 1.59 Second interaction mean 0.01 Second interaction std. dev. 1.01

A30 Table A10: Average effects of information treatments on beliefs about incumbent performance, intention to vote for the incumbent, and requests from the incumbent (baseline survey)—alternative index

Benchmark - Incumbent 0.076** 0.089** 0.066** 0.005 -0.002 0.061** -0.022 -0.008 -0.033 (0.04) (0.041) (0.035) (0.01) (0.013) (0.032) (0.013) (0.014) (0.029) Observations 3,942 3,932 3,928 3,999 3,998 3,999 3,999 3,998 3,999 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1} [-2.4,2.0] {0,1}{0,1} [-1.5,0.7] Control outcome mean 2.83 3.20 3.15 0.59 0.67 0.01 0.70 0.70 0.01 Control outcome std. dev. 1.07 0.90 1.09 0.49 0.47 1.01 0.46 0.46 1.00

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects. All specifications include a lagged dependent variable as a control; in columns (5)-(9), pre-treatment incumbent vote is used as a proxy. All observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis.

A31 Table A11: Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (baseline survey)—alternative index

A32 Benchmark × Relevant performance (alpha) 0.156*** 0.130*** 0.125*** 0.022** 0.012 0.118** * -0.009 0.015 0.007 (0.038) (0.037) (0.032) (0.012) (0.015) (0.030) (0.015) (0.016) (0.033) Panel C: Heterogeneity by (standardized) local and national reported performance level Incumbent 0.357*** 0.285*** 0.256*** 0.031*** 0.033*** 0.278*** 0.046*** 0.033*** 0.088*** (0.046) (0.042) (0.039) (0.009) (0.015) (0.032) (0.013) (0.013) (0.028) Incumbent × National performance (alpha) 0.119*** 0.141*** 0.060* 0.005 0.000 0.053* -0.001 0.012 0.012 (0.055) (0.055) (0.045) (0.008) (0.016) (0.037) (0.016) (0.016) (0.034) Incumbent × Local performance (alpha) 0.316*** 0.268*** 0.247*** 0.024*** 0.002 0.209*** 0.008 0.007 0.017 (0.047) (0.039) (0.043) (0.011) (0.015) (0.035) (0.014) (0.014) (0.030) Benchmark 0.428*** 0.373*** 0.319*** 0.036*** 0.032*** 0.337*** 0.023** 0.026** 0.054** (0.050) (0.054) (0.042) (0.011) (0.013) (0.038) (0.013) (0.013) (0.028) Benchmark × National performance (alpha) 0.061 0.005 -0.006 0.012 -0.006 0.002 -0.001 0.002 0.001 (0.068) (0.093) (0.056) (0.010) (0.015) (0.058) (0.015) (0.014) (0.030) Benchmark × Local performance (alpha) 0.197*** 0.167*** 0.163*** 0.037*** 0.010 0.163*** 0.000 0.016 0.018 (0.047) (0.037) (0.041) (0.013) (0.014) (0.033) (0.014) (0.015) (0.032) Observations 3,942 3,932 3,928 3,999 3,998 3,890 3,999 3,998 3,998 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1} [-2.4,1.9] {0,1}{0,1} [-1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.59 0.67 0.02 0.70 0.70 0.01 Control outcome std. dev. 1.07 0.90 1.09 0.49 0.47 1.01 0.46 0.46 1.01 Overall performance (alpha) range [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] Relevant performance (alpha) range [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] National performance (alpha) range [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] Local performance (alpha) range [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36]

Notes: See Table A10. Lower-order interaction terms are included but not shown. Panel B also includes the interaction between incumbent and benchmark treatments and an indicator for respondents that did not regard local or national performance as one of the top three most important factors in determining their vote choice. (Continued...) Table A11 (continued): Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (baseline survey)—alternative index

