Together We Will: Evidence from a Field Experiment on Female Voter Turnout in Pakistan Xavier Gine and Ghazala Mansuri Abstract July 2010 (PRELIMINARY DRAFT) Women in many emerging democracies are less likely to participate in the electoral process as voters and more likely to vote according to the wishes of household males or clan heads when they do vote. In this paper we assess the impact of a nonpartisan door-to-door pre-election voter awareness campaign on female turnout and candidate choice in rural Pakistan. The campaign, conducted just before the 2008 national elections, provided information to women on the importance of voting and the secrecy of the ballot. We varied treatment intensity across geographic clusters in order to measure information externalities. We find that treated women are 12 percentage points more likely to vote than women in untreated clusters and are also significantly more likely to exercise independence in candidate choice. We also find large spillovers: untreated women in treated clusters behave as treated women. Further, using polling station turnout data, we show that treating 10 women resulted in 7 additional votes. Interestingly, the impact of the campaign on turnout was highest in more contested polling stations where fears of voter intimidation were substantial. Gine: Development Research Group, The World Bank (e-mail:xgine@worldbank.org). Mansuri: Development Research Group, The World Bank (e-mail: gmansuri@worldbank.org). We thank Jishnu Das, Stuti Khemani and Berk Ozler for valuable discussions and advice. We are especially grateful to the following for their help and support in organizing the experiment: Shahnaz Kapadia, at ECI Islamabad, for her help with designing and pilot testing the information campaign; Irfan Ahmad at RCONs, Lahore, for managing all field operations and data collection; Sughra and Ashiq Solangi of The Marvi Rural Development Organization, Sukkur, for their tremendous support to the team throughout; Qazi Azmat Isa at the World Bank Office in Islamabad and Kamran Akbar at the Pakistan Poverty Alleviation Program (PPAF) in Islamabad for their support and encouragement. This project was jointly funded by the Development Research Group and the PPAF. Paloma Acevedo de Alameda, Paola de Baldomero Zazo and Santhosh Srinivasan provided outstanding research assistance. The views expressed herein are those of the authors and should not be attributed to the World Bank, its executive directors, or the countries they represent.
Introduction A basic premise of representative democracy is that those who are subject to policy should have a voice in its making. In recognition of this, suffrage was extended to women in most western democracies in the early years of the 20 th century. The new democracies that emerged after the Second World War followed suit and extended de jure rights to women to participate in all democratic institutions. 1 However, in the early years of the 21 st century, women are still far less likely than men to stand for public office, even in the older democracies of the West. Things are much worse in most developing countries. Not only are women still greatly underrepresented in public office, they are also much less likely to participate in the electoral process as voters or to exercise independence in candidate choice when they do vote. Instead, women report voting in accordance with the preferences of the caste, clan or household head in contrast to men of all ages. While women s relative absence from elected public office has received considerable policy attention in recent years 2, there have been no systematic attempts to reduce barriers to women s participation as voters, and even less attention has been paid to the use of women as passive vote banks, when they do participate. If preferences over the allocation of public resources vary significantly by gender, this neglect could have implications for public policy, in addition to any equity related concerns. 3 1 These rights were also brought into international law by several important agreements to which most countries are signatories. These include the Universal Declaration of Human Rights (1948), the Convention on the Political Rights of Women (1952); the International Covenant on Civil and Political Rights (1966); and the Convention on the Elimination of All Forms of Discrimination against Women (1979). 2 A number of countries have passed legislation requiring fixed quotas for women. In South Asia, for example, India, Pakistan and Bangladesh have all instituted quotas for women in both local and national assemblies. 3 Chattopadhyay & Duflo (2004) for example exploit the quota introduced for women in the Indian Gram Panchayats and find that elected women leaders are more likely to provide public goods preferred by women. Lott & Kenny (1999) find that women s suffrage in the US increased overall government revenues and expenditures and has led to more liberal voting patterns. Edlund and Pande (2002) find that the decline in marriage has contributed to the shift by women voters in the US towards the democratic party. We also know, from several studies of intrahousehold resource allocation, that women tend to make different, and welfare enhancing, choices over the allocation of household budgets. For example several studies have shown that women invest more in the health and education of children. 1
Women in emerging democracies could face two distinct barriers in exercising their right to vote for a candidate of their choice. On the one hand, costs of participation may simply be too high. Traditions or cultural stereotypes may discourage the exercise of own preferences, mobility constraints may limit participation and, if there are expectations of voter intimidation or violence, personal security concerns may also have a larger dampening effect on female participation 4. On the other hand, women may simply have fewer and/or poorer sources of information about the significance of political participation or the balloting process, perhaps due in part to illiteracy and limited mobility. Lack of information may also reinforce stereotypes that further disengage women from public life. Voter education and awareness campaigns, common in older democracies, are increasingly being used as an important instrument for promoting greater citizen engagement in the democratic process. The goal of this paper is to assess the role of information and social interactions in enhancing turnout and independence of candidate choice among women. Our experiment was conducted in rural Pakistan which presents a particularly interesting context for two reasons. First, there have been significant legislative reforms aimed at enhancing women s participation in public life, and second, women still face significant cultural barriers to effective participation, including restrictions on their mobility and clan based voting 5. Political parties also view them largely as passive vote banks, following the dictates of male family and clan heads. The argument for providing information in this context is simple. While attitudes and social mores change slowly, improvements in access to information can occur relatively quickly. If lack of awareness limits participation, then access to information could enhance both equity and allocative efficiency as women select candidates that best reflect their preferences. At the same time, the mere act of participation may serve to weaken pejorative perceptions about female efficacy that limit women s engagement in public life. 6 4 5 6 See, in particular, Beaman et al, 2007. 2
The experiment was conducted just before the 2008 national elections. Each study village was divided into geographical clusters, essentially contiguous streets. A subset of non contiguous clusters were then randomly selected for treatment while others were left as controls. Within treated clusters, a subset of sample households were randomly assigned to receive the awareness campaign. This allows us to assess the size and significance of information spillovers within treated clusters without confronting the usual set of identification problems (Manski, 1993 and Manski, 1995). The campaign was developed as a set of simple visual aids with two different messages. The importance of voting (T1) which focused on the relationship between the electoral process and policy, and the significance of secret balloting (T2) which focused on the actual balloting process. Treated women received either T1 or T1 plus T2, allowing us to test whether knowledge of secret balloting is important for female participation. The campaign was delivered door-to-door, and was only attended by the women in each household. No men were present during the sessions. The information campaign never mentioned a political party or candidate by name. While pre-election voter motivation and persuasion campaigns (Get Out the Vote) are commonplace in the US and other developed democracies and have been extensively studied, 7 voter information campaigns are arguably more important in newer democracies where voters often have poorer access to relevant information on the political process and where institutionalized party structures which disseminate such information are also less well developed, as are partisan preferences. One could argue, therefore that prima facie our campaign would have a larger effect on turnout, as compared to studies from the US and this does seem to be the case. We find that turnout increases by about 12% for women in treated households when compared to women in control clusters, with larger effects for women exposed to T2. In contrast, GOTV experiments in the US find that non-partisan canvassing boosts turnout rates by between 7 and 9 percent. 7 More than a 100 Get Out the Vote (GOTV) field experiments have now been done in the US alone. 3
