Entry, Exit and Candidate Selection: Evidence from India

Entry, Exit and Candidate Selection: Evidence from India Abhijit V Banerjee Esther Duflo Clément Imbert Rohini Pande August 18, 2017 Abstract What motivates candidates to run in low income democracies? We examine the electoral impact of a voter awareness campaign which emphasized the responsibilities of local leaders in implementing a nation-wide public employment program. The intervention led to exit by the worse performing incumbents, substitution by incumbent family members and entry by challengers from traditionally disadvantaged groups. We contrast these effects with those of previous reservation for women, which weakens the incumbent who faces greater social pressure to not re-run. Previous reservation had similar effects on incumbency, accountability, and challenger entry, but did not change the composition of the candidate pool. This suggests that the wider political representation due to the campaign is not a mechanical effect from weakening the incumbent but reflects the motivation of candidates from disadvantaged groups to be in charge of the program. We do not find that this motivation led to better performance once in office. We do however find persistent effects on political representation of women and lower caste in the next election round. Keywords: Elections, Candidate entry, Information campaign JEL: We would like to thank Aparna Krishnan, Vipin Awatramani, Dinesh Bhatt, Nikhil Dev, Avnish Gungardurdoss and Harsh Gupta for field work and Jeff McManus for exceptional research assistance. Thanks to Sharun Mukhand and Patrick Francois for helpful comments. Abhijit Banerjee and Esther Duflo: MIT; Clément Imbert: Warwick University; Rohini Pande: Harvard University. Protocol 091003477). The project was approved by the Institutional Review Boards at Harvard, and MIT (COUHES 1

1 Introduction There is broad agreement that the effectiveness of democratic governance rests, to a large extent, on whether high quality citizens enter politics (Myerson (2011)) and on whether citizens reward quality and performance at election time (Besley, 2007). There is also a concern that entrenchment of the economic elite in politics may prevent this process in many low-income democracies (Acemoglu and Robinson, 2008). A growing body of evidence demonstrates the importance of entry costs in affecting candidate entry (Chattopadhyay and Duflo (2004)) and of voter information (awareness) for candidate selection (Ferraz and Finan, 2008; Banerjee et al., 2011). However, our evidence base on how these factors influence incumbent decisions to re-run and resulting strategic interactions across candidates is much more limited. We also know less about the implications for the quality of governance. In this paper we examine these issues in the context of elections for local village council head in the Indian state of Rajasthan in 2010. In our sample of 382 village councils, two exogenous factors influenced citizens entry decision and subsequent voter behavior. First, a random sample of councils had been reserved in the previous electoral cycle for women. To the extent that these women had been first time entrants into politics and faced greater family pressure to not rerun, challengers could anticipate a lower probability that the incumbent would run for election. Second, voters in a random sample of councils were exposed to a pre-election voter awareness campaign (now on, PEVAC) in which they were informed about the responsibilities of the council head in implementing a large public works program and average incumbent performance (across a large set of village councils, which included theirs) in implementing this program. In a subsample of PEVAC councils, the campaign also emphasized the fact that the performance of female incumbents was similar (if not better) than male incumbents. The PEVAC treatment occurred in the month prior to election (when candidates still made entry decisions). It, thus, both provided potential challengers with more information on the returns to political office and depending on incumbent performance may have either strengthened or weakened the incumbency advantage and, therefore, the incumbent decision on whether to run. What are the likely impacts of these two exogenous shocks? The seminal papers on citizen entry into politics, which allow for strategic choices, are the citizen candidate models of political 2

competition (Besley and Coate, 1997; Osborne and Slivinski, 1996). These papers identify policy preferences of other candidates and probability of electoral success (running is assumed to be costly) as determinants of a candidate s decision to run. A natural enrichment of this model is to assume that candidates also enjoy private rents to office. Caselli and Morelli (2004) show that if several candidates are running, but voters have limited ability to distinguish between good and bad candidates, then the winning probabilities are small and highly skilled citizens may prefer to not enter politics. That said, if candidate entry aids learning about candidate type then increased candidate entry may improve outcomes in a dynamic setting. 1 On the other hand, more candidate entry also increases the challenge for voter coordination. To summarize, these models predict that reductions in entry costs (in our case, due to reduced incumbency advantage) should encourage candidate entry. Who enters is less clear, and will depend on the heterogeneity of preferences in the population, on the private returns to political office and on the quality distribution among potential candidates. The implications for the quality of elected officials, in turn, depends on voter information about candidate quality and their ability to coordinate their voting decisions. From our experiments, the first finding is that both the end of mandated representation for women and greater citizen awareness about the returns to political office and average quality of incumbents (from the PEVAC) reduced the likelihood that the incumbent chose to run for re-election. Specifically, (female) incumbents in previously reserved councils are less likely to re-run than the (typically male) incumbents in previously un-reserved councils (14.8% versus 46.2%). Second, the PEVAC reduced the probability that a (typically male) incumbent (in a previously unreserved seat) would run again by 24 percentage points (more than a half), and reduced the vote share going to a typically male incumbent by almost a half. The latter finding 1 For example, we could think of a dynamic extension of the Caselli and Morelli framework, where candidates need to enter for voters to learn their types and the presence of a very good incumbent makes it so unlikely that any new entrant would win that essentially no one enters and no one learns their type, making the eventual transition after the incumbent retires less efficient. Or there may be voter inertia: Myerson (1993) argues that good candidates may not join the race because there is a small group of established but not necessarily good candidates who have traditionally attracted all the votes, and no voter will (rationally) want to be the only one switching her vote to a new candidate. In other words there may be a coordination failure, with too little entry. Caselli and Morelli (2004) rule this out by assuming that voters vote sincerely rather than strategically. 3