A33 (0.039) (0.033) (0.035) (0.012) (0.016) (0.025) (0.017) (0.016) (0.036) Panel F: Heterogeneity by importance of performance in determining vote choice Incumbent 0.355*** 0.284*** 0.261*** 0.031*** 0.032*** 0.281*** 0.039*** 0.029*** 0.076*** (0.048) (0.043) (0.040) (0.009) (0.014) (0.034) (0.013) (0.013) (0.028) Incumbent × Performance most important -0.007 -0.008 0.009 0.020*** 0.013 0.007 0.021** 0.019* 0.045** (0.033) (0.030) (0.030) (0.009) (0.013) (0.023) (0.012) (0.012) (0.026) Benchmark 0.431*** 0.373*** 0.328*** 0.036*** 0.030*** 0.341*** 0.017 0.022* 0.043* (0.051) (0.052) (0.043) (0.012) (0.013) (0.039) (0.013) (0.013) (0.028) Benchmark × Performance most important -0.001 -0.023 -0.010 0.016** 0.030*** 0.020 0.024** 0.019* 0.047** (0.031) (0.028) (0.029) (0.009) (0.014) (0.021) (0.012) (0.013) (0.027) Panel G: Heterogeneity by preference for locally-oriented deputies Incumbent 0.370*** 0.192*** 0.198*** -0.018 0.039* 0.235*** 0.041* 0.035 0.085* (0.079) (0.069) (0.065) (0.024) (0.029) (0.058) (0.027) (0.028) (0.059) Incumbent × Prefer locally-oriented deputies -0.019 0.130** 0.089* 0.068*** -0.011 0.065 -0.002 -0.009 -0.012 (0.075) (0.069) (0.066) (0.028) (0.034) (0.062) (0.031) (0.032) (0.068) Benchmark 0.502*** 0.286*** 0.302*** 0.014 0.040 0.335*** 0.028 0.028 0.062 (0.077) (0.088) (0.067) (0.023) (0.031) (0.069) (0.030) (0.031) (0.065) Benchmark × Prefer locally-oriented deputies -0.098 0.123* 0.036 0.029 -0.015 0.009 -0.016 -0.010 -0.028 (0.073) (0.080) (0.067) (0.024) (0.037) (0.064) (0.033) (0.033) (0.071) Observations 3,908 3,906 3,905 3,922 3,921 3,890 3,922 3,921 3,921 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1} [-2.4,1.9] {0,1}{0,1} [-1.6,0.7] Control outcome mean 2.83 3.20 3.15 0.60 0.68 0.02 0.70 0.70 0.02 Control outcome std. dev. 1.08 0.90 1.10 0.49 0.47 1.01 0.46 0.46 1.00 Prior index (alpha) range [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] Prior precision (alpha) range [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] Performance most important range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Prefer locally-oriented deputies range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}

Notes: See Table A10. Lower-order interaction terms are included but not shown. Table A12: Average effects of information treatments on beliefs about incumbent performance, reported vote for the incumbent, and requests from the incumbent (endline survey)—alternative index

Benchmark - Incumbent 0.085*** 0.102*** 0.026* 0.024* 0.123*** 0.009** 0.080 0.000 0.010 0.079** (0.032) (0.034) (0.02) (0.018) (0.038) (0.004) (0.066) (0.007) (0.015) (0.044) Observations 3,834 3,825 3,781 3,781 3,876 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5}{1,...,5}{0,1}{0,1} [-2.4,2.0] {0,1}{0,1}{0,1}{0,1} [-1.5,0.7] Control outcome mean 3.08 3.46 0.64 0.53 -0.00 0.98 0.98 0.95 0.11 -0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 0.99 0.14 0.14 0.21 0.32 1.01

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects. All specifications include a lagged dependent variable as a control; in columns (5)-(10), pre-treatment incumbent vote is used as a proxy. All observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis.

A34 Table A13: Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (endline survey)—alternative index

A35 Benchmark × Relevant performance (alpha) 0.008 0.011 0.005 0.008 0.008 -0.003 -0.061 -0.002 0.026** 0.002 (0.033) (0.032) (0.020) (0.018) (0.036) (0.005) (0.046) (0.006) (0.014) (0.040) Panel C: Heterogeneity by (standardized) local and national reported performance level Incumbent 0.147*** 0.140*** -0.022 -0.021 0.111*** -0.001 -0.015 0.003 0.015 0.023 (0.035) (0.034) (0.021) (0.020) (0.037) (0.005) (0.028) (0.007) (0.014) (0.037) Incumbent × National performance (alpha) -0.039 -0.031 0.004 -0.018 -0.052 -0.009 -0.005 0.004 0.040*** 0.029 (0.044) (0.041) (0.023) (0.022) (0.044) (0.007) (0.022) (0.009) (0.018) (0.054) Incumbent × Local performance (alpha) 0.055** 0.090*** 0.004 0.008 0.055* 0.005* 0.011 0.015*** 0.024* 0.104*** (0.032) (0.032) (0.021) (0.021) (0.034) (0.004) (0.024) (0.007) (0.015) (0.035) Benchmark 0.228*** 0.240*** 0.005 0.003 0.231*** 0.008** 0.064 0.004 0.025* 0.103*** (0.034) (0.035) (0.021) (0.020) (0.039) (0.004) (0.056) (0.007) (0.015) (0.043) Benchmark × National performance (alpha) -0.035 -0.075 0.013 -0.001 -0.061 -0.017+++ -0.073 -0.021++ 0.023 -0.118+ (0.040) (0.048) (0.022) (0.022) (0.045) (0.006) (0.055) (0.010) (0.020) (0.064) Benchmark × Local performance (alpha) 0.002 0.035 0.018 0.026 0.024 -0.000 -0.068 0.009* 0.015 0.024 (0.032) (0.030) (0.022) (0.022) (0.037) (0.004) (0.056) (0.007) (0.017) (0.041) Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5}{1,...,5}{0,1}{0,1} [-2.8,1.9] {0,1}{0,1}{0,1}{0,1} [-7.8,1.6] Control outcome mean 3.08 3.46 0.64 0.53 -0.00 0.98 0.98 0.95 0.11 -0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 0.99 0.14 0.14 0.21 0.32 1.01 Overall performance (alpha) range [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] [-1.54,1.05] Relevant performance (alpha) range [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] [-1.76,2.80] National performance (alpha) range [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] [-1.27,2.03] Local performance (alpha) range [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36] [-0.90,1.36]

