Path-Breakers: How Does Women s Political Participation Respond to Electoral Success? *

Transcription

1 Path-Breakers: How Does Women s Political Participation Respond to Electoral Success? * Sonia Bhalotra University of Bristol Irma Clots-Figueras Universidad Carlos III de Madrid Lakshmi Iyer Harvard Business School May 19, 2013 Abstract We examine whether women s electoral success induces greater female political participation in subsequent elections. Using the regression discontinuity afforded by close elections between women and men, and constituency level data on India s state elections for , we show that electoral victory for a woman leads to no change in female or male voter turnout but to a large and significant increase in the share of female candidates from major political parties in the subsequent election. Approximately half of this increase is attributable to the entry of new female candidates. Empirically testing the implications of a stylized model of political candidacy suggests than an increase in the supply of potential women candidates is the main mechanism driving increased candidacy, rather than a reduction in voter bias against women. * We thank Rafael di Tella and Anandi Mani and several seminar participants for invaluable feedback. Damian Clark, Maya Shivakumar and the staff of Paradigm Data Services provided excellent research assistance. 1

2 1. Introduction Women are under-represented in political office across the world. Women comprise only 11% of India s national legislators, 18% of the members of the United States Congress and 22% of the United Kingdom s House of Commons. Recent empirical work has documented that the political representation of women has a substantive influence on changing policy outcomes towards the interests of women (Chattopadhyay and Duflo 2004, Miller 2008, Iyer et al. 2012, Bhalotra and Clots 2012, Brollo and Troiano 2012). In this paper, we examine women s participation in Indian politics and find that they lag behind men in their participation both as political candidates and as voters. Over the last three decades, women comprised 5.4% of India s state legislators, but only 4.3% of candidates for office, suggesting that candidacy is an important barrier to women holding office. Voter turnout among women was 59% compared to 66% for men, perhaps because of the lack of women candidates on the ballot. 1 In this paper, we examine the extent to which the victory of women in competitive elections results in an increase in female political participation in subsequent elections, both in terms of candidacy and voter turnout. If it does, then we are likely to see spontaneous and sustained increases in women s political inclusion. To account for the fact that the electoral success of women is not randomly allocated, we use a regression discontinuity design that takes advantage of the fact that some women won in close elections against men (Lee 2008, Clots-Figueras 2011, Broockman 2012). Using constituency level data on elections to state legislatures in India during , we find that a female electoral victory results in a significant increase in the fraction of female candidates in the next election in the same constituency. This increase is primarily from major parties, and is not accompanied by any decline in political competitiveness. Approximately half of the increase in the number of women candidates can be attributed to the entry of new candidates, that is, women who did not contest the previous election. There is no impact on female or male voter turnout despite the noted increase in women s candidacy. Our paper makes three substantive advances to the literature on women s political participation. First, while there has been a lot of recent work on the effects of electoral quotas in India, we present the first results in a non-quota setting. 2 This is of substantive importance, both 1 Other studies have shown that voters respond to descriptive representatives on the ballot. For instance, both black and white turnout increase in response to black candidates on the ballot (Washington 2006). 2 Using the random assignment of women s quotas for leadership positions in India s village councils in the state of West Bengal, Beaman et al. (2009) find that women are significantly more likely to stand for and win elected positions in village councils if they had a female chief councilor in the previous two elections. They find evidence of one likely 2