is consistent with the fact that the average incumbent performance in the public works program was relatively poor. If we use a broader measure of incumbency whether the incumbent or a member of her family stood for election then the impacts of reservation and PEVAC on the probability of running are much smaller, and become insignificant. In the case of reservation this is consistent with the idea that the spouse of the female incumbent is the long-run political actor in the household, and replaces the female incumbent when reservation ends. The fact that PEVAC leads to intra-family substitution suggests a strategic response; our finding resonates with Querubin (2011) finding on term limits in the Philippines. A key feature of the PEVAC campaign was the focus on performance in the public employment program. We construct an index of incumbent performance using publicly available statistics on program implementation (number of man days generated), combined with measures from a household survey which we conducted for the purpose of this study. The anti-incumbent impact of both PEVAC and becoming unreserved are much more muted for better performing incumbents. Thus, it appears that the anti-incumbent impact of the campaign was strongest for the worst performing incumbents. Using a broader definition of incumbency which includes the family members only strengthens this result. Alongside, we observe candidate entry. Reservation increased the number of candidates by one (on a basis of about seven) and the number of literate candidates. The PEVAC did not have any significant effect on the number of candidates. However, it did have important effects on the composition of the pool of candidates. It increased the number of female candidates, the number of candidates with political experience in lower offices (councilor position) and the number of candidates belonging to traditionally less advantaged lower caste groups (scheduled caste, scheduled tribes, or other backward castes). Finally, reservation and PEVAC do not seem to have had a cumulative impact on any of these dimensions. The only additive effect of these interventions is on encouraging more landless candidates to enter the race. In terms of who won, we see that PEVAC increased the likelihood that a non-general caste candidate won. It also increased the likelihood that the winning candidate had previous experience as a councilor. Turning to the impact of PEVAC on quality of newly elected council head, we find no evidence that performance as measured by the number of jobs generated under the 4

employment guarantee scheme or their allocation to disadvantaged groups did not improve as reported in the official statistics, 2. We also find no evidence that leaders elected in treated councils at the time of the PEVAC (in 2010) were no more likely to be re-elected in the next election round (in 2015). 3 We do however find evidence of long term effects on the candidate pool: there were still more female and lower caste candidates in 2015 in treated councils. Overall, these results demonstrate both the importance of strategic behavior in politicians decisions to run, and the role of incumbency advantage in influencing the costs of entry faced by potential challengers. On the one hand if we were to measure the success of reservation and PEVAC by representation, then both interventions led to entry and greater electoral success for historically disadvantaged groups women and lower castes. On the other hand, if we measure the campaign success by incumbent performance then it does not seem that entry by new entrants (with likely less political experience) necessarily leads to better performance while in office. There is a rich theoretical (Barro, 1973; Ferejohn, 1986; Persson et al., 1997) and more recently empirical (Ferraz and Finan, 2008; Banerjee et al., 2011) litterature on the effect of political information on governance outcomes. One key insight of this literature is that information about the incumbent s performance while in office will be used at the time of the elections by voters to sanction bad politicians. However, even if in the long run more information should help select better politicians, the short term implications of exposing the incumbent on the quality of the challengers who enter the race and on the performance of the elected candidate are less clear. Related empirical work includes Diermeier et al. (2005) and Keane and Merlo (2010) who estimate structural models of entry decisions of US legislators where the emphasis, as in Caselli and Morelli (2004) is on the role of the outside option. They find evidence that highly skilled people are more likely to exit politics, especially when congressional salaries are low. In Ferraz and Finan (2009), higher wages both increase the number of candidates and improves the quality 2 Official reports on MG NREGS employment are often inflated by corrupt officials (Niehaus and Sukhtankar, 2013; Imbert and Papp, 2011; Banerjee et al., 2016). 3 Because of rotating reservation and new electoral rules which mandated that candidates should have at least eight years of education, few incumbents could actually run. 5