Notes: See Table A12. Lower-order interaction terms are included but not shown. Panel B also includes the interaction between incumbent and benchmark treatments and an indicator for respondents that did not regard local or national performance as one of the top three most important factors in determining their vote choice. (Continued...) Table A13 (continued): Heterogeneous effects of information treatments by leaﬂet content, priors beliefs, and importance of performance information for vote choice (endline survey)—alternative index

A36 (0.038) (0.038) (0.020) (0.021) (0.043) (0.004) (0.041) (0.009) (0.017) (0.045) Panel F: Heterogeneity by importance of performance in determining vote choice Incumbent 0.161*** 0.144*** -0.024 -0.018 0.120*** -0.000 -0.016 0.006 0.011 0.026 (0.034) (0.034) (0.021) (0.020) (0.036) (0.004) (0.028) (0.008) (0.015) (0.036) Incumbent × Performance most important 0.021 -0.007 0.001 0.011 0.009 -0.006 0.007 -0.001 0.003 -0.021 (0.031) (0.032) (0.019) (0.019) (0.034) (0.004) (0.013) (0.008) (0.013) (0.040) Benchmark 0.246*** 0.246*** 0.002 0.006 0.243*** 0.009*** 0.063 0.006 0.021* 0.104*** (0.033) (0.035) (0.021) (0.020) (0.037) (0.004) (0.054) (0.007) (0.016) (0.041) Benchmark × Performance most important -0.018 0.002 0.025* 0.031** 0.015 -0.005 0.060 -0.003 0.028*** 0.033 (0.032) (0.031) (0.019) (0.018) (0.032) (0.004) (0.052) (0.007) (0.013) (0.040) Panel G: Heterogeneity by preference for locally-oriented deputies Incumbent 0.104** 0.103** -0.095+++ -0.073++ 0.029 -0.018++ -0.004 0.006 0.018 -0.036 (0.062) (0.056) (0.033) (0.035) (0.062) (0.008) (0.038) (0.015) (0.026) (0.076) Incumbent × Prefer locally-oriented deputies 0.079 0.057 0.099*** 0.077** 0.127** 0.025*** -0.015 -0.001 -0.009 0.088 (0.075) (0.063) (0.039) (0.043) (0.071) (0.011) (0.046) (0.017) (0.028) (0.092) Benchmark 0.191*** 0.227*** -0.048 -0.045 0.174*** -0.003 0.283 0.002 0.050** 0.157* (0.057) (0.056) (0.036) (0.035) (0.062) (0.005) (0.282) (0.014) (0.029) (0.104) Benchmark × Prefer locally-oriented deputies 0.077 0.025 0.070* 0.072* 0.096* 0.016*** -0.306 0.005 -0.041 -0.074 (0.067) (0.066) (0.043) (0.045) (0.073) (0.007) (0.324) (0.015) (0.032) (0.118) Observations 3,834 3,825 3,781 3,781 3,708 3,876 3,876 3,876 3,876 3,876 Outcome range {1,...,5}{1,...,5}{0,1}{0,1} [-2.8,1.9] {0,1}{0,1}{0,1}{0,1} [-7.8,1.6] Control outcome mean 3.08 3.46 0.64 0.53 -0.00 0.98 0.98 0.95 0.11 -0.00 Control outcome std. dev. 0.93 0.95 0.48 0.50 0.99 0.14 0.14 0.21 0.32 1.01 Prior index (alpha) range [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] [-2.37,2.19] Prior precision (alpha) range [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] [-2.73,1.38] Performance most important range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Prefer locally-oriented deputies range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}

