Repairing Information Underload: The Effects on Vote Choice of Information Provided by Mobile Phone on Politician Performance in Uganda

Transcription

1 Repairing Information Underload: The Effects on Vote Choice of Information Provided by Mobile Phone on Politician Performance in Uganda 1 Mark T. Buntaine University of California, Santa Barbara buntaine@bren.ucsb.edu Sarah Bush Temple University sarah.bush@temple.edu Ryan Jablonski London School of Economics and Political Science R.S.Jablonski@lse.ac.uk Daniel Nielson Brigham Young University dan_nielson@byu.edu Paula Pickering College of William & Mary pmpick@wm.edu 6 May 2016 Prepared for presentation at the Evidence in Governance and Politics Conference, Santiago, Chile, May The authors thank Jacob Skaggs, Catherine Tabingwa, and Immaculate Apio Ayado for contributions to the design and implementation of this research, as well as Twaweza staff for their contributions to research design and for sharing data. This project was funded by an anonymous donor as part of the EGAP Information and Accountability Metaketa Initiative.

2 Abstract For government by the people to work, citizens must be able to discern politicians actions, attribute credit or blame, and then vote for leaders who promote the people s interests and against those who do not. Can this accountability mechanism work in low information and semi democratic environments? Can mobile technologies help resolve the informational constraints in such environments? We undertook a large ( n =16,083) randomized control trial during the February and March 2016 Ugandan local elections. We compiled information on irregularities in district budgets and independently audited local public services in education, health, roads, and water access. We then randomly assigned and transmitted the information to subjects in a nationally representative sample of polling stations in the run up to local elections via approximately 500,000 SMS text messages. We ind some evidence that voters are responsive to positive information about public services. Information that public services in given villages were better than the district norm increased votes for incumbent sub county councillors, but not for other of ices. News that budget discrepancies are worse than normal signi icantly decreased votes for incumbent district chairs and councillors, though these results are not robust to all pre registered speci ications. Contrary to others studies, we ind little evidence that information treatments affect turnout. These results suggest that information on politician performance can sometimes alter voters choices in local elections, though not always in the expected or consistent ways. The effects of factual messages on voting for local of icials are small substantively, but could be large enough to swing many elections if the intervention were scaled up and enhanced efforts were made to capture voters attention. 2

3 Introduction Government by the people relies on citizens learning what of icials do with the public s trust and then rewarding good performance with longer tenure or punishing bad outcomes with ouster. But are citizens able and willing to hold politicians accountable and thus encourage better performance by politicians? There are several reasons why the answer could be no. First, citizens may lack adequate information on politician performance, in part due to leaders attempts to conceal it. Second, citizens may care more about receiving patronage from public of icials (i.e., clientelism) than politician performance. Third, poor performance by politicians may demobilize the public over time, discouraging citizens from exercising their right to hold of icials accountable. And fourth, government typically takes the form of layers of of icials at different levels of geographic speci icity, with each more local government nested in larger jurisdictions. Voters may have trouble attributing credit or blame for the outcomes of policies and programs to politicians at the appropriate level. To better understand these dynamics, this study investigates the effects of political information on citizen behavior in Uganda, a competitive authoritarian regime that held local elections in February and March Factual, nonpartisan information about budget irregularities and local public services could affect whether citizens turn out to vote in the irst place and how they choose to vote for local politicians. Moreover, these effects could vary according to individuals prior beliefs about politician performance, previous interactions with political of icials, and the importance and novelty of the information. 3

4 Below, we develop these hypotheses in more detail, placing them in the context of a rich literature that examines the relationship between political information and accountability. In keeping with emerging best practices in the social sciences, these hypotheses and our plan for testing them were fully speci ied and pre registered in advance of our 2 intervention and analysis via the Evidence in Governance and Politics registry. Our research design involved a randomized control trial in which we compiled information on budget irregularities, audited local public services, and then transmitted the information we gathered to potential voters through SMS text messages. The informational interventions focused on budget discrepancies and the public service health, education, roads or water access that individual citizens had previously identi ied as most important to them when deciding how to vote. The experiment was executed in the run up to elections for several local of ices. This design permits causal identi ication of the electoral effects of information about budget discrepancies and public services. In particular, it allows us to answer this question: if citizens learn that their local government performs better or worse than other locales, will the information increase votes for incumbents compared to placebo messages that are devoid of information about politician performance? Rigorously studying the information accountability relationship would be an important task anywhere, but it is particularly relevant in Uganda, where the quality of governance is seriously hampered by corruption. Many Ugandan citizens lack knowledge of the basic functions of local governments, which hinders their ability to participate in decision making and hold local of icials accountable (Natamba et al. 2010, p. 16). Since 2 4

