AN EMPIRICAL ANATOMY OF POLITICAL ACCOUNTABILITY: EXPERIMENTAL EVIDENCE FROM A PRE-ELECTION

Transcription

1 AN EMPIRICAL ANATOMY OF POLITICAL ACCOUNTABILITY: EXPERIMENTAL EVIDENCE FROM A PRE-ELECTION INFORMATION DISSEMINATION CAMPAIGN IN SENEGAL ABHIT BHANDARI HORACIO LARREGUY JOHN MARSHALL FEBRUARY 2018 The mixed effects of information campaigns on electoral accountability suggest complex interactions between information content, voters, and politicians. We conducted a field experiment around Senegal s 2017 parliamentary elections to probe the steps linking information dissemination to electoral accountability, as well as non-electoral voter requests from incumbents. We randomized whether leaflets detailed deputies duties, incumbents local and national performance, and benchmarks comparing current to previous incumbent performance. We find that voters immediately processed information as Bayesians, and their updated beliefs persisted for at least a month. Consistent with current incumbents typically outperforming previous incumbents, individual requests of high-performing politicians persistently increase among treated voters, especially following benchmarked information. In contrast, the link from immediate intention to vote for better-performing incumbents to actual vote choice weeks later is mediated by interactions among voters and between voters and parties. Nevertheless, better-forming performing incumbents are still rewards electorally. In spite of complex political equilibria, these findings indicate that voters are willing and able to use information to hold incumbent politicians to account. We thank Fodé Sarr and his team of enumerators for invaluable research assistance, and Elimane Kane and LEGS- Africa for partnering with us to implement this project. We thank conference participants at 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 Harvard Committee on the Use of Human Subjects (IRB ). 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. ab3901@columbia.edu. Department of Government, Harvard University. hlarreguy@fas.harvard.edu. Department of Political Science, Columbia University. jm4401@columbia.edu. 1

2 1 Introduction As a cornerstone of effective democracy, political accountability has received substantial attention from scholars and policy-makers (see Ashworth 2012; Khemani et al. 2016; Manin, Przeworski and Stokes 1999). It is often argued that providing voters with information about their incumbent s performance in office will help them to select high-quality politicians (Fearon 1999). However, a wealth of recent studies examining the effects of informational campaigns on electoral accountability and collective action yield mixed findings (e.g. Banerjee et al. 2011; Björkman Nyqvist, de Walque and Svensson 2017; Chong et al. 2015; de Figueiredo, Hidalgo and Kasahara 2013; Dunning et al. forthcoming; Ferraz and Finan 2008; Humphreys and Weinstein 2012; Larreguy, Marshall and Snyder 2018; Lieberman, Posner and Tsai 2014). Given the complex chain of conditions linking the provision of information to better governance (Lieberman, Posner and Tsai 2014), it is often hard to know where accountability breaks down and under what conditions it can succeed. Moreover, little is known about the extent to which information s effects also carry over to non-electoral forms of political accountability. In this paper, we extensively examine the anatomy of political accountability, and where attempts to stimulate greater accountability break down. We focus on three key links between the provision of information about the performance of parliamentary deputies and electoral accountability, as well as pre- and post-election voter attempts to make requests from their representatives. First, we examine in detail the extent to which voters can process, retain, and ultimately act on incumbent performance information. Second, we vary the content of the information provided to understand what information is most relevant to voters. In particular, we focus on informing voters about the responsibilities of their deputies and comparing incumbent performance to previous incumbents within the same department. Third, we leverage our panel survey structure to study how equilibrium responses to information provision that occur after information dissemination but before elections, both among voters and between voters and local political actors, can alter vote 2

3 intentions derived purely from updated beliefs. To study these core mechanisms underpinning political accountability, we designed a field experiment and sequence of panel surveys in rural Senegal around the 2017 parliamentary elections. We consider this a hard test for the broader theory, given the popular belief that the low levels of education and political knowledge in rural Senegal might prohibit effective accountability (e.g. Gomez and Wilson 2006), the prevalence of bloc voting led by chiefs and other local leaders (Gottlieb and Larreguy 2018; Koter 2013), and the traditionally low levels of parliamentary accountability in Senegal s hyper-presidential system (Thomas and Sissokho 2005). We partnered with LEGS-Africa, a Dakar-based civic organization working to increase civic engagement. Across 450 villages from five of Senegal s departments, our experiment entailed distributing informational leaflets to nine young adults (aged 20-38) in treated villages before the election. Our factorial design varied whether respondents were: (1) informed about the duties of their parliamentary deputies; (2) received performance information pertaining to their incumbent s activity in the legislative, as well as the number and value of projects and transfers received by their district; and/or (3) received incumbent performance information benchmarked against the previous incumbent. In most departments, incumbent performance exceeded voter expectations and the performance of previous incumbents. To separate immediate and longer-term effects of these treatments, we measured the beliefs and (intended and actual) behavioral responses of voters both immediately after treatment and after the election (around a month later). To understand how voters process the information provided, we elicited detailed pre-treatment prior beliefs and intended voting behavior. Our findings first show that rural Senegalese process incumbent performance information in a sophisticated manner. Immediately after receiving the information, voters demonstrate that they are able to retain the leaflets information and update their beliefs accordingly. Voter updating is consistent 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. Moreover, while information about deputy responsibilities does not affect 3

4 beliefs, we find that benchmarked information substantially influences the extent of voter updating and increases the precision of posterior beliefs. Perhaps most remarkably, we observe similar results one month after information was provided. We further show that voters generally 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 the incumbent s performance became three percentage points more likely to intend to vote for the incumbent. Heterogeneity in such rewards is closely tied to the extent to which voters updated relative to their priors, and the extent to which the performance information is the most important factor in determining vote choices. These intentions are backed up by behavioral outcomes indicating voters greater willingness to request posters and visits from incumbent candidates about whom they update positively, as well as opportunities to express their views to those candidates. In spite of voters updated beliefs persisting after the election and many voters proclaiming a willingness to hold politicians to account on the basis of exactly the type of information provided, electoral accountability materializes to a lesser extent than intentions suggested. Specifically, selfreported and polling station voting behavior suggest that, on average, our leaflets ultimately did not affect vote choice. Nevertheless, heterogeneous effects indicate that those caring most about information still rewarded the incumbent andthat the treatments increased the vote shares of the best-performing incumbents, even though we only directly treated less than 2% of registered voters in the electoral precincts containing our villages. In contrast, non-electoral behavioral outcomes show that respondents that updated favorably about incumbent candidates are more likely to make more requests from those candidates, including when having to pay to send a message to the election victor the incumbent in each of our five departments using a hotline that LEGS-Africa set up for this project. This is particularly pronounced among those receiving benchmarked information. Together, these findings suggest that election-oriented responses to the provision of information 4

5 alter voting behavior without substantially changing voters beliefs. Exploring such channels, we document substantial information diffusion, voter coordination, and incumbent and challenger political responses, even to our small-scale intervention. Although deciphering the causal effects of these responses is notoriously challenging (e.g. Imai et al. 2011), our detailed descriptions of these events suggest that equilibrium responses can play a critical role in mediating the effects of beliefbased electoral accountability mechanisms. At minimum, our results show that such responses have the potential to produce polling station-level effects. Ultimately, our findings suggest that accountability failures are more likely to reflect complex political responses than cognitive constraints afflicting voters. We thus make several main contributions. First, the combination of behavioral outcomes with our endline and pre- and post-treatment baseline survey data enable us to unpack the mechanisms underpinning electoral accountability in unprecedented detail. We thus build on the framework expounded by Lieberman, Posner and Tsai (2014) by specifically designing an experiment to assess the key links in the accountability chain. Contrary to concerns that political accountability relies on highly sophisticated voters (Gomez and Wilson 2006), we show that voters with low levels of education and exposure to political news approximate Bayesian information processors. Rather than unstable beliefs, the imperfect mapping from initial intentions to voting behavior appears to reflect interactions precipitated by the dissemination of information. Second, our findings chime with larger recent interventions reporting significant political pushback (Arias et al. 2018; Banerjee et al. 2011; Bidwell, Casey and Glennerster 2016; Chauchard and Sircar forthcoming; Cruz, Keefer and Labonne 2017) and engagement with the information across voters (Arias, Balán, Larreguy, Marshall and Querubín 2017). Moving beyond such studies, we describe in detail the nature and effectiveness of events that occur in between information dissemination and election day, including voter coordination and responses from local political operatives and village chiefs. Our findings suggest that political pushback mediates, rather than fully inhibits, the effects of information dissemination in a context where incumbent performance generally ex- 5

6 ceeded previous incumbents performance. Third, we find that performance information also influences non-electoral forms of political accountability. In contrast with non-our mixed extant findings regarding widely-studied electoral accountability, our original behavioral measures of voter requests from the preferred candidates and winning incumbents suggest that changes in beliefs induced by information campaigns also affect voters willingness to make requests of incumbents. Although we do not examine the types or success of requests of politicians, our findings suggest that creating the sense that politicians are responsive could help hold politicians to account outside of electoral periods. Although she lacks behavioral measures, Gottlieb (2016) also finds that voters expectations of increase upon learning that they are performing comparatively well. Fourth, we further illuminate the types of information that facilitate political accountability. Most notably, we provide the clearest evidence yet that benchmarking can support political accountability by helping voters to separate elements of an individual incumbent s performance from elements affecting all incumbents performance (e.g. Meyer and Vickers 1997). Our stronger findings may reflect our focus on a temporal benchmark, rather than the spatial benchmarks that previous studies have failed to document clear effects of (Arias, Larreguy, Marshall and Querubín 2017; Gottlieb 2016), although it remains possible that our findings are context-specific. Furthermore, we find that Senegalese voters prioritize politicians bringing projects and higher-value transfers to their department. In contrast, greater involvement in parliamentary activities is if anything punished by voters; this adds weight to Adida et al. s (2017) suggestion that Beninese voters seek to remove nationally-oriented politicians. However, we find little evidence that information about incumbent responsibilities influences voter appraisals on its own, or that this substitutes or complements performance information. This suggests that any accountability-enhancing effects of civic education programs (e.g. Gottlieb 2016) are likely to operate through components of the program beyond information about incumbent responsibilities. The paper is structured as follows. Section 2 clarifies the chain of events hypothesized to 6

7 be required for information about incumbent performance to increase electoral and non-electoral accountability. Section 3 provides a brief overview of the Senegalese legislative context. Section 4 explains how our research design enables us to capture the mechanisms of accountability. Section 5 presents the main findings, while Section 6 concludes. 2 Incumbent performance information and political accountability The canonical selection model of electoral accountability reflects an agency interaction between agents (politicians) and their principals (voters), whereby information helps forward-looking voters to select competent or policy-aligned politicians. In its simplest formulation, voters use incumbent performance information to update their beliefs about the likelihood that the incumbent is of highquality, and thus likely to continue to perform well if re-elected (Fearon 1999). Voters re-elect the incumbent if their posterior belief exceeds their expectation of a challenger s quality. In essence, information enables voters to better identify, and thus retain, high-quality incumbents. Voters may similarly be more willing to take costly action to engage beyond the ballot box with incumbents believed to be responsive to their interests. Although the core accountability insight remains, more complex models overlay other features such as (often inefficient) incumbent signaling of quality through their actions in office (e.g. Ashworth and Bueno de Mesquita 2014; Rogoff 1990), the presence of multiple agents (e.g. Buisseret and Prato 2017), and endogenous challenger actions while the incumbent is in office (e.g. Ashworth and Bueno de Mesquita 2008). By providing district-specific information that incumbents could not have anticipated being released just before elections, our design sidesteps these interesting dynamics to focus on whether information can induce voters to sanction and reward incumbents on the basis of their achievements in office. Moreover, by controlling the provision of information, we also abstract from the process through which information is supplied and con- 7

8 sumed (see Besley and Prat 2006; Chan and Suen 2008; Marshall 2017, forthcoming). The extent to which the canonical model captures electoral and off-election behavior rests upon the validity of the model s assumptions regarding the presence of relevant information, voters capacity to process such information, and ultimately voters willingness to act on their updated beliefs about the incumbent s quality. To motivate the hypotheses underpinning our empirical analysis, the following sections describe the anatomy of political accountability by outlining when each of these necessary conditions for political accountability is expected to hold. 2.1 Internalization of novel information The first requirement for information to influence electoral accountability is that voters possess the capacity to comprehend that information and durably update their beliefs about incumbent quality accordingly (at least until an election). In practice, this entails voters absorbing information that they read, hear, or observe. Political information is often complex, and unsophisticated voters may only exhibit a superficial understanding of such information that prohibits them from meaningfully updating their beliefs (e.g. Alt, Lassen and Marshall 2016; Gomez and Wilson 2001; Weitz-Shapiro and Winters 2014). The durable updating of beliefs about incumbent quality is also not trivial over longer periods. For example, Zaller (1992) suggests opinions may reflect a sample of recently encountered proposition, rather than an internalized set of beliefs based on information encountered over sustained periods. Moreover, voters may reject novel information that challenges their pre-existing beliefs (Zaller 1992). A second dimension of retention is regarding novel information as sufficiently credible to be worth internalizing. From a Bayesian perspective, this means believing that the information emanates from a signal generating process that correlates with the unknown state about which voters seek to learn. This may not occur where the media is known (or believed) to be biased (e.g. Besley and Prat 2006; Gentzkow and Shapiro 2006), or where voters perceive a source as politically motivated (e.g. Bullock 2011). Conversely, sources with a reputation for impartiality or incentives to 8

9 report opposite views are more credible (e.g. Alt, Lassen and Marshall 2016; Chiang and Knight 2011). Consequently, the source of incumbent performance information must be perceived to be sufficiently expert and trustworthy that voters are willing to use the source s information to update their beliefs. Even credible information is only likely to significantly alter behavior to the extent that it alters voter s posterior beliefs about the relative suitability of the incumbent for office. Bayesian voters should alter the position and precision of their posterior beliefs about incumbent suitability most in response to a credible signal when the signal is precise, it deviates from voters priors beliefs, and voters prior beliefs are relatively imprecise (e.g. Arias et al. 2018; Banerjee et al. 2011; Kendall, Nannicini and Trebbi 2014). 1 To the extent that voters in developing contexts are poorly informed, we may thus expect to observe larger changes in beliefs in such contexts. Importantly, the direction in which any behavior changes as a consequence of information provision will depend on how the information relates to voters prior beliefs. As Arias et al. (2018) and Banerjee et al. (2011) note, it is essential to understand the position of prior beliefs with respect to the signal in order to formulate hypotheses regarding a signal s effects. 2.2 The relevance of novel information Voters are likely to update most from precise signals that convey information relevant for appraising an incumbent s suitability for office. In this study, we focus on two features about which little is known that could enhance the effects of information provision: information about incumbent responsibilities, and performance benchmarks. First, information about incumbent responsibilities may be required for voters to be able to attribute outcomes to those responsible. Where decision-making responsibility is diffuse or unclear, voters may struggle to assign sanctions or rewards upon receiving information (e.g. Powell 1 Although incumbent performance metrics most directly pertain to incumbents, voters may also update about challengers to the extent that they believe that such signals are correlated (e.g. Kendall, Nannicini and Trebbi 2014). 9