of legislators. However, these papers assume that the likelihood of winning is a candidate characteristic, i.e. they do not allow for strategic elements in the entry/exit decision. Section 2 describes the context of our study, the electoral interventions and identifies their likely impacts. Section 3.1 describes the data and empirical strategy. Section 4 reports our experimental findings and Section 5 concludes. 2 Background 2.1 Electoral context Our focus is on village council or Gram Panchayat (hereafter GP) elections. This is the lowest tier of elected representation in India, which is responsible for the provision of village infrastructure (public buildings, water, roads, etc.) and for identifying welfare recipients within the villages. Since 2006, the GP is also responsible for the administration of the federal government s flagship anti-poverty program, the Mahatma Gandhi Rural Employment Guarantee Act (MGN- REGS). The act entitles every rural household to 100 days of work per year at the state-statutory minimum wage. The GP is in charge of registering potential beneficiaries, planning and executing public works 4, providing employment upon request and insuring timely payment of wages to the workers. In Rajasthan, the setting for our study, the typical Gram Panchayat (GP) has an electorate of roughly 3,000 spread across multiple villages. The GP is divided into electoral wards, and a ward councilor is directly elected from each ward by plurality rule every five years. In addition, in Rajasthan, the Sarpanch (head councilor) is also elected by direct vote. GP council decision-making is by majority voting. However, while the Sarpanch does not have formal veto power, he or she is the only full-time member of the council and exercises significant control over the final council decisions (Besley et al., 2012). Our experiment therefore, focused on the Sarpanch election. Our evaluation occurred in the context of the 2010 Sarpanch elections which were held between 22nd January and 4th February 2010. It focuses on 382 GP, which constituted all 4 The village general assembly, or Gram Sabha decides every year on a shelf of projects. The head of the Gram Panchayat selects and executes works depending on the demand for work. 6

Sarpanch seats not reserved for women in three districts of Eastern Rajasthan: Bharatpur, Dholpur and Karauli. Below, we describe the two cross-cutting interventions which we exploit in our analysis. 2.2 Reservation The introduction of elected GP in 1993 was accompanied by electoral quotas for disadvantaged minorities and for women, a policy which has led to a dramatic rise in local female leadership across Indian villages. Each Indian state is required to reserve at least 1/3 of elected Sarpanch positions for women, but each state separately decides the rules for implementing political reservation. In Rajasthan, GP to be reserved for women are randomly selected at each election. At the time of our evaluation, two electoral cycles had been completed with reservation: in 2000 and 2005. For these two elections, reserved GP were selected by rotation; in principle the villages that had been reserved in 2000 were excluded from randomization in 2005, although there is a couple of exceptions in our data. For 2010 elections, 50% of the GP were randomly selected to be reserved for women regardless of their previous reservation status; our evaluation focuses on the remaining 50%. By law each Indian state is also required to reserve some GP for disadvantaged minorities (low caste and tribals). In each district of Rajasthan, the percentage of GP reserved for each minority was fixed to reflect their share in the total population, and GP with the highest population share of each group were selected for reservation. In 2005 and 2010, reservation was rotated in a non transparent way. The reservation for caste and the reservation for women were done independently for each election, so that the probability of being reserved for women in 2010 or having been reserved in 2000 and 2005 should be the same for all GP, regardless of their caste reservation status in 2010 or in previous elections. In our sample of GP not reserved for women, 16% were reserved for Other Backward Castes (OBC) in 2010, 23% for Scheduled Castes (SC) and 11% for Scheduled Tribes (ST) so that the incumbent could only run in 64% of the GP In our analysis, we take advantage of the random assignment of GP to be reserved for women to identify the impact of past reservation for women. However, because of the endogenous 7

selection of GP reserved for minorities, we control for caste reservation status in 2010, but we cannot identify the impact of caste reservation per se. 2.3 Voter Awareness campaigns The second intervention exploited in our analysis is an awareness campaign, which used calendars, door-to-door canvassing and street theater to convey information about Sarpanch responsibility in providing public goods and employment to the villages. The main objectives were to prompt voters and potential candidates to participate to the elections and to focus on performance, instead of caste or gender. There were two versions of the campaign, which used two different sets of calendars and two different scripts for the plays, which we will refer to as the general campaign and the gender campaign. The general campaign emphasized that beyond caste or gender only performance mattered to choose the right candidate. The street play illustrated the importance of electing a wellperforming leader by comparing one village where public infrastructure (roads or hand pump for drinking water) was well maintained with another village where public infrastructure was poor. District average statistics on MGNREGS implementation were printed on the calendar and read out during the play to highlight the fact that the Sarpanch played an important role in providing employment to the community. 5 The campaign did not disclose information on the level of MGNREGS implementation in each GP along with district averages. The reason was to avoid being unfair vis-a-vis the incumbent, because for example lower than average MGNREGS expenditures could reflect low demand for work in the GP rather than low supply of public employment. Honest and effective citizens were encouraged to run as candidates and voters were encouraged to take such candidates seriously. The gender campaign incorporated the elements of the general treatment. In addition, it provided the message that both genders could run in unreserved councils, and that women Sarpanch (from reserved GP) typically performed as well as men. To illustrate the first message, 5 The statistics used were expenditures per GP, share of participant households and number of working days per participant household, which we downloaded in December 2009 for each district from the official website http:\\nrega.nic.in. They were hence publicly available information, in contrast with the data from the household and asset surveys which we use for the purpose of the analysis. 8