To our knowledge, this study is among the first to use a field experiment to assess the impact of a voter information campaign in a developing country and it is the only study which can cleanly assess the impact of information externalities on turnout and candidate choice. 8 In this respect, our study also contributes significantly to the nascent but growing literature on information spillovers. In this respect, our work relates most closely with Duflo & Saez (2003). The peer group in their study is a university department. In their experiment, a random subset of employees from some departments of a large university were encouraged to attend a benefits information fair about a tax deferred retirement plan offered by the university. The encouragement included a small monetary reward for attendance. They find evidence of significant spillover effects in both meeting attendance and subsequent enrollment in the advertised retirement plan. Like Duflo and Saez (2003), the peer group in our case is fixed by location and a subset of the peer group in a treatment cluster is treated. In contrast, for example, other studies look at peer effects by randomly assigning individuals to peer groups (see, for example, Katz et al. (2001) and Sacerdote (2001)). Overall, we find larger spillover effects than Duflo and Saez: control women in treated clusters behave as if they have been directly treated since their turnout rate is comparable to that of directly treated women in the same clusters. Thus, within treated clusters, there is no difference between treated women and their untreated close neighbors. We go beyond Duflo and Saez by collecting information on the closest friend/confidant of each sample woman and subsequently verifying their decision to vote. We therefore assess whether peer effects are larger for close friends than for close neighbors. We find comparable results, which may not be surprising given that most close friends reside in the same cluster as the study women. Using the GPS location of study households, we estimate information spillovers in two additional ways. First, we compare treated women to control women but drop control women located within a given distance of any treated household. Second we estimate the specification 8 There is a small experimental literature that has focused on electoral violence, clientelism and vote buying in the context of developing countries. Collier and Vicente (2007), for example, study the effect of an information campaign against electoral violence in Nigeria. Wantchekon (2003) has examined the effectiveness of clientelist messages in Benin and Vicente (2007) analyzes an information campaign against vote buying in Sao Tome and Principe. 4
used by Kremer and Miguel (2001) which relies on exogenous variation in the local density of treated households generated by the cluster level randomization to assess how the number of treated women within a given radius of each household affects turnout. Both strategies yield even larger geographical spillover effects. In the context of awareness campaigns in the US, Gerber et al (2008) find substantially higher turnout amongst individuals who had received postcards promising to publicize their turnout records to other members of their household or neighborhood. 9 Nickerson (2008) conducts a door-to-door canvassing experiment targeted at households with two registered voters and finds that the member that did not answer the door is nearly 60% as likely to vote as the treated member. 10 Finally, we use administrative data from all polling stations in the study and find that for every 10 women, or 4 households, treated, female turnout increases by 7 additional votes. This translates into a cost of $3.4 or Rs 231 per additional vote. If voting were habit forming (Gerber, Green and Shachar, 2003; Plutzer, 2002), then this cost could be seen as a lower bound as the campaign could have a sustained impact on women s political engagement. We also assess secrecy of candidate choice by examining cross reports of the male head and other women in the household regarding whom a given woman voted for. We find that male heads in treated households are significantly less likely to predict a woman s candidate choice. Also, consistently with our earlier results, male heads in control households in treated clusters correctly predict a woman s candidate choice about as often as do male heads in directly treated households. Finally, we study the impact of contestation at the polling both level on turnout using the vote share of candidates. Electoral competition is generally expected to increase voter turnout because of voter perceptions about the value of the marginal vote or party and candidate 9 The field experiment was conducted just before the August 2006 primary election in Michigan. The authors find that turnout increased by about 8 % in households which were shown their voting records as well as the voting records of their neighbors. This is about as large as the impact of direct personal canvassing in getting the vote out. 10 However, Nickerson interjects a note of caution by stating that results from two member households may not be applicable to broader social settings or even to households in other settings. 5
efforts to increase mobilization. 11 However, much of this empirical literature comes from the United States and other established democracies where there are few concerns about voter intimidation or violence The few studies that do assess the impact of electoral competition on turnout in a politically volatile environment find that turnout is significantly lower when there is more competition. Collier and Vicente (2007), for example, find that voter intimidation reduced turnout in the 2008 Nigerian elections by at least 9 percentage points. Bratton (2008) uses Afro- Barometer data along with electoral returns also for Nigerian elections and shows that people who were threatened during election campaigns were significantly less likely to vote. Consistent with these findings, we also find that turnout was significantly depressed in more contested polling stations due to security concerns but this effect is overturned for women in treated clusters, irrespective of whether or not they were directly treated, once again confirming the significance of social interaction effects. 12 While competing candidates obtained similar shares in some polling stations, at the constituency level, there was no close election The remainder of the paper is organized as follows. Section 2 describes the context, the 2008 election, the design of our experiment and the data. Section 3 discusses the impact of the information campaign on turnout and assesses the size and significance of information spillovers. Section 4 discusses the evidence on independence of candidate choice. Section 5 examines the interaction between electoral competition and information provision. Section 6 provides a cost-benefit analysis of the intervention and discusses some broader impacts from a follow-up survey. Section 7 concludes. 2. Context and Experiment Design 2.1 Context The experiment was carried out in collaboration with the Marvi Rural Development Organization (MRDO), Research Consultants (R-Cons), ECI and the World Bank. MRDO is a non- 11 See, in particular, Franklin (2004). 12 Vicente (2007) also finds that a randomized anti-vote buying information campaign conducted during the 2006 presidential elections in Sao Tome and Principe dampened the effect of vote-buying on candidate choice. 6
partisan NGO that mobilizes women using a community based approach. R-Cons is the survey firm that helped MRDO in implementing the awareness campaign and collected the baseline and follow-up data. MRDO also helped us get post election voting data for all sample polling stations. ECI is a local training firm that collaborated in the design of the campaign. ECI has developed many training materials for the Pakistani government, NGOs and private firms, including visual aids and pamphlets used to train local election officers prior to elections. The campaign was carried out in the districts of Sukkur and Khairpur in Sindh, selected because of sharp electoral competition between the PPP, which has a secular-left leaning platform and the Pakistan Muslim League Functional (PMLF), which was allied with the military regime. PMLF candidates in the study area, in addition to being large landowners, also lay claim to being the spiritual leaders or pirs of their constituents. The initial sample included 12 villages, 6 villages from each district, and 24 polling stations. Villages were chosen to ensure variation in the expected political competition at the polling station level and because MRDO was either active there or about to start operations. However, given the context of the 2008 elections particularly in Sindh, 3 villages (3 polling stations) that were a PMLF stronghold had to be dropped because the safety of the canvassing teams could not be guaranteed. The polling stations in these 3 villages were relatively more contested than those in our final sample of 9 villages and 21 polling stations. Indeed, the 2008 national elections were held in an environment that was politically charged. After seven years of military rule and due to widespread opposition, the government had declared emergency rule, the sitting judges of the supreme court had been dismissed and there were fears that the incumbent government would engage in massive rigging. Scheduled initially for January 8 th, 2008, the elections were postponed to 18 th February 2008 because of the assassination of Benazir Bhutto on December 27 th, 2007, the leader of the Pakistan People s Party (PPP) who had been elected twice as Prime Minister in the past. In Sindh, traditionally a PPP stronghold, concerns about electoral rigging and voter intimidation by the incumbent military government increased after Bhutto s assassination, but a PPP 7