10 and Whitten 1993; Duch and Stevenson 2008). Similarly, Gottlieb (2016) finds that voters systematically underestimate government capacity in Mali, requiring civics training in order to better identify poorly-performing incumbents. While responsibilities are often implicit when incumbent performance scorecards are provided, or provided alongside performance information, we explicitly separate between providing information about responsibilities and performance information by examining the interaction between the two, in order to understand what pieces of information are required to increase political accountability. Specifically, we hypothesize that providing information about incumbent duties should alter how voters update their beliefs about incumbent quality, by helping them to evaluate whether the incumbent has performed well given their responsibilities. To the extent that voters underestimate incumbent capabilities, we expect voters to become more critical. The same force may temper reports of strong performance and accentuate concerns about poor performance. Second, performance benchmarks can facilitate greater belief updating. A relevant comparison against which performance can be gauged helps voters to filter out common factors that influence the performance of all agents in a given period or location (Meyer and Vickers 1997). In the context of political accountability, this represents an increase in the precision of the signal pertaining to the current incumbent, and may also provide a credible signal about challengers. However, a key challenge is defining the relevant benchmark. In Mexico, Arias, Larreguy, Marshall and Querubín (2017) and Marshall (2017) find that cross-municipality comparisons do not influence voter beliefs and voting behavior. Gottlieb (2016) does find that comparatively greater performance can raise voters expectations of politicians in Senegal, although her design does not distinguish benchmarked from non-benchmarked information. We instead focus on a temporal comparison within the same electoral unit, comparing the current incumbent s performance to their predecessor. We thus expect favorably benchmarked performance indicators to update voter beliefs more than receiving the same information about only the current incumbent. From the perspective of our hypotheses, the performance benchmark also provides us with a basis for predicting the direction 10

11 in which incumbent performance information will alter voters beliefs. 2.3 Acting on novel information Even if voters meaningfully update their beliefs on the basis of incumbent performance information, accountability requires that they ultimately act on them. In our empirical application, we focus on two forms of resulting action that could reflect such belief updating: voting to re-elect the incumbent and seeking to make requests of the incumbent (both before and after elections). The former captures the standard electoral accountability models, while the latter captures the extent to which voters believe that their incumbent will be responsive to their requests. Voters are most likely to act on their beliefs under several key conditions. First, voters must be able to connect their beliefs about the incumbent to their available actions. Gomez and Wilson (2006) argue that relatively unsophisticated voters may lack the cognitive capacity to translate their beliefs into political action. In developing contexts, voters may lack the civic skills to draw these connections without basic education (Gottlieb 2016; Weitz-Shapiro and Winters 2014). Second, voters must be willing to vote on the basis of incumbent competence or policy alignment. However, beliefs about competence or alignment are likely to be one among many factors that voters take into account when considering how to vote. If previous incumbent performance carries receives a low weight in voters calculus, then even large changes in beliefs may not influence voting behavior. For example, voters in India engage in limited sanctioning of suspicious financial disclosures because they care more about performance in office and are unwilling to sanction copartisans (Chauchard, Klašnja and Harish 2017). Similarly, voters may be unwilling to sacrifice clientelistic benefits associated with opposing inferior candidates (Adida et al. 2017), or unwilling to deviate from bloc voting agreements (Koter 2013). However, such weights may not be fixed: it is also possible that informational interventions, especially those outlining incumbent responsibilities or documenting extremely high or low levels of performance, could prime voters 11

12 to upweight incumbent performance in office. Durante and Gutierrez (2014) argue for such a logic with respect to campaign ads in Mexico. Third, the process of providing information could set in motion other (potentially countervailing) forces that override the influence of voter beliefs on vote choice. At the voter level, information dissemination campaigns particularly those that create public signals, e.g. through the media could induce tacit or explicit voter coordination around particular candidates (Chwe 2000; Morris and Shin 2002). Tacit coordination would entail voters altering their vote choice on the basis that they expect others also to do so. Explicit coordination might involve meetings among voters where collective agreement is reached on which candidate to vote for. While either notion of coordination may often reinforce the effect of voter updating, it could sometimes counteract it (Arias, Balán, Larreguy, Marshall and Querubín 2017). 2 At the political elite level, political parties may seek to counteract or adapt their electoral strategies in response to adverse performance revelations. Recent studies in India, Mexico, Sierra Leone, and the Philippines suggest that incumbents may respond to information campaigns by confiscating leaflets, threatening those disseminating information, increasing their campaign spending, or escalating vote buying campaigns (Arias et al. 2018; Banerjee et al. 2011; Bidwell, Casey and Glennerster 2016; Chauchard and Sircar forthcoming; Cruz, Keefer and Labonne 2017). To the extent that such equilibrium responses affect vote choices, changes in the individual beliefs of voters may not ultimately translate into accountability-seeking behaviors. 3 Accountability in Senegalese parliamentary elections Senegal represents an excellent location to examine electoral accountability. It is one of Africa s oldest and strongest democracies, and has generally experienced robust multi-party political competition since It is one of the few African nations that has never experienced a coup d état, 2 Arias, Balán, Larreguy, Marshall and Querubín (2017) provide an example where public information dissemination led voters to update favorably about incumbents on the margin, but enabled voters in densely connected precinct to coordinate around challenger candidates still regarded as objectively superior. 12

13 and is known for its vibrant civil society and freedom of press and expression. Having seen two incumbents removed from office through competitive and fair democratic elections Abdou Diouf in 2000 and Abdoulaye Wade in 2012 Senegal remains a model of peaceful democratic transition, and thus represents a developing context where voters might given credible and relevant information feasibly hold governments to account. 3.1 Senegal s Assemblée Nationale While Senegal enjoys well-functioning electoral institutions in many respects, the country is characterized by hyper-presidentialism (Larreguy and Liu 2018; Thomas and Sissokho 2005). Senegal s Assemblée Nationale (Parliament) still plays a limited role in democratic representation, which has prevented full democratic consolidation (Beck 2012; Khagram, Fung and De Renzio 2013; O Donnell 1994). Focusing on legislators has the potential to inform how they can become more active and accountable to voters, and potentially promote socioeconomic development. Representatives to the Assemblée Nationale are elected by a mixed system, wherein 60% of deputies are selected by majoritarian voting at the department level, and the remaining 40% are selected by national-level proportional representation. All competing political coalitions form departmental lists for each of the 45 departments in Senegal, as well as an overall national list. At the department level, the coalition that receives the most votes wins all the seats allotted to that particular department s majoritarian list, the number of which is decided in advance by the Ministry of Interior. In 2017, there were 12 single and 33 multi-member departments. Proportional seats, by contrast, are allotted in proportion to a list s national vote share. In the legislative elections of 2012, the president s coalition Benno Bokk Yakkar (BBY) won almost all department-level contests, taking 87 of the 90 departmental seats based on the majoritarian voting system. At the national level, the president s coalition won approximately half of the proportionally allocated seats. The present project focuses on deputies elected from departmental majoritarian lists, as these politicians have distinct constituencies and thus stronger ties of accountability relative to 13

14 deputies elected from national lists. The primary task of deputies in the Assemblée Nationale is amending and voting on laws drafted by government ministries. In theory, deputies have the ability to initiate new laws themselves, though this is extremely rare in practice (Thomas and Sissokho 2005). As noted above, Senegal is characterized by hyper-presidentialism and consolidated executive power, and this influence shows in parliament. Very few laws are rejected by the Assemblée Nationale, and its role as a check on executive power is often called into question by civil society (as well as in our interviews with parliamentary deputies). Nonetheless, deputies have several important roles as part of their parliamentary duties. First, they can serve on one or several of the 11 technical parliamentary commissions. When a ministry proposes a law and it is brought to the Assemblée Nationale, it first goes to the technical commission that is paired with that particular ministry. These commissions meet in closed sessions and propose recommendations and amendments to the law, which are eventually debated in open plenary sessions. Leaders of commissions are considered influential in the amending of laws. The finance commission is considered to be particularly influential, as it plays a significant role in the deliberation over the national budget (Thomas and Sissokho 2005). Second, deputies can also submit written or oral questions to the executive government, to be answered by the relevant ministers in open sessions. Participating in these debates, as well as in ordinary plenary sessions, is one of the ways in which deputies can defend and publicize the interests of their constituents. Third, deputies are widely believed to be capable of influencing the allocation of local projects and government transfers. Deputies are not given any specific funds for local development projects in their home districts, nor is this a responsibility of deputies to the Assemblée Nationale. Nevertheless, despite this reality, constituents often see local development as the primary responsibility of deputies. Accordingly, many deputies make efforts to increase transfers to their districts for local development projects by lobbying government ministers. 14

15 3.2 Voter engagement with parliamentary deputies Voter turnout in Senegal generally lies around 60% for presidential elections, and reached 53% in the 2017 parliamentary elections. Confidence in the quality of democracy within Senegal trends upward as well. In the latest Afrobarometer round in Senegal, approximately 87% respondents stated that they consider Senegal to be a democracy, with 64% reporting that they are either fairly or very satisfied with the way democracy functions in Senegal. Nevertheless, voters often have limited access to information. In our rural sample, voters primarily rely on radio broadcasts; 82% and 44% of voters in our sample never obtain news via newspapers and television respectively. Senegalese parties are relatively strong, and are the source for much of the political information distributed to rural communities. This is particularly the case for parliamentary elections, for which available information on candidates is otherwise low. As the ballots during the voting process in legislative elections are grouped by party, it is common for citizens to be unfamiliar with the names of the candidate for whom they are voting, and defer instead to familiarity with the particular party or coalition. Only 35% of our rural sample in the present project knew the name of at least one of their parliamentary representatives. Citizen interactions with deputies in Senegal are rare, and there is significant variation in the amount of time deputies spend in their home districts despite having a specific budget to make such visits. Citizen interactions with party officials and brokers, however, are much more common. Voters are generally pessimistic about the extent to which deputies listen to voters and respond to requests, while only 9% reported interacting directly with a representative in the last year. Considerations of potential clientelistic returns as a function of the home village of the candidate as well as their ethnicity and religion influence voter choices heavily in rural Senegal. The Sufi Muslim networks, to which the majority of Senegalese citizens belong, intimately shape the political process across the country and moderate access to political resources (Villalón 1995). Political parties attempt to influence rural vote choice by using religious leaders as well as village chiefs 15

16 as intermediaries (Gottlieb and Larreguy 2018; Koter 2013). Group meetings, rallies, and parties organized by parties and their intermediaries in rural villages are common, as is the distribution of gifts to encourage voting. 4 Research design To examine whether and how Senegalese voters hold elected deputies to account for their performance in office, we implemented an information dissemination campaign prior to the 2017 parliamentary elections held on Sunday 30th July. Our field experiment randomized several key components of the information s content across villages, and deployed a panel study that permitted us to track voter beliefs and (intended) actions before the treatments were delivered, immediately after their delivery, and a month after the election. We also implemented a follow-up survey among a subset of respondents four months later. This design enables us to trace in rare detail the process relating the receipt of (different types of) information about incumbent performance to political accountability. 4.1 Sample selection We conducted our study in the five departments shown in Figure 1: Fatick, Foundiougne, Kanel, Oussouye, and Ranérou Ferlo. These departments were selected because they satisfied five criteria that prior theoretical arguments suggest would increase the likelihood that performance information would help voters to hold politicians to account: (1) only a single incumbent was seeking reelection through the majoritarian vote (with the exception of Kanel where two were standing); (2) there were no incumbents from the proportional list (with the exception of Kanel), who face different career incentives and have different accountability links to their constituents; (3) the incumbent could be compared with previous incumbent(s) because no incumbent was seeking re-election for a second time; (4) the department is not a newly-created administrative unit, and thus has a history of 16

17 Figure 1: Distribution of departments across Senegal incumbent performance against which the current incumbent could be benchmarked; and (5) given the preceding criteria, the selected departments have the lowest number of deputies representing the department. Oussouye and Ranérou Ferlo had 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, both on the majoritarian list. Within these five departments, we selected 450 rural villages for our sample. Starting from the 859 possible villages in these departments, we excluded all villages with fewer than 200 people and all villages with more than 4,000 people. Logistical concerns and access to newly-constructed 17

18 schools further restricted the set of potential villages. 3 Within each village, we aimed to survey nine registered voters aged (stratified by three age groupings) that grew up in each village. 4 The only further restrictions were that the respondents must have a cellphone number that they could be contacted on, that they were eligible to vote, and that they were born in the village or have lived there prior to the age of primary school enrollment (in Senegal, this is typically at 6 years of age). 4.2 Treatment conditions We disseminated information pertaining to legislator duties, as well as current and previous incumbent legislator performance prior to the 2017 election. Regarding legislator duties, we highlighted that legislators can: (1) serve on one or several of the 11 technical parliamentary commissions, where they can influence legislation and budgets in addition to taking on influential leadership conditions; (2) represent constituent interests in parliamentary debates; and (3) lobby government ministers to increase transfers to their districts for local specific projects. Regarding incumbent legislator performance, we focused on five measures of performance in office that relate to deputies primary duties: (1) the commissions they are members of; (2) positions of parliamentary leadership; (3) the number of parliamentary debates participated in; (4) the number of local projects budgeted for their department; and (5) the number and (inflationadjusted) per capita per year value of ministry transfers received by the department, decomposed by type of transfer. 5 The latter two measures are department-level variables, and thus do not vary 3 Of logistical concerns, 19 villages were dropped because they were too expensive to reach, e.g. because they are located on islands. We had also set out to explore the interaction between information and education by also leveraging variation across cohorts in access to schooling originating from a secondary-school construction program started in 2002 (Larreguy and Liu 2018). As our pre-analysis plan explains in greater detail, we excluded villages where the first post-2002 school was built between 2006 and 2010 to facilitate this design (Bhandari, Larreguy and Marshall 2017). We dropped this component of the analysis because access to new schools did not ultimately predict educational attainment among our survey respondents. 4 This restriction also arose from our intention to also examine interaction effects with access to newlyconstructed secondary schools across cohorts (Bhandari, Larreguy and Marshall 2017). 5 Annual transfer data from the government is only available from 2010 to To ensure comparability 18