Notes: See Table A12. Lower-order interaction terms are included but not shown. Table A14: Effects of information treatments on posterior beliefs and reported vote for the incumbent, among those that turned out in 2012 (baseline and endline survey)—alternative index

Incumbent overall Incumbent vote Incumbent vote performance (endline) intention (validated) (1) (2) (3) (4) (5) (6) (7) (8) (9) Incumbent 0.134** 0.137** 0.117** 0.015 0.019 0.012 0.034 0.040 0.013 (0.059) (0.059) (0.056) (0.014) (0.015) (0.015) (0.031) (0.031) (0.031) Benchmark 0.235*** 0.236*** 0.211*** 0.022 0.021 0.019 0.034 0.041 0.020 (0.057) (0.056) (0.059) (0.015) (0.016) (0.016) (0.028) (0.028) (0.027) Incumbent × Relevant performance (alpha) 0.039 0.033** 0.044 (0.062) (0.016) (0.033) Benchmark × Relevant performance (alpha) 0.029 0.006 0.045 (0.053) (0.015) (0.032) Incumbent × National performance (alpha) -0.035 -0.024* -0.086*** (0.058) (0.014) (0.029) Incumbent × Local performance (alpha) 0.050 0.016 0.026 A37 (0.054) (0.017) (0.034) Benchmark × National performance (alpha) -0.041 -0.008 -0.049 (0.081) (0.016) (0.031) Benchmark × Local performance (alpha) -0.028 0.024 0.078*** (0.057) (0.017) (0.028)

Observations 1,469 1,469 1,469 1,528 1,528 1,528 1,435 1,435 1,435 Outcome range {1,...,5}{1,...,5}{1,...,5}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Control outcome mean 3.10 3.10 3.10 0.63 0.63 0.63 0.59 0.59 0.59 Control outcome std. dev. 0.93 0.93 0.93 0.48 0.48 0.48 0.49 0.49 0.49 Interaction mean 0.02 -0.14 0.03 -0.14 0.02 -0.17 Interaction std. dev. 1.02 0.78 1.02 0.77 1.0 2 0.75 Second interaction mean 0.14 0.15 0.14 Second interaction std. dev. 0.95 0.95 0.95

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects. Lower-order interaction terms are included but not shown. All specifications include a lagged dependent variable as a control; in columns (5)-(9), pre-treatment incumbent vote is used as a proxy. All observations are inversely weighted by the number of respondents surveyed in the village. Standard errors are clustered by village. Given that these hypotheses were not pre-specified, * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. Table A15: Effects of information treatments on polling station-level incumbent vote share, by leaflet content (polling station data)—alternative index

Incumbent vote share Incumbent vote share (proportion of turnout) (proportion of registered voters) (1) (2) (3) (4) (5) (6) Incumbent 0.001 -0.002 -0.022 0.006 0.003 -0.017 (0.026) (0.026) (0.026) (0.019) (0.018) (0.018) Benchmark -0.003 -0.007 -0.026 -0.003 -0.007 -0.024* (0.024) (0.024) (0.026) (0.016) (0.015) (0.017) Incumbent × Relevant performance (alpha) 0.026* 0.036*** (0.019) (0.013) Benchmark × Relevant performance (alpha) 0.031* 0.037*** (0.021) (0.012) Incumbent × National performance (alpha) 0.003 0.008 (0.038) (0.022) Incumbent × Local performance (alpha) 0.047** 0.041*** (0.024) (0.018) Benchmark × National performance (alpha) 0.015 -0.003 (0.034) (0.022) Benchmark × Local performance (alpha) 0.037* 0.033*** (0.024) (0.016)

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects. Lower-order interaction terms are included but not shown. All observations are inversely weighted by the registered voters in the experimental village as a share of those at the polling station. Standard errors are clustered by village. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. fact that directly providing leaflets to less than 2% of registered voters still resulted in some dis- cernible polling station-level effects.