landslide was also being anticipated due to a strong sympathy vote for Bhutto. The net effect of these two tendencies on turnout, particularly for women, was uncertain. The information campaign was designed as a set of simple visual aids accompanied by a well rehearsed and limited text and included two nonpartisan messages (see appendix B). The first focused on the importance of voting the relationship between the electoral process and policy, including village development outcomes (T1), the second focused on the actual balloting process (the structure of a typical voting station and booth, the fact male and female booths the secrecy of the ballot and the basic appearance of the ballot paper (T2). We decided to implement the campaign via door-to-door visits for two reasons. First, household visits provide a high degree of control over which households receive the campaign and which do not, which is critical to measure information externalities. Second, we followed studies from the US and other developed countries that have found strong support for door-todoor information campaigns over other strategies like phone calls and direct mailings. 13 It is important to distinguish our awareness campaign designed to inform women about their rights in the electoral process and how to exercise them from a GOTV campaign whose objective is to increase turnout by persuading subjects to vote. GOTV may be partisan or not, although a number of experimental studies have shown that partisan messages are less successful in motivating turnout. 14 13 Gerber and Green (2000a) reports on a randomized GOTV campaign conducted in New Haven, Connecticut, just prior to the 1998 election. The campaign delivered non-partisan messages through personal canvassing, direct mailings and telephone calls. The study found that personal canvassing had a substantially greater impact on voter turnout as compared with other modes of contact. Green et al., (2003) and Michelson (2003) find similar results. Michelson also finds that direct canvassing has a particularly strong impact on voter turnout when the canvasser and the targeted voter hail from a similar ethnic background. In her study Latino activists were especially effective in mobilizing the Latino vote. 14 Cardy (2005), for example, finds that neither partisan direct mail nor partisan phone calls - used independently or together - managed to garner a significant voter response. In a similar vein, Gerber and Green (2000b) find that non-partisan messages are particularly effective in mobilizing unaffiliated past voters. The authors hypothesize that partisan voters may already receive adequate encouragement from their respective political parties while unaffiliated voters do not. Moreover, they speculate that politically unattached voters may also have been impressed by the non-partisan appeal to civic responsibility. Horiuchi et al. (2005) also find that voters are less likely to abstain when they receive policy information about both ruling and opposition parties through their official party websites. Indeed, viewing party manifestos online increased voter turnout by more than 4% on average. The information effects are larger among those voters who were planning to vote, but were undecided about which party to vote for. Desposato (2007) assesses the impact of campaign tone on turnout and vote choice 8
Our argument for information provision is based on the hypothesis that women may have both fewer and poorer sources of information than men. This may be in part due to low levels of literacy and mobility or due to cultural values which disengage women from public life. Our survey indicates that female literacy rates are indeed very low in our study areas (less than 20% of adult women have any schooling), as they are all over rural Sindh. Women also have rather limited mobility even within their own villages. In the survey we defined mobility using a 0 to 3 scale, where women with 0 mobility cannot travel alone within their own settlement while women with a mobility score of 3 can travel outside the village by themselves. Most women can travel within the village on their own or accompanied by other females but not outside the village, where the presence of a male is required. Women are also far less likely to listen to either local, national or international news channels (see Appendix table A3) and are far less informed about any political issue (see Appendix Table A4) including major events such as the imposition of emergency rule in the country only 6 percent of women knew compared to 82 percent among men. Interestingly this is not due to differential access to popular media like radio and TV. Instead it appears that men and women use media resources very differently. 15 Women are also far less engaged with any aspect of village public life. They are far less likely, for example, to attend community meetings related to village development, attend demonstrations or contact their local councilor or local party official for any matter. Interestingly, though, when they do engage, women tend to avoid formal authority and reach out to traditional or religious leaders (See Appendix table A5). These findings are also consistent with Pakistan s rather dismal performance on a range of development indices. According to the 1998 Human Development Report, for example, Pakistan ranked 138 out of 174 on the Human Development Index (HDI), 131 out of 163 on the Gender Development Index (GDI), and 100 out of 102 on the Gender Empowerment Measure(GEM). using a lab experiment conducted in Brazil and finds substantially larger effects of negative advertising in Mexico on voter participation, candidate support and regime support than what studies have reported for the United States. 15 Our survey result indicates that both radio and TV are widely available and there is no gender difference in access. Approximately, 50% of men and women had access to a radio and about 70% had access to a TV. 9
2.2 Experiment Design and Survey Data The timeline of the study is shown in Figure 1. The information campaign was carried out two weeks prior to the elections (from February 5 th -15 th 2008) by 8 teams consisting of one MRDO staff and one female enumerator from RCON each. Each sample village was covered in approximately one day and was first divided into contiguous geographical clusters of approximately 40 households. A cluster was typically one or two contiguous streets in the village (see appendix C). Each cluster was then randomly assigned to get T1, T1 plus T2 or nothing as follows: the canvassing team selected one cluster in each village at random and began there. T1 was delivered in this cluster. Next, a gap cluster was left between two selected clusters. In the second selected cluster a coin toss determined whether T1 plus T2 was delivered or all selected households were left as controls. The third selected cluster was then given the opposite treatment of the second cluster. A typical sample village had about 7 sample clusters and 11 clusters in all, including gap clusters. Within each selected cluster, irrespective of the specific treatment, households were selected in the same way. Starting on either end of the cluster, every 4 th household was selected and surveyed. In T1 and T1 plus T2 clusters, every 5 th selected household was also left as a control. This generated 2 to 4 controls in each T1 and T1 plus T2 clusters in addition to the control households selected in the control clusters. Treatment was therefore randomized at the household level, and in addition to the variation in treatment type (T1 or T1 plus T2), we also randomly varied treatment intensity in order to measure peer effects. During the visit, some basic data was also collected on each sample household, including household GPS location; a basic roster of all adult women, with their past voting record and the name and address of their closest friend or confidant in the village. None of the households that were visited refused to be interviewed or given the awareness session. A local woman, usually a school teacher, was also identified in each village during the awareness campaign. Her task was to assist the canvassing team with the verification of voting post-election by checking the ink stain on each woman s hand. 16 This woman was provided the 16 Harbaugh (1996) observes that the importance of social conformity often leads to widespread over-reporting of voting in surveys. He estimates that approximately 10% of eligible voters in presidential elections will lie to survey takers that they voted in the election, even though they actually did not. Further, the frequency of lying, 10
list of women whose thumb mark needed to be verified the day before the election. This list included one confidant from each household who was selected randomly from amongst the woman who were eligible to vote in the household. Elections took place on February 18 th. Voting verification was done starting February 18 th evening and continued through the 19 th. The canvassing teams visited each village on the 19 th to check the validity of the verifier s assignment. They did a 10% random check. During the initial visit, we reached 2735 women in 1019 households and 64 clusters, roughly two-thirds of which were treated (given either T1 or T2). During the verification exercise, we were unable to locate 98 women (and 27 households) because of temporary or permanent migration. This left us with a sample of 2637 women and 992 households in our 64 clusters. We were also able to reach 727 confidants. This number is smaller than 992 because not all households yielded at least one eligible woman. Attrition was unrelated to treatment assignment. In addition, 135 women claimed to have cast a vote but they did not have the requisite Ink mark. We err on the safe side by treating these women as not having voted. Verification was followed by a post-election survey, which took place in March 2008. This survey collected information on household demographics, including the household s zaat/caste affiliation; verification of visits by canvassing teams to enable an intervention check; information on mobility constraints; access to and use of various media; knowledge regarding the location of polling stations and the balloting process; the election day environment; knowledge of candidates, party platforms, recent political events, and election outcomes; knowledge of whether other household members voted and for whom. Finally, we collected official data from each of our sample polling stations which provided the breakdown of votes cast by gender, candidate and political party. conditional on not voting increases with education. Similarly, in a study that analyzed voter validation studies in 1964, 1976 and 1980, Silver, Anderson and Abramson find that the more educated have a greater propensity to lie about having voted (Silver et al., 1986). These studies attribute the conditional increase in lying to the existence of a social norm in favor of voting, a norm which increases with education; having decided not to vote, more educated voters are less likely to admit to having broken that norm (Silver et al., 1986; Harbaugh, 1996). Moreover, social psychological literature also contends that individuals often modify their behavior depending on whether people perceive their actions to be in the public good (Lerner and Tetlock, 1999; Cialdini and Goldstein, 2004). 11