19 Table 1: Factorial treatment conditions Performance information provided: None Incumbent Benchmark Duties information provided: None 75 villages [pure control] 75 villages 75 villages Duties 75 villages 75 villages 75 villages across deputies in multi-member departments. 6 All deputy performance data was obtained from either the Assemblée Nationale or a government department. The relevance of the information on deputy duties and performance was validated by the head of legislative services at the Assemblée Nationale, the librarians and archivists at the Assemblée Nationale, and several active and former deputies. Based on the performance metrics just described, we used a factorial design to randomly assign villages to one of the six experimental conditions shown in Table 1. 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 informs voters of the main functions that their parliamentary deputies can perform relating to the incumbent performance metrics described above. In contrast, the second dimension varies two types of information relating to the/a incumbent representative s performance over the preceding five years. The incumbent information reports the following information covering the current incumbent s performance over the previous five-year electoral cycle: committee membership, committee leadership, debates participated in, projects budgeted for the legislator s department, and ministry transfers to the department. In addition to such information, the benchmark variant also provides the same information pertaining to the performance of the department s previous incumbent representative. To maximize the between current and previous incumbents we convert ministry transfers to a per year basis and adjust for inflation. We further adjust for population size using the 2013 Census. Where transfers affect multiple departments, transfers are distributed in accordance with the share of the 2013 population in each department. 6 The information on the leaflet makes clear which departments have singular or multiple deputies. 19

20 comparability of the previous incumbent to the current incumbent, voters only received performance information about one incumbent drawn from the 2007 majoritarian list. In departments previously represented by multiple deputies, we randomize (across randomization blocks) which previous incumbent s performance is reported. Figure 2 shows the distribution of the main performance metrics provided, where each dot represents a current incumbent-previous incumbent pairing. All dots above the 45 o line are cases where the current incumbent outperformed the previous incumbent. Although there are eight pairings across our five departments, some comparisons are exact duplicates. The figure shows that, in general, the current incumbent outperformed preceding incumbents within the same department. 7 If we assume that voters prior expectations are broadly in line with previous incumbents performance, then we expect that our information will increase voters favorability towards current incumbents, on average, across departments. Each information treatment was distributed to voters through a leaflet like the one shown in Figure 3. 8 The leaflets, which were created specifically for this project by a graphic artist, were designed in partnership with our partnering civil association, LEGS-Africa. At the top of each leaflet variant is the LEGS-Africa logo together with a statement that the organization is non-partisan, while contact information and a description of data sources was provided at the bottom. The particular example in Figure 3 from Oussouye depicts the duties and benchmark treatment variant containing all information. 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, while the performance information on the left and right was provided to participants receiving the 7 This is particularly true with respect to debates, projects, and transfers. Moreover, the current incumbents in Fatick and Kanel are performing consistently better across all dimensions, while the current incumbent in Oussouye is generally performing worse. The difference in performance between current and previous incumbents is positively correlated with the performance of current incumbents. This positive association makes it difficult to separate the effect of information content across incumbent and benchmark performance treatments. 8 All leaflet variants are shown in the Appendix. 20

21 Current incumbent Fatick Committee memberships Kanel Kanel Fatick Foundiougne Ranérou Ferlo Previous incumbent Oussouye Current incumbent Leadership positions Kanel Kanel Ranérou Ferlo Fatick Foundiougne Foundiougne Previous incumbent Oussouye Debates participated in Local projects Current incumbent Kanel Fatick Kanel Oussouye Foundiougne Fatick Ranérou Ferlo Kanel Kanel Previous incumbent Current incumbent Foundiougne Kanel Ranérou Ferlo Oussouye Fatick Previous incumbent Current incumbent Number of transfers Kanel Ranérou Ferlo Foundiougne Oussouye Fatick Previous incumbent Current incumbent Transfer value per capita-year Kanel Ranérou Ferlo Foundiougne Fatick Previous incumbent Oussouye Figure 2: Distribution of treatment information across departments (45 degree line in gray) Note: Cases within departments where previous incumbents performed identically are not duplicated in this figure. 21

22 Figure 3: Example of duties + benchmark treatment in Oussouye benchmarked incumbent performance information. To ensure factual accuracy and impartiality, our leaflets were pre-tested with the head of legislative services at the Assemblée Nationale, the librarian at the Assemblée Nationale, and several active and former deputies. We also piloted the leaflet with voters to ensure that the information was comprehensible. The leaflet was delivered in person to respondents as part of our baseline survey on behalf of LEGS-Africa. 9 Our enumerators gave each voter a few minutes to read the leaflet and walked through the meaning of each component in the respondent s local language (since many rural Senegalese do not speak French). A key component of our training was to ensuring that enumerators, 9 We are not aware of other similar campaigns occurring in our sample. 22

23 who were generally recent university graduates, themselves understood and could clearly explain the leaflets content. 10 After receiving treatment, 82% of treated respondents reported that the leaflet came from an NGO, while only 19% believed it originated from the current deputy or the national government. 4.3 Information provision randomization Leaflet treatment conditions were randomly assigned at the village level to mitigate potential within-village spillovers and John Henry effects. Specifically, we constructed 75 blocks, each containing six similar villages. Village similarity was determined by being in the same department (and thus facing the same parliamentary election), as well as similarity along 24 pre-treatment covariates (including latitude, longitude, prior education levels, population, socioeconomic development indicators, local language, and incumbent presidential vote share). 11 Within each block, we used complete randomization to assign villages to either the control or one of the five treatment conditions. All voters in a village received the same treatment. In departments where there are multiple incumbents or possible benchmarked comparisons, we used complete randomization to assign an incumbent-previous incumbent pair to each block. Consequently, while information content varies across blocks, all villages within a block received the same information. 4.4 Data collection To assess the mechanisms linking incumbent performance information to electoral accountability, we designed a two-wave panel survey and collected administrative electoral returns. The baseline survey of nine young registered voters per village was conducted in person between 4th July and 29th July, and our treatments were administered toward the end of this baseline survey, af- 10 Depending on the department in which they worked, enumerators were extensively trained prior to implementation on translating the survey into Diola, Peulh, Serer, or Wolof. 11 We used the R package blocktools to assign blocks, department by department, based on a greedy algorithm using Mahalanobis distance. 23

24 ter baseline beliefs, behaviors, and intentions had been elicited. The shorter post-election survey was conducted by telephone around a month later, between 4th August and 26th August. We also collected polling station-level electoral returns, and mapped each village to its associated polling station based on the electoral register. This combination of data sources allows us to capture selfexpressed beliefs and attitudes, voting intentions, and behavioral responses, and examine how such responses vary across time and by theoretically-relevant pre-treatment respondent characteristics. We next describe our main outcome variables, while pre-treatment covariates used to estimate heterogeneous effects are introduced as the analysis progresses. The full survey is provided in our pre-analysis plan (Bhandari, Larreguy and Marshall 2017). 4.5 Measurement of primary outcomes Our primary outcomes focus on voter beliefs about previous and future incumbent performance in office, willingness to hold incumbents to account, and the extent to which such willingness translates into action. We thus distinguish self-reported attitudinal or intended outcomes from behavioral outcomes. Crucially, we measure both types of outcome immediately after the treatment was administered before the election, as well as after the election. Before treatment delivery, enumerators elicited baseline respondent characteristics, prior beliefs, and past and intended behavior. We focus on two main classes of attitudinal and intentional outcomes measured in our panel surveys. First, we measured voter beliefs about the performance of their current incumbent standing for re-election, previous incumbents, and challenger candidates on five-point scales. Specifically, this entailed eliciting responses to four questions: how well the incumbent has done overall since they were elected in 2012 ( incumbent overall performance ); how such performance compares to the previous incumbent ( relative performance (v. previous) ); how respondents think that current incumbent seeking re-election would do if re-elected ( prospective incumbent performance ); and how candidates from other parties would do ( prospective challenger perfor- 24

25 mance ). 12 Furthermore, 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). 13 Each belief, and its associated certainty, was elicited in the baseline survey before and after any information treatments were delivered for both treated and control respondents. The first two questions were asked again at endline. 14 Summary statistics in the control group are reported at the bottom of each column in results tables. Second, we also elicited self-reported (intended) vote choices. This entailed asking respondents whether they intended to vote (before and after treatment in the baseline survey) or actually did vote (at endline), and (if so) for which party or candidate. We code Vote incumbent as an indicator for respondents that stated they would or did vote for the incumbent candidate or the BBY coalition that all incumbents were part of. Around 96% of respondents registered an intention to vote, despite turnout in the polling stations containing our villages ultimately only reaching 60% on average. 15 Since 73% of respondents claimed to have actually voted at endline, we adopt a validated vote approach, in addition to examining purely self-reported votes. We register valid votes for the incumbent among the 53% of respondents that 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 allow for a nuanced understanding of voter beliefs and actions, selfreported behaviors especially for voting outcomes could be subject to social desirability bias. Such biases could arise from a desire among respondents to report that they supported the winner after the event (the incumbent in each case), attempts to please the enumerator or respond in the right way to the leaflet provided, or a fear that incumbents would receive survey results (in spite 12 Don t know responses are coded at the mid-level of the scale. 13 Don t know responses to scales eliciting voter beliefs are coded as the lowest level of certainty. 14 The latter two prospective questions made less sense after the election had occurred. 15 Although our sample of younger voters could vote at higher rates, intentions undoubtedly overstate actual behavior, even as evidenced by self-reported turnout after the election. 25

26 of our efforts to explain that the survey was anonymous and conducted by a non-partisan NGO). Accordingly, we also directly measured behavioral outcomes. First, we obtained polling station-level electoral returns from electoral officials, which enables us to calculate turnout and the incumbent party vote share both as a share of turnout and registered voters at the polling station corresponding to each village in our sample. Since less than 2% of voters at a polling station were actually treated, on average, we expected these estimates to be noisy. Nevertheless, it is possible that votes change enough to manifest at the polling station level, especially if there is information diffusion or coordinated responses within villages. As discussed in detail below, the endline survey also asked about general equilibrium responses namely, the extent to which voters discussed the information with others, coordinated their responses, and reported party responses. We thus also examine effect heterogeneity by the proportion of registered voters at a polling station from villages in our sample. Second, we created several opportunities for voters to make costly requests of the incumbent that can reflect the belief that the incumbent is responsive to voter interests (and may also indicate support). 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 incumbent poster ); request a visit from the party or candidate they intended to vote for, and/or other parties or candidates ( request incumbent visit ); 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 ( request incumbent conversation ). In the endline survey after the election had occurred, 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. For each action, the voter s name and village would be shared with the party. 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 voic s (costing around US$0.25) requesting to be contacted 26

27 by the winning candidate. 16 We measure requests for the associated telephone number ( request hotline number ) and actual calls to this number from telephone numbers linked to the respondent ( called hotline ). Voters were informed that all requests and messages sent to the hotline would be submitted to candidates on behalf of LEGS-Africa. 17 For all types of outcome, we increase the precision of our estimates by considering the difference between pre- and post-treatment responses as our main outcomes. Where this was not possible we include a pre-treatment proxy for the outcome as a control, e.g. controlling for pre-treatment vote intention in the baseline survey. Given the large number of outcomes which engender concerns about multiple comparisons and noise in specific variables we also pool across similar individual-level outcomes using indexes. Separately across baseline and endline panel waves, we created two indexes: self-reported incumbent support (i.e. all attitudinal and (intended) vote choice outcomes, except beliefs about challengers); and behavioral indicators of support for and requests from incumbents. We constructed two types of index in each case: the average across each standardized item (Kling, Liebman and Katz 2007) and the inverse covariance weighted (ICW) average across items (Anderson 2008). The former measure captures the mean across items, while the latter is preferred because it also accounts for correlation among items. Given that it is not obvious whether voters implicitly account for such correlations, we also report robustness checks for the main results using the Kling, Liebman and Katz (2007) index in the Appendix. Both measures were standardized with respect to the control group, such that differences between groups reflect standard deviation changes in control group outcomes. 16 Phone credit in Senegal is expensive per minute used, and thus especially costly in poor rural areas. Text messages are comparably cheaper. 17 No deception was involved because each commitment was honored. This of course does not guarantee that politicians will respond to voter requests. 27

28 4.6 Estimation In accordance with our pre-analysis plan, our fully-saturated treatment specification estimates the within-block average treatment effect of each component of the information leaflet using baseline regressions of the form: Y ivb = β 1 duties v + β 2 incumbent v + β 3 benchmark v + β 4 (incumbent v duties v ) +β 5 (benchmark v duties v ) + αy baseline ivb + η b + θ e + ε ivb, (1) where Y ivb is an outcome for individual i in village v in randomization block b, η b are randomization block fixed effects, and θ e are survey enumerator fixed effects (at baseline and, where relevant, also at endline). Wherever possible, the outcome s pre-treatment baseline counterpart Y baseline ivb included as a control to increase efficiency. 18 For polling station-level outcomes, we substitute the i for a p subscript. To weight each village equally, all survey-based regressions are weighted by the inverse of the number of respondents in the corresponding baseline or endline survey. 19 Standards errors are clustered by randomization block and significance tests are based on one-sided t tests for pre-specified hypotheses. To test many of the specific hypotheses underpinning the accountability process, we are also interested in how the effects of the treatment vary with the content of the information provided and other predetermined covariates such as voters prior beliefs. Taking the simpler case that pools across the duties treatment dimension, which we will do frequently in the analysis, we estimate 18 Our pre-analysis plan proposed both first-differencing and a controlling for a lagged dependent variable. We favor the latter due to its efficiency benefits (McKenzie 2012) and because only baseline proxies are available for some outcomes. 19 This ensures that treatment is uncorrelated with blocks containing more observations, which would prevent estimation of average treatment effects under by-block effect heterogeneity. is 28