5 mobile phone access is nearly universal in Uganda (Hellström and Karefelt 2012) and SMS text messaging is a cost effective and scalable way to deliver personalized information to citizens our experimental design for improving political information is highly relevant for the many government agencies, international donors, and activists that are pursuing this goal in Uganda and beyond. Indeed, the information transmitted in this experiment could be gathered and transmitted by civil society organizations or by concerned citizens interested in improving governance. In this large scale experiment ( n = 16,083), we ind evidence that informing citizens that district budget discrepancies are worse than they expected causes signi icantly decreased votes for incumbent parties candidates for district chair and councillor, though this result is not robust to all speci ications, especially those pre speci ied in our analysis plan. However, information that local budget discrepancies are better than subjects expectations has no signi icant effect on votes for incumbents. Additionally, informing citizens that local public services are better than the district average causes a signi icant increase in votes for candidates from incumbent parties for sub county councillor, but not for district chairs, district councillors, or sub county chairs. Against expectations, some evidence suggests that informing subjects that public services are worse than their expectations appears to cause an increase in votes for district chairs and councillors. In what follows, we develop our argument about the relationship between information on politician performance and voter behavior, describe the research design, present the results, and discuss their implications. 5

6 Information and Politician Performance In many developing countries including Uganda knowledge about the behavior of politicians and their performance in of ice is poor (Natamba et al. 2010, p. 16). This likely contributes to a low level of programmatic voting and to a preponderance of clientelistic and identity voting. If voters cannot observe the behavior of politicians, they are more likely to take shortcuts by voting for their co ethnics or in response to transactional remunerations, such as vote buying or employment (Besley and Burgess 2002; Keefer and Khemani 2005; Casey 2013). This dynamic undermines development by reducing incentives for politicians to deliver on programmatic promises and increasing incentives for politicians to amass rents from corruption (Banerjee and Pande 2007). However, in environments with greater transparency, voters should be more likely to reduce reliance on informational shortcuts, for example to the tribe or family of the councillor (Keefer and Vlaicu 2008; Golden and Min 2013) and instead take into account the programmatic performance of politicians when deciding how to vote. They should also be less sensitive to private inducements such as vote buying. Information should improve democratic accountability by empowering voters to hold governments accountable and make more informed choices during elections; it should also compel government of icials to perform more effectively (Persson and Tabellini 2000; Pande 2011). While often posited, these links have only recently been subjected to experimental tests. The compelling evidence that does exist suggests that information about performance and the provision of public goods affects voting behavior under certain 6

7 conditions. Ferraz and Finan (2008) ind in Brazil that public audits of local government of icials released before elections signi icantly reduced the probability of reelection for politicians who engage in above average corruption. Report cards on politician performance in India s slums induced higher turnout and higher vote shares for incumbents who were rated favorably on public spending (Banerjee et al. 2010). In Uganda, voters were signi icantly more likely to express support for members of parliament who performed above their peers in a comprehensive activity scorecard (Humphreys and Weinstein 2013). Information that is directly linked to voters own well being, such as information on the comparative quality of public services, is particularly important in driving the behavior of voters (Banerjee et al. 2010). Other studies ind similarly compelling evidence (Bjorkman and Svensson 2009; Gottlieb 2014). Yet we remain ignorant about the mechanisms underlying these indings. Transparency treatments are not ef icacious across all contexts and protocols. Several studies have found mixed or null effects of transparency treatments on political behavior (Olken 2007; Humphreys and Weinstein 2013; Lieberman, Posner and Tsai 2013). Humphreys and Weinstein (2013), for instance, found no impact of scorecards on voter choices in Uganda s national parliamentary elections or on parliamentarians responsiveness to increased information. Other studies suggest that transparency can have unintended effects such as inducing incumbents to hide their activity (Malesky, Schuler, and Tran 2012). As Ashworth and Bueno de Mesquita (2014) note, transparency affects both voter behavior and politician behavior, and untangling one treatment effect from the other can be dif icult. In corrupt or authoritarian environments, transparency might discourage 7