Table 2 provides differences by treatment status for a number of household and woman characteristics. By and large, there is little difference in household characteristics by treatment status. Treatment households have a little more land than control households in some comparisons, but no difference in assets or housing quality. Women in treated households are a little younger in some comparisons and have more young kids as a result and also appear to have less access to cable TV, perhaps due to their lower mobility. In the analysis, we control for the household and woman characteristics that we lack balance on as well as the total number of women registered to vote in a polling station. We also control for whether the woman had a national id card (NIC), which is needed to cast a ballot. This is sensible since, in practice, some women had cast a vote without an NIC and were verified as having done so. Also, younger women are also less likely to have an NIC or to have voted in the past. These constitute our full set of woman, household and polling station level controls. Table 1 provides further summary statistics about the characteristics of households, polling stations and women in our sample. There are approximately 3 women per household. Of these, about 70 percent have national identity cards (NIC) required to be on the voter list or to cast a ballot. The intervention also appears to have been implemented well overall. Only treatment households report an information visit prior to the elections and far more women in treatment clusters report talking to a neighbor (Appendix A: Table A2). 3. Empirical Strategy and Results 3.1 Impact of Information on Turnout and Information Spillovers Because the treatment is assigned randomly at the geographical cluster level, its impact on the various outcomes of interest (say, whether the women voted or not) can be estimated via the following regression equation: Y ihnpv = βt hnpv + γx + (ε v + ε ihnpv ) (1) where Y ihnpv indicates whether woman i in household h in cluster n in polling station p in village v voted (1=Yes) based on verification, T hnpv is the treatment indicator (1 if women in 12
household h in cluster n in polling station p in village v were given the voting awareness campaign), and X is a vector of polling station, household and individual woman characteristics collected at baseline prior to the campaign. The first term in the composite error, ε v + ε ihnpv, includes unobserved village characteristics that could affect the impact of treatment on turnout. The second is a mean-zero error term. β, which captures the impact of treatment on turnout is the main coefficient of interest. All reported specification include village fixed effects to remove the impact of village specific unobservables. Since treatment assignment is at the geographical cluster level, we also cluster standard errors at this level to deal with potential spatial correlation among women living in the same cluster (Moulton 1986). Inclusion of the vector X of baseline characteristics, discussed above, increases efficiency by absorbing residual variation. In order to capture the importance of within cluster spillovers, we run Equation (1) under four different assumptions about control women living in treated clusters. First, we simply compare treated women to all control women, regardless of where control women live (comparison T-C) ignoring spillovers altogether. Second, we compare treated clusters with control clusters (comparison T N -C N ). Third, we compare treated women in treated clusters with women in control clusters (comparison T-C N ). Finally, we compare control women in treated clusters to women in control clusters, that is, we drop treated women from the analysis (comparison C TN - C N ). If spillover effects are important, we would expect the coefficient of interest in the first naïve comparison to be smaller since we are treating as controls a set of women that are likely to have been influenced by treated women. Table 3 reports the results. The naïve estimate (T-C) in column 1 indicates that treatment increased turnout by about 6%. There is a substantial difference also between the T1 and T2, however, with the latter having almost 3 times the effect as the former. However, we are unable to detect statistically different effects of T1 and T2 due to the relatively small number of geographic clusters in our sample. Column 2, 3 and 4 account for within cluster spillovers by comparing women in control clusters to the treated group using three different samples. In column 2, the treated group is the treatment cluster (T N -C N ), in column 3, it is treated women 13
(T- C N ) and in column 4 it is control women in treated clusters. The size of the coefficient almost doubles for all three comparisons but it achieves significance at the 10% level only in columns 3 and 4. The key issue is that accounting for within cluster spillovers increases impact and it appears that control women in treated clusters respond to treatment assignment about as much as the women directly treated. While this strategy allows us to assess spillovers within the treatment cluster effectively, it does not allow an assessment of spillovers beyond the treatment cluster. The worry is that information externalities could benefit the comparison group beyond the treated cluster. If so, turnout differences between treated and comparison clusters, even accounting for within cluster spillovers, will understate the benefits of treatment on the treated. We use two strategies to assess potential geographical spillover effects beyond the cluster. First, we compare treated women to all control women but drop control women that reside within a given distance of any treated household. Next we use a strategy inspired by Kremer and Miguel (2001). We rely on exogenous variation in the local density of treated households, generated by the cluster level randomization, to assess how treatment density within a given radius of each woman affects turnout. 17 The regression specification in this case is Y ihnpv = βt hnpv + dd (α dd NT dd + τ dd N dd ) + γx + (ε v + ε ihnpv ) (2) where NT dd is the number of treated households between distance d and D from each sample household, and N dd is the number of households between distance d and D from each sample household. The rest of the terms are as defined in Equation (1). Standard errors are clustered at the geographical cluster level, as before. The results using the first strategy are presented in table 4. In columns 1,2 and 3 we successively drop control women residing within 50, 100 and 200 meters of the any treated 17 Kremer and Miguel (2001) assess cross-school externalities using exogenous variation in the local density of treatment school pupils generated by the school-level randomization. 14
household, while controlling for overall population density within each radius. Now we get robust treatment effects which range from 13 to 16 percent and are again considerably larger for T2. Once again, however, we cannot detect a significant difference between T1 and T2 given standard errors. Table 5 presents the results from the 2 nd strategy. Specifically, we construct non-overlapping concentric rings of 200 meters around each sample woman. In each ring or band we compute the total number of treated women within the band. Since the median distance between any two women in the village is about 1.2 kilometers, the bands start at 0-200 meters and extend up to 1200 meters. The results indicate strong spatial spillovers up to a distance of 400-600 meters. In this range, having more treated women generates strong positive effects on the likelihood of a woman voting. An increase in the number of treated women increases the odds of a woman voting by 12 to 16 percentage points, with coefficients that are highly significant. Beyond this distance, however, the effects taper away very quickly. Next we explore whether there is evidence of larger peer effects for the close friends of treated women. We do not have an exact measure of distance for the households these women reside in, but we do know which cluster they live in and whether they cast a vote. Not surprisingly, the vast majority reside in same cluster as the woman who identified them as a close confidant. Table 6 presents the results. Interestingly, we find no additional treatment impact from friendship. However, the impact of information on turnout for friends is in the same range as that for other control women in treated clusters, ranging from 10 to 12 percent. Clearly, social interactions among women are dictated largely by geographic proximity. Talk is easy and apparently plentiful when it can happen over laundry, cooking and childcare and preferably requires little movement away from home. Finally, we assess the size of information spillovers at the polling station level, using official data on votes cast by gender. The outcome of interest in this case is, Y pv, the number of votes cast by women in polling station p in village v. The impact of treatment is measured by the number of women treated in polling station p in village v, NT pv, controlling for a vector of polling station 15