29 heterogeneous effect specifications of the form: Y ivb = β 1 incumbent v + β 2 benchmark v + β 3 (incumbent v X iv ) +β 4 (benchmark v X iv ) + γx iv + αy baseline ivb + η b + θ e + ε ivb, (2) where X iv is a predetermined covariate. 4.7 Experimental validation checks Before turning to our main results, we estimate equation (1) to demonstrate that treatment assignment was indeed random. Consistent with chance, Table A1 shows that 91 predetermined individual- and village-level covariates are well-balanced across treatment conditions at baseline: a two-sided joint test of the constraint that each treatment group is indistinguishable from the other was only rejected at the 10% level in seven cases. Analysis of endline data is more complex, since estimates using endline data could be confounded by selective attrition in response to treatment. Fortunately, we were able to successfully re-interview 96% of the baseline sample. This remarkably high success rate for a telephone followup survey may have reflected the low frequency with which rural Senegalese (with access to mobile phones) have opportunities to express their views to survey teams, especially those that offered to pass on requests to politicians at baseline. 20 Unsurprisingly, given the low rates of attrition, Table A3 shows that there is no significant difference in attrition rates across treatment groups. Moreover, Table A2 shows that the balance we observed at baseline continues to hold within the endline response sample: for only six of 113 predetermined variables do we observe significant differences across treatment conditions. 20 In some cases, participants even called us back after an enumerator s initial called went unanswered. 29

30 5 Results To test the mechanisms linking, or failing to link, incumbent performance information and electoral and off-election accountability, we proceed by first examining immediate responses to the treatment before turning to post-election responses. 5.1 Immediate effects of information provision on beliefs, intentions, and requests We start by examining post-treatment voter beliefs, intentions, and behavior in the baseline survey. A key virtue of eliciting responses immediately after our leaflets were provided is that we are able to examine responses uncontaminated by the subsequent interactions with other voters or political actors that we discuss below Voters understand the information contained in the leaflets The first step linking information provision to accountability is that voters are able to comprehend the information provided. To assess this, we asked four multi-choice factual questions of treated and control respondents about the information contained in the leaflets. 21 Each question refers to a particular component of the information the number of commissions in the Assemblée Nationale (four choices), the lack of specific funds available to deputies for development projects in their department (true or false), the number of projects undertaken 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). Although these questions could, in theory, have been answered by knowledgeable control respondents, the control group means at the bottom 21 After the leaflet was provided and explained, respondents were then asked about its provenance before we tested leaflet comprehension. 30

31 Table 2: Effects of information treatments on leaflet comprehension (baseline survey) Correctly state......number of...deputies...number of...number of parliamentary lack own incumbent s previous commissions fund pots local incumbent s projects debates (1) (2) (3) (4) Duties 0.788*** 0.545*** (0.028) (0.041) (0.031) (0.024) Incumbent 0.204*** 0.122*** 0.704*** 0.054** (0.038) (0.039) (0.041) (0.027) Incumbent Duties *** ** (0.045) (0.053) (0.046) (0.040) Benchmark 0.204*** 0.119*** 0.699*** 0.468*** (0.035) (0.034) (0.043) (0.048) Benchmark Duties *** *** (0.048) (0.053) (0.050) (0.054) Observations 4,051 4,051 4,051 4,051 Outcome range {0,1} {0,1} {0,1} {0,1} Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block and baseline enumerator fixed effects. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. of Table 2 indicate that few respondents provided the correct answer. 22 Perhaps unsurprisingly given the written and verbal explanations, the results in Table 2 demonstrate that immediate comprehension was very high. With respect to the information about deputies duties, columns (1) and (2) indicate that receiving the duties-only information leaflet increased the proportion of respondents correctly identifying the number of parliamentary commissions from 5% to 84% and that deputies lack individual funds for department projects from 14% to 58%. This increase is similar when combined with incumbent or benchmarked performance information, while 22 The average is likely below random guesswork because many respondents answered don t know. 31

32 either type of performance information on their own increased the likelihood that the respondent attempted to answer the question. 23 Column (3) demonstrates that incumbent performance information (whether on its own or with a benchmark), but not the information about duties, increases the proportion of respondents that know the number of local projects received under the current incumbent from 7% to 77%. Similarly, column (4) shows that the benchmark leaflet is the only one to substantially increase correct answers regarding the number of debates that the previous incumbent participated in, increasing the share of correct responses by around 0.5. Collectively, these results demonstrate that our respondents were able to immediately recall the information contained in the leaflet shown and explained to them. As in previous studies from across the world (e.g. Alt, Lassen and Marshall 2016; Dunning et al. forthcoming; Lieberman, Posner and Tsai 2014), voters inability to comprehend the information provided does not seem to be a bottleneck in the political accountability process Voters update their beliefs in a Bayesian manner The preceding results indicate that voters understood the information that they received. However, given that information may not be perceived as credible or useful, this does not necessarily imply that it would alter their beliefs about the incumbent or their intention to vote for them. Moreover, there may be a subtle but important difference between regurgitating and internalizing new information. Although belief updating may not be strictly necessary for information provision to influence electoral accountability (since information provision could stimulate other actions without affecting beliefs), it is generally regarded as the primary mechanism. As columns (1)-(3) in Table 3 illustrate, voters did indeed update their posterior beliefs on average after receiving the leaflet. If voters prior beliefs were not already centered on the truth, we expected that voters would positively update from our performance indicators on the basis that 23 Inspection of don t know responses suggests that a greater willingness to answer questions after receiving the leaflet primarily accounts for this increase. 32

33 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) Self-reported outcomes Behavioral outcomes Incumbent Relative Prospective Incumbent Index Index Prospective Request Request Request Index Index overall performance incumbent vote (average) (ICW) challenger incumbent incumbent incumbent (average) (ICW) performance (v. previous) performance performance poster visit conversation (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: All information treatment conditions Duties * (0.061) (0.051) (0.048) (0.012) (0.041) (0.035) (0.048) (0.019) (0.020) (0.018) (0.038) (0.039) Incumbent 0.361*** 0.229*** 0.248*** 0.029** 0.297*** 0.240*** (0.076) (0.059) (0.068) (0.015) (0.059) (0.052) (0.053) (0.018) (0.020) (0.019) (0.038) (0.038) Incumbent Duties * * 0.062*** 0.046* 0.113** 0.111** (0.092) (0.075) (0.085) (0.023) (0.074) (0.067) (0.072) (0.029) (0.029) (0.029) (0.061) (0.061) Benchmark 0.467*** 0.368*** 0.394*** 0.038*** 0.426*** 0.355*** 0.093*** (0.073) (0.068) (0.062) (0.016) (0.061) (0.054) (0.042) (0.021) (0.021) (0.020) (0.043) (0.043) Benchmark Duties ** * 0.084* 0.086* (0.086) (0.081) (0.079) (0.020) (0.065) (0.059) (0.065) (0.026) (0.026) (0.026) (0.053) (0.054) Panel B: Pooling duties treatment conditions Incumbent 0.355*** 0.289*** 0.266*** 0.031*** 0.320*** 0.266*** 0.057* 0.032*** 0.039*** 0.029** 0.076*** 0.076*** (0.055) (0.048) (0.048) (0.011) (0.045) (0.038) (0.039) (0.015) (0.017) (0.017) (0.034) (0.034) Benchmark 0.436*** 0.379*** 0.338*** 0.037*** 0.401*** 0.334*** 0.119*** 0.031*** * 0.053* 0.056** (0.054) (0.057) (0.043) (0.012) (0.049) (0.044) (0.033) (0.015) (0.016) (0.015) (0.032) (0.032) Null: Incumbent Benchmark (p value) Observations 3,994 3,984 3,980 4,051 4,051 3,943 3,939 4,050 4,051 4,050 4,051 4,049 Outcome range {1,...,5} {1,...,5} {1,...,5} {0,1} [-2.4,2.2] [-2.3,2.0 {1,...,5}] {0,1} {0,1} {0,1} [-1.6,0.7] [-1.6,0.7] Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block and baseline enumerator fixed effects. All specifications include a lagged dependent variable as a control; in columns (8)-(12), pre-treatment incumbent vote is used as a proxy. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. 33

34 they exceed previous incumbents performance. Indeed, panel A shows that voters who received incumbent and benchmark performance information became substantially more positive about their incumbent deputy across all assessments of their suitability for office. Although column (7) shows that voters receiving the benchmark also updated somewhat positively about challengers, suggesting a perceived positive correlation across all types of politician (as in Arias et al. 2018), these changes are notably smaller than updating relating to the incumbent. The results thus suggest that voters regarded the performance information as credible and accordingly favorably updated their posterior beliefs about the incumbent. 24 Conversely, information about deputies duties did not systematically affect voter evaluations whether on its own on in conjunction with performance indicators. This suggests that voters may not need additional information about incumbent duties to feel confident in using incumbent performance indicators to evaluate them. The low importance of information about duties is a consistent pattern throughout this study. We thus focus primarily on the comparison between benchmarked and non-benchmarked incumbent performance information (as in panel B). Pooling the control and duty conditions, panel B further demonstrates that voters process within-department benchmarks in a rational manner. In line with theoretical expectations, based on the current incumbent almost always outperforming the previous incumbent (see Figure 2), the F tests at the bottom of column (1)-(3) in panel B consistently reject the one-sided null hypothesis that the benchmarked information is no more likely to increase incumbent support. In contrast with extant findings that spatial benchmarks are relatively unimportant (e.g. Arias, Larreguy, Marshall and Querubín 2017), we find that temporal benchmarks can meaningfully influence beliefs by providing a relevant yardstick for appraising incumbent performance. Furthermore, information provision increases the precision of voters beliefs. Consistent with 24 There are several reasons to believe that these results do not simply reflect social desirability bias. First, subsequent behavioral measures show similar results. Second, it is hard to believe that those that learn most, based on the hetereogeneous effects below, are also most susceptible to social desirability bias. Third, Tables A11 and A12 show that the results are robust to removing any treated respondents that did not believe that the leaflets were delivered by an NGO. 34

35 voters possessing relatively imprecise prior beliefs, Appendix Table A4 shows that the leaflets increased 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. The benchmark was particularly important in increasing the belief precision. However, voters certainty about voting for the incumbent choice was not significantly altered by performance information. Finally, the interaction between information provision, information content, and voter prior beliefs is consistent with Bayesian updating. We illustrate this by examining how the effects of our treatments on posterior beliefs vary with the performance indicators provided and pre-treatment measures of prior beliefs in Table 4. First, panel A shows that voters updated significantly more positively about the incumbent when the leaflet indicated higher performance (captured by an ICW scale averaging across our six reported performance indicators). 25 Aggregated across measures, the scales in columns (5) and (6) show that this was true of both the incumbent only and benchmarked information, although the level effect of the latter is greater. Panel B reports similar results when reported performance is weighted by the factors voters regard as most important. 26 Furthermore, panel C shows that voters increased favorability toward the incumbent almost entirely reflects the projects and transfers components of the performance index, suggesting that voters are more concerned about local projects than national processes. As in Adida et al. (2017), appearing to work on national or parliamentary issues if anything sends a negative signal. 27 Second, panel B similarly shows that voters updated most favorably about the incumbent when they had the least favorable prior beliefs. Third, panel C shows that favorable updating was lower among respondents with more precise prior beliefs. Together, these findings paint a clear picture of sophisticated updating, and suggest that beliefs may have changed enough to affect vote choices. 25 The results are robust to using the standardized average scale. 26 In practice, this entailed constructing a relevance-weighted performance index where local and national performance are only included if the respondent regarded either as one of the top three most important factors in determining their vote choice. 27 These coefficients are not statistically significant due to the one-tailed test, but are statistically significant under two-tailed tests in some cases. 35

36 Table 4: Heterogeneous effects of information treatments on beliefs about incumbent performance, intention to vote for the incumbent, and requests from the incumbent, by leaflet content, priors beliefs, and importance of performance information for vote choice (baseline survey) Self-reported outcomes Behavioral outcomes Incumbent Relative Prospective Incumbent Index Index Prospective Request Request Request Index Index overall performance incumbent vote (average) (ICW) challenger incumbent incumbent incumbent (average) (ICW) performance (v. previous) performance performance poster visit conversation (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Heterogeneity by reported performance level Incumbent 0.366*** 0.301*** 0.273*** 0.031*** 0.327*** 0.272*** 0.061* 0.033*** 0.041*** 0.030** 0.080*** 0.080*** (0.048) (0.041) (0.044) (0.011) (0.039) (0.036) (0.039) (0.015) (0.017) (0.017) (0.034) (0.034) Incumbent Overall performance (ICW) 0.245*** 0.201*** 0.186*** 0.022*** 0.188*** 0.137*** 0.051** (0.045) (0.035) (0.035) (0.007) (0.033) (0.028) (0.028) (0.011) (0.015) (0.018) (0.028) (0.027) Benchmark 0.445*** 0.386*** 0.343*** 0.039*** 0.407*** 0.342*** 0.117*** 0.031*** * 0.055** 0.058** (0.050) (0.056) (0.042) (0.011) (0.047) (0.042) (0.033) (0.015) (0.016) (0.015) (0.032) (0.032) Benchmark Overall performance (ICW) 0.214*** 0.153*** 0.153*** 0.035*** 0.171*** 0.147*** (0.039) (0.046) (0.035) (0.008) (0.038) (0.033) (0.026) (0.013) (0.016) (0.015) (0.028) (0.027) Panel B: Heterogeneity by relevance-weighted reported performance level Incumbent 0.390*** 0.324*** 0.304*** 0.038*** 0.356*** 0.300*** 0.058* 0.042*** 0.048*** 0.041*** 0.100*** 0.100*** (0.050) (0.043) (0.043) (0.012) (0.040) (0.037) (0.038) (0.017) (0.018) (0.019) (0.038) (0.037) Incumbent Relevant performance (ICW) 0.265*** 0.246*** 0.208*** 0.033*** 0.228*** 0.185*** (0.052) (0.048) (0.046) (0.014) (0.043) (0.041) (0.041) (0.018) (0.020) (0.020) (0.040) (0.040) Benchmark 0.479*** 0.407*** 0.360*** 0.043*** 0.428*** 0.358*** 0.126*** 0.040*** 0.025* 0.030** 0.071** 0.075*** (0.054) (0.060) (0.043) (0.013) (0.049) (0.045) (0.035) (0.016) (0.018) (0.017) (0.037) (0.037) Benchmark Relevant performance (ICW) 0.189*** 0.147*** 0.123*** 0.029** 0.154*** 0.135*** (0.048) (0.047) (0.040) (0.015) (0.042) (0.040) (0.040) (0.020) (0.022) (0.021) (0.045) (0.045) Panel C: Heterogeneity by local and national reported performance level Incumbent 0.392*** 0.314*** 0.291*** 0.034*** 0.346*** 0.291*** *** 0.047*** 0.032** 0.086*** 0.085*** (0.050) (0.043) (0.047) (0.012) (0.040) (0.036) (0.039) (0.016) (0.018) (0.018) (0.036) (0.036) Incumbent National performance (ICW) * (0.071) (0.059) (0.049) (0.010) (0.048) (0.037) (0.042) (0.017) (0.017) (0.017) (0.032) (0.032) Incumbent Local performance (ICW) 0.295*** 0.217*** 0.242*** 0.023* 0.254*** 0.216*** (0.066) (0.058) (0.057) (0.015) (0.052) (0.047) (0.045) (0.020) (0.022) (0.021) (0.043) (0.043) Benchmark 0.457*** 0.407*** 0.355*** 0.041*** 0.422*** 0.357*** 0.109*** 0.035*** 0.024* 0.027** 0.065** 0.068*** (0.052) (0.050) (0.040) (0.012) (0.043) (0.039) (0.034) (0.016) (0.017) (0.016) (0.034) (0.034) Benchmark National performance (ICW) (0.071) (0.100) (0.056) (0.011) (0.071) (0.063) (0.036) (0.012) (0.019) (0.017) (0.032) (0.031) Benchmark Local performance (ICW) 0.209*** 0.259*** 0.211*** 0.028** 0.244*** 0.216*** (0.063) (0.080) (0.049) (0.015) (0.059) (0.054) (0.040) (0.017) (0.020) (0.019) (0.040) (0.039) Observations 3,994 3,984 3,980 4,051 4,051 3,943 3,939 4,050 4,051 4,050 4,051 4,049 Outcome range {1,...,5} {1,...,5} {1,...,5} {0,1} [-2.4,2.2] [-2.3,2.0 {1,...,5}] {0,1} {0,1} {0,1} [-1.6,0.7] [-1.6,0.7] Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block and baseline enumerator fixed effects. All specifications include a lagged dependent variable as a control; in columns (8)-(12), pre-treatment incumbent vote is used as a proxy. 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. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. (Continued...) 36