8 voters from turning out even if it decreases their support for corrupt incumbents (Chong et al. 2013). Our starting point is therefore a theory of rational voting in which citizens decide whether to vote and for whom to vote after evaluating the past behavior of candidates for of ice (Ferejohn 1986; Coate and Conlin 2004; Besley 2006). According to this theory, citizens hold prior beliefs about politicians. These priors vary in terms of their content (i.e., are the candidates of my preferred type?) as well as their strength (i.e., how certain am I that the candidates are or are not of my preferred type?). The content and strength of individuals priors are important factors that determine how receptive they will be to new information about politician performance. According to a model of rational voting, information is most likely to impact behavior when it is new and when it reduces uncertainty in voters evaluations of politicians. Information Effects on Voting According to our theory, citizens voting decisions depend on their assessments of politicians. Citizens seek to select candidates that best match their preferences, whether in terms of policy, patronage, or personal characteristics like trustworthiness. Unfortunately, and as noted above, voters often lack quality information about politicians. There are a number of reasons for this problem, including that the government is not transparent, the media is not free, politicians seek to conceal information about their performance, and voters do not seek out quality information. As a consequence of poor information, voters 8

9 may hold prior beliefs about politician performance that are incorrect and their priors about politicians may also be uncertain. If voters prior beliefs are incorrect or uncertain, then we would expect new information about politician performance to affect their choices when voting. This proposition is central to the idea that, in a democracy, the public holds elected of icials accountable for their performance in of ice. We propose two testable hypotheses. First, if people receive positive information about incumbent politicians relative performance as compared to their prior (i.e., good news ), then they will be more likely to support incumbents (H1a). Second, if people receive negative information about incumbent politicians relative performance as compared to their prior (i.e., bad news ), then they will be less likely to support incumbents (H1b). The effects of credible information will, however, depend on whether it provides new information or greater certainty about whether politicians are of the voter s preferred type. In some cases, information will simply con irm what voters already believed and little updating with regards to the content of individuals evaluations of candidates ought to occur. Thus, we expect information to have a larger effect on the choices of voters when it differs from individuals prior beliefs about politicians. To put this argument in the form of testable hypotheses, we propose irst that the effect of good news on voter support for incumbent politicians will be greater when the gap between voters prior beliefs about candidates and the information provided is larger (H2a). Moreover, the effect of bad news on voter support for incumbent politicians will be stronger when the gap between voters prior beliefs about candidates and the information provided is larger (H2b). 9

10 The effects of information also depend on whether information affects individuals certainty about whether politicians are of the voter s preferred type. As noted above, in some cases, information will con irm what voters already believed about politicians. Although we therefore do not expect new information to cause updating with regards to the content of individuals prior beliefs, it can update the certainty of individuals prior beliefs. A voter may, for example, suspect that an incumbent politician is performing poorly. Credible information that con irms that the politician is performing poorly may strengthen the voter s belief that the incumbent is performing poorly. Thus, we expect information to have a larger effect on the choice of voters when they are more uncertain. Again, to put this argument in testable terms, we propose two hypotheses. First, the effect of good news on the support of voters for incumbent politicians will be greater when the voter is more uncertain about candidate quality (H3a). Second, the effect of bad news on the support of voters for incumbent politicians is stronger when the voter is more uncertain about 3 candidate quality (H3b). Our inal hypotheses about the effect of information on vote choice re lect the idea that not all information affects equally voters evaluations of political candidates. New information about politician performance may lead voters to update their beliefs about a particular kind of information, without leading to changes in voting behavior if the information that is provided is not salient to the individual. Imagine that a voter who holds negative prior beliefs about an incumbent receives information that the incumbent is 3 In our pre analysis plan (PAP), these hypotheses are labeled H5a and H5b. For the purposes of this paper, we have renumbered some of our hypotheses to promote greater clarity. Other hypotheses from our PAP that are not discussed in this paper pertaining to beliefs about candidates, clientelism, the extent of common knowledge in an area, and election violence will be analyzed separately. 10

11 performing relatively well in terms of maintaining public parks. The voter might update her beliefs about the incumbent s performance in providing public parks in light of the new information without changing her likelihood of voting for the politician because she cares much more about his performance providing health services. Thus, we hypothesize that the effect of good news on voter support for incumbent politicians will be greater when the voter attaches greater importance to the type of information provided (H4a). Similarly, the negative effect of bad news on voter support for incumbent politicians will be stronger 4 when the voter attaches greater importance to the type of information provided (H4b). Effects on Turnout In addition to studying the effects of information on the choices of voters, we consider the effects of information about politician performance and public services on turnout. Some research suggests that people will become more likely to vote when they are better informed. The logic is that when people have more information about political candidates, they will be less likely to allow their fellow citizens, who may have less information, to select political leaders (e.g., Lassen 2005). The recent literature on the link between information and voting, however, suggests that individuals willingness to turn out to vote depends on whether the information they hold about the political system is positive or negative. In this paper, we focus on testing hypotheses drawn from this recent literature. Speci ically, previous research suggests that bad news about politicians causes people to lose con idence in the political system (Bowler and Karp 2004). If voters doubt 4 Hypotheses 6a and 6b in the PAP. 11