level variables, including the number of registered women, X pv. This yields the regression specification Y pv = βnt pv + γx pv + ε pv (3) Table 7 present the results. A marginal increase in the number of treated women increases turnout in the polling station by between.64 and.73 percent, depending on whether we control for polling station characteristics. Thus for every 10 women or 4 households treated, 7 additional votes are cast. 3.2 Impact of Information on Candidate Choice We examine the impact of the voter awareness campaign on candidate choice by using the cross reports of family members. During the follow-up survey we asked every woman in the household as well as the male head to report their choice of political party as well as that for all members interviewed in the household, who had also cast a vote. We then construct a measure of agreement that equals 1 if the reporter s assessment of the reportee s choice is correct and zero otherwise. To do this we use only the sample of women who voted and retain only reports made by family members, including the male head, about each woman in the household. For each sample woman, then we have a set of observations, each of which is an indicator of agreement between the woman own choice and the report of each family member about her choice. The number of observations per woman thus vary by household size. This final sample includes 2825 cross reports. If the PEVAC successfully conveyed information on the secrecy of the ballot and the significance of voting one s conscience, then we would expect fewer matches in treated households, particularly in the match between the male head s report about women in the household and the women s own reports. Put differently, in treated households we would expect women, and specially the head of the household, to make more mistakes about who the other members voted for. We test this with the following regression M ij hnpv = βt hnpv + τh ihnpv + δ(t hnpv H ihnpv ) + γx + (ε v + ε ihnpv ) (4) 16
where M ij hnpv is an indicator that takes the value 1 if the report of individual i on individual j s choice of candidate is correct (according to j s self-report); H ihnpv is an indicator that takes the value 1 if reporter i is the male head; and X is reporter i s vector of polling station, household and individual characteristics. Since this is a dyadic regression, the error term is likely to be correlated across all observations with the same reporter i and reportee j. We cluster at the geographical cluster level, which is a more conservative approach. The results are presented in table 9. The coefficient of interest is δ which captures the differential effect of treatment on the quality on male reports about the candidate choice of women in the household. The results indicate that treatment reduces male knowledge about women s chosen candidates by about 8 percentage points. Interestingly, however, in this case the effects are larger for directly treated women, signaling perhaps that the more nuanced message about the secrecy of the ballot was not as easily shared as the more direct message regarding the importance of participation. 3.3 Electoral Competition and Information Our sample design allows us to examine the extent to which response to the information campaign was conditioned by the underlying political environment. Our major concern going in was whether perceived levels of voter intimidation or expectations of violence could dampen the impact of the information campaign in more contested polling stations. Of course the reverse is also possible, where information provision, particularly about the voting process and the secrecy of the ballot serves to alleviate fears and boosts turnout. The post intervention follow up survey does indicate higher levels of concern about safety and incidents of violence in the more contested polling stations (see Appendix Table A9). It is also clear from the official data that turnout was significantly lower in the more contested polling stations. There is essentially a 20 percentage point dip in turnout in polling stations which are above the median level of contestation (see Appendix Table A7). Using sample data, on the other hand, 17
we find a dip in turnout of about 10 percent. This already suggests that treatment may have served to alleviate fears of voter intimidation. We use official data on votes cast for each candidate/political party to construct two measures of electoral competition. The first is a measure of concentration, the Herfindahl index of vote shares, which ranges from.02 to.72 with the median at.43. The second is simply the percentage of votes obtained by PMLF. Our final sample includes 21 polling stations with very different levels of local electoral contestation. While these measures of electoral competition are arguably orthogonal to our treatment since they are at the polling station level, we nonetheless check whether any household, village or polling station level variables we have, including the number of women treated in the polling station and the total number of women registered to vote in the polling station influences our measures of contestation. The results are presented in Appendix Table A6. The number of treated women has no impact on electoral competition. Further, only two polling station level variables matter. Polling stations that have more wealth dispersion and a more dispersed population, tend to be more contested. We control for these variables when we examine the effect of contestation and treatment on turnout rates. The regression specification we use is Y ihnpv = βt hnpv + τc hnpv + δ(t hnpv C hnpv ) + γx + (ε v + ε ihnpv ) (5) where C hnpv is a variable representing either contestation at the polling station or the percentage of votes obtained by PMLF in polling station p. The coefficient δ on the interaction term T hnpv C hnpv reveals the extent to which the impact of the awareness campaign on turnout varies according to the underlying political environment. Table 9 reports the results. Columns 1, 2 and 3 allow for within cluster spillovers as before. The results indicate that residing in a contested polling station had a very large negative impact on turnout. However, for treated women as well as control women in treated clusters, the dampening effect of contestation is essentially fully overcome so that, on balance, they are somewhat more likely to vote than women in less contested polling stations. 4.4. Cost-Benefit Analysis and Discussion 18
We can use our estimate of impact at the polling station level to evaluate the cost effectiveness of the information campaign. This cost is likely an overestimate for two reasons. First, the development of the information campaign constitutes a third of the cost. If such an effort were scaled up, this cost would be distributed over a much larger base. Second the cost of delivering the information campaign includes the costs of collecting basic data for each household and conducting the verification visits. These costs would not be incurred in a scaled up intervention. The initial development of the information campaign cost $1500. The training of field enumerators, the delivering of the information campaign, the collection of basic information on all sample households and post election verification cost $4204. Since roughly two-thirds of sample households were treated we include 2/3 rd of this latter amount in the intervention cost. This gives us a total intervention cost of $4275. Since we have about 673 treated households, that implies a per household cost of $6.35. We have 2.69 women per household, on average, so for every 3.72 households treated we treat 10 women. The cost of treating 3.72 households is $23.6 and the cost per vote, given 7 additional votes for every 10 treated women, is $3.4 (or Rs. 231) 5. Conclusions This paper examines the role of pre-election voter information campaigns in inducing broader participation in new democracies. Our interest is specifically to assess lack of information as a barrier to political participation by women and to understand the extent to which social interactions among women could be instrumental in boosting participation beyond those directly targeted by the information campaign. While voter awareness campaigns are a staple feature of developed democracies they are relatively rare in younger and emerging democracies and we are not aware of any other studies that systematically examine the prospects of such interventions for increasing women s engagement in public life in a developing country context. Arguably, though, the value of information is likely to be far greater in a context where voters have less knowledge about the policy positions taken by candidates or parties and are engaged in various clientelist 19
relationships that influence voting decisions. Access to reliable information is likely to be an even greater barrier for women, who are generally more constrained by lack of education, lower levels of mobility and less exposure to public spaces in which political ideas can be developed. We find evidence of large spillover effects. Control women in treated clusters respond to cluster treatment assignment about as much as do directly treated women. Moving beyond clusters we examine spatial spillovers more generally and find even larger peer effects. On average, the information campaign appears to have increased turnout among treated women by about 12% which amounts to little more than an additional female vote for every 10 women or about 4 households treated. Spillovers are even higher at the polling station level where for every 10 women treated, there are almost 7 additional votes. Further, the information campaign had an effect not just on turnout but also on independence in candidate choice. Specifically, in treated households men have significantly less knowledge about women s candidate choice. We also find that information on electoral rights may be more valuable where differences in preferences over candidates are larger. The information campaign increased turnout in more contested polling stations despite the fact that contestation tended to substantially depress turnout. Finally, a simple cost benefit analysis suggests that an additional vote costs about US$3.4. This makes the information campaign a relatively cost effective way of reaching poor rural women. Bibliography Cardy, Emily A. 2005. An Experimental Field Study of the GOTV and Persuasion Effects of Partisan Direct Mail and Phone Calls. Annals of the American Academy of Political and Social Science 601: 28 40. Cialdini, Robert B., and Noah J. Goldstein. 2004. Social Influence: Compliance and Conformity. Annual Review of Psychology 55: 591 621. Desposato, Scott. 2007. The Impact of Campaign Messages in New Democracies: Results From An Experiment in Brazil. Manuscript, University of California, San Diego. 20