3 because candidates for reserved seats tend to be negatively selected (Chattopadhyay and Duflo 2004, Deininger et al. 2011) and because it is of interest to identify the extent to which a spontaneous dynamic operates in launching women into the political sphere where quotas are not available to kick-start the process. Second, we explicitly bring political parties into the picture. We focus upon elections to state legislative assemblies in which political parties nominate candidates and party bias against women may come into play. 3 This is relevant because even if voters update their perceptions of women, their chances of holding office may be held back by parties limiting their candidacy. Even in the quota setting, party bias against women has been shown to result in limited representation for women in executive positions and therefore a limited policy impact (Casas-Arce and Saiz, 2011; Campa, 2012). Third, we construct a simple theoretical framework closely based on Casas-Arce and Saiz (2011) to distinguish between several different mechanisms which could generate increased female candidacy following a woman s electoral victory. We find that the increases in candidacy are limited to the party from which the woman won the previous election. Testing the implications of a simple theoretical framework helps us to verify that the key mechanism driving increased candidacy is changes in the candidate selection dynamics of political parties, rather than changes in voter perceptions per se. In particular, the most likely channel appears to be a greater willingness of women to come forward as potential candidates within the party where the woman won. Other studies have highlighted similar evidence regarding the reluctance of women to enter politics due to their perception of politics as being extremely competitive or a political career not being attractive to women (Lawless and Fox, 2010; Lawless, 2012; Pande, 2007). Our paper is most similar in empirical approach to that of Broockman (2012), who examines the consequences for turnout and female candidacy, but in neighboring constituencies, not in the same constituency, in US elections. 4 The analysis of the effects of female electoral success on female mechanism, which is that voters update their priors about the acceptability and the effectiveness of women leaders. In a similar vein, Bhavnani (2009) finds large increases in women s candidacy and their chances of winning in reserved constituencies after quotas are lifted, although a subsequent re-analysis shows that the results are much weaker when the sample of wards for which the outcomes are measured include both reserved and unreserved wards, most likely because of discouragement of women candidates in areas that were not subject to quotas (Sekhon and Titiunik, 2012). Deininger et al. (2011) find that people report a lower quality of service delivery in villages reserved for women chairpersons, but a greater ability to hold officials accountable and greater participation in village assemblies; they do not study candidacy, winning chances or voting behavior. 3 This is also a key difference between our analysis and the previous literature on local elections in India. Political party participation in local elections is explicitly restricted in many states. 4 Broockman (2012) was written independently and in parallel with our paper. 3

4 candidacy in the same electoral constituency makes it easier to understand whether the effect is due to a reduction in voter bias, a reduction in party bias, or an increase in the number of women willing to contest for election. Our paper also differs from this paper because we assess the extent to which increases in candidacy reflect the entry of new women, and we investigate in more detail the political viability of female candidates and the relevance of party affiliation. The rest of the paper is structured as follows: Section 2 provides relevant background on the Indian political system, Sections 3 and 4 describe our data and empirical strategy, and Section 5 presents our main results on female political participation. Section 6 outlines a simple theoretical framework to distinguish between different possible mechanisms driving greater female candidacy, Section 7 empirically tests the implications of the model and Section 8 concludes. 2. Indian Electoral Politics India is the world s largest democracy, with a parliamentary system of government at both the central and state levels. Elections are held every five years, on a first-past-the-post system in singlemember constituencies. States may hold midterm elections, before the five-year term of the government expires if the governing coalition loses the confidence of the majority of the state legislature and an alternative government cannot be formed. Elections are very competitive in India, with more than 100 parties participating in the 2009 national elections. We focus on elections at the state level. In India s federal system, state governments have the responsibility for several major development policy areas including law and order, health and education. State level parties play a major role in forming governing coalitions at the center, and previous research has shown that state-level voting behavior is highly correlated with voting in national elections (Ravishankar, 2009). Empirically, the fact that state election dates vary across states, and do not always coincide with national elections, enables us to control for time trends in our regressions. There are currently no quotas for women in state or national level elections. 5 A bill proposing to enact a one-third quota for women at this level has been passed by the upper house of parliament and is expected to be voted on in the lower house as well. Our paper examines the dynamics of female political representation in a non-quota world. 5 A constitutional amendment in 1993 mandated a one-third quota for women in elections to district and village level political institutions. The impact of this reform has been examined in several recent papers including Chattopadhyay and Duflo (2004), Beaman et al (2009) and Iyer et al (2012). Political quotas exist at state and national level for members of the Scheduled Castes and Scheduled Tribes; Pande (2003) and Krishnan (2008) examine the policy impact of these quotas. 4