A.8.7 Cross-village informational spillovers

Another possibility is that information crossed from treated to control villages. Such cross-village spillovers could account for the lack of average effects on incumbent vote choice if control villages similarly become more positive about the incumbent. To estimate spillovers to the 75 villages in

A38 Table A16: Effects of information treatments on self-reported importance of performance in making vote choice (endline survey)

Performance is one of Performance is the the three most important most important factors in vote choice factor in vote choice (1) (2) (3) (4) (5) (6) Duties 0.005 0.001 (0.014) (0.020) Incumbent -0.018 -0.017* -0.009 -0.024* (0.014) (0.010) (0.020) (0.014) Incumbent × Duties 0.001 -0.029 (0.022) (0.031) Benchmark 0.010 0.002 0.001 -0.006 (0.014) (0.010) (0.021) (0.016) Benchmark × Duties -0.015 -0.013 (0.019) (0.030) Performance -0.008 -0.015 (0.008) (0.012)

Null: Incumbent≥Benchmark (p value) 0.03 0.15 Observations 3,876 3,876 3,876 3,876 3,876 3,876 Outcome range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Control outcome mean 0.89 0.89 0.89 0.50 0.50 0.50 Control outcome std. dev. 0.31 0.31 0.31 0.50 0.50 0.50

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects and a lagged dependent variable. All observations are inversely weighted by the endline number of respondents surveyed in the village. Standard errors are clustered by village. Given that these hypotheses were not pre- specified, * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests.

A39 Table A17: Effects of information treatments on information diffusion (endline survey)

Discussed leaﬂet with others Incumbent 0.372*** (0.028) Benchmark 0.390*** (0.027)

One-sided null: Incumbent ≥ Benchmark (p value) 0.13 Observations 3,875 Outcome range {0,1} Control outcome mean 0.00 Control outcome std. dev. 0.07

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects. All observations are inversely weighted by the baseline number of respondents surveyed in the village. Standard errors are clustered by village. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis. the pure control group, we restrict the sample to this subset of villages and define spillovers as the number of villages within xkm of a treated village receiving performance (incumbent or benchmark) information. Panels A, B, and C of Appendix Table A18 indicate that for treated villages respectively within 1km, 2.5km, and 5km of a village, there is no consistent evidence to suggest that proximity to performance information significantly affected endline voter beliefs or voting behavior, conditional on the number of villages within our sample within the same distance.20 This applies both on average, as well as by the level of reported performance. Although the final four columns in panels A and C tentatively suggest positive effects, a spillover explanation is hard to substantiate given the lack of change in self-reported beliefs and behaviors—unless effects were concentrated entirely among older voters. 20Unreported results also show that leaflet recall is also unaffected by treatment assignment.

A40 Table A18: Effects of information spillovers on voter beliefs, self-reported vote choices, and precinct-level vote choices (endline survey and polling station data)

Incumbent overall Relative performance Incumbent Incumbent Incumbent vote Incumbent vote performance (v. previous) vote vote (validated) share (turnout) share (registered) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Spillovers within 1km Performance spillover 0.017 0.022 -0.065 0.002 -0.169* -0.084 -0.070 0.022 0.157 0.097 0.054 0.063 (0.174) (0.176) (0.138) (0.152) (0.098) (0.103) (0.079) (0.072) (0.116) (0.173) (0.078) (0.126) Performance spillover × Performance index (ICW) -0.387 -0.467* -0.283 -0.339 0.282 -0.024 (0.301) (0.277) (0.231) (0.250) (0.645) (0.423)

Spillover mean 0.12 0.12 0.12 0.12 0.12 0.12 0.12 0.12 0.13 0.13 0.13 0.13 Panel B: Spillovers within 2.5km Performance spillover -0.006 0.055 -0.099 -0.207 -0.103* -0.095 -0.091 -0.039 -0.050 0.086 -0.033 0.043 (0.125) (0.152) (0.127) (0.169) (0.060) (0.070) (0.063) (0.070) (0.102) (0.086) (0.059) (0.076) A41 Performance spillover × Performance index (ICW) -0.207 0.308 -0.026 -0.165 -0.201 -0.119 (0.303) (0.287) (0.141) (0.158) (0.181) (0.105)

Spillover mean 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.56 0.56 0.56 0.56 Panel C: Spillovers within 5km Performance spillover -0.147 -0.165 -0.161 -0.188 0.136 0.125 0.161* 0.152 0.040 0.072 0.013 0.058 (0.144) (0.183) (0.137) (0.150) (0.104) (0.107) (0.087) (0.094) (0.076) (0.083) (0.049) (0.039) Performance spillover × Performance index (ICW) -0.006 -0.064 -0.039 -0.029 0.025 0.051** (0.084) (0.088) (0.058) (0.050) (0.044) (0.023)

Spillover mean 0.85 0.85 0.85 0.85 0.85 0.85 0.85 0.85 0.86 0.86 0.86 0.86 Observations 635 635 635 635 632 632 632 632 72 72 72 72 Control outcome mean 3.08 3.08 3.46 3.46 0.64 0.64 0.53 0.53 0.69 0.69 0.41 0.41 Control outcome std. dev. 0.93 0.93 0.95 0.95 0.48 0.48 0.50 0.50 0.18 0.18 0.13 0.13