Duflo, Esther and Emmanuel Saez, 2003, The Role of Information and Social Interactions in Retirement Plan Decisions: Evidence From a Randomized Experiment, Quarterly Journal of Economics, Vol. 118, No. 3, Pages 815-842 Edlund, Lena and Rohini Pande, 2002, Why have Women Become Left-Wing? The Political Gender Gap and the Decline in Marriage, Quarterly Journal of Economics, Vol. 117, No. 3, Pages 917-961 Gerber, Alan S., and Donald P. Green. 2000a. The Effects of Canvassing, Telephone Calls, and Direct Mail on Voter Turnout: A Field Experiment. American Political Science Review 94(3): 653 63. Gerber, Alan S., and Donald P. Green. 2000b. The Effect of a Nonpartisan Get-Out-the-Vote Drive: An Experimental Study of Leafletting. Journal of Politics 62(3):846-857. Gerber, Alan S., and Donald P. Green. 2001. Do Phone Calls Increase Voter Turnout?: A Field Experiment. Public Opinion Quarterly 65(1):75-85. Gerber, Alan S., and Donald P. Green. 2005. Do Phone Calls Increase Voter Turnout? An Update. Annals of the American Academy of Political and Social Science 601: 142-154. Gerber, Alan S., Donald P. Green, and Ron Shachar. 2003. Voting May Be Habit-Forming: Evidence from a Randomized Field Experiment. American Journal of Political Science 47 (3): 540 50. Green, Donald P., Alan S. Gerber, and David W. Nickerson. 2003. Getting Out the Vote in Local Elections: Results from Six Door-to-Door Canvassing Experiments. Journal of Politics 65 (4): 1083 96. Gerber, Alan S., Donald P. Green and Christopher W. Larimer. 2008. Social Pressure and Voter Turnout: Evidence from a Large-scale Field Experiment. American Political Science Review 102 (1): 33-48. Harbaugh, W. T. 1996. If People Vote because they Like to, then Why do so Many of them Lie? Public Choice 89 (October):63 76. Horiuchi, Yusaku, Kosuke Imai, and Naoko Taniguchi. 2005. "Designing and Analyzing Randomized Experiments." Technical Report, http://www.princeton.edu/~ kimai/research/ Horiuchi, Yusaku, Kosuke Imai and Naoko Taniguchi. 2007. Designing and analyzing randomized experiments: application to a Japanese election survey experiment. American Journal of Political Science. 51(3):669 687. 21
Imai, Kosuke. 2005. "Do Get-Out-the-Vote Calls Reduce Turnout? The Importance of Statistical Methods for Field Experiments." American Political Science Review, 99 (2): 283-300. Kremer, Michael and Edward Miguel, 2001, Worms: Identifying Impacts On Education And Health In The Presence Of Treatment Externalities, Econometrica, Vol. 72, No. 1, 159 217 Michelson, Melissa R. 2003. "Getting Out the Latino Vote: How Door-to-Door Canvassing Influences Voter Turnout in Rural Central California." Political Behavior 25(3): 247-63. Michelson, Melissa R. 2005. Meeting the Challenge of Latino Voter Mobilization Annals of the American Academy of Political and Social Science, 601 (September): 85 101. Milbrath, Lester W. 1965. Political Participation: How and Why Do People Get Involved in Politics? Chicago: Rand McNally. Nickerson, David. "Friends Don't Make Friends Vote: Selection and Reputation in Voter Mobilization" Paper presented at the annual meeting of the American Political Science Association, Hyatt Regency Chicago and the Sheraton Chicago Hotel and Towers, Chicago, IL, Aug 30, 2007 Nickerson, David W. 2008. Is Voting Contagious? Evidence from Two Field Experiments. American Political Science Review, 102(1): 49 57. Nickerson, David W. 2006b. Volunteer Phone Calls Can Increase Turnout: Evidence from Eight Field Experiments. American Politics Research 34 (May): 271 92. Plutzer, Eric. 2002. Becoming a Habitual Voter: Inertia, Resources, and Growth in Young Adulthood. American Political Science Review 96(2): 41 56. Sacerdote, Bruce, 2001, Peer Effects with Random Assignment: Results for Dartmouth Roommates," Quarterly Journal of Economics, 116(2), 681-704. 22
Table 1: Summary Statistics N. Obs Mean St. Dev Percentile 10 Percentile 50 Percentile 90 Panel A: Household Characteristics Household size 963 10.2 5.17 5 9 16 Number of women in the household (*) 991 2.69 1.48 1 2 5 Asset Index 963 0.00 1.85-2.03-0.49 2.66 Total owned land (in acres) 963 2.58 7.55 0.01 0.04 7.02 Average monthly expenditure (in thousands) 963 8.80 4.71 3.00 9.00 12.50 House quality index 963 0.00 1.38-1.62-0.32 1.97 Distance to polling station (Km.) 991 0.42 0.94 0 0 2 Low Zaat Status 963 0.26 0.44 0 0 1 Panel B: Woman Characteristics Age 2,637 37.76 16.09 20 35 60 Woman has formal schooling (1=Yes) 2,637 0.18 0.39 0 0 1 Woman is married (1=Yes) 2,622 0.80 0.40 0 1 1 Number of children under 5 years old 2,637 0.86 1.19 0 0 3 Woman has a National Identity Card (1=Yes) 2,637 0.70 0.46 0 1 1 Woman voted in last local level elections (1=Yes) (*) 2,735 0.70 0.46 0 1 1 Access to radio (1=Yes) 2,637 0.48 0.50 0 0 1 Access to TV (1=Yes) 2,637 0.70 0.46 0 1 1 Access to cable (1=Yes) 2,637 0.30 0.46 0 0 1 Mobility (0 to 3) 2,637 2.17 0.42 2 2 3 Woman allowed to join a village organization (1=Yes) 2,637 0.73 0.44 0 1 1 Woman is a member of MRDO, an NGO in the village (1=Yes) (*) 2,735 0.11 0.31 0 0 1 Woman gets advice from a religious leader or "Pir" (1=Yes) 2,479 0.64 0.48 0 1 1 Panel C: Polling Station Characteristics Number of women registered in each polling station 21 433.95 196.71 195 464 656 Percentage of women with access to cable in the polling station 21 0.34 0.26 0.06 0.23 0.75 Percentage of women voting for PMLF party in the polling station 21 0.15 0.18 0 0.05 0.48 St. Dev of asset index 21 1.76 0.30 1.46 1.72 2.09 St. Dev of distance index 21 0.79 0.52 0.16 0.69 1.31 Index of Contestation (for each polling station) 21 0.37 0.18 0.18 0.43 0.43 High Contestation (dummy=1 if contestation index above median) 21 0.48 0.51 0 0 1 Notes: The symbol * indicates that the variable is created using only the sample from the pre-election visit. Variables are defined in Table A1 in the Appendix. 23
Table 2: Differences by treatment status Treatment vs control households Treatment 1 vs control households Treatment 2 vs control households Treated clusters vs control clusters Treated households only vs control clusters Control households in treated clusters vs households in control clusters Comparison: T-C T 1 -C T 2 -C T N -C N T-C N C TN -C N (1) (2) (3) (4) (5) (6) Panel A: Household Characteristics Household size 0.271 0.247-0.002 0.533 0.519 0.598 [0.337] [0.338] [0.365] [0.421] [0.397] [0.597] Number of women in the household (*) 0.099 0.072 0.019 0.01 0.11 0.03 [0.103] [0.108] [0.112] [0.150] [0.138] [0.173] Asset index 0.014 0.08-0.077-0.042-0.031-0.099 [0.131] [0.158] [0.187] [0.199] [0.193] [0.207] Total owned land (in acres) 0.954** 0.973* -0.137 0.783 0.939* -0.004 [0.403] [0.513] [0.515] [0.515] [0.536] [0.391] Average monthly expenditure 475.107 235.091 220.944 325.878 414.563-140.57 [400.192] [377.321] [366.025] [569.725] [591.562] [486.687] House quality index -0.031-0.058 0.034-0.171-0.152-0.286** [0.099] [0.119] [0.120] [0.124] [0.123] [0.136] Distance to polling station (Km) -0.033-0.056 0.03 0.137 0.102 0.318* [0.080] [0.089] [0.109] [0.135] [0.126] [0.165] Low Zaat status 0.027-0.002 0.03 0.063 0.061 0.087 [0.059] [0.067] [0.733] [0.099] [0.097] [0.069] N. Obs 952 952 952 952 826 295 Panel B: Woman Characteristics Age -0.763-0.506-0.17-1.410** -1.392** -1.478** [0.516] [0.539] [0.565] [0.627] [0.638] [0.700] Woman has some schooling (1=yes) 0.008-0.016 0.026 0.015 0.016 0.018 [0.019] [0.022] [0.021] [0.031] [0.029] [0.037] Woman is married (1=yes) -0.009-0.023 0.018-0.017-0.017-0.017 [0.015] [0.017] [0.019] [0.015] [0.015] [0.022] Number of children under 5 years old 0.087* 0.099* -0.028 0.147*** 0.150*** 0.139* [0.046] [0.054] [0.061] [0.050] [0.048] [0.083] Woman has a NIC or CNIC (1=yes) 0.028 0.002 0.025 0.049 0.048 0.048 [0.026] [0.030] [0.024] [0.035] [0.035] [0.032] Voted last year (1=yes) (*) 0.021-0.022 0.045* 0.036 0.036 0.036 [0.023] [0.028] [0.026] [0.030] [0.030] [0.033] Access to radio (1=yes) 0.012 0.037-0.03-0.014-0.008-0.046 [0.033] [0.031] [0.034] [0.045] [0.046] [0.045] Access to TV (1=yes) 0.022 0.044-0.028 0.026 0.028 0.028 [0.034] [0.033] [0.042] [0.053] [0.053] [0.053] Access to cable (1=yes) -0.059-0.036-0.016-0.118* -0.116* -0.116** [0.043] [0.049] [0.051] [0.066] [0.065] [0.051] Allowed to move outside settlement (0 to 3) 0.033 0.051-0.025 0.028 0.033 0.009 [0.043] [0.035] [0.041] [0.043] [0.041] [0.046] Woman allowed to join a NGO (1=yes) -0.004 0.016-0.022-0.022-0.019-0.031 [0.026] [0.025] [0.027] [0.032] [0.033] [0.037] MRDO membership (*) -0.004 0.035-0.044 0.03 0.023 0.073 [0.025] [0.029] [0.034] [0.036] [0.035] [0.044] Get advice from the Pir -0.052 0.012-0.063-0.057-0.062-0.042 [0.033] [0.041] [0.044] [0.048] [0.049] [0.053] N. Obs 2637 2637 2637 2637 2304 767 Notes: T refers to the sample of treated households, C control households, CTN control households in treated clusters, TN households in treated clusters (including both treated and control households) and CN households in control clusters (all are control households). The symbol * indicates that the variable is created using only the sample from the pre-election visit. Variables are defined in Table A1 in the Appendix. 24
Table 3: Average Effect of the Information Campaign on Turnout Treatment vs control households Allowing for spillovers within clusters Treated clusters vs control clusters Treated households only vs control clusters Control households in treated clusters vs households in control clusters Comparison: T-C T N -C N T-C N C TN -C N (1) (2) (3) (4) Panel A: Treatment Treatment (T) 0.06 0.118 0.120* 0.121* [0.045] [0.073] [0.071] [0.062] R-squared 0.18 0.19 0.19 0.21 Panel B: T 1 vs T 2 Importance of voting (T 1 ) 0.034 0.095 0.094 0.109 [0.052] [0.077] [0.075] [0.070] Importance of voting & secret balloting (T 2 ) 0.093* 0.145* 0.152** 0.135* [0.048] [0.077] [0.074] [0.079] R-squared 0.18 0.19 0.2 0.21 Observations 2637 2637 2304 767 Mean dependent variable 0.59 0.59 0.58 0.56 P-value (T 1 = T 2 ) 0.22 0.31 0.23 0.75 P-value (F-test for joint significance of T 1 & T 2 ) 0.15 0.16 0.11 0.15 Note: The dependent variable takes the value 1 if a woman reports having voted in the February 2008 elections and had a verifiable ink mark on her thumb. All specifications include village fixed effects and woman, household and polling station level controls. Standard errors (reported in brackets below the coefficient) are corrected for clustering within geographic clusters. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 25