12 that voting will bring about good outcomes, perhaps because the system is inevitably corrupt or anti democratic, then they will be less likely to participate. In line with this argument, Chong et al. (2015) ind that new information about corruption de mobilizes voters over time and makes them less likely to identify with political parties. By the same logic, receiving good news about the political system may increase individuals sense of political ef icacy and partisan attachments, and thus their likelihood of voting. Consequently, we hypothesize that good news will increase voter turnout (H5a) and that 5 bad news will decrease voter turnout (H5b). Research Design This project evaluates the effects of two types of factual, nonpartisan information on vote choice and turnout, including how effects are moderated by priors, uncertainty, and importance. Our irst treatment arm evaluates the effect of information about the inancial accountability of district of icials on voter behavior. District governments in Uganda are required to submit to an annual audit by the Uganda Of ice of the Auditor General. Among other things, this audit tracks and validates that the district council followed procurement rules, completed projects as speci ied, and accounted for expenditures. We use these audits to create relative measures of fraud and mismanagement at the district level. We then provide factual information about how the percentage of unaccounted for funds in a voter s district compares with unaccounted for funds in other districts as our irst treatment arm. 5 Hypotheses 7a and 7b in the PAP. 12

13 Building upon previous foundational work like that of Ferraz and Finan (2008), we expect that such information will be salient to voter preferences and behavior. Here we also build on a key results from Humphreys and Weinstein (2013): they found strong, immediate treatment effects on voter support for politicians when the information was provided privately in a survey experiment, but weak evidence of an effect on support when the same information was given during public political processes in which politicians could spin the information to their favor. Text messaging via mobile phones in the week prior to a local election is a form of private communication. This makes it less likely that our results will be confounded by the strategic behavior of incumbents (Ashworth and Bueno de Mesquita 2014). Our second treatment arm evaluates the effects of information regarding public services on voter behavior. We expect that information on the comparative quality of public services will in luence voting because information is directly linked to voters own well being (Banerjee et al. 2010). We expect this to be particularly true for local politicians, from whom citizens expect quality provision of frontline public services (Pickering 2010). We conducted local audits of public services and presented voters via SMS with factual, nonpartisan information about how education, healthcare, water access, and road quality compare to average quality in each district (see Implementation Report for audit protocols). Our treatment protocol for this arm contains several innovations as compared to past research. First, we consider the effects of information on citizens votes for local of icials, which has not been the subject of most experiments to date. Arguably, we are more 13

14 likely to see an effect in such an environment. In local elections, voters may be better able to attribute the management of government and the quality of public services to the actions of local politicians (Putnam 1993; McNeil et al. 2009). They are also more likely to vote on issues of public services (IRCBP 2008; Casey 2015). However, we do note that multiple layers of local government are sometimes jointly responsible for providing local services, so, depending on how voters attribute credit or blame, voters may not always reward or punish speci ic local politicians based on new information. Second, we provide our treatment information via mobile phones, which makes treatments and measurement more timely and precise. Unlike more public distribution mechanisms, politicians are less likely to strategically change their behavior in response to information that is held privately and not subject to immediate public discussion and debate. This makes it easier to attribute the treatment effect to voters rather than politicians. The use of mobile phones also allows us to more easily evaluate effects on vote choice and turnout, since follow up with subjects is relatively inexpensive and instant. Finally, and perhaps most importantly, we provide each subject with information about the public service that they say is most important when deciding how to vote from the menu of services that we audited. One reason why previous information campaigns may have had a negative or null effect on vote choice is that the information provided may not be salient, such as whether an MP regularly attends parliament (Humphreys and Weinstein 2013). By allowing each respondent to receive information about the public service that is most salient to them, any updating that does occur is more likely to translate into action. 14

15 Setting Uganda is a semi democratic country that is prone to problems of corruption and poor governance. The country s relative representativeness of many African countries in terms of poverty levels, government type, and corruption makes it a useful site for the study. Though Uganda s regular elections are typically not judged free and fair by external monitors, elites appear to take elections seriously. Patronage and corruption extend down to local government levels, undermining good governance (Tripp 2010, p. 25; Greene 2010). In 1992, Uganda engaged in decentralization and began enacting legal reforms, establishing districts and sub counties as pillars of local government through which service delivery and local governance could be provided (Tumushabe et al. 2010). The Local Governments Act of 1997 speci ies that local governments are headed by an elected chairperson who is supported by an executive committee and by elected local government councilors. Though the law implies some shared responsibilities between districts and sub counties, among the top responsibilities of sub counties are water supply, sub county roads, and health clinics. Among the top responsibilities of districts are primary schools (Local Governments Act 1997, Part 4). Most citizens are likely to encounter corruption at the local level, given that important responsibilities and capacities sit with local governments. Following local practice, we will refer to district and sub county councils as LC Vs and LC IIIs respectively. Research has found that many Ugandan citizens evince relatively high levels of political awareness at the national level (Findley et al. 2014) but lack knowledge of the 15