the play depicted a male candidate who attempted to forbid to a woman to run against him, but was forced by village elder to admit that she had the right to run. To illustrate the second message, statistics on MGNREGS implementation in GP reserved for women in 2005 and in unreserved GP were printed on the calendars and compared during the play. The modus operandi was similar to other pre-election information campaigns done by NGOs or by the government in Rajasthan. Street plays are frequently used to raise awareness about particular issues (e.g. HIV) in rural areas, where the reach of newspapers, radio and television is still very limited. 6 It was implemented by Prayatn, a NGO based in Jaipur (Rajasthan), which hired teams of actors, dancers and singers for the purpose of the campaign. Prayatn had been hired by the government to raise awareness about the National Rural Employment Guarantee Scheme when it was launched in Karauli in 2006. 7 The intervention took place in a randomly selected set of 119 GP (out of our full sample of 382 GP). The campaign schedule was organized in order to minimize the time to the elections; in practice the teams visited the GP between one and 20 days (10 days on average) before the elections. Close as it was to the election, the campaign still happened before candidates registered: candidate registration in every GP occurred the day before the elections. In the treatment GP, the campaign teams visited all villages with more than 300 inhabitants, 439 villages in total. A team would typically go to the main village in the morning to distribute calendars and advertize the evening show. It would then travel to a second, smaller village, where the distribution of calendars and the show would happen in a 2 hours sequence around noon. After covering a third, smaller village in the afternoon, the team would go back to the main village and make an evening show. Villages with more than 1000 inhabitants had two shows, set in different neighborhoods. In total, a team of 60 members performed in 439 villages (all villages of more than 300 inhabitants were covered, a Panchayat generally has 4 large villages). Monitors were sent with each team to check if the protocol was followed and to record the number of calendars distributed and the number of male and female adults who attended the shows. The campaign was well implemented and popular with villagers. In total, 472 shows took 6 The campaign drew from a vast repertoire of songs written in Rajasthani dialects for awareness campaigns, including songs about elections for get-out-the vote campaigns, and songs about the right to employment for the launch of the MGNREGS. 7 Website: http://prayatn.org/ 9

place (the bigger villages had 2 shows), on average 234 people (32% of the adult population) attended the show in each village, and 130,000 calendars were distributed. 8 Among the adults who attended the show, the proportion of women was 41% overall, with some variation depending on the time of the day when the show was held (only 35% of women after 6pm). Finally, there was substantial variation of attendance rates across GP, with less than 12% of the adults who attended the show in the bottom 10% of GP and more than 55% of adults who attended in the top 10% of the GP. 3 Data 3.1 Data sources This study relies on five different data sources: two administrative sources and three surveys. The first administrative source is the State Election Commission (SEC) from which we collected 2010 electoral results for all 382 GP. In each GP, the SEC records the names, caste and votes received by each candidate, in addition to the name of the winner. Based on this list, we construct measures of the number of candidates in total, by gender, and by caste, as well as their share of votes and their probability of winning. The second administrative source is the official website of the MGNREGS program, which gives information on employment provided, wages paid and expenditures made under the MGN- REGS for each GP, for each financial year, from April 2009 onwards. 9 We use these data to construct indicators of MGNREGS implementation for the financial year April 2009-March 2010, which covers the last year of the incumbent s mandate and the election month, and for the two financial years from April 2010 to March 2012, when the new Sarpanch is in power. In the first survey, 4541 households were interviewed in March-April 2010, i.e. one to two months after the elections in 240 GP (119 treatment GP and 121 control GP). The household survey included questions about households participation in MGNREGS. Each respondent was 8 To obtain these figures we completed monitors reports with census 2001 and 2011 information for each village. Using 2001 census figures, we assumed that 50% of the population was adult, and that each household had 6 members. Finally, we applied to the 2001 village population totals a 19% of decadal population growth rate, which matches 2011 census results in the sample districts. 9 http://nrega.nic.in 10

asked if she had wanted work during the past 12 months and whether or not she was provided employment by the administration of the GP. We use the respondents answers to corroborate official figures on MGNREGS employment during the incumbent mandate. 10 The survey also included knowledge questions about MGNREGS rules and the role of the Sarpanch in MGN- REGS. The second survey, implemented in October 2011, covered all Sarpanch candidates in the 240 GP from the household survey sample. The survey provided both a cross-check on the accuracy of the SEC data on candidates gender and caste and information on candidate education, profession, wealth and political career. In addition, we constructed a family tree for each candidate to identify whether any family member was the incumbent, and more broadly to measure the political experience in the family. Finally, alongside the Sarpanch survey we also implemented an asset survey (in October 2011) in the same sample of GP. We sampled approximately 3200 assets (roads, irrigation canals, pavements, information center) built between April 2009 and March 2011 under MGNREGS from the official website. Once in the villages, enumerators tried to locate the infrastructures, and if found, to code their degree of completion. We use this survey to measure the number of assets completed by the incumbent (April 2009-March 2010) and the new Sarpanch (April 2010-March 2011). 3.2 Sample Depending on the outcome of interest, the delimitation of our sample varies. Some of our outcomes are measured for the whole sample of 382 GP: the number of candidates, by gender and caste and the performance of the new Sarpanch in implementing MGNREGS according to administrative sources. However, most of our outcomes are measured only for the 240 GP in which the three surveys of households, Sarpanch candidates and MGNREGS assets were implemented. Finally, because of rotating reservation for minorities, incumbents could only run in 64% of the sample GP. Hence, our sample is further reduced to 152 GP for outcomes such as the probability that the incumbent re-runs or incumbent s success in the elections. 10 Official numbers are known to be inflated because of widespread corruption in the scheme (Niehaus and Sukhtankar, 2013; Imbert and Papp, 2011) 11