37 Table 5 (continued): Heterogeneous effects of information treatments on beliefs about incumbent performance, intention to vote for the incumbent, and requests from the incumbent, by leaflet content, priors beliefs, and importance of performance information for vote choice (baseline survey) Self-reported outcomes Behavioral outcomes Incumbent Relative Prospective Incumbent Index Index Prospective Request Request Request Index Index overall performance incumbent vote (average) (ICW) challenger incumbent incumbent incumbent (average) (ICW) performance (v. previous) performance performance poster visit conversation (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel D: Heterogeneity by prior belief level Incumbent 0.353*** 0.294*** 0.267*** 0.032*** 0.312*** 0.266*** 0.063* 0.033*** 0.041*** 0.031** 0.080*** 0.080*** (0.055) (0.047) (0.048) (0.011) (0.046) (0.040) (0.038) (0.014) (0.016) (0.017) (0.033) (0.032) Incumbent Prior index (ICW) *** *** *** *** *** *** *** ** * *** *** (0.041) (0.031) (0.037) (0.012) (0.034) (0.033) (0.036) (0.014) (0.016) (0.017) (0.033) (0.032) Benchmark 0.439*** 0.382*** 0.339*** 0.040*** 0.397*** 0.336*** 0.124*** 0.030*** * 0.056** 0.058** (0.055) (0.057) (0.045) (0.012) (0.051) (0.045) (0.033) (0.015) (0.016) (0.015) (0.032) (0.032) Benchmark Prior index (ICW) *** ** * *** *** *** * ** (0.038) (0.042) (0.036) (0.011) (0.036) (0.034) (0.035) (0.014) (0.016) (0.016) (0.032) (0.032) Panel E: Heterogeneity by prior belief precision Incumbent 0.362*** 0.300*** 0.284*** 0.039*** 0.334*** 0.282*** 0.070** 0.027** 0.036*** 0.034*** 0.074*** 0.073*** (0.055) (0.049) (0.049) (0.011) (0.046) (0.039) (0.040) (0.015) (0.017) (0.017) (0.033) (0.033) Incumbent Prior precision index (ICW) *** ** *** * (0.036) (0.030) (0.040) (0.012) (0.030) (0.027) (0.025) (0.016) (0.015) (0.015) (0.031) (0.031) Benchmark 0.445*** 0.392*** 0.361*** 0.045*** 0.420*** 0.352*** 0.132*** 0.028** ** 0.058** 0.060** (0.056) (0.059) (0.045) (0.012) (0.051) (0.046) (0.033) (0.015) (0.016) (0.015) (0.033) (0.033) Benchmark Prior precision index (ICW) ** ** ** *** *** *** * * ** (0.043) (0.032) (0.040) (0.013) (0.033) (0.030) (0.031) (0.015) (0.015) (0.014) (0.030) (0.030) Panel F: Heterogeneity by importance of performance in determining vote choice Incumbent 0.354*** 0.288*** 0.265*** 0.031*** 0.320*** 0.265*** 0.057* 0.033*** 0.040*** 0.029** 0.077*** 0.077*** (0.055) (0.047) (0.048) (0.011) (0.044) (0.038) (0.039) (0.015) (0.017) (0.017) (0.034) (0.034) Incumbent Performance most important *** * 0.021* 0.041* 0.040* (0.036) (0.035) (0.031) (0.009) (0.026) (0.024) (0.027) (0.014) (0.013) (0.013) (0.026) (0.027) Benchmark 0.435*** 0.378*** 0.338*** 0.038*** 0.400*** 0.334*** 0.119*** 0.031*** * 0.054** 0.057** (0.054) (0.057) (0.043) (0.012) (0.049) (0.044) (0.033) (0.015) (0.016) (0.015) (0.032) (0.032) Benchmark Performance most important * * 0.030*** 0.024** 0.019* 0.055*** 0.057*** (0.029) (0.028) (0.031) (0.010) (0.023) (0.021) (0.025) (0.012) (0.014) (0.012) (0.026) (0.026) Panel G: Heterogeneity by preference for locally-oriented deputies Incumbent 0.368*** 0.195*** 0.201*** *** 0.183*** * * 0.090* (0.083) (0.082) (0.071) (0.026) (0.066) (0.060) (0.048) (0.028) (0.033) (0.035) (0.065) (0.064) Incumbent Prefer locally-oriented deputies * *** ** 0.083* (0.075) (0.082) (0.075) (0.029) (0.066) (0.065) (0.061) (0.033) (0.034) (0.037) (0.071) (0.071) Benchmark 0.498*** 0.285*** 0.303*** *** 0.293*** 0.062* 0.044* * 0.082* (0.089) (0.094) (0.070) (0.027) (0.078) (0.074) (0.045) (0.030) (0.031) (0.029) (0.060) (0.060) Benchmark Prefer locally-oriented deputies ** * (0.079) (0.076) (0.073) (0.028) (0.067) (0.066) (0.055) (0.034) (0.033) (0.032) (0.066) (0.066) Observations 3,994 3,984 3,980 4,051 4,051 3,943 3,939 4,050 4,051 4,050 4,051 4,049 Outcome range {1,...,5} {1,...,5} {1,...,5} {0,1} [-2.4,2.2] [-2.3,2.0 {1,...,5}] {0,1} {0,1} {0,1} [-1.6,0.7] [-1.6,0.7] Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block and baseline enumerator fixed effects. All specifications include a lagged dependent variable as a control; in columns (8)-(12), pre-treatment incumbent vote is used as a proxy. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. 37

38 5.1.3 Performance information alters vote intentions We next examine the extent to which changes in beliefs translate into intended vote choice. Column (4) of Table 3 indicates that higher appraisals of the incumbent are reflected in around a 3 percentage point increase in the probability of reporting an intention to vote for the incumbent. We again find no evidence to suggest that information about incumbent duties moderates the effects of performance information, but some tentative (rather than statistically significant) evidence to suggest that benchmarked information induces greater incumbent support. This increase in incumbent vote intention also varies with several key predictions of the accountability logic. First, the heterogeneous effects in column (4) of panels A-E of Table 4 show that changes in vote intention mirror changes in voter beliefs, such that those updating most positively are also most likely to express an intention to vote for the incumbent. Second, panel F further demonstrates that the 54% of respondents who (before treatment) stated that incumbent performance on metrics like those provided are the most important determinant of their vote choice were also significantly more likely to shift their vote intention in favor of the incumbent. Similar results hold in panel G among voters that expressed a preference for a locally-oriented, as opposed to nationally-oriented, politician in a pre-treatment vignette. Exactly as predicted, increased intention to vote for the incumbent is driven by those induced to update most positively on issues they regard as important Performance information increases requests from incumbents The change in self-reported vote intention is generally reinforced by our behavioral outcomes capturing willingness to make forward-looking requests from the incumbent. Specifically, columns (8)-(12) of Table 4 show that those respondents induced to support the incumbent by performance information also become significantly more willing make requests of the incumbents. However, panel B indicates that the benchmarked information does not differentially influence such requests. 38

39 To the extent that requests can serve as a behavioral proxy for baseline vote choice, these findings suggest that changes in vote intentions do not simply reflect social desirability bias. Requests represent the one area where information about deputies duties complements performance indicators. Columns (8)-(12) of panel A in Table 4 indicate that learning that the incumbent is generally better than expected primarily translates into requests when voters are aware of what the incumbent could do. These estimates suggest that an understanding of what politicians can do could help voters to make requests from relatively effective politicians. However, as Table 7 below shows, such findings do not persist into the endline. Broadly in line with vote intentions, Table 4 shows that increased incumbent requests also reflect incumbent performance levels, prior beliefs, and the importance of performance to voters. While the interactions with such beliefs are generally weaker, there remains clear evidence that those caring most about performance become significantly more likely to make requests upon learning that the incumbent performed better than expected Summary The preceding findings strongly suggest that credible performance information, at least on issues that matter to voters, have the potential to play a key role in electoral and non-electoral accountability. Contrary to concerns that voters lack the sophistication to hold politicians to account, even relatively uneducated rural voters comprehend incumbent performance leaflets and correspondingly update their beliefs in a Bayesian manner. Furthermore, this translates into an intention to hold the incumbent to account at the ballot box, and a willingness to make somewhat costly requests from incumbents. In this particular context, information favored the incumbent. However, while such responses demonstrate the potential effects of information dissemination campaigns, the key question is whether immediate changes in voter intentions are ultimately acted upon at the ballot box. 39

40 5.2 Longer-term effects on beliefs, voting behavior, and requests To understand electoral behavior several weeks after information was provided and requests of newly re-elected incumbents, we now turn to our endline survey and polling station electoral returns Voters correctly recalled treatment information after the election While voters clearly changed their beliefs and vote intentions immediately, such changes could dissipate and thus not affect voting behavior. The endline examines this possibility by asking respondents if voters recall receiving our leaflet, and if so what type of information the leaflet contained. To guard against enumerators replicating baseline answers, no enumerator was permitted to re-interview a respondent that they interviewed at baseline and no enumerator knew the treatment status of their endline survey respondents. The results in Table 5 indicate high leaflet recall around a month after our leaflet was disseminated. Column (1) shows that virtually all (around 98%) treated respondents correctly recall receiving the LEGS-Africa leaflet, while only 7% of control respondents incorrectly recall receiving the leaflet. In comparison with other contexts, these large numbers may reflect the lack of leaflets being distributed in the remote areas in which we worked. Moreover, columns (2)-(4) demonstrate that almost as many respondents correctly remember the differentiating features of the leaflet s content. Clearly, voter recall of the information is remarkably high, and thus does not represent an impediment to electoral accountability in a context The relative importance of incumbent legislative performance in making vote choices did not change While voter appraisals of their incumbents were immediately altered, the provision of incumbent performance information could also change the relative importance attached to incumbent legisla- 40

41 Table 5: Effects of information treatments on leaflet recall (endline survey) Received Received Received Received leaflet duties incumbent previous information information incumbent information (1) (2) (3) (4) Duties 0.916*** 0.939*** 0.050*** 0.027*** (0.014) (0.012) (0.015) (0.009) Incumbent 0.913*** 0.071*** 0.934*** 0.020*** (0.014) (0.015) (0.012) (0.008) Incumbent Duties *** *** * (0.018) (0.020) (0.020) (0.015) Benchmark 0.915*** 0.038*** 0.961*** 0.948*** (0.013) (0.011) (0.009) (0.010) Benchmark Duties *** *** *** *** (0.015) (0.019) (0.016) (0.015) Observations 3,925 3,925 3,925 3,925 Outcome range {0,1} {0,1} {0,1} {0,1} Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects and baseline and endline enumerator fixed effects. All observations are inversely weighted by the endline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. tive 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 akin to at baseline before treatment what were the three most factors in determining their vote choice in the 2017 election. The results in Table 6 indicate that our leaflets did not increase the weight attached to incumbent performance in determining vote choices. Specifically, columns (1)-(3) show that our treatments did not change the probability of ranking performance in the top three most important factors, while columns (4)-(6) similarly report no change in incumbent performance being the most important. This suggests that voters utility functions are fixed in the face of information dissemination. 41

42 Table 6: 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 (0.014) (0.022) Incumbent (0.017) (0.013) (0.021) (0.016) Incumbent Duties (0.024) (0.034) Benchmark (0.015) (0.011) (0.025) (0.019) Benchmark Duties (0.019) (0.034) Performance (0.010) (0.015) Null: Incumbent Benchmark (p value) Observations 3,926 3,926 3,926 3,926 3,926 3,926 Outcome range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects and baseline and endline enumerator fixed effects and a lagged dependent variable. All observations are inversely weighted by the endline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests Beliefs about incumbent performance persist after the election While the salience of incumbent performance on the dimension we provide information about did not change, beliefs about incumbent performance did persist. Reinforcing the findings regarding information recollection, columns (1) and (2) of Table 7 show that a month after the leaflet was provided voters receiving performance information continue to register higher performance rating of the incumbent, and believe that the incumbent performed better than previous incumbents. A 42