16 basic functions of local governments, which has hindered citizens ability to participate in local government decision making and hold local of icials accountable (Natamba et al. 2010, p. 16). This lack of understanding likely contributes to what Uganda s Of ice of the Auditor General (2013, pp ) has described as poor service delivery. Many Ugandan citizens agree with that description, with less than half (46%) of Afrobarometer ( ) respondents in Uganda describing the government provision of services as good. One half of respondents in that survey reported having paid a bribe to obtain water or sanitation services, for example. We expect that bridging citizens gap in information on the quality of salient public services through private means should encourage citizens to hold local of icials accountable through their votes. Information dissemination by mobile phones in this study takes advantage of several useful aspects of mobile communication technology in Uganda. Mobile phones are nearly ubiquitous in the country. The International Telecommunications Union estimated that cell service covered nearly 100 percent of the Ugandan population as early as 2009 (see ITU 2009). A 2015 survey by the Pew Charitable Trust reported that mobile phone penetration in Uganda was 65 percent of the adult population, and all indications suggest that mobile phone ownership continues to grow rapidly (Pew Research Center 2015). Data for Treatments Just prior to the local district and sub county elections, we compiled information on budget irregularities and the quality of primary schools, as well as audited local public services. We gathered information on district level budget irregularities from the most 16

17 recent, publicly available, annual report produced by the Ugandan Auditor General on Local Authorities, 2013/2014. Data on primary school education was provided by Twaweza, a respected and independent East African initiative, from its 2013/2014 Uwezo study. The Uwezo initiative assesses basic literacy and numeracy levels and attendance rates of a nationally representative sample of children aged 5 16 years in districts across Uganda using a household survey that is carried out by volunteers. Because theory predicts information about the quality of that public service to in luence political participation more when this information is valuable, we used Afrobarometer data to identify other local public services most valued by Ugandans and developed independent audit protocols for these public services. From those priority problems, we then used data from ive focus groups in Kampala and Mukono to select the priority public services to audit: water, roads, and health clinics. Local enumerators conducted audits of these services all of which are under the primary or partial 6 competence of sub county of icials at each polling station using a standard protocol. Concurrent to the public services audits, a team recruited subjects for participation in this experiment after obtaining informed consent. Experimental Conditions and Randomization We provided several pieces of information about budget irregularities or public services via mobile phone SMS messages to randomly selected voters in each polling 6 See Repairing Information Underload: Implementation Report, Appendices I Q for protocols for auditing public services, and Section 4, Conducting Public Service Audits, for a discussion of challenges enumerators encountered while conducting audits, often due to the rainy season in and/or the remoteness of areas audited. 17

18 station in our experimental sample. Using a full factorial design, we provided subjects assigned to treatment with the following information within each arm: 1. Budget Accountability Treatment: The treatment provided details about irregularities in the district council (LC V) budget. Using the 2013/2014 audit of Local Authorities of the Uganda Auditor General, we developed the budget irregularity proportion by calculating the total sum in shillings of funds not accounted for, procurement issues, and payroll anomalies. This total sum was divided by the district council s budget so that we could compare inancial irregularities across districts. In addition to providing comparative information about the budget mismanagement of subjects district based upon this index in the SMS messages, we also provided two messages per district that illustrated one way in which their LC V performed better or worse than other LC Vs in in Salient Public Service Treatment: Information in this treatment arm focused on the quality of public services that were most important to individual subjects, as revealed in the pre treatment baseline survey. Treated respondents received factual information about the quality of the public service (primary education, water access, local roads, or health clinics) that they deemed most important when deciding how to vote. They were provided a series of messages about the overall score for that service and different components of the audit that they selected as part of the 18