Notes: Each specification is estimated using OLS, and includes department and enumerator fixed effects. Lower-order interaction terms are included but not shown. Observations in columns (1)-(8) are inversely weighted by the baseline number of respondents surveyed in the village. Standard errors are clustered by village. Given that these hypotheses were not pre-specified, * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. A.8.8 Party responses to information dissemination

Politicians rarely stand by when potentially influential information is released (e.g. Arias et al. 2018b; Bidwell, Casey and Glennerster 2016). Consequently, a possible explanation for the lack of a persistent average treatment effect on incumbent electoral support, but positive effects when interacted with the information content, is that challenger parties were particularly effective at counteracting information that generally increased favorability toward the incumbent. Incumbents may also respond by highlighting positive information, although—to the extent that it is effective— this should reinforce the favorable immediate updating of voters. Another channel through which strategic responses could explain our findings is if incumbents (challengers) reallocate resources from treatment (control) to control (treatment) villages upon learning that favorable information had already been disseminated. We investigated such equilibrium campaign responses to information dissemination by using our endline survey to gauge two types of party or candidate action. First, we asked respondents if, and how, the incumbent or challenger parties (or their agents) responded specifically to the leaflet’s provision. Second, we used a list experiment to measure the extent of vote buying, in order to assess whether party electoral strategies change, even without explicitly mentioning the leaflets.21 As shown in columns (1) and (5) of Table A19, challengers and especially incumbents responded directly to the intervention. As the almost-zero control group mean indicates, responses were concentrated in treated villages. Decomposing candidate responses by type, the vast majority of incumbent responses involved a community meeting or talking with the village chief, while challenger parties held community meetings or had party operatives visit voters. If incumbent responses are at least as effective as challenger responses, it is hard to account for the zero average effects observed at the individual and polling station levels. To better understand what parties 21Half the sample was subject to a list experiment including incumbent vote buying as the omitted option from the list; vote buying by a challenger party was omitted for the other half of the sample.

A42 Table A19: Effects of information treatments on incumbent and challenger responses (endline survey)

Incumbent response Challenger response (1) (2) (3) (4) (5) (6) (7) (8) Incumbent 0.066*** 0.067*** 0.067*** 0.067*** 0.042*** 0.042*** 0.042*** 0.047*** (0.012) (0.012) (0.012) (0.012) (0.009) (0.009) (0.009) (0.009) Benchmark 0.078*** 0.077*** 0.078*** 0.079*** 0.047*** 0.046*** 0.046*** 0.051*** (0.011) (0.010) (0.010) (0.011) (0.009) (0.009) (0.009) (0.009) Incumbent × Overall performance (ICW) -0.007 -0.004 (0.013) (0.007) Benchmark × Overall performance (ICW) -0.023** -0.021** (0.011) (0.009) Incumbent × Relevant performance (ICW) 0.014 0.000 (0.016) (0.012) Benchmark × Relevant performance (ICW) 0.001 -0.012 A43 (0.013) (0.012) Incumbent × National performance (ICW) -0.044*** -0.028*** (0.011) (0.008) Benchmark × National performance (ICW) -0.051*** -0.037*** (0.011) (0.008) Incumbent × Local performance (ICW) 0.039*** 0.022** (0.012) (0.009) Benchmark × Local performance (ICW) 0.040*** 0.018** (0.012) (0.008)

Observations 3,875 3,875 3,875 3,875 3,875 3,875 3,875 3,875 Control outcome mean 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 Control outcome std. dev. 0.08 0.08 0.08 0.08 0.08 0.08 0.08 0.08

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects. Lower-order interaction terms are included but not shown. All observations are inversely weighted by the baseline number of respondents surveyed in the village. Lower order interactions between the list experiment and performance indexes are omitted. Standard errors are clustered by village. Given that these hypotheses were not pre-specified, * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. did, we followed up with respondents in December 2017 to ask about what actions parties took and whether they were effective. Voters that reported incumbent-held community meetings or dis- cussions with the chief were convinced to vote for the incumbent 70-80% of the time, while the less-frequent challenger community meetings and party visits rarely convinced or even encouraged voters to support them. The interactions with national and local incumbent performance, in column (4), suggest that incumbents capitalized on positive local performance information. Given such responses were compelling to voters, and likely reached a broader electorate that was more likely to turn out and which Table8 found to be more receptive to local performance than our survey respondents, this could explain the positive effects of treatment on incumbent vote share at the polling station level where local performance was strongest. The lack of an effect on average at the polling station could then reflect effective but relatively sparse incumbent responses. Column (8) indicates that challengers sought to counteract such efforts, but—as noted above—these were rarely seen as effective. In contrast, both incumbents and challengers respond more to national performance information when they performed poorly, although this is not a major factor determining vote choices. Although vote buying is prevalent, we were not able to detect a systematic indirect response to information dissemination through vote buying. Table A20 uses a list experiment to examine the effects of the information treatments on vote buying: half the sample received the control list containing three items, while one quarter of the sample received a 4-item list either containing incumbent vote buying or challenger vote buying. The results of the list experiment in columns (1) and (5) A20 indicate that 22% of voters reported receiving a gift from the incumbent, while another 22% reported receiving a gift from a challenger. Although such vote buying was a little lower in treated villages, especially among challengers where local incumbent performance was strong, the estimates are too imprecise to be able to conclude that the substitution of vote buying across villages can account for our findings.