4 Results 4.1 Empirical Strategy and Randomization Balance Given the randomization of both the reservation for women and the voter awareness campaign, the empirical strategy is straightforward. We estimate the following specification for most outcomes: y i = α + βr i + γt i + δt i R i + λ X i + ɛ i (1) where y i is an outcome of interest (e.g. number of candidates in GP i) R i is a dummy for having been reserved for women in 2005, T i is a dummy for receiving the campaign. We also allow for the additional interaction between women s reservation and exposure to treatment. X i is a vector of control variables, which include GP population, total number of voters, caste reservation (OBC, SC, ST) in 2010 elections, and reservation for women in 2000 elections. Table 1 reports village-level descriptive statistics, taken from the State Election Commission and from the 2001 census. Because of the randomization inherent both to our treatment and to the reservation policy, there are few significant differences between the villages 11. In the analysis below, we always include both the 2005 reservation dummy and the campaign dummies, and we also control for 2000 gender reservation and current reservation for OBC, SC and ST. The reservation was well implemented. All incumbents in GP reserved between 2005 and 2010 are female, while only 8% of incumbents in GP not reserved in 2005 were female. As we described above, the voters mobilization campaign was also well implemented. An important question is how voters received the information disseminated in the campaign. The household survey data provides us with some indirect measures of this with questions on voters knowledge of MNREGA rules and of the role of the Sarpanch in implementing the scheme. Table A.1 shows the effect of the campaign : awareness seems rather high (60-65%) in the control and does not seem to increase with the campaign. Optimism about the future benefits of development schemes in general to the village, or whether a good Sarpanch could make a difference, even in a corrupt environment, is also high (76% and 67% respectively). The 11 The very significant negative correlation between reservation for women in 2000 and 2005 elections is due to the rotation of reservation decided by the State Election Commission. 12

average rating of Sarpanch performance in NREGA implementation and public goods provision was also not affected by the campaign. Hence the PEVAC may not have brought knowledge about what the Sarpanch is responsible for in MGNREGS, changed expectectations about the Sarpanch or made voters more critical of Sarpanch in general. As we will see, it may have increased salience of MGNREGS implementation in the elections and induced voters to benchmark the performance of their Sarpanch with that of other Sarpanch, and to judge candidates based on their (potential) ability to deliver MGNREGS employment. 4.2 Incumbency Advantage In Table 2 we start by examining the likelihood that incumbents ran for election, and their electoral success. Column (1) shows that the incumbent was less likely to run in both previously reserved GP and PEVAC GP. In GP with no PEVAC, incumbents from previously reserved GP were 31% less likely to run than their counterparts in un-reserved GP (their probability of running again is less than half of that of the incumbent from unreserved GP, 15% versus 46%). PEVACs were (perhaps) surprisingly effective at inducing incumbents to run: the decline due to the campaign in the probability that the incumbent ran in GP that were not previously reserved is 24%! The campaign did not further discourage women incumbents from running (the interaction between campaign and previous reservation is positive and of similar magnitude as the negative main effect), which may be because incumbent women were not always perceived as established candidates. The incumbent s vote share did go down as a result of the campaign but it is possible that this just resulted in a change of names. The incumbent could have just substituted a family member for himself Querubin (2011) finds evidence of this in the Philippines. To be able to test for this in this context, we collected data on family relationships of each candidate with the past incumbent. We replace the variable did the incumbent run with did the incumbent or any member of his or her family run, and run the same regressions as for the incumbent (Table 2, columns 4 to 6). The impact of reservation on the decision to run and on vote share disappears almost entirely, suggesting that women were replaced by a (typically male) family 13

member. 12 The impact of the campaign on the decision to run is reduced by more than a half, and once again the impacts on vote share and winning are much more muted. 4.3 Incumbent Selection Next, we examine whether incumbent performance influenced her decision of whether or not to re-run. In particular, we may expect greater accountability of incumbents (or incumbent dynasties, since citizens should be able to take into consideration the family strategies). This would imply that both the reservations and the anti-incumbent campaign should lead to a greater correlation between performance and the incumbent (and incumbent plus dynasty) vote share or decision to run. To test this hypothesis, we use survey interview on how many households received employment and whether people got jobs when they wanted them.we then run a regression where we include interactions of the performance index with the campaign: y i = α + βr i + γt i + δt i R i + ai i + br i I i + ct i I i +dt i R i I i + X i λ + ɛ i where I i is the performance index. The results are presented in Table 3. In GP that were neither reserved nor targeted by the PEVAC campaign, there is no positive correlation between the performance index and the decision to run or the vote share of the incumbent: the coefficient is even negative and significant. However, we find that past reservation and the campaign seem to have increased political accountability: both in reserved GP and in GP where PEVAC occurred, there is a significant positive relationship between the decision to run and the employment and the asset index. Better incumbents are less affected by reservation and the campaign: moving from the average performance level to one standard deviation above the mean, the effect of the PEVAC on the probability of running for incumbents shrinks from -24% to virtually zero. In previously reserved places, it goes from -34% to -5%. Conversely, the probability that a bad incumbent with performance one standard deviation below the mean runs goes down to zero in GP previously reserved for women or where the campaign was held. We find 12 We will see below that reservations affects which dynasties get to continue to run and receive votes. 14