43 comparison with the estimates in Table 3 suggests some decay around a half of the immediate effect of the information provision. Since the endline occurred after the election, there is good reason to believe that treated voters were more likely to believe that the incumbent performed well while in office at the time of the election. The tests at the bottom of panel B indicate that this was still especially the case for the benchmark information. Moreover, Appendix Table A5 similarly shows that treated voters continue to express greater certainty about their beliefs. The heterogeneous effects in Table 8 generally point in the same direction, but suggest that the exact details of incumbent performance start to become fuzzier over time. The estimates in panel C, however, provide clear evidence though that voters continue to favorably update about incumbents with higher levels of local performance. Again, national-level performance may even be sanctioned Performance information influences vote choices We next examine whether the vote intentions registered at baseline were ultimately acted upon. Given the persistence of beliefs favorable to the incumbents, there is reason to believe that this key component of the electoral accountability dynamic could hold. To test this, we first study self-reported voting for the incumbent before analyzing polling station-level returns. Our individual-level endline survey data provide initially weak evidence that providing performance information translates into holding incumbents to account at the ballot box. On average, columns (3) and (4) in panels A and B of Table 7 do not show significant increases in incumbent voting whether entirely self-reported or after adjusting for our approach to validated voting. Nevertheless, consistent with voter beliefs, the tests in panel B tentatively suggest a larger effect of the benchmark treatment in comparison with the incumbent-only performance information. The heterogeneous effects in Table 8 provide more mixed evidence: while the leaflet s performance content and voters prior beliefs did not influence the self-reported voting of the average respondent, panels F and G show that support for the incumbent was greater among treated respondents receiving the benchmarked indicators that also regard performance information as the most impor- 43

44 Table 7: Average effects of information treatments on beliefs about incumbent performance, reported vote for the incumbent, and requests from the incumbent (endline survey) Self-reported outcomes Behavioral outcomes Incumbent Relative Incumbent Incumbent Index Index Request Request Request Called Index Index overall performance vote vote (average) (ICW) incumbent incumbent hotline hotline (average) (ICW) performance (v. previous) (valid) visit conversation number (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: All information treatment conditions Duties * * 0.087* (0.060) (0.062) (0.036) (0.036) (0.071) (0.075) (0.007) (0.021) (0.010) (0.025) (0.060) (0.061) Incumbent 0.149*** 0.121*** ** (0.047) (0.044) (0.033) (0.031) (0.049) (0.054) (0.007) (0.023) (0.012) (0.019) (0.057) (0.058) Incumbent Duties * (0.071) (0.078) (0.050) (0.050) (0.084) (0.092) (0.010) (0.024) (0.016) (0.031) (0.083) (0.085) Benchmark 0.236*** 0.262*** *** 0.206*** 0.014*** *** 0.143*** 0.156*** (0.055) (0.048) (0.036) (0.036) (0.059) (0.064) (0.006) (0.035) (0.012) (0.024) (0.058) (0.055) Benchmark Duties * *** ** (0.077) (0.073) (0.050) (0.049) (0.087) (0.093) (0.008) (0.177) (0.014) (0.035) (0.102) (0.077) Panel B: Pooling duties treatment conditions Incumbent 0.161*** 0.150*** *** 0.119*** (0.037) (0.037) (0.022) (0.023) (0.041) (0.042) (0.005) (0.021) (0.008) (0.014) (0.041) (0.041) Benchmark 0.245*** 0.251*** *** 0.221*** 0.008* * 0.102*** 0.088*** (0.037) (0.036) (0.023) (0.024) (0.041) (0.042) (0.005) (0.056) (0.008) (0.016) (0.045) (0.042) Null: Incumbent Benchmark (p value) Observations 3,884 3,875 3,830 3,830 3,926 3,757 3,926 3,926 3,926 3,926 3,926 3,926 Outcome range {1,...,5} {1,...,5} {0,1} {0,1} [-3.4,2.0] [-2.8,1.9 {0,1}] {0,1} {0,1} {0,1} [-7.3,1.5] [-7.8,1.6] Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects, baseline and endline enumerator fixed effects, and a control for the corresponding pretreatment outcomes. All observations are inversely weighted by the endline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. 44

45 Table 8: Heterogeneous effects of information treatments on beliefs about incumbent performance, reported vote for the incumbent, and requests from the incumbent, by leaflet content, priors beliefs, and importance of performance information for vote choice (endline survey) Self-reported outcomes Behavioral outcomes Incumbent Relative Incumbent Incumbent Index Index Request Request Request Called Index Index overall performance vote vote (average) (ICW) incumbent incumbent hotline hotline (average) (ICW) performance (v. previous) (valid) visit conversation number (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Heterogeneity by reported performance level Incumbent 0.165*** 0.154*** *** 0.124*** (0.038) (0.037) (0.022) (0.022) (0.041) (0.042) (0.005) (0.021) (0.007) (0.013) (0.040) (0.041) Incumbent Overall performance (ICW) *** 0.023* 0.051* (0.030) (0.036) (0.027) (0.028) (0.040) (0.045) (0.005) (0.008) (0.006) (0.017) (0.034) (0.034) Benchmark 0.247*** 0.253*** *** 0.226*** 0.008* ** 0.101*** 0.086*** (0.037) (0.036) (0.023) (0.023) (0.039) (0.041) (0.005) (0.056) (0.008) (0.014) (0.044) (0.042) Benchmark Overall performance (ICW) * (0.036) (0.037) (0.031) (0.033) (0.048) (0.054) (0.005) (0.015) (0.008) (0.020) (0.049) (0.047) Panel B: Heterogeneity by relevance-weighted reported performance level Incumbent 0.185*** 0.144*** *** 0.117*** * (0.045) (0.043) (0.025) (0.025) (0.050) (0.051) (0.006) (0.017) (0.008) (0.014) (0.045) (0.046) Incumbent Relevant performance (ICW) 0.071* 0.063* *** 0.029** 0.097*** 0.085** (0.045) (0.048) (0.029) (0.029) (0.054) (0.059) (0.005) (0.019) (0.008) (0.016) (0.043) (0.044) Benchmark 0.250*** 0.247*** *** 0.236*** *** 0.108*** 0.087** (0.041) (0.039) (0.026) (0.026) (0.046) (0.049) (0.006) (0.067) (0.008) (0.015) (0.051) (0.048) Benchmark Relevant performance (ICW) (0.042) (0.040) (0.031) (0.032) (0.053) (0.058) (0.007) (0.022) (0.008) (0.017) (0.049) (0.050) Panel C: Heterogeneity by local and national reported performance level Incumbent 0.160*** 0.163*** *** 0.129*** (0.038) (0.037) (0.024) (0.024) (0.042) (0.044) (0.005) (0.017) (0.008) (0.014) (0.039) (0.039) Incumbent National performance (ICW) (0.043) (0.039) (0.026) (0.029) (0.049) (0.055) (0.010) (0.019) (0.009) (0.016) (0.055) (0.057) Incumbent Local performance (ICW) 0.107*** 0.136*** *** 0.105*** 0.015** * *** 0.120*** (0.044) (0.041) (0.026) (0.029) (0.048) (0.051) (0.008) (0.021) (0.010) (0.015) (0.053) (0.054) Benchmark 0.234*** 0.259*** *** 0.227*** 0.009** ** 0.110*** 0.096*** (0.037) (0.033) (0.023) (0.023) (0.039) (0.041) (0.005) (0.055) (0.008) (0.015) (0.043) (0.041) Benchmark National performance (ICW) (0.046) (0.046) (0.041) (0.042) (0.056) (0.067) (0.011) (0.060) (0.009) (0.017) (0.075) (0.077) Benchmark Local performance (ICW) 0.059* 0.120*** ** 0.093** 0.013* *** ** 0.125** (0.045) (0.042) (0.034) (0.034) (0.049) (0.056) (0.009) (0.024) (0.008) (0.015) (0.059) (0.063) Observations 3,884 3,875 3,830 3,830 3,926 3,827 3,926 3,926 3,926 3,926 3,926 3,926 Outcome range {1,...,5} {1,...,5} {0,1} {0,1} [-3.4,2.0] [-2.8,1.9 {0,1}] {0,1} {0,1} {0,1} [-7.3,1.5] [-7.8,1.6] Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects, baseline and endline enumerator fixed effects, and a control for the corresponding pretreatment outcome. All observations are inversely weighted by the endline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. (Continued...) 45

46 Table 8 (continued): Heterogeneous effects of information treatments on beliefs about incumbent performance, reported vote for the incumbent, and requests from the incumbent, by leaflet content, priors beliefs, and importance of performance information for vote choice (endline survey) Self-reported outcomes Behavioral outcomes Incumbent Relative Incumbent Incumbent Index Index Request Request Request Called Index Index overall performance vote vote (average) (ICW) incumbent incumbent hotline hotline (average) (ICW) performance (v. previous) (valid) visit conversation number (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel D: Heterogeneity by prior belief level Incumbent 0.166*** 0.158*** *** 0.119*** (0.037) (0.036) (0.022) (0.023) (0.042) (0.042) (0.005) (0.023) (0.008) (0.015) (0.042) (0.042) Incumbent Prior index (ICW) (0.041) (0.039) (0.019) (0.018) (0.042) (0.043) (0.006) (0.026) (0.006) (0.016) (0.046) (0.046) Benchmark 0.241*** 0.254*** *** 0.222*** 0.008* * 0.103*** 0.090*** (0.038) (0.037) (0.023) (0.024) (0.041) (0.043) (0.005) (0.053) (0.008) (0.016) (0.046) (0.043) Benchmark Prior index (ICW) (0.042) (0.039) (0.019) (0.020) (0.041) (0.042) (0.005) (0.012) (0.007) (0.015) (0.042) (0.042) Panel E: Heterogeneity by prior belief precision Incumbent 0.164*** 0.161*** *** 0.124*** (0.037) (0.037) (0.024) (0.024) (0.042) (0.043) (0.005) (0.006) (0.008) (0.016) (0.042) (0.043) Incumbent Prior precision index (ICW) 0.053* (0.040) (0.038) (0.024) (0.024) (0.046) (0.049) (0.006) (0.007) (0.008) (0.014) (0.046) (0.047) Benchmark 0.244*** 0.272*** *** 0.234*** 0.008* * 0.086** 0.089** (0.038) (0.037) (0.023) (0.024) (0.041) (0.043) (0.005) (0.005) (0.008) (0.016) (0.045) (0.045) Benchmark Prior precision index (ICW) ** (0.035) (0.036) (0.022) (0.024) (0.040) (0.042) (0.006) (0.006) (0.008) (0.018) (0.052) (0.053) Panel F: Heterogeneity by importance of performance in determining vote choice Incumbent 0.161*** 0.150*** *** 0.119*** (0.037) (0.037) (0.022) (0.022) (0.041) (0.042) (0.005) (0.021) (0.008) (0.015) (0.041) (0.041) Incumbent Performance most important (0.036) (0.031) (0.021) (0.020) (0.035) (0.038) (0.005) (0.013) (0.007) (0.013) (0.041) (0.041) Benchmark 0.245*** 0.251*** *** 0.221*** 0.008* * 0.102*** 0.087*** (0.037) (0.036) (0.023) (0.024) (0.041) (0.043) (0.005) (0.056) (0.008) (0.015) (0.044) (0.042) Benchmark Performance most important * 0.030** * * (0.036) (0.034) (0.018) (0.018) (0.034) (0.035) (0.004) (0.058) (0.008) (0.016) (0.044) (0.038) Panel G: Heterogeneity by preference for locally-oriented deputies Incumbent 0.107* 0.107* *** ** *** (0.069) (0.071) (0.037) (0.042) (0.078) (0.083) (0.007) (0.022) (0.015) (0.026) (0.073) (0.072) Incumbent Prefer locally-oriented deputies *** 0.077* 0.125* 0.166** 0.025*** (0.080) (0.079) (0.041) (0.046) (0.083) (0.091) (0.009) (0.039) (0.018) (0.029) (0.091) (0.091) Benchmark 0.184*** 0.225*** *** 0.117* * ** (0.061) (0.060) (0.041) (0.045) (0.071) (0.080) (0.006) (0.292) (0.015) (0.028) (0.111) (0.065) Benchmark Prefer locally-oriented deputies * 0.071* * 0.021*** (0.075) (0.074) (0.044) (0.047) (0.083) (0.091) (0.008) (0.330) (0.015) (0.032) (0.126) (0.073) Observations 3,884 3,875 3,830 3,830 3,926 3,827 3,926 3,926 3,926 3,926 3,926 3,926 Outcome range {1,...,5} {1,...,5} {0,1} {0,1} [-3.4,2.0] [-2.8,1.9 {0,1}] {0,1} {0,1} {0,1} [-7.3,1.5] [-7.8,1.6] Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects, baseline and endline enumerator fixed effects, and a control for the corresponding pretreatment outcome. All observations are inversely weighted by the endline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. 46

47 tant factor in determining their vote choice or preferred locally-oriented deputies. Appendix Table A9 shows that turnout was not affected. However, our focus on voters aged does not necessarily capture the response of a broader electorate that may also receive the information provided. Differences could arise because younger voters are less likely to vote, and may respond to information differently. To better approximate polling station-level electoral outcomes, we subset our sample to the set of voters that turned out in the 2012 election a strong predictor of validated turnout in Table 9 suggests that the self-reported vote choices of such voters was more likely to reflect the information provided. While not statistically significant in this smaller sample, columns (1) and (5) show that voters initial increase in support for the incumbent persists and if anything increases in magnitude by election day. More definitively, column (8) shows that valid votes in favor of the incumbent significantly increase with local performance. The points estimates in columns (2)-(3) and (6)-(7) also suggest persisting from vote intentions to actual vote choices. To the extent that such behavior also proxies for the behavior of older voters that might learn of the information contained in our leaflet, it is possible that the incumbent vote share could change at the polling station level. We now turn to the polling station returns. Such administrative data is a more reliable measure of voting behavior, particularly given the over-reporting of incumbent voting noted above. 29 However, the results may be noisy because only nine voters received a leaflet in each treated village. A key factor is thus the extent to which information provision precipitated other actions, such as information dissemination and coordination among voters; below, we show that such activities were substantial. To account for these dynamics, we primarily focus on heterogeneous treatment effects reflecting the share of the polling station electorate that is registered in the village in our sample ( share ). 30 In our sample, the average village comprises 65% of registered voters at its polling 28 This eliminates the first-time voters aged 23 and below that were not eligible to vote in For four several villages, we were unable to obtain complete electoral returns. 30 Unreported analyses indicate similar results when examining the actual number of respondents directly treated, and various multipliers of this number. These interactions were not pre-specified, but follow naturally from the discussion of noisy polling station data in our pre-analysis plan. 47