Table 4: Spillover Effects using distance - I 75m 100m 200m (1) (2) (3) Panel A: Treatment Treatment (T) 0.127* 0.158** 0.131* [0.065] [0.074] [0.078] Number of households within radius 0.005 0.005* 0.003* [0.004] [0.003] [0.002] R-squared 0.2 0.21 0.21 Panel B: T 1 vs T 2 Importance of voting (T 1 ) 0.103 0.137* 0.112 [0.070] [0.079] [0.082] Importance of voting & secret balloting (T 2 ) 0.156** 0.183** 0.155* [0.067] [0.075] [0.079] Number of households within radius 0.005 0.005 0.003* [0.004] [0.003] [0.002] R-squared 0.2 0.21 0.21 Observations 2207 2128 2049 Mean dependent variable 0.58 0.58 0.58 P-value (T 1 = T 2 ) 0.25 0.31 0.33 P-value (F-test for joint significance of T 1 & T 2 ) 0.06 0.05 0.14 Note: The dependent variable takes the value 1 if a woman reports having voted in the February 2008 elections and had a verifiable ink mark on her thumb. In each specification, women in control households located within the indicated radius of a treated household are dropped from the sample. All specifications include village fixed effects and woman, household and polling station level controls. Standard errors (reported in brackets below the coefficient) are corrected 26
Table 5: Spillover Effects using distance - II Treatment (T) 0.027 [0.031] Number of treated households within 0-200 radius 0.017*** [0.004] Number of treated households within 200-400 radius 0.022*** [0.004] Number of treated households within 400-600 radius 0.017*** [0.005] Number of treated households within 600-800 radius 0.008 [0.006] Number of treated households within 800-1000 radius 0.008 [0.008] Number of treated households within 1000-1,200 radius 0.004 [0.007] Number of households within 0-200 radius -0.008*** [0.003] Number of households within 200-400 radius -0.012*** [0.003] Number of households within 400-600 radius -0.013*** [0.004] Number of households within 600-800 radius -0.004 [0.004] Number of households within 800-1000 radius -0.008 [0.005] Number of households within 1000-1,200 radius 0.001 R-squared 0.23 Mean dependent variable 0.59 Observations 2637 Note: The dependent variable takes the value 1 if a woman reports having voted in the February 2008 elections and had a verifiable ink mark on her thumb. All specifications include village fixed effects and woman, household and polling station level controls. Standard errors (reported in brackets below the coefficient) are corrected for clustering within geographic clusters. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 27
Table 6: Spillover Effects via Friendship Friends of Women in Treated Clusters vs. Friends of Women in Control Clusters Friends of Treated Women in Treated Clusters vs. Friends of Women in Control Clusters Friends of Control Women in Treated Clusters vs. Friends of Women in Control Clusters T-C T-C N C TN -C N Treatment 0.107 0.104 0.124* [0.078] [0.075] [0.070] Treatment controlling for the characteristics of sample women 0.12 0.117 0.124* [0.075] [0.071] [0.068] N. obs 797 692 245 Mean dependent variable 0.6 0.6 0.58 Note: The dependent variable takes the value 1 if a friend reports having voted in the February 2008 elections. All specifications include village fixed effects. The specification in the second row control for the characteristics of the sample woman, her household and the polling station. Standard errors (reported in brackets below the coefficient) are corrected for clustering at the geographic cluster level. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 28
Table 7: Spillovers at the Polling Station Level (1) (2) Number of Treated Women 0.635* 0.726** [0.322] [0.290] Percentage of Women with Access to Cable TV 93.293 [57.213] SD of Asset Index 131.578** [49.900] SD of Distance to the Polling Station 30.733 [28.834] Number of Women Registered to Vote 0.19** 0.17** [0.089] [0.078] Constant 86.318* -200.262* [44.510] [100.015] Observations 21 21 R-squared 0.39 0.64 Note: The dependent variable is the number of valid votes cast by women in the polling stations of our sample villages. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. Robust standard errors in parentheses. 29
Table 8: Effect on Candidate Choice Using Cross Reports from Family Members Treatment vs control households Treated clusters vs control clusters Spillovers within Clusters Treated households only vs control clusters Control households in treated clusters vs households in control clusters 75 Meter Radius Geographic Spillovers 100 Meter Radius 200 Meter Radius T-C T N -C N T-C N C TN -C N Treatment -0.026-0.022-0.022-0.011-0.011 0.012 0.019 [0.024] [0.030] [0.030] [0.026] [0.032] [0.036] [0.047] Man reporting about woman -0.066*** -0.031-0.024-0.048* -0.027-0.04-0.025 [0.024] [0.024] [0.024] [0.025] [0.027] [0.034] [0.024] Interaction between treatment & man reporting -0.043-0.074** -0.084** -0.043-0.082** -0.069-0.083** [0.030] [0.033] [0.034] [0.042] [0.038] [0.044] [0.040] Number of households within radius 0.002 0.002 0 [0.003] [0.002] [0.001] Mean dependent variable 0.91 0.91 0.91 0.94 0.07 0.08 0.07 N.obs 2825 2825 2454 672 2408 2349 2289 R-Sq 0.068 0.067 0.072 0.12 0.905 0.903 0.901 Note: The dependent variable takes the value 1 if a woman voted and her report regarding candidate choice matches with the report of another woman in the household or the male head. Each observation is therefore a pair with several observation for each reportee. All specifications include village fixed effects and the controls for the reportee woman, her household and her polling station. Standard errors (reported in brackets below the coefficient) are corrected for clustering at the geographic cluster level. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 30
Table 9: Contestation and Information Treated clusters vs control clusters Spillovers within Clusters Treated households only vs control clusters Control households in treated clusters vs households in control clusters T N -C N T-C N C TN -C N Panel A: Herfindahl for the Share of Votes Obtained by the Major Political Parties (Contestation I) Treatment -0.117-0.114-0.151 [0.121] [0.116] [0.125] Contestation I -0.615* -0.654** -0.477 [0.337] [0.327] [0.313] Treated Woman X Contestation I 0.657* 0.654* 0.732** [0.357] [0.346] [0.340] R-Sq 0.19 0.2 0.23 Panel B: Share of Votes Obtained by PML-F (Contestation II) Treatment -0.065-0.058-0.109 [0.084] [0.082] [0.075] Contestation II -1.449-1.438-1.445 [1.146] [1.110] [1.135] Treated Woman X Contestation II 2.102* 2.058* 2.285** [1.125] [1.090] [0.995] R-Sq 0.22 0.23 0.25 Mean Dependent Variable 0.59 0.58 0.56 N.obs 2637 2304 767 Note: Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. Robust standard errors in parentheses are clustered at the geographic cluster level. 31
Appendix 32
Table A2: Intervention check C with at least 1 T within 75m C with at least 1 T within 100m C with no T within 75m C with no T within 100m N. Obs All T T1 T2 1 2 3 4 5 6 7 8 9 Visit before elections (1=Yes) 2505 0.71 1.00 1.00 1.00 0.00 0.00 0.00 0.00 Neighbors joined during visit (1=Yes) 1862 0.11 0.11 0.08 0.15 0.10 0.10 0.00 0.00 Neighbor talked to woman (1=Yes) 2505 0.44 0.55 0.50 0.62 0.29 0.25 0.04 0.03 Issues raised Importance of voting 1867 0.64 0.64 0.98 0.19 0.58 0.58 0.20 0.00 Importance of secret voting 1867 0.06 0.06 0.02 0.12 0.04 0.04 0.00 0.00 Both 1867 0.30 0.30 0.01 0.69 0.38 0.38 0.80 1.00 33
Table A3: Gender Differences in Radio and TV Access and Exposure to News N. Obs Mean Dependant Variable P-value Female Male Access to Radio 1923 0.469 0.417 0.061 Number of hours of radio listened to in an average 852 9.739 9.633 0.867 Access to TV 1923 0.668 0.629 0.103 Number of hours of TV watched in an average week 1222 15.294 11.033 0.000 Access to cable TV 1951 0.289 0.224 0.004 Listen Watch to cable BBC TV for channels world news for 847 0.095 0.483 0.000 national news 492 0.19 0.272 0.095 Note: P-values are from regressions with standard errors clustered at the geographic cluster level. 34
Table A4: Gender Differences in Knowledge about Current Political Issues and the Results of the Election N. Obs Mean Dependant P-value Variable Respondent Female Male Does not know the meaning of democracy 1923 0.972 0.694 0 Was aware of the imposition of emergency and the position of parties position on the 1951 0.058 0.822 0 removal of judges Able to identify political party signs correctly (proportion identifed-national assembly) 1951 0.283 0.406 0 Able to identify political party names correctly (proportion identifed-national assembly) 1951 0.858 0.934 0 Able to identify political party signs correctly (proportion identifed-provincial assembly) 1951 0.28 0.412 0 Able to identify political party names correctly (proportion identifed-provincial assembly) 1951 0.849 0.929 0 Knows the gender of main candidates (national assembly) 1951 0.948 0.957 0.483 Recalled the names of the candidates correctly (proportion identified-national assembly) 1951 0.823 0.84 0.488 Knows the gender of main candidates (provincial assembly) 1951 0.952 0.951 0.967 Recalled the names of the candidates correctly (proportion identified-provincial assembly) 1951 0.818 0.836 0.519 Able to recall the winning party (national assembly) 1951 0.964 0.944 0.093 Able to recall the winning candidate (national assembly) 1951 0.902 0.924 0.219 Able to recall the winning party (provincial assembly) 1951 0.96 0.949 0.318 Able to recall the winning candidate (provincial assembly) 1951 0.909 0.934 0.119 Note: P-values are from regressions with standard errors clustered at the geographic cluster level. 35