similar results when looking at vote shares of incumbents. As we saw for the incumbent results, the impact of reservation and the campaign are not cumulative: the triple interaction between the index, reservation, and the general campaign is negative, offsetting one of the two positive interaction such that overall, there is no additional effect of the campaign on accountability in reserved GP. The results are similar, but slightly muted when we include both the incumbent and the incumbent s family (see Columns 4 to 6 of Table 3). Both previous reservation and the PEVAC campaign encouraged the family of the best incumbents to run. The probability of running increases from 54% to about 80% for family members of incumbents whose performance was one standard deviation above the mean and their average vote share goes up from 11 to 15%. The probability that there was a winner from an incumbent family also goes up significantly with incumbent s performance. 4.4 Candidate Entry Table 4 examines candidate entry excluding the incumbent. Column (1) shows that having been reserved or being exposed to a PEVAC lead to an increase in the number of candidates. Excluding the incumbent, roughly 7 candidates entered the election in GP which were not previously reserved for women. We observe roughly one additional candidate entering the fray in GP which were previously reserved. As Column 4 shows, half of the increase in the total number of candidates comes from candidates who decided to run less than 15 days before the elections. This reflects reflect uncertainty about the process of reservation itself: the reservation status of each GP was announced not more than a month before the elections. Turning to who was induced to enter, as shown in Column (3) the PEVACs attracted a different group of traditional political outsiders: lower castes and tribal candidates (Other Backward Castes, Scheduled Castes and Scheduled Tribes). While political outsiders, these groups are not a minority: in the control group they represent the majority of candidates. There is more muted evidence that the combination of treatment and a previous history of reservation increased entry by landless candidates (Column 5). Both reservation and PEVAC increased the number of literate candidates by about one (Column 6); however, they had no effect on the number of candidates with secondary education 15

(results not shown here). This suggests that they attracted candidates with not more than primary education, as one would expect for members of disadvantaged social groups: female candidates in councils previously reserved for women and lower caste candidates where the PEVAC was implemented. A different measure of political advantage is political experience. While we find no evidence that PEVAC treatment increased entry by candidates who had contested the same position in 2005 (column 7) we do find that the PEVAC increased entry by candidates who were previously elected for the lower position of ward councilor (column 9). As shown in Table 5, the campaign increases the vote share and the probability of winning of non-general candidates (Columns 3-4), of literate candidates (Columns 7-8) and of previous ward councillor (Columns 13-14). As before, the combination of reservation and PEVAC seems to have favored landless candidates, with a higher probability of winning. To summarize, it seems that reservation and the PEVAC had different effects. Reservation widened the pool of candidates, by weakening the incumbent and motivating more candidates to run. By contrast, the PEVAC had no significant effect on the number of candidates, but changed the pool of candidates. The PEVAC attracted candidates from lower caste. They seem to have had a minimum level of education, and some political experience in the lower position of ward councilor. The changing nature of candidate entry raises the question of whether governance improved as a result of this campaign, to which we now turn. 4.5 Post Election Performance Here, we use the official statistics on employment to construct an employment index for each financial year since the elections (from 2010 to 2015) The results are presented in table 6. We find that the campaign had no effect on MGNREGS performance in terms of number of days provided. We find no impact on the number of households participating nor on the allocation of MGNREGS work to SC/ST households. There is no impact of past reservations on performance, and, in keeping with all the previous results, there is no impact of the campaign in previously reserved GP either (once again the interaction is of the opposite sign a magnitude suffcient to undo the negative effect of the campaign in previously 16

unreserved areas). Thus, there is no evidence that broadening the representation base through the PEVAC campaign led to a change in performance. This may be due to a lack of experience or political connections of the newly elected Sarpanch to get things done by the administration in charge of NREGA. This finding needs to be interpreted with caution. For the post election period we only have official statistics of employment, and no household survey data. In the 2009-2010 period, for which we have both sources of information, the two measures are strongly correlated. However, it is widely recognized that there are significant gaps between official figures and household data on MGNREGS employment, due to inflation of person-days by corrupt officials (Niehaus and Sukhtankar, 2013; Imbert and Papp, 2011; Banerjee et al., 2016). 13 Hence we cannot exclude that the absence of effect in the official statistics in fact reflect a decline in corruption and an increase in actual employment. 4.6 Long term Electoral Outcomes We now provide evidence on the long term effects of our intervention on outcomes in the next election, in 2015. The estimating equation is: y i = α + βr i + γt i + δt i R i + λ X i + ɛ i (2) where y i is an outcome of interest (e.g. number of candidates in GP i) R i is a dummy for being reserved for women in 2015, T i is a dummy for receiving the campaign. X i is a vector of control variables, which include GP population, total number of voters, caste reservation (OBC, SC, ST) in 2010 and 2015 elections, and reservation for women in 2000 elections. and regress each outcome on a dummy for current (2015) woman reservation, a treatment dummy and their interaction, controlling for past caste and gender reservation status. The treatment effects on 2015 electoral results are presented in table 7 and 8. The number of candidates was comparable to 2010, but declined in GP that were reserved for women (from eight to seven). The low incumbency rate we observed in 2010 is also present in 2015, it declined even further. This is partly due to rotating women reservation: the state election commission 13 Comparisons across Indian states suggest that the gap is relatively small in Rajasthan (Imbert and Papp, 2011). 17