48 Table 9: Effects of information treatments on reported vote for the incumbent among those that turned out in 2012 (baseline and endline survey) Incumbent vote (intention) Incumbent vote (valid) (1) (2) (3) (4) (5) (6) (7) (8) Incumbent (0.018) (0.018) (0.019) (0.019) (0.036) (0.035) (0.036) (0.038) Benchmark (0.017) (0.017) (0.017) (0.017) (0.032) (0.032) (0.034) (0.032) Incumbent Overall performance (ICW) 0.020* (0.011) (0.044) Benchmark Overall performance (ICW) 0.032** (0.014) (0.034) Incumbent Relevant performance (ICW) 0.041** (0.020) (0.048) Benchmark Relevant performance (ICW) (0.021) (0.048) Incumbent National performance (ICW) * ** (0.018) (0.045) Incumbent Local performance (ICW) * (0.025) (0.047) Benchmark National performance (ICW) (0.022) (0.050) Benchmark Local performance (ICW) ** (0.023) (0.047) Observations 1,550 1,550 1,550 1,550 1,457 1,457 1,457 1,457 Outcome range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} Control outcome mean Control outcome std. dev Interaction mean Interaction std. dev Second interaction mean Second interaction std. dev Notes: Each specification is estimated using OLS, and includes randomization block and baseline enumerator fixed effects and pre-treatment incumbent vote intention. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. 48

49 station. The polling station returns suggest significant increases in incumbent voting where the incumbent performed best. The first column for each outcome in Table 10 reports that our intervention did not affect turnout or incumbent party vote share on average, while the second column finds little evidence of greater average effects where the village represents a greater share of the polling station electorate. However, the third, fourth, and fifth columns provide significant evidence of the anticipated heterogeneous effects: as the positive triple interactions show, the marginal effect of the incumbent and especially benchmark treatments on the incumbent s vote share increases with reported primarily local performance when the village represents a large fraction of voters at the polling station. 31 The estimates imply that a standard deviation increase in performance increases the incumbent s vote share by around five percentage points for a village with an above-median share of the polling station electorate. This magnitude of this effect far exceeds the number of survey respondents (the average polling station has over 500 registered voters), suggesting that our leaflets induced a response among voters beyond our panel participants. Appendix Table A8 provides stronger evidence of such heterogeneous effects when instead using the Kling, Liebman and Katz (2007) index. In sum, these findings indicate that while performance information did not affect incumbent vote share on average, villages did ultimately reward better-performing incumbents at the polls 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 somewhat different form of accountability. The requests that we study are less likely to be the source of election campaign interactions, and thus less subject to interference between survey 31 Subsetting the sample to villages with above-median shares of the polling station electorate provides similar results. 49

50 Table 10: Heterogeneous effects of information treatments on polling station-level turnout and incumbent vote share, by village contribution to polling station and leaflet content (polling station data) Turnout Incumbent vote share Incumbent vote share (proportion of turnout) (proportion of registered voters) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Incumbent ** * ** *** ** (0.013) (0.027) (0.031) (0.026) (0.022) (0.020) (0.040) (0.045) (0.037) (0.035) (0.016) (0.031) (0.034) (0.029) (0.026) Benchmark (0.013) (0.020) (0.027) (0.024) (0.026) (0.018) (0.043) (0.049) (0.044) (0.040) (0.013) (0.034) (0.037) (0.033) (0.030) Incumbent Share 0.051* * (0.035) (0.042) (0.034) (0.030) (0.059) (0.061) (0.057) (0.055) (0.044) (0.046) (0.043) (0.041) Benchmark Share * * (0.024) (0.030) (0.029) (0.027) (0.065) (0.071) (0.067) (0.064) (0.047) (0.050) (0.047) (0.044) Incumbent Overall performance (ICW) (0.017) (0.034) (0.025) Incumbent Share Overall performance (ICW) 0.075* * (0.053) (0.064) (0.044) Benchmark Overall performance (ICW) (0.025) (0.028) (0.023) Benchmark Share Overall performance (ICW) 0.080* ** (0.060) (0.050) (0.036) Incumbent Relative performance (ICW) (0.027) (0.044) (0.032) Incumbent Share Relative performance (ICW) (0.035) (0.064) (0.047) Benchmark Relative performance (ICW) * (0.036) (0.048) (0.037) Benchmark Share Relative performance (ICW) *** (0.043) (0.066) (0.049) Incumbent National performance (ICW) (0.016) (0.063) (0.035) Incumbent Share National performance (ICW) (0.023) (0.103) (0.060) Benchmark National performance (ICW) * (0.037) (0.061) (0.039) Benchmark Share National performance (ICW) (0.046) (0.086) (0.053) Incumbent Local performance (ICW) (0.027) (0.058) (0.035) Incumbent Share Local performance (ICW) * (0.040) (0.088) (0.053) Benchmark Local performance (ICW) * (0.031) (0.056) (0.037) Benchmark Share Local performance (ICW) 0.069* (0.043) (0.078) (0.050) Observations Outcome range [.15,1.5] [.15,1.5] [.15,1.5] [.15,1.5] [.15,1.5] [.06,.96] [.06,.96] [.06,.96] [.06,.96] [.06,.96] [.02,.66] [.02,.66] [.02,.66] [.02,.66] [.02,.66] Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block and baseline enumerator fixed effects and a lagged dependent variable. Observations are not weighted. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. 50

51 waves. We next evaluate the extent to which such accountability dissipates a month after treatment by examining requests from the newly-elected deputy (always the incumbent). Unlike with voting behavior, treated voters continue to make more requests from incumbents shown to be more responsive on average. Columns (7)-(10) of Table 7 indicate that the benchmark information, in particular, generally increased requests from the government. Aggregated as an index, columns (11) and (12) of panel B report a 0.1 standard deviation increase in such behaviors. Since almost all voters sought a visit or conversation with the incumbent, calling our hotline is perhaps the most useful measure. Despite the financial costs involved, column (10) shows that the benchmark treatment increased attempts to engage with the newly-reelected incumbent by 2.5 percentage points, or around a quarter of the control group mean. The heterogeneous effects in Table 8 also report that these increases in hotline usage were greatest in departments where the incumbent was performing best, particularly on local issues. Moreover, at least in the case of calling the hotline, the benchmark treatment had its largest effects on respondents with imprecise prior expectations and who care most about incumbent performance on the dimensions that information was provided about Summary These findings suggest that information provision has lasting effects: voters remember the information, to a significant degree retain their updated beliefs about incumbents, and continue to make requests of incumbents that they learn performed well. Each feature indicates that information could help hold politicians accountable on the basis of their performance. Given that posterior beliefs and vote intentions suggested that a sharp shift toward the incumbent would occur, the most notable difference from voters immediate responses in the baseline survey is the relatively small change in voting behavior on average in our sample. This echoes findings from Mexico and the Philippines, where information treatments produced lower average effects than expected (Arias et al. 2018; Cruz, Keefer and Labonne 2017). Nevertheless, rather than 51

52 voters neglecting the information entirely, polling station returns suggest that the best-performing incumbents were still rewarded. This raises several key questions: given that treated voters continue to perceive the incumbent more favorably, what altered their average pre-election voting intentions? And why do changes in voting behavior show up at an aggregate level that exceeds any feasible effects on our sample? Since respondent requests persisted into the endline survey, we focus on political responses relating to the election itself. 5.3 General equilibrium influences on voting behavior We focus on three potential explanations for the relatively low average effect, but non-negligible heterogeneous effects, of providing incumbent performance information at the polling station level: village-level social interactions, cross-village spillovers, and politician responses. By also affecting the behavior of other voters, each account could plausibly explain the polling station level outcomes and the disjuncture on the level between immediate vote intentions and ultimate vote choices. The unusually rich data we collected from followup surveys to address the first and last potential mediators more broadly highlights the importance of election-based responses to information dissemination campaigns Voter social networks within villages As suggested in Arias, Balán, Larreguy, Marshall and Querubín (2017), social networks could amplify or alter the effect of information provision by respectively diffusing information or inducing voter coordination on a preferred (set of) candidate(s) within villages. Both rely on our treated voters interacting with fellow villagers, either by disseminating the leaflet s information or inducing others to change their voting behavior, although only the former requires that novel information actually be transferred. We use our endline survey to consider whether these explanations could account for our findings by exploring the extent to which the leaflets generated social interactions. The endline survey provides clear evidence of substantial voter engagement with the leaflets 52

53 Table 11: Effects of information treatments on voter interactions with the leaflets (endline survey) Discussed leaflet with others Discussion led to vote coordination Coordinated vote on incumbent (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Incumbent 0.377*** 0.375*** 0.378*** 0.341*** 0.242*** 0.241*** 0.241*** 0.218*** 0.213*** 0.212*** 0.213*** 0.194*** (0.016) (0.016) (0.016) (0.019) (0.019) (0.019) (0.019) (0.019) (0.017) (0.017) (0.017) (0.016) Benchmark 0.397*** 0.395*** 0.398*** 0.362*** 0.269*** 0.269*** 0.270*** 0.248*** 0.239*** 0.240*** 0.240*** 0.220*** (0.020) (0.021) (0.021) (0.022) (0.020) (0.020) (0.019) (0.020) (0.018) (0.018) (0.017) (0.018) Incumbent Overall performance (ICW) (0.015) (0.021) (0.018) Benchmark Overall performance (ICW) (0.018) (0.020) (0.019) Incumbent Relevant performance (ICW) (0.024) (0.024) (0.021) Benchmark Relevant performance (ICW) (0.027) (0.022) (0.020) Incumbent National performance (ICW) *** (0.031) (0.025) (0.019) Benchmark National performance (ICW) (0.027) (0.028) (0.028) Incumbent Local performance (ICW) * (0.029) (0.023) (0.020) Benchmark Local performance (ICW) (0.029) (0.027) (0.026) Null: Incumbent Benchmark (p value) Observations 3,925 3,925 3,925 3,925 3,924 3,924 3,924 3,924 3,924 3,924 3,924 3,924 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} Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects and baseline and endline enumerator fixed effects. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from pre-specified one-sided t tests. 53

54 within their village. Column (1) of Table 11 demonstrates that nearly 40% of treated respondents discussed the leaflet with others. This suggests that significant information diffusion occurred, which may account for the fact that directly providing leaflets to less than 2% of registered voters still resulted in some discernible polling station-level effects. Furthermore, column (5) indicates that more than half of these people also coordinated their vote as a consequence of their discussion. In both cases, these activities were slightly higher following a benchmarked leaflet. Of those that coordinated, more than half reported coordinating with six or more people; 10% reported coordinating with the entire village. Column (2)-(4) and (6)-(8) suggest that such social interactions were orthogonal to the content of the leaflet, even when focusing on the local information that interested respondents most. Such social interaction around the leaflet s provision could potentially account for the polling station-level electoral shifts. In order to understand the nature of voter coordination, we conducted a followup survey in December 2017 among those that reported coordinating. The vast majority of coordination involved informal interactions rather than organized meetings between voters, although in around one half of cases it also involved interacting with candidates (or the operatives) or local traditional leaders that typically approached voters first. Much of the discussion among voters concerned projects and transfers. Although our respondents generally updated their individual beliefs positively, group discussions were equally split between regarding the incumbent s performance as higher or lower than expected. While activity levels in parliament were discussed much less frequently, they were more favorable. Where politicians got involved, they generally promised to improve local project delivery rather than enhance national policies. Ultimately, when voters coordinated, they generally reported siding with the incumbent as shown in column (9) of Table 11. Large majorities agreed that the incumbent would work harder in the chamber and bring their department more projects. Of those that coordinated, 87% reported that their group coordinated on voting for the incumbent. Since such coordination overwhelmingly favored incumbents, but was not systematically cor- 54

55 related with incumbent performance (as columns (10)-(12) in Table 11 illustrate), the coordination induced by this intervention does not appear to have a significantly altered polling station level outcomes. To the extent that social networks might explain our findings, it thus appears more likely that information diffused within the village. However, this would likely require that the information became less favorable toward the incumbent as it passed between voters 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 the pure control group, we restrict the sample to this subset of village 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 A10 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. 32 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 Party responses to information dissemination Politicians rarely stand by when potentially influential information is released (see e.g. Arias et al. 2018; Bidwell, Casey and Glennerster 2016; Cruz, Keefer and Labonne 2017). Consequently, another possible explanation for the lack of a persistent average treatment effect on incumbent 32 Although information could travel without reference to our leaflets, unreported results also show that leaflet recall is also unaffected. 55

56 electoral support, but positive effects when interacted with the information content, is that challenger parties were particularly effective at counteracting information than 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 plausible channel through which incumbent responses could explain our findings is if incumbents reallocate resources from treatment to control 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. 33 As shown in Table 12, 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 did, we used our December followup survey to ask about what actions parties took and whether they were effective. Voters that reported incumbent-held community meetings or discussions 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. 33 Half the sample were 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. 56

57 Table 12: Effects of information treatments on incumbent and challenger responses (endline survey) Incumbent response Challenger response (1) (2) (3) (4) (5) (6) Incumbent 0.067*** 0.067*** 0.068*** 0.042*** 0.044*** 0.043*** (0.012) (0.012) (0.012) (0.008) (0.008) (0.008) Benchmark 0.083*** 0.082*** 0.082*** 0.048*** 0.049*** 0.047*** (0.012) (0.012) (0.012) (0.010) (0.010) (0.010) Incumbent Performance index (average) (0.012) (0.009) Benchmark Performance index (average) (0.013) (0.011) Incumbent Performance index (ICW) (0.011) (0.007) Benchmark Performance * ** index (ICW) (0.013) (0.008) Observations 3,925 3,925 3,925 3,925 3,925 3,925 Outcome range {0,1} {0,1} {0,1} {0,1} {0,1} {0,1} Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects and baseline and endline enumerator fixed effects. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Lower order interactions between the list experiment and performance indexes are omitted. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. The lower response to more positive information among both incumbents and challengers, in columns (3) and (6), respectively counteracts, to the extent that incumbent campaigns efforts are (more) effective, our finding that the incumbent vote share increases with performance in treated villages. These results are reinforced by an unreported negative correlation between performance and how convincing respondents found the incumbent response. Together, such evidence suggests that electoral accountability may have been greater in the absence of political responses. Although vote buying is prevalent, we were not able to detect a systematic indirect response to information dissemination through vote buying. The results of the list experiment in columns (1) and (5) of Table 13 indicate that 22% of voters reported receiving a gift from the incumbent, 57