5 In India s political system, party leaders decide who will be their candidate in every constituency. There are no primaries as in the United States, and the process of choosing candidates is not transparent. Our interviews with several different party leaders indicate that the main consideration is the candidate s ability to win the election (termed winnability in Indian politics). To this end, several metrics including name recognition, party service, financial resources, caste identity and internal party support are considered relevant. Interestingly, very few party leaders felt that women voters were more likely to vote for women candidates, though several expressed the view that women were less interested in politics than men, and that a political career was not particularly attractive to women. State elections in India are fairly competitive. In our data (described in more detail in Section 3), the median number of candidates in each constituency is eight. Of these, however, only 34% (2.7 candidates on average) are from major political parties, which we define as any party which won more than 5 percent of the seats in the state in any year. Only 3 candidates on average manage to obtain more than 5 percent of the total votes, suggesting that the majority of candidates in India s electoral races are not politically viable. In the analysis to follow, we will therefore consistently present results separately for major party candidates and competitive candidates in order to focus on candidates who are politically meaningful Data and Descriptive Statistics We obtained data on state legislative elections from the Election Commission of India over the period We have information on the name, gender, party affiliation and votes obtained by every candidate in every state election held in India, as well as gender-specific voter turnout in each constituency. Over this period, electoral constituency boundaries remained fixed, and therefore we do not have to worry about concerns such as gerrymandering which might differentially affect the electoral prospects for women. We use data for 16 major states of India which account for over 95% of the total population. 7 We have obtained some demographic data (literacy, urbanization) from the 2001 census at the constituency level. For other regional census 6 There is a strong overlap between these categories. The vast majority (87%) of major party candidates are competitive in the sense of obtaining at least 5% of total votes cast. Similarly, about 71% of competitive candidates belong to a major party. 7 In 2001, three new states of Jharkhand, Chhattisgarh and Uttarakhand were carved out of the larger states of Bihar, Madhya Pradesh and Uttar Pradesh respectively. For the first two states, electoral constituency boundaries remained fixed over time; we drop the data from the Uttarakhand state elections of 2002 and 2007 since we are unable to match the electoral constituencies over time. 5

6 characteristics we have matched state assembly constituencies to the census at the level of administrative districts. Women in India are significantly disadvantaged in most domains. In 2007, India was ranked 114 out of 182 countries on the Gender Development Index of the UNDP and had only 940 women for every 1000 men in the 2011 census. Only 65% of women in India were literate in 2011, compared with 82% of men. Women s political participation as voters also lags behind that of men: female voter turnout over was 59%, compared to 66% for men. Women are dramatically under-represented in political office. Our data show that over the period , only 5.4% of state legislators on average were women. However, over this period, only 4.3% of candidates were women and 70% of electoral races had no female candidates at all. This suggests that barriers to candidacy may be a major determinant of the low representation of women in Indian politics. There is a positive secular trend in female political representation, with the share of female legislators rising from 4% at the beginning of the sample period ( ) to 8% in the post-2004 period. The secular trend is visible across all the states of India, though to different extents. In the period, the share of female legislators varied from nearly 10% in the state of Andhra Pradesh to only 4% in the state of Bihar (Figure 1). Only four out of forty major parties in India are headed by women (the Indian National Congress, the AIADMK, the Bahujan Samaj Party and the Trinamool Congress). These parties have a slightly higher share of female candidates (6.67%) compared to the 5.5% for major parties as a whole. Comparing the characteristics of electoral constituencies where female candidates are present to those where there are no female candidates, we find that constituencies with female candidates have larger electorates, a greater number of candidates and a greater likelihood of the incumbent contesting (results available upon request). There is therefore no obvious pattern of women avoiding more competitive races per se. There is also no significant difference in female voter turnout in the previous election across constituencies with more or fewer women candidates, suggesting that potential support from female voters is not a major factor in inducing women to contest elections. The share of female candidates is not significantly correlated with demographic characteristics such as average literacy levels, female-male population ratios or urbanization, but women candidates are more likely to be present in areas with a lower female-male literacy differential. Data on candidate attributes are available only after 2004, when the Election Commission made it mandatory for all candidates to file affidavits giving details such as their educational levels, asset ownership and any pending criminal charges. We have this information for candidates in 14 6