mandated that one third of all GP should be reserved for women in 2015, but that GP should not be reserved for women for two consecutive elections. As a result, in our sample of GP who were not reserved for women in 2010, two thirds were reserved for women in 2015. This is also due to a new electoral rule that mandated that candidates should have at least eight years of education. Only 2% of incumbents re-ran in GP not reserved for women in 2015. The PEVAC seems to have had long run effects on the candidate pool. Candidates who ran in 2010 were more likely to run again in 2015 in treated councils (table 7 column 3). As column 4 and 5 show, there were also more female candidates and more candidates from lower caste in treated councils. These effects are concentrated in GP not reserved for women in 2015. The increase in the number of female candidates doubled their voteshare in non reserved GP but had no significant effect on the probability that a woman was elected (table 8 column 4 and 5). There is no significant effect on electoral outcomes of low caste candidates (columns 6 and 7). 5 Conclusion This paper studies the effect of two interventions which weakened the incumbent in the context of local elections in Rajasthan, India. We find that both reservation for women in the previous electoral cycle and a voter awareness campaign focused on the village head responsibilities decreased the probability that incumbents runs and wins the election. We measure performance of the incumbent in implementing a government program who provides employment to the poor and show that the worse incumbents were pushed out of the race. We find evidence that the incumbents fought back by letting another family member to run instead, but this did not help the worst performing incumbents. Both interventions attracted new candidates, but while reservation just increased the number of candidates, the PEVAC attracted different candidates, coming from traditionally disadvantaged groups (non-general caste). These candidates were more likely to be literate and to have some political experience acquired in the lower position of councilor. We find no evidence that the newly elected leaders selected in constituencies which had the campaign performed better in terms of employment generated under the same government program. These elected leaders 18

were also no more likely to run and win in the next elections. However, women and lower caste candidates were still more likely to run five years after the intervention. Our results suggest that the PEVAC did not bring about a new generation of successful leaders, but did improve political participation of disadvantaged groups in the long run. 19

References Acemoglu, D. and J. A. Robinson (2008, March). Persistence of power, elites, and institutions. American Economic Review 98 (1), 267 93. Banerjee, A., E. Duflo, C. Imbert, S. Mathew, and R. Pande (2016, November). E-governance, Accountability, and Leakage in Public Programs: Experimental Evidence from a Financial Management Reform in India. NBER Working Papers 22803, National Bureau of Economic Research, Inc. Banerjee, A. V., S. Kumar, R. Pande, and F. Su (2011). Do informed voters make better choices? experimental evidence from urban india. Barro, R. J. (1973). The control of politicians: An economic model. Public Choice 14, pp. 19 42. Besley, T. (2007). Principled Agents?: The Political Economy of Good Government. Number 9780199283910 in OUP Catalogue. Oxford University Press. Besley, T. and S. Coate (1997, February). An economic model of representative democracy. The Quarterly Journal of Economics 112 (1), pp. 85 114. Besley, T., R. Pande, and V. Rao (2012). Just rewards? local politics and public resource allocation in south india. The World Bank Economic Review 26 (2), 191 216. Caselli, F. and M. Morelli (2004, March). Bad politicians. Journal of Public Economics 88 (3-4), 759 782. Chattopadhyay, R. and E. Duflo (2004). Women as policy makers: Evidence from a randomized policy experiment in india. Econometrica 72 (5), pp. 1409 1443. Diermeier, D., M. Keane, and A. Merlo (2005, March). A political economy model of congressional careers. American Economic Review 95 (1), 347 373. Ferejohn, J. (1986). Incumbent performance and electoral control. Public Choice 50 (1/3), pp. 5 25. 20

Ferraz, C. and F. Finan (2008, 05). Exposing corrupt politicians: The effects of brazil s publicly released audits on electoral outcomes. The Quarterly Journal of Economics 123 (2), 703 745. Ferraz, C. and F. Finan (2009, April). Motivating politicians: The impacts of monetary incentives on quality and performance. NBER Working Papers 14906, National Bureau of Economic Research, Inc. Imbert, C. and J. Papp (2011). Estimating Leakages in India s Employment Guarantee. In R. Khera (Ed.), Battle for Employment Guarantee. Oxford University Press. Keane, M. P. and A. Merlo (2010, August). Money, political ambition, and the career decisions of politicians. American Economic Journal: Microeconomics 2 (3), 186 215. Myerson, R. B. (1993, December). Incentives to cultivate favored minorities under alternative electoral systems. The American Political Science Review 87 (4), pp. 856 869. Myerson, R. B. (2011, December). Toward a theory of leadership and state-building. Volume 108, pp. pp.21297 21301. Niehaus, P. and S. Sukhtankar (2013, November). Corruption Dynamics: The Golden Goose Effect. American Economic Journal: Economic Policy 5 (4), 230 69. Osborne, M. J. and A. Slivinski (1996, February). A model of political competition with citizencandidates. The Quarterly Journal of Economics 111 (1), pp. 65 96. Persson, T., G. Roland, and G. Tabellini (1997). Separation of powers and political accountability. The Quarterly Journal of Economics 112 (4), pp. 1163 1202. Querubin, P. (2011). Political Reform and Elite Persistence: Term Limits and Political Dynasties in the Philippines. Unpublished. 21