58 Table 13: Effects of information treatments on vote buying (endline survey) Incumbent list experiment Challengers list experiment Items Items Items Items Items Items listed listed listed listed listed listed (1) (2) (3) (4) (5) (6) Incumbent list 0.219*** 0.231*** 0.233*** (0.026) (0.047) (0.047) Incumbent list Incumbent (0.067) (0.067) Incumbent list Benchmark (0.069) (0.069) Incumbent list Incumbent Overall performance (ICW) (0.063) Incumbent list Benchmark Overall performance (ICW) (0.062) Opposition list 0.219*** 0.248*** 0.244*** (0.030) (0.051) (0.051) Opposition list Incumbent (0.080) (0.080) Opposition list Benchmark (0.070) (0.070) Opposition list Incumbent Overall performance (ICW) (0.057) Opposition list Benchmark Overall performance (ICW) (0.055) Observations 2,933 2,933 2,970 2,970 2,970 2,970 Outcome range {0,...,4} {0,...,4} {0,...,4} {0,...,4} {0,...,4} {0,...,4} Control outcome mean Control outcome std. dev Notes: Each specification is estimated using OLS, and includes randomization block fixed effects and baseline and endline enumerator fixed effects. All observations are inversely weighted by the baseline number of respondent surveyed in the village. Lower order interactions between the list experiment and performance indexes are omitted. Standard errors are clustered by randomization block. * denotes p < 0.1, ** denotes p < 0.05, *** denotes p < 0.01 from two-sided t tests. 58

59 while another 22% reported receiving a gift from a challenger. Although such vote buying was somewhat lower in treated villages, our estimates are too imprecise to be able to conclude that the substitution of vote buying across villages can account for our findings. Furthermore, widespread vote buying (which may also operate at the village bloc level) could also potentially explain the lack of electoral response to information in the face of belief updating. This is also consistent with persistent requests from better-performing politicians. 6 Conclusion In light of the mixed evidence that information campaigns can support electoral accountability, we illuminate step-by-step the process linking incumbent performance information to both voting behavior and requests from politicians among among rural Senegalese voters. We observe that votersprocess new information in a Bayesian manner, retain such information at least one month later, and make more requests from and at least initially express an intention to vote for better-performing incumbents. Moreover, such sophisticated responses did not depend on educating voters about the responsibilities of parliamentary deputies. Given the relatively low levels of education in our sample, these findings clearly suggest that information may be sufficient to enable most of the world s voters to hold politicians to account. However, the extent to which politicians are held to account also depends on other forces stimulated by information dissemination. In contrast with voter requests from politicians, which persistently follow the textbook accountability story and are augmented by benchmarked performance information, vote choices appear to reflect a more complex electoral equilibrium that demands further analysis. In highlighting the importance of complex responses to information provision before elections, our findings accord with various recent studies documenting strategic responses to information campaigns (Arias et al. 2018; Banerjee et al. 2011; Besley and Burgess 2002; Bidwell, Casey and Glennerster 2016; Cruz, Keefer and Labonne 2017). We build on these studies by demonstrating 59

60 that the mixed effects of information campaigns in the extant literature did not reflect faults on the part of voters, but likely reflected electoral interactions that did not change voter beliefs but dissuaded voters from sanctioning or rewarding incumbents. Furthermore, we describe in detail the nature of these responses, finding that meetings to discuss the information and voter coordination around the information cannot fully account for voting behavior. However, such analysis remains tentative in light of the difficulties of estimating such causal mediation effects. This suggests that future researchers should focus on more directly unpacking political responses. For example, this could involve factorial designs crossing information dissemination with informing incumbent and challenger politicians such campaigns. Less demanding experiments could seek to inform only politicians of events that they may be unaware of, but which could influence their electoral prospects, before measuring their responses. A key question regarding our anatomy of political accountability is the extent to which the findings extend beyond rural Senegal. In particular, such rural areas may respond more to information campaigns, given their comparatively low normal levels of exposure to political information and NGO activity. Nevertheless, there are several reasons to believe that our results should inform the broader understanding of political accountability. First, by documenting sophisticated responses among poorly educated voters with limited interaction with the type of information that we provided, we believe that this study highlights that all types of voters possess the capacity to process simple incumbent performance indicators. While perhaps not surprising to some, this mitigates concerned raised in the American politics literature (e.g. Gomez and Wilson 2001, 2006). 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 to campaigns like ours. In more developed contexts, it is possible that different types of campaign strategy would interact with localized information provision differently. Third, because our leaflets are similar in design to previous studies (e.g. Adida et al. 2017; Arias et al. 2018; Chong et al. 2015; Gottlieb 2016; Humphreys and Weinstein 2012), our findings may help to direct 60

61 researchers in other contexts toward the types of impediments to electoral accountability that we highlight here. Similarly, our findings suggest to information campaign designers that temporally benchmarking locally-relevant incumbent performance information may be an important way of enhancing accountability, while detailed information about duties and national behaviors may have limited value. 61

62 References Adida, Claire, Jessica Gottlieb, Gwyneth McClendon and Eric Kramon Breaking the Clientelistic Voting Equilibrium: The Joint Importance of Salience and Coordination. AidData Working Paper #48. Alt, James E., David D. Lassen and John Marshall Information sources, belief updating, and the politics of economic expectations: Evidence from a Danish survey experiment. Journal of Politics 78(2): Anderson, Michael L Multiple inference and gender differences in the effects of early intervention: A reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects. Journal of the American Statistical Association 103(484): Arias, Eric, Horacio A. Larreguy, John Marshall and Pablo Querubín 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 Querubín Priors rule: When do malfeasance revelations help and hurt incumbent parties? Working paper. Arias, Eric, Pablo Balán, Horacio A. Larreguy, John Marshall and Pablo Querubín Malfeasance Revelations, Coordination Through Social Networks and Electoral Sanctioning: Experimental Evidence from Mexico. Working paper. Ashworth, Scott Electoral accountability: recent theoretical and empirical work. Annual Review of Political Science 15: Ashworth, Scott and Ethan Bueno de Mesquita Electoral selection, strategic challenger entry, and the incumbency advantage. Journal of Politics 70(4):

63 Ashworth, Scott and Ethan Bueno de Mesquita Is Voter Competence Good for Voters? Information, Rationality, and Democratic Performance. American Political Science Review 108(3): Banerjee, Abhijit V., Selvan Kumar, Rohini Pande and Felix Su Do Informed Voters Make Better Choices? Experimental Evidence from Urban India. Working paper. Beck, Linda J Senegal. In Countries at the Crossroads 2011: An Analysis of Democratic Governance, ed. Jake Dizard, Christopher Walker and Vanessa Tucker. Rowman & Littlefield. Besley, Timothy and Andrea Prat Handcuffs for the grabbing hand? Media capture and government accountability. American Economic Review 96(3): Besley, Timothy and Robin Burgess The political economy of government responsiveness: Theory and evidence from India. Quarterly Journal of Economics 117(4): Bhandari, Abhit, Horacio A. Larreguy and John Marshall Learning accountability: Can formal education help hold politicians to account? Pre-analysis plan, Social Science Registry. Bidwell, Kelly, Katherine Casey and Rachel Glennerster Debates: Voting and Expenditure Responses to Political Communication. Working Paper. Björkman Nyqvist, Martina, Damien de Walque and Jakob Svensson Experimental Evidence on the Long-Run Impact of Community-Based Monitoring. American Economic Journal: Applied Economics 9(1): Buisseret, Peter and Carlo Prato Legislative Representation in Multi-member Districts. Working paper. Bullock, John G Elite influence on public opinion in an informed electorate. American Political Science Review 105(3):

64 Chan, Jimmy and Wing Suen A spatial theory of news consumption and electoral competition. Review of Economic Studies 75(3): Chauchard, Simon, Marko Klašnja and S.P. Harish The Limited Impact of Information on Political Selection: An Experiment on Asset Disclosures in India. Working paper. Chauchard, Simon and Neelan Sircar. forthcoming. Dilemmas and Challenges of Citizen Information Campaigns: Lessons from a Failed Experiment in India. In Information and Political Accountability: A New Method for Cumulative Learning, ed. Thad Dunning, Guy Grossman, Marcatan Humphreys, Susan Hyde and Craig McIntosh. Cambridge University Press. Chiang, Chun-Fang and Brian Knight Media bias and influence: Evidence from newspaper endorsements. Review of Economic Studies 78(3): Chong, Alberto, Ana De La O, Dean Karlan and Leonard Wantchekon Does Corruption Information Inspire the Fight or Quash the Hope? A Field Experiment in Mexico on Voter Turnout, Choice and Party Identification. Journal of Politics 77(1): Chwe, Michael Suk-Young Communication and coordination in social networks. Review of Economic Studies 67(1):1 16. Cruz, Cesi, Philip Keefer and Julien Labonne Incumbent Advantage, Voter Information and Vote Buying. Working paper. de Figueiredo, Miguel F.P., F. Daniel Hidalgo and Yuri Kasahara When Do Voters Punish Corrupt Politicians? Experimental Evidence from Brazil. Working paper. Duch, Raymond M. and Randolph T. Stevenson The Economic Vote: How Political and Economic Institutions Condition Election Results. Cambridge University Press. Dunning, Thad, Guy Grossman, Macartan Humphreys, Susan Hyde and Craig McIntosh. forthcoming. Metaketa I: The Limits of Electoral Accountability. Cambridge University Press. 64

65 Durante, Ruben and Emilio Gutierrez Political Advertising and Voting Intentions: Evidence from Exogenous Variation in Ads Viewership. Working paper. Fearon, James D 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 Exposing Corrupt Politicians: The Effects of Brazil s Publicly Released Audits on Electoral Outcomes. Quarterly Journal of Economics 123(2): Gentzkow, Matthew and Jesse M. Shapiro Media bias and reputation. Journal of Political Economy 114(2): Gomez, Brad T. and J. Matthew Wilson Political Sophistication and Economic Voting in the American Electorate: A Theory of Heterogeneous Attribution. American Journal of Political Science 45(4): Gomez, Brad T. and J. Matthew Wilson Cognitive heterogeneity and economic voting: A comparative analysis of four democratic electorates. American Journal of Political Science 50(1): Gottlieb, Jessica Greater Expectations: A Field Experiment to Improve Accountability in Mali. American Journal of Political Science 60(1): Gottlieb, Jessica and Horacio Larreguy An Informational Theory of Electoral Targeting in Young Clientelistic Democracies: Evidence from Senegal. Working Paper. Humphreys, Macartan and Jeremy Weinstein Policing Politicians: Citizen Empowerment and Political Accountability in Uganda Preliminary Analysis. Working paper. 65

66 Imai, Kosuke, Luke Keele, Dustin Tingley and Teppei Yamamoto Unpacking the Black Box of Causality: Learning about Causal Mechanisms from Experimental and Observational Studies. American Political Science Review 105(4): Kendall, Chad, Tommaso Nannicini and Francesco Trebbi How do voters respond to information? Evidence from a randomized campaign. American Economic Review 105(1): Khagram, Sanjeev, Archon Fung and Paolo De Renzio Open budgets: The Political Economy of Transparency, Participation, and Accountability. Brookings Institution Press. Khemani, Stuti, Ernesto Dal Bó, Claudio Ferraz, Frederico S. Finan, Johnson Stephenson, Louise Corinne, Adesinaola M. Odugbemi, Dikshya Thapa and Scott D. Abrahams Making politics work for development: Harnessing transparency and citizen engagement. World Bank Policy Research Report Kling, Jeffrey R., Jeffrey B. Liebman and Lawrence F. Katz Experimental analysis of neighborhood effects. Econometrica 75(1): Koter, Dominika King makers: Local leaders and ethnic politics in Africa. World Politics 65(2): Larreguy, Horacio, John Marshall and Jr. Snyder, James M Publicizing malfeasance: When Media Facilitates Electoral Accountability in Mexico,.. Larreguy, Horacio and Shelley Liu The Effect of Education on Political Participation: Evidence from a Competitive Consolidating Democracy.. Lieberman, Evan S., Daniel N. Posner and Lily L. Tsai Does information lead to more active citizenship? Evidence from an education intervention in rural Kenya. World Development 60:

67 Manin, Bernard, Adam Przeworski and Susan C. Stokes Introduction. In Democracy, Accountability, and Representation, ed. Adam Przeworski, Susan C. Stokes and Bernard Manin. Cambridge University Press pp Marshall, John Political information cycles: When do voters sanction incumbent parties for high homicide rates? Working paper. Marshall, John. forthcoming. Signaling sophistication: How social expectations can increase political information acquisition. Journal of Politics. McKenzie, David Beyond baseline and follow-up: The case for more T in experiments. Journal of Development Economics 99(2): Meyer, Margaret A. and John Vickers Performance comparisons and dynamic incentives. Journal of Political Economy 105(3): Morris, Stephen and Hyun Song Shin Social value of public information. American Economic Review 92(5): O Donnell, Guillermo A Delegative democracy. Journal of Democracy 5(1): Powell, G. Bingham, Jr. and Guy D. Whitten A Cross-National Analysis of Economic Voting: Taking Account of the Political Context. American Journal of Political Science 37(2): Rogoff, Kenneth Equilibrium Political Budget Cycles. American Economic Review 80(1): Thomas, Melissa A and Oumar Sissokho Liaison Legislature: The Role of the National Assembly in Senegal. The Journal of Modern African Studies 43(1):

68 Villalón, Leonardo A Islamic Society and State Power in Senegal: Disciples and Citizens in Fatick. Cambridge University Press. Weitz-Shapiro, Rebecca and Matthew S. Winters Discerning corruption: Credible Accusations and the Punishment of Politicians in Brazil. Working paper. Zaller, John R The Nature and Origins of Mass Opinion. Cambridge University Press. 68

69 A Online appendix Contents A.1 All leaflet configurations A1 A.2 Experimental validation tests A1 A.3 Additional results A1 A.1 All leaflet configurations In addition to the duties + benchmark leaflet show in Figure 3, Figures A1-A4 show our other leaflet configurations. A.2 Experimental validation tests Tables A1 and A2 respectively report individual-level and village-level balance across information treatment in the baseline and endline samples. The F test in the final column tests the restriction that all treatment conditions are indistinguishable from the control group, reporting significant differences consistent with chance in both the baseline and endline samples. This reflects the low attrition rate, and the lack of differential attrition demonstrated in Table A3. A.3 Additional results Tables A4 and A5 respectively show that our information treatments increased the precision of respondent beliefs about the incumbent at both baseline and endline. Tables A6 and A7 report our estimates of the survey-level heterogeneous when using the Kling, Liebman and Katz (2007) approach to creating a summative rating scale, instead of our preferred A1

70 LEAFLET - 35.pdf 1 26/06/17 20:08 C M Y CM MY CY CMY K Figure A1: Example of duties treatment A2

71 Figure A2: Example of incumbent treatment in Oussouye A3

72 Figure A3: Example of duties + incumbent treatment in Oussouye A4

73 Figure A4: Example of benchmark treatment in Oussouye A5