7 out of our 16 states, which held elections between 2004 and Compared to male candidates, women candidates are on average three years younger, significantly less likely to have completed high school, and significantly less likely to have any criminal charges filed against them (results available upon request). In examining the validity of our regression discontinuity strategy, we will verify that none of these characteristics vary across the winners and losers in close elections. 4. Empirical Methodology We are interested in how indicators of women s political participation at the constituency level respond to a woman having won the previous election. The identification problem is that a woman s political victory might be correlated with unobserved characteristics such as voter preferences, which might directly determine women s participation as voters and candidates in subsequent elections. We address this problem using a regression discontinuity (RD) estimator, exploiting the fact that there is a discontinuity in the vote margin at zero, since the candidate with the most votes wins, irrespective of how close the runner up stands. The identifying assumption is that constituencies in which women win in close elections against men are very similar to constituencies in which women lose such close elections, except for the gender of the winning candidate. 8 The estimated model is of the form: (1) Y ist = a + b WomanWon is,t-1 + f(m is,t-1 ) + e ist where Y ist is a measure of female political participation for constituency i in state s in year t. Our main measures of participation will focus on the share of women candidates (overall and among major parties) and female voter turnout. We restrict our sample to elections where a man and a woman were among the top two vote-getters and define M is,t-1 as the margin of victory between the female and the male politician in the previous election in the constituency. WomanWon is,t-1 is a dummy which equals one if M is,t-1 >0. So as to limit the possibility that the discontinuity arises from an underlying non-linearity in the relationship between the indicator of participation (Y) and the vote margin (M), we fit a flexible function of M, f(m is,t-1 ), on either side of the discontinuity. We conduct RD analysis with different parameterizations of the function f. We use polynomial functions 8 See Lee (2008) for the seminal use of the regression discontinuity design using electoral data. Studies which use close elections between men and women include Rehavi (2010), Clots-Figueras (2012a,b), Bhalotra and Clots-Figueras (2012) and Broockman (2012). 7

8 (up to fourth order) on either side of the discontinuity, along the lines of Lee, Moretti and Butler (2004). We also use local linear regressions (Hahn et al, 2001; Imbens and Lemieux, 2007), restricting the sample to an optimal bandwidth around the discontinuity, where the optimal bandwidth is selected by applying the method in Imbens and Kalyanaraman (2011). Finally, we show a robustness check restricting the sample to a very small bandwidth around the discontinuity (Angrist and Lavy, 1999). After presenting the RD estimates using the close elections sample, we present OLS estimates using the entire sample and a typical panel data specification as follows: (2) Y ist = a is + b t + cwomanwon is,t-1 + d X ist + u ist Where a is is a fixed effect for constituency i in state s, b t is a fixed effect for the election cycle, 9 and X ist are other constituency- and time-varying variables. This is a useful way of assessing the external validity of the RD estimates. It also indicates the size of any bias arising from omitted voter preferences. In different specifications, X ist will include state-specific trends, district-specific trends, lagged dependent variables and other time-varying electoral variables (such as the number of candidates or the vote margin in the previous election). 5. Does Women s Political Participation Respond to Prior Electoral Success? 5.1 Regression Discontinuity Estimates We find that a woman s electoral victory leads to a large and statistically significant increase in the share of female candidates in subsequent elections. Table 2, Panel A shows the RD estimates obtained using a quadratic control in the victory margin on either side of the discontinuity, and with controls for state*year fixed effects 10. The share of women candidates increases by 2.4 percentage points after a woman s electoral victory (column 1). These effects are primarily due to the large increase in the share of women among major parties (column 2), where the share of woman candidates increases by 9.2 percentage points (on a mean of 22.7%). These increases in the female candidate share are not driven by an increase in the overall number of candidates. In supplementary regressions, we have verified that the total number of candidates does not increase significantly 9 Since elections are held every five years on average, we use the five year intervals ( , etc) to denote the election cycle. Our results remain unchanged if we include annual year dummies instead. 10 Results are very similar without state-year fixed effects, we will show this in Table 3. 8

9 following a woman s electoral victory, while the number of female candidates does. For the major parties, we clearly see a pattern of female candidates taking the place of male candidates: the number of female candidates increases by 0.17 and the number of male candidates decreases by 0.21, with both of these coefficients being statistically significant at the 1% level (results available upon request). The rise of female candidates does not come with any significant decline in their political viability. The share of women among competitive candidates (those who get at least 5% of total votes cast) also increases significantly (column 3). These effects can also be seen in the graphs in Figure 2, where a clear difference is visible between the female candidate share to the left of the zero point (where a woman lost the previous election) and to the right of the zero point (where a woman won the previous election). We next examine how much of this rise in candidacy comes from the appearance of new female candidates i.e. those who were not candidates in the previous election. 11 In our data, we find that new female candidates account for about half of the observed increase in female candidacy among major parties. The RD estimate for the new female share of major party candidates is (table 2, panel A, column 4), which is approximately half of the percentage points overall increase in female candidacy. Several previous studies have documented a significant degree of incumbency disadvantage in India s national and state level elections (Linden, 2005; Uppal, 2009). This means that a woman winner being significantly more likely to contest the next election is a non-trivial margin in Indian politics. Our interviews with political party leaders also confirmed that the incumbent is never presumed to be the default candidate; a full selection process is conducted before each election with the incumbent usually (but not always) being one of the contenders for the party nomination. In supplementary regressions, we have also examined whether these results are generalizable to longer electoral lags. Using a similar RD specification, we find that a woman s electoral victory two elections ago also results in a statistically significant increase in the female share of major party candidates. The coefficient value, at 0.051, is smaller than for the single lag. This is encouraging in that it suggests long-term changes to the political perceptions of women. However, at even longer lags (i.e. three electoral cycles), the effect reduces to a statistically insignificant 0.022, though we 11 To do this, we manually matched candidates by name across consecutive election cycles, since Indian politician names are often spelt differently in different years of electoral data. In general, about XX% of previous winners contest the next election, but this fraction is only XX% among women. 9