A44 Table A20: Effects of information treatments on vote buying (endline survey)

Incumbent list experiment Challengers list experiment Items Items Items Items Items Items Items Items listed listed listed listed listed listed listed listed (1) (2) (3) (4) (5) (6) (7) (8) Treatment list 0.215*** 0.223*** 0.225*** 0.230*** 0.219*** 0.247*** 0.244*** 0.248*** (0.031) (0.052) (0.051) (0.051) (0.031) (0.051) (0.051) (0.049) Treatment list × Incumbent -0.040 -0.039 -0.047 -0.079 -0.075 -0.088 (0.067) (0.066) (0.067) (0.079) (0.078) (0.077) Treatment list × Benchmark 0.016 0.019 0.005 -0.005 0.004 0.000 (0.076) (0.076) (0.076) (0.070) (0.070) (0.072) Treatment list × Incumbent × Overall performance (ICW) 0.031 0.005 (0.067) (0.066) Treatment list × Benchmark × Overall performance (ICW) 0.069 0.048 A45 (0.068) (0.052) Treatment list × Incumbent × National performance (ICW) 0.000 -0.011 (0.086) (0.102) Treatment list × Benchmark × National performance (ICW) 0.016 0.062 (0.104) (0.080) Treatment list × Incumbent × Local performance (ICW) -0.039 -0.123 (0.086) (0.099) Treatment list × Benchmark × Local performance (ICW) -0.037 -0.065 (0.102) (0.092)

Observations 2,893 2,893 2,893 2,893 2,933 2,933 2,933 2,933 Control outcome mean 1.59 1.59 1.59 1.59 1.62 1.62 1.62 1.62 Control outcome std. dev. 0.72 0.72 0.72 0.72 0.73 0.73 0.73 0.73

Table A21 reports the polling station level results for the sample of experimental villages, weighting observations by the fraction of the registered voter pool residing in an experimental village.22 The results are qualitatively similar to those in Table9, but are smaller in magnitude. The smaller coefﬁcient values likely reﬂect adding greater than zero weight to observations where very few voters could have been exposed to treatment within their village.

Table A21: Effects of information treatments on polling station-level incumbent vote share, by leaﬂet content and weight by the share of registered voters within a polling station’s experimental village (polling station data)

Incumbent vote share Incumbent vote share (proportion of turnout) (proportion of registered voters) (1) (2) (3) (4) (5) (6) Incumbent -0.001 -0.002 -0.006 0.003 0.001 -0.006 (0.021) (0.021) (0.020) (0.016) (0.015) (0.014) Benchmark 0.004 0.002 -0.005 0.003 0.000 -0.006 (0.020) (0.020) (0.019) (0.013) (0.013) (0.012) Incumbent × Relevant performance (ICW) 0.027 0.037*** (0.022) (0.016) Benchmark × Relevant performance (ICW) 0.025 0.032*** (0.022) (0.014) Incumbent × National performance (ICW) -0.033 -0.016 (0.028) (0.015) Benchmark × National performance (ICW) -0.003 -0.014 (0.028) (0.021) Incumbent × Local performance (ICW) 0.054*** 0.043*** (0.026) (0.016) Benchmark × Local performance (ICW) 0.021 0.029* (0.026) (0.018)

Observations 440 440 440 440 440 440 Control outcome mean 0.70 0.70 0.70 0.41 0.41 0.41 Control outcome std. dev. 0.17 0.17 0.17 0.13 0.13 0.13

Notes: Each specification is estimated using OLS, and includes randomization block fixed effects and a lagged dependent variable. Lower- order interaction terms are included but not shown. Observations are weighted by the share of registered voters at the polling station that are registered in the associated experimental village. Robust standard errors are in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests; + denotes p < 0.1, ++ denotes p < 0.05, +++ denotes p < 0.01 from two-sided tests when coefficients point in the opposite direction to the pre-specified hypothesis.