Table 1: Summary Statistics Incumbent Incumbent or family member Runs Voteshare Wins Runs Voteshare Wins 1 2 3 4 5 6 Women Reserved in 2005 (WR) -0.363*** -10.14*** -0.0421-0.0927-5.102-0.0610 (0.107) (2.722) (0.0551) (0.122) (3.416) (0.0692) Treatment -0.244** -4.795** -0.000240-0.0593-2.594 0.0535 (0.0936) (2.396) (0.0483) (0.107) (3.006) (0.0606) Treatment*WR 0.305** 7.390** 0.00917 0.0127 4.726 0.00808 (0.143) (3.660) (0.0736) (0.163) (4.592) (0.0924) Mean in Control (not WR 2005) 0.462 10.10 0.0577 0.538 11.55 0.0769 Mean in Control ( WR 2005) 0.148 2.505 0.0370 0.481 8.652 0.0370 Observations 152 149 152 152 149 152 District FE Yes Yes Yes Yes Yes Yes GP Controls Yes Yes Yes Yes Yes Yes NOTES: 1.The sample includes all 152 GPs which were surveyed and where the incumbent could re-run in 2010. Information on voteshare is missing for three GPs. 2.GP controls include GP population, literacy rate,, number of registered voters, Gender reservation status of GP for 2000, Caste (OBC SC and ST) reservation status for 2010. 3.Columns 2 and 5 are missing 3 observations since vote data does not exist for these GPs. 22

Table 2: Incumbent success Incumbent Incumbent or family member Runs Voteshare Wins Runs Voteshare Wins 1 2 3 4 5 6 Women Reserved in 2005 (WR) -0.363*** -10.14*** -0.0421-0.0927-5.102-0.0610 (0.107) (2.722) (0.0551) (0.122) (3.416) (0.0692) Treatment -0.244** -4.795** -0.000240-0.0593-2.594 0.0535 (0.0936) (2.396) (0.0483) (0.107) (3.006) (0.0606) Treatment*WR 0.305** 7.390** 0.00917 0.0127 4.726 0.00808 (0.143) (3.660) (0.0736) (0.163) (4.592) (0.0924) Mean in Control (not WR 2005) 0.462 10.10 0.0577 0.538 11.55 0.0769 Mean in Control ( WR 2005) 0.148 2.505 0.0370 0.481 8.652 0.0370 Observations 152 149 152 152 149 152 District FE Yes Yes Yes Yes Yes Yes GP Controls No No No No No No NOTES: 1.The sample includes all 152 GPs which were surveyed and where the incumbent could re-run in 2010. Information on voteshare is missing for three GPs. 2.GP controls include GP population, literacy rate,, number of registered voters, Gender reservation status of GP for 2000, Caste (OBC SC and ST) reservation status for 2010. 3.Columns 2 and 5 are missing 3 observations since vote data does not exist for these GPs.

Table 3: Incumbent Success Depending on Incumbent Performance Incumbent Incumbent or family member Runs Voteshare Wins Runs Voteshare Wins 1 2 3 4 5 6 NREGA Performance Index -0.109* -3.373** -0.038-0.079-3.199-0.052 [0.063] [1.606] [0.033] [0.073] [2.032] [0.042] WR in 2005-0.354*** -10.104*** -0.035-0.069-4.432-0.057 [0.107] [2.719] [0.056] [0.123] [3.440] [0.071] Treatment -0.240** -4.673* -0.003-0.054-2.433 0.052 [0.092] [2.363] [0.048] [0.107] [2.990] [0.061] Treatment*WR in 2005 0.308** 7.499** 0.000-0.043 3.460-0.000 [0.144] [3.690] [0.076] [0.167] [4.669] [0.096] WR in 2005*Index 0.226* 5.189 0.094 0.271* 8.954** 0.097 [0.131] [3.319] [0.069] [0.151] [4.199] [0.087] Treatment*Index 0.242** 6.728** -0.012 0.174 6.025* 0.043 [0.105] [2.672] [0.055] [0.121] [3.381] [0.070] Treatment*WR in 2005*Index -0.475*** -10.898** -0.047-0.297-11.567** -0.087 [0.179] [4.550] [0.094] [0.207] [5.757] [0.118] Mean in Control (not WR 2005) 0.462 10.10 0.0577 0.538 11.55 0.0769 Observations 152 149 152 152 149 152 District FE Yes Yes Yes Yes Yes Yes GP controls Yes Yes Yes Yes Yes Yes Notes: 1.The sample includes all 152 GPs which were surveyed and where the incumbent could re-run in 2010. Information on asset quality is missing in one GP and voteshare is not observed for three GPs. 2. The Performance Index is based on the fraction of household who worked in MGNREGS in the last 12 months and the fraction of men and women who would have liked to work in MG-NREGS in the last 12 months and could get work (household survey). Both measures are standardized and summed up to build the Index. 3.GP controls include GP population, number of villages, fraction of the population who is sc, st, literatem number of registered voters, Gender reservation status of GP for 2000, Caste (OBC SC and ST) reservation status for 2010. 4.The specification is described in Section 4.3