19 7 baseline survey. For example, if they chose water access as most important, we provided information about the overall water access score and its components: walk time to water point, number of people waiting to draw water, and fees charged per jerrycan. The audit information for each locale was contrasted with district averages to enable a salient comparison point for subjects. To avoid instrumentation bias or priming effects, we provided subjects not assigned to treatment with placebo messages. Placebo respondents received only general information in the form of general public service announcements about the importance of quality public services or clean budgets without any information about the performance of their politicians. This design helps us to isolate the speci ic effect of new information, rather than the increased salience that subjects might attach to a certain kind of information when they are prompted to think about it. This approach also helped avoid unequal attrition between treatment and control conditions. We sent approximately 500,000 text messages to subjects over the course of this study. We employed experienced enumerators to translate messages into the 11 primary languages spoken by subjects and delivered each subject messages in their preferred language. Messages were sent using a bulk SMS delivery service developed by SMSOne, a communications technology company based in Kampala. Each day we sent two to four text messages to subjects over one week periods leading up to both the LC V and LC III 7 See Repairing Information Underload Implementation Report, Appendices W Z for the protocols for calculating treatment scores for public services and Appendix AA for treatment and placebo message templates. 19

20 elections. Public service treatment and placebo messages were sent out twice, before both the LC V and then then LC III elections. Similar to the Internet, mobile phone messaging is highly scaleable, possessing the ability to reach very large populations with lower unit costs as volume increases. Sampling Our research design was informed by collaboration with our local partner, the East African non governmental organization, Twaweza. We employed a nationally representative sample of districts (LC V) and polling stations using a hierarchical sampling procedure. We endeavored to use the sample of 28 districts selected for Twaweza s 2014 Uwezo project. The Uwezo sample of districts is nationally representative and strati ied by region of the country. However, challenges encountered in the ield meant that we were 8 only able to conduct research in 27 of the 28 Uwezo districts. Within each of these 27 districts, we identi ied approximately 30 polling stations that fell within the randomly selected enumeration areas in the 2014 Uwezo study. These enumeration areas were based on blocks used for the 2014 census. Within villages we varied the density of the budget treatment both to learn about the effects of more common knowledge and to test for positive spillover effects. For the 8 In addition to obtaining approval for our research through of icial Ugandan bodies (Mildmay REC, UNCST and the President s Of ice), we also sought research approvals from Resident District Commissioners (RDCs). Our local research teams met personally with and obtained approval from 26 RDCs in the 2014 Uwezo districts sampled. We had time to replace one district where we were unable to obtain RDC approval Namutumba with Kamuli district, from whose RDC we obtained approval. These districts share similar linguistic and demographic patterns The 27 districts in which we conducted research are Mpigi, Lyantonde, Butambala, Mityana, Buikwe, Nakasongola, Buyende, Iganga, Pallisa, Bulambuli, Sironko, Kumi, Kampala, Kaabong, Nakapiripirit, Gulu, Agago, Amolatar, Kiruhura, Bushenyi, Kisoro, Arua, Zombo, Hoima, Kasese, and Kyegegwa. 20

21 budget accountability treatment, within each district, polling stations were partitioned into those that had at least 15 participating subjects in the post recruitment, baseline survey and those that did not. Polling stations with at least 15 participating subjects were formed into paired blocks based on an ordered sorting of the number of participating subjects. Within each pair, we assigned one polling station to have 80% of subjects treated and the other polling station to have 20% of subjects treated. Within each polling station, individual subjects were then randomly assigned to treatment according to the polling station density assignment. We will report these results elsewhere, but it is worth noting here that we have no clear evidence that the density of treatment moderates treatment effects, so we leave aside this complexity for now. For the salient public services treatment, for all polling stations in the sample, half of subjects were randomly assigned to receive information about the public service that they individually deemed most important when they decide how to vote, as stated in their responses to pre treatment survey questions. The other subjects received a placebo message with general, public service style messages about budgets and public services. Surveys Local enumerators conducted baseline and endline phone surveys to evaluate the 9 effects of information on turnout and vote choice. We used these surveys to measure (1) prior beliefs about budget management and public goods provision; (2) pre treatment 9 For survey instruments, see appendix of Repairing Information Underload Pre Analysis Plan, registered February 17, Baseline and Endline survey questionnaires incorporated questions contained in Political Information and Electoral Choices: a Pre Meta Analysis Plan (Adida et al., 2015) which described a plan for a meta analysis of seven studies including this study on the impact of information on voting behavior in developing countries. 21

22 intentions to vote for incumbents and turnout in local elections; (3) post treatment votes for sub county and district council chairpersons and turnout, and (4) subjective evaluations of politicians, demographic characteristics, and other measures of political participation. Before assigning any of the subjects to treatment, we conducted a 25 question baseline survey in early January Of the 31,310 subjects recruited in the ield, we were able to re contact 16,083 for our baseline survey by call center. These subjects form our experimental sample, since we were able to con irm their use of the mobile phone that they provided at recruitment, which should increase rates of compliance. The baseline survey collected background information about the subjects prior political participation, intended participation at the upcoming elections, and demographic characteristics. In addition, we inquired about how subjects believed services in their areas compared to national and regional averages. We also probed subjects perceptions of the inancial accountability of politicians. This enables the assessment of how information affected people who receive data that differed from their priors. Other questions were designed to assess subjects access to reliable information about public services and inancial accountability, their political knowledge, and the salience of public services and inancial accountability to them. The pre treatment measures serve several purposes. First, stated intentions for the vote choice and turnout variables are used to transform the outcome measure, as outlined in the estimating equations below. This should increase power to detect treatment effects. Second, they enable us to divide the sample of subjects into those eligible for good news and bad news. Third, we used them to create subgroups for testing heterogeneous treatment effects. Fourth, we used them to assess balance across treated and control units 22