22We were unable to obtain complete electoral returns in four villages.

A46 A.8.10 Effects on electoral turnout

Table A22 examines the effects of the information treatments on self-reported turnout, finding little evidence to suggest that turnout decisions were influenced by the information. Consistent with the aggregate findings in Table A23, turnout is if anything a little higher in treated villages where the incumbent performed better.

A47 Table A22: Effects of information treatments on turnout (endline survey)

Turnout (self-reported) Turnout (validated) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) Incumbent -0.014 -0.011 -0.023 -0.010 -0.016 -0.017 -0.001 -0.016 -0.014 -0.003 -0.014 -0.017 -0.012 -0.032 (0.019) (0.018) (0.020) (0.020) (0.019) (0.021) (0.027) (0.020) (0.019) (0.023) (0.020) (0.020) (0.021) (0.028) Benchmark 0.005 0.008 0.013 0.007 0.003 0.009 -0.009 0.009 0.011 0.029 0.007 0.004 0.017 -0.025 (0.016) (0.016) (0.018) (0.017) (0.016) (0.017) (0.024) (0.019) (0.019) (0.022) (0.020) (0.019) (0.019) (0.027) Incumbent × Overall performance (ICW) 0.015 0.006 (0.016) (0.024) Benchmark × Overall performance (ICW) 0.016 0.015 (0.013) (0.019) Incumbent × Relevant performance (ICW) 0.011 0.027 (0.022) (0.025) Benchmark × Relevant performance (ICW) 0.001 -0.004 (0.019) (0.024) Incumbent × National performance (ICW) 0.003 -0.019 (0.018) (0.023) Benchmark × National performance (ICW) 0.012 -0.001 (0.016) (0.020) Incumbent × Local performance (ICW) 0.007 0.040*

A48 (0.021) (0.021) Benchmark × Local performance (ICW) -0.002 0.018 (0.020) (0.022) Incumbent × Prior index (ICW) -0.019 0.020 (0.018) (0.019) Benchmark × Prior index (ICW) -0.009 0.021 (0.016) (0.020) Incumbent × Prior precision index (ICW) -0.020 -0.011 (0.020) (0.021) Benchmark × Prior precision index (ICW) 0.002 0.003 (0.019) (0.021) Incumbent × Performance most important -0.023 0.030 (0.034) (0.037) Benchmark × Performance most important 0.026 0.062* (0.033) (0.036)

Observations 3,874 3,874 3,874 3,874 3,801 3,551 3,874 3,876 3,876 3,876 3,876 3,803 3,553 3,876 Outcome range {0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1}{0,1} Control outcome mean 0.74 0.74 0.74 0.74 0.74 0.74 0.74 0.53 0.53 0.53 0.53 0.53 0.53 0.53 Control outcome std. dev. 0.44 0.44 0.44 0.44 0.44 0.44 0.44 0.50 0.50 0.50 0.50 0.50 0.50 0.50

Notes: Each specification is estimated using OLS, and includes randomization block and enumerator fixed effects and a control for the corresponding pre-treatment outcome. Lower-order interaction terms are included but not shown. All observations are inversely weighted by the baseline number of respondents surveyed in the village. Standard errors are clustered by village. Given that these hypotheses were not pre-specified, * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. Table A23: Effects of information treatments on polling station-level turnout, by leaflet content (polling station data)

Turnout (1) (2) (3) Incumbent 0.001 -0.004 -0.012 (0.012) (0.013) (0.016) Benchmark -0.009 -0.012 -0.018 (0.011) (0.013) (0.016) Incumbent × Relevant performance (ICW) 0.036*** (0.018) Benchmark × Relevant performance (ICW) 0.025* (0.018) Incumbent × National performance (ICW) 0.003 (0.020) Benchmark × National performance (ICW) -0.006 (0.023) Incumbent × Local performance (ICW) 0.021 (0.019) Benchmark × Local performance (ICW) 0.011 (0.021) Observations 284 284 284 Outcome range {0.15,1.67}{0.15,1.67}{0.15,1.67} Control outcome mean 0.58 0.58 0.58 Control outcome std. dev. 0.11 0.11 0.11

Notes: Each specification is estimated using OLS, and includes randomization block fixed effects and a lagged dependent variable. Lower-order interaction terms are included but not shown. Observations are not weighted, and polling stations where the village in our sample comprises less than 50% of registered voters at the polling station are excluded. Robust standard errors are in parentheses. Given that these hypotheses were not pre-specified, * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests.

A49