Table A5: Gender Differences in Participation in Village Political and Social Events N. Obs Mean Dependant Variable P-value Female Male Attend community meetings 1921 0.179 0.52 0.000 Get together to raise issues 1921 0.243 0.506 0.000 Attend demonstrations 1921 0.121 0.233 0.000 Take action to rectify election official missing name in voter list 1921 0.762 0.924 0.000 if police arrest family member wrongly 1921 0.929 0.98 0.000 if someone seized family land 1921 0.924 0.982 0.000 Index of community action taken 1921-0.292 0.686 0.000 Contact local councilor 1921 0.252 0.372 0.000 Contact a local political party official 1921 0.224 0.378 0.000 Index of contacting formal authority 1921-0.102 0.336 0.000 Contact religious leader 1921 0.66 0.49 0.000 Contact traditional ruler 1921 0.445 0.32 0.000 Index of contacting informal authority 1921 0.124-0.303 0.000 Note: p values were calculated from regressions that were clustered at the neighborhood level. 36
Table A6: Treatment Check for Measures of Political Contestation at the Polling Station Level Contestation-I Contestation-II (1) (2) (3) (4) Number of Treated Women 0 0 0 0 [0.001] [0.001] [0.001] [0.001] Percentage of Women with Access to Cable TV 0.211-0.114 [0.146] [0.177] SD of Asset Index 0.268* 0.176 [0.127] [0.155] SD of Distance to the Polling Station 0.141* 0.009 [0.073] [0.089] Number of Women Registered to Vote 0 0 0 0 [0.000] [0.000] [0.000] [0.000] Constant 0.327*** -0.329 0.051-0.217 [0.111] [0.255] [0.109] [0.310] Observations 21 21 21 21 R-squared 0.01 0.38 0.04 0.18 Note: The dependent variable is the Herfindahl on the share of votes obtained by the two major political parties, in the first two specifications; and is the share of votes obtained by PML-F in the last two specifications. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. Robust standard errors in parentheses. 37
Table A7: Contestation and Voter Turnout All Low contestationhigh contestation Panel A: Polling Station Turnout Total votes cast 4501 2051 2450 As percentage of registered voters 76.6 88.2 68.7 For PPP 66.4 81.8 53.6 For PML 32.1 17.1 44.7 Others 1.4 1.1 1.7 Panel B: Sample Turnout Total votes cast 1543 831 712 As a percentage of women with NIC 84.0 89.0 78.8 For PPP 73.68 91.53 57.34 For PML 26.23 8.47 42.5 Note: Sample turnout rates (in Panel B) are calculated over women who could be verified as having voted. 38
Table A8: Impact of Contestation on Women's Participation and Candidate Choice All High Contestation Low Contestation Percentage of women who T C p-value T C p-value T C p-value Voted for the same party as head 44.7 44.3 0.92 46.9 44.3 0.73 42.7 44.3 0.91 Voted for different party from head 10.1 5.1 0.01*** 14.8 8.0 0.01*** 6.0 2.9 0.13 Voted but head did not 6.1 5.0 0.62 8.2 6.5 0.65 4.3 3.9 0.81 Did not vote 38.6 45.1 0.07* 29.6 41.2 0.03** 46.4 47.9 0.56 TN CN p-value TN CN p-value TN CN p-value Voted for the same party as head 45.1 41.7 0.37 47.6 38.1 0.04** 42.9 44.2 0.86 Voted for different party from head 9.8 2.8 0.01*** 14.3 5.1 0.02** 5.9 1.2 0.02** Voted but head did not 6.2 3.7 0.19 8.4 4.0 0.25 4.3 3.5 0.81 Did not vote 38.4 50.9 0.01*** 29.3 52.8 0.00*** 46.3 49.6 0.33 T CN p-value T CN p-value T CN p-value Voted for the same party as head 44.6 41.7 0.5 46.9 38.1 0.07* 42.7 44.2 0.87 Voted for different party from head 10.1 2.8 0.01*** 14.8 5.1 0.01*** 6 1.2 0.03** Voted but head did not 6.1 3.7 0.22 8.2 4 0.09* 4.3 3.5 0.98 Did not vote 38.6 50.9 0.01*** 29.6 52.9 0.00*** 46.4 49.6 0.32 CTN CN p-value CTN CN p-value CTN CN p-value Voted for the same party as head 47.8 41.7 0.08* 51.7 45.4 0.23 44.6 43 0.63 Voted for different party from head 8.1 2.8 0.14 11.4 13.2 0.45 5.4 5 0.98 Voted but head did not 6.6 3.7 0.16 9.4 7.5 0.66 4.4 4.1 0.55 Did not vote 37.5 50.9 0.00*** 27.5 33.5 0.09* 45.7 47 0.62 Note: P-values are from regressions with village fixed effects, woman characteristics as controls and robust standard errors clustered at the geographic cluster level. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 39
Table A9: Contestation & Sample Women's Reports Regarding Election Day Contestation I Woman Witnessed/Heard Acts of Violence In/Near Election was Fair Village Environment was Safe the Village Election was Fair Village Environment was Safe Panel A: T-C (Treatment vs Control Households) (1) (2) (3) (4) (5) (6) 40 Contestation II Woman Witnessed/Heard Acts of Violence In/Near the Village Panel B: TN-CN (Treatment Clusters vs Control Clusters) Treatment 0.05 0.087-0.203-0.027-0.079-0.01 [0.063] [0.095] [0.133] [0.041] [0.066] [0.094] Contestation -0.185-0.397 0.11-1.502** -2.901** 1.981** [0.183] [0.328] [0.303] [0.685] [1.146] [0.822] Treated Woman X Contestation 0.035-0.104 0.313 1.472** 2.545** -1.681** [0.200] [0.371] [0.347] [0.688] [1.156] [0.836] Mean Dependent Variable 0.88 0.82 0.24 0.88 0.82 0.24 N.obs 2637 2637 2637 2637 2637 2637 R-Sq 0.01 0.05 0.03 0.02 0.05 0.03 Panel C: T-CN (Treated Households vs Control Clusters) Treatment 0.061 0.089-0.208-0.026-0.085-0.01 [0.064] [0.096] [0.133] [0.042] [0.070] [0.096] Contestation -0.185-0.397 0.11-1.502** -2.901** 1.981** [0.183] [0.328] [0.303] [0.685] [1.146] [0.822] Treated Woman X Contestation 0.019-0.129 0.337 1.494** 2.540** -1.668* [0.204] [0.376] [0.349] [0.688] [1.159] [0.837] Mean Dependent Variable 0.89 0.81 0.25 0.89 0.81 0.25 N.obs 2304 2304 2304 2304 2304 2304 R-Sq 0.01 0.05 0.03 0.02 0.05 0.03 Panel D: CTN-CN (Controls in Treated Clusters vs Control Clusters) Treatment -0.013 0.072-0.163-0.032-0.042-0.014 [0.074] [0.094] [0.143] [0.049] [0.054] [0.094] Contestation -0.185-0.396 0.11-1.502** -2.901** 1.981** [0.184] [0.328] [0.303] [0.686] [1.147] [0.823] Treated Woman X Contestation 0.136 0.045 0.165 1.347* 2.572** -1.757** [0.219] [0.361] [0.358] [0.699] [1.155] [0.839] Mean Dependent Variable 0.85 0.82 0.27 0.85 0.82 0.27 N.obs 767 767 767 767 767 767 R-Sq 0.01 0.04 0.44 0.03 0.11 0.05 Note: All specifications include village fixed effects and woman, household and polling station level controls. Standard errors (reported in brackets below the coefficient) are corrected for clustering within geographic clusters. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence.
Table A10: Impact of treatment on women's voting report All High Contestation Low Contestation Panel A: Treatment vs Control Households T C p-value T C p-value T C p-value Self reports voting but verified as not voted 27.16 28.81 0.48 24.22 31.58 0.24 30.48 26.52 0.77 Self reports not voting but verified as voted 38.62 37.46 0.29 32.05 28.89 0.04** 42.27 42.92 0.84 Panel B: Treatment Clusters vs Control Clusters TN CN p-value TN CN p-value TN CN p-value Self reports voting but verified as not voted 26.61 33.96 0.21 23.64 46.34 0.04** 29.94 26.15 0.84 Self reports not voting but verified as voted 38 39.19 0.30 32 27.66 0.003*** 41.27 47.66 0.72 Panel C: Treated Households vs Control Clusters T CN p-value T CN p-value T CN p-value Self reports voting but verified as not voted 27.16 33.96 0.18 24.22 46.34 0.03** 30.48 26.15 0.90 Self reports not voting but verified as voted 38.62 39.19 0.32 32.05 27.66 0.006*** 42.27 47.66 0.69 Note: P-values are from regressions with village fixed effects, woman characteristics as controls and robust standard errors clustered at the geographic cluster level. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 41
Table A11: Effect of the Information Campaign by Woman Characteristics Woman Characteristic (WC) Panel A: TN-CN (Treatment Clusters vs Control Clusters) Mobility Voting History Literacy Access to TV (1) (2) (3) (4) TN 0.426*** 0.017 0.085 0.056 [0.123] [0.111] [0.091] [0.113] WC 0.189*** 0.16 0.028 0.064 [0.056] [0.106] [0.081] [0.075] Interaction between TN x WC -0.161** 0.091 0.034 0.042 [0.064] [0.117] [0.089] [0.086] Mean dependent variable 0.59 0.59 0.59 0.59 Observations 2637 2637 2637 2637 R-squared 0.12 0.15 0.11 0.12 Panel B: T-CN (Treated Households vs Control Clusters) Treatment excluding controls in treated neighborhoods 0.399*** 0.021 0.086 0.054 [0.122] [0.110] [0.089] [0.111] Variable 0.187*** 0.162 0.019 0.062 [0.055] [0.105] [0.079] [0.071] Interaction (Treatment excluding contaminated controls x Variable) -0.148** 0.086 0.034 0.045 [0.064] [0.117] [0.088] [0.084] Mean dependent variable 0.58 0.58 0.58 0.58 Observations 2304 2304 2304 2304 R-squared 0.13 0.15 0.12 0.12 Panel c: CTN-CN (Controls in Treated Clusters vs Control Clusters) Controls in treatment neighborhood 0.153*** -0.016 0.093 0.063 [0.053] [0.098] [0.066] [0.087] Variable 0.470*** 0.186* 0.034 0.065 [0.150] [0.099] [0.057] [0.052] Interaction (Controls in treated neighborhood x Variable) -0.173** 0.164 0.1 0.062 [0.077] [0.123] [0.102] [0.092] Mean dependent variable 0.56 0.56 0.56 0.56 Observations 767 767 767 767 R-squared 0.17 0.19 0.15 0.15 Note: The dependent variable takes the value 1 if a woman reports having voted in the February 2008 elections and had a verifiable ink mark in her finger. All specifications include village fixed effects and woman, household and polling station level controls. Robust standard errors clustered at the geographic cluster level. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 42
Table A12: Effect of the Information Campaign on Behavior N. Obs Female knowledge index Female opinion index Woman checked voter list after intervention Woman expresses political opinion Woman thinks election was fair Treatment vs control households (T-C) 2637 0.024 0.174** 0.033 0.027 0.057** [0.059] [0.085] [0.024] [0.020] [0.025] Treated clusters vs control clusters (TN-CN) 2637 0.005 0.203** 0.051* 0.005 0.070** [0.072] [0.092] [0.030] [0.025] [0.034] Treated households only vs control clusters (T-CN) 2304 0.012 0.212** 0.053* 0.011 0.075** [0.070] [0.097] [0.031] [0.024] [0.033] Control households in treated clusters vs households in control clusters 767-0.025 0.167 0.048-0.036 0.06 [0.094] [0.116] [0.035] [0.036] [0.037] Mean dependent variable (CTN-CN) 0 0 0.61 0.23 0.88 Note: All specifications include village fixed effects and woman, household and polling station level controls. Standard errors (reported in brackets below the coefficient) are corrected for clustering within geographic clusters. Significantly different from zero at 99 (***), 95 (**), and 90 (*) percent confidence. 43
Appendix B 44
45
46
Appendix C 47