23 in our experimental conditions, and potentially to re weight our sample to estimate population average treatment effects in future work. We conducted two endline surveys, one immediately following the February district/lc V elections and another following the March 9 sub county/lc III elections. Of the 16,083 subjects in our experimental sample, we were able to re contact 12,972 at endline. Approximately 20% of the 12 question endline surveys were completed before the of icial LC V election results were announced and approximately 35% of the endline surveys were completed prior to the release of the of icial results of the LC III elections, with subjects contacted in random order at endline. In the future, we will use the timing of recontact to adjust for the effect of the release of results on self reported voting behavior. Of respondents who completed the LC V endline survey, only 26% indicated that they did not see the messages we sent to them. Many of these subjects reported that they had issues distinguishing our messages from election related messages sent from political parties and other organizations during the same period. Of respondents who completed the LC III endline survey, 17% indicated that they did not see the messages we sent to them. In the future, we will use these data to calculate complier average causal effects. To maximize institutional memory and retention of respondents, we retained 80% 11 of the enumerators throughout the recruitment, baseline, and endline phases. Enumerators worked to build trust with subjects by striving to interface with the same subjects throughout all phases of the study. Refusal rates were higher in Kampala, where 10 We conducted the sub county, LC III endline one week earlier in Kampala, where city government elections followed a different schedule than other sub counties. 11 To further encourage respondents to continue participation, we provided 1,000 Ugandan Shillings of airtime for each survey completed. We also had the same RA speaking with the respondent through each round of communication. 23

24 support for opposition parties is strong. Respondents in the capital were suspicious that our study was af iliated with the ruling party or meant to undermine the election, suspicions that our enumerators worked hard to dispel in communications with subjects. Analytical Methods Given that this experiment is based on an information treatment, the sample average treatment effect unconditional on the type of information is not relevant. Information can bring either good or bad news, which theoretically should drive behavioral and attitudinal outcomes in different directions. Additionally, the degree to which individuals perceive and respond to information as good or bad news depends on their prior beliefs about the content of the information that they receive. Thus, we split our sample into two subsets for the analysis of each information treatment arm, de ined by whether an individual subject is eligible to receive good or bad news given their priors. We de ine P i as each subject i s pre treatment belief about the quality of the incumbent candidate j, where the true value is Q j, measured on the same scale for a particular type of information, for example whether budget management is better than other districts. Additionally, we de ine Q as the median value of quality for the relevant comparison group of districts or sub counties. We then de ine the subset of subjects eligible for good news (L + ) as all subjects where P i > Q j or where P i = Q j and Q j Q. We de ine the subset of subjects eligible for bad news (L ) as all subject where P i < Q j or where P i = Q j and Q j < Q. Estimation is carried out separately for subsets de ined for each information arm. 24

25 To estimate the overall effect of good news or bad news (H1 on vote choice / H5 on turnout), averaging over the different strengths of good or bad news as de ined by distance from a subject s prior, we collapse all types of good news irrespective of distance from individual priors into a single treatment indicator N + that is positive when t i = 1 and subject i is part of the relevant subset L +. Likewise, we collapse all types of bad news irrespective of distance from individual priors into a single treatment indicator N that is positive when t i = 1 and subject i is part of the relevant subset L. We then use the following estimating equation, here with the notation for the L + group, where y ijh, post is a binary variable indicating voting for the incumbent party or turnout, y ijh, pre is a binary variable indicating pre treatment intention to vote for the incumbent party or turnout, j is a polling station ixed effect implemented via demeaning the outcome by polling station, and kh is the error term clustered at the relevant politician k within district h : y ijh, post y ijh, pre = + 1 N + + j + kh (1) To test for heterogeneous effects with respect to the prior beliefs of individual subjects about candidate quality (H2 on priors), we create a categorical variable Q j P i that indicates the number of categories different (e.g., much worse, a little worse, better, much better ) the actual treatment information diverged from subjects prior. We then use the following estimating equation, where t i is the treatment assignment of subject i, and j is a polling station ixed effect, and kh is the error term clustered at the relevant politician k within district h : 25

26 y ijh, post y ijh, pre = + 0 P i + 1 (Q j P i ) + 1 t i + 2,3,4 (t i (Q j P i )) + j + kh (2) To estimate heterogeneous treatment effects with respect to other covariates M i in each relevant subset (H3 on certainty / H4 on importance), we collapse all types of good news irrespective of distance from individual priors into a single treatment indicator N + that is positive when t i = 1 and subject i is part of the relevant subset L + and then interact the treatment indicator with the relevant covariate: y ijh, post y ijh, pre = + 1 N M i + 2 N + M i + j + kh (3) It should be noted that each of these estimating equations is only relevant for estimating treatment effects of the experimental treatments. Since we rescale the outcome behavior by pre treatment intention to perform the behavior, the coef icients on any observational covariates that are correlated with intentions will pick up that correlation. In the tables below, we report only the ATEs identi ied by. In the future, we plan to derive all standard errors on treatment effects from randomization inference based on these estimating equations, rather than model assumptions. We also plan to include observational covariates on demographics, trust in different sources of information, and past participation in elections in these estimating equations as pre registered, though we have not yet had time to clean those data. As an extension and because priors did not always in luence outcomes as expected, for 26

27 difference in means tests and model speci ications reported in the appendix, we also de ine the positive information subgroup I+ that contains all subjects eligible to receive positive information (Q j Q ) and the negative information subgroup I that contains all subjects eligible to receive negative information (Q j < Q ), irrespective of the priors of subjects. Findings We now present the evidence re lecting on of each of our hypotheses in turn. H1a posited that if people receive good news about incumbent politicians relative performance, then they will be more likely to vote for them. In contrast, H1b posited that if people receive bad news about incumbent politicians relative performance, then they will be less likely to vote for them. Recall that we provided two types of information: information about local budgets and information about local public services. Moreover, this information could have been relevant for how people evaluated incumbents at two types of elections: district (LC V) or sub county (LC III) elections. Finally, this information could have been relevant for how people evaluated two types of incumbents at both the LC V and LC III levels of government: chairs and councillors. As speci ied by the Metaketa Pre Analysis Plan, we irst report the mean incumbent vote shares across experimental conditions. We list vote proportions for two sets of subgroups: good news or bad news, which re lect the differences between treatment information and subjects stated beliefs about public services in the baseline survey (L+ / L subgroups) and positive or negative information unadjusted by prior beliefs (I+ / I subgroups). For example, in the positive information subgroup, subjects are eligible to 27

28 received messages informing them that budget irregularities or the salient public service was better or much better than the norm; in the negative information subgroup, subjects are eligible to receive messages reporting that budget discrepancies or the salient public service was a little worse or much worse than average. The good news subgroup includes subjects who thought budget discrepancies were much worse than average at baseline, but were eligible to receive information that they were only a little worse ; the bad news subgroup includes subjects who thought budget discrepancies were much better than average, but were eligible to receive information that they were only better. In all subgroups, treatment is the transmission of information via SMS text messaging. We report differences between treatment and control conditions and p values for 12 simple difference in means tests (Table 1). These tests suggest that both negative information (subgroup I, unadjusted by subjects priors) and bad news (subgroup L, re lecting subjects priors) about budget irregularities both lead to decreases in vote shares for incumbent district chairs and councillors. However, both positive information and good news about public services signi icantly increase votes for sub county councillor. All other effects are not signi icant statistically. 12 The difference in means tests are not strictly prescribed by the Metaketa Pre Analysis plan. However, reporting the mean proportions across all experimental conditions is required. The simple differences therefore provide the most basic test of the core Metaketa hypotheses that good news increases voter support and bad news decreases votes. 28

29 Table 1 Subgroup Of icial Treatment Control Difference p (one tailed) Budget Treatment: Proportion Voting for LC V Incumbent Good News Chair Good News Councillor Positive Info Chair Positive Info Councillor Bad News Chair Bad News Councillor Negative Info Chair Negative Info Councillor Public Services Treatment: Proportion Voting for LC V Incumbent Good News Chair Good News Councillor Positive Info Chair Positive Info Councillor Bad News Chair Bad News Councillor Negative Info Chair Negative Info Councillor Public Services Treatment: Proportion Voting for LC III Incumbent Good News Chair Good News Councillor Positive Info Chair Positive Info Councillor Bad News Chair Bad News Councillor Negative Info Chair Negative Info Councillor However, when we turn to the rest of the statistical tests speci ied in our pre analysis plan, the difference in means results attenuate. Here, we do not ind consistent 29