10 should note that the sample size is dramatically restricted in this specification (we have on average 6 elections per constituency). Finally, we examine women s participation as voters by estimating the effects of female electoral victory on female voter turnout. We do not find any changes in female voter turnout in response to women winning elections (column 5); male voter turnout also does not change significantly in the wake of a woman s electoral victory (column 6). We further verify that voter turnout does not change significantly in response to the presence of female candidates in the current election (results not shown). This contrasts with the findings of Washington (2006), who shows that both black and white voters are more likely to turnout to vote when blacks are on the ballot in U.S. elections. 5.2 OLS Estimates We find that the OLS estimates, obtained using the panel data specification (2), are quite close to the RD estimates, though the OLS estimates tend to be slightly larger in magnitude. This serves as a partial test of external validity in the sense that our results on the positive impact of electoral success on future female candidacy hold beyond the sample used for the RD estimation. Table 2, Panel B shows the OLS specification with constituency and election-period fixed effects and districtspecific trends. 12 The OLS regression finds that a woman s electoral victory leads to an increase of 3.6 percentage points in the overall share of female candidates, compared with 2.4 percentage points in RD (table 2, panel B, column 1). Again, most of the effect is concentrated among major parties where the increase is 10.5 percentage points, very close to the RD estimate of 9.2 percentage points (column 2). As in the RD results, the increase in female candidacy is accompanied by an increase in the share of competitive female candidates (column 3). In the OLS specification, we find a smaller effect (0.016) on new female candidates, accounting for only 15% of the overall increase in female candidacy (column 4). Finally, as with the RD results, we find no effect on voter turnout of female or male voters (columns 5 and 6). In Table A.1, we show the robustness of the OLS results for female candidacy in major parties to varying specifications. The results do not vary much when we include state-specific trends instead of district-specific ones (comparing columns 1 and 2), or when we directly control for timevarying characteristics such as the lagged vote margin, the number of candidates in the previous election, and lagged voter turnout (columns 3). Further verification that the results are not driven by omitted characteristics is provided in column 4, when we estimate the equation with state*year fixed 12 An administrative district usually comprises 9-10 electoral constituencies. 10

11 effects, in essence comparing constituencies within each state-election-cycle; in fact, the coefficients are much larger in magnitude in this specification. Results are also robust to controlling for the lagged dependent variable, confirming that the results are not simply a function of persistence in female candidacy (column 5). The coefficients remain very similar when we control for the fraction of female winners in the administrative district as a whole (column 6), suggesting that there are no spillovers of these effects across constituency boundaries. This is consistent with our politician interviews, in which respondents described the candidate selection process as being constituency specific in the sense that local name recognition or local resources were a major determining factor, and parachuting in candidates from outside the constituency happened only for very high-profile candidates. This kind of process also suggests that candidacy dynamics in one constituency might be less likely to be affected by events in neighboring constituencies. Finally, we verify that our results remain strong and robust if we simply use as our dependent variable a dummy for the presence of any female candidate (table A.1, column 7). 5.3 Validity of the Regression Discontinuity Design Robustness to different functional forms We show a series of robustness checks for the RD results on the female share of major party candidates (table 3). Columns 1-4 show the results when we do not include state*year fixed effects and control successively for polynomials up to the fourth order. The estimated coefficient decreases in magnitude as the order of the polynomial increases, but remains statistically significant and positive throughout. Column 5 shows that the results are robust to running local linear regressions restricting the sample to an optimal bandwidth around the discontinuity, as outlined in Imbens and Lemieux (2007). 13 The coefficients also remain strongly significant when we restrict the sample to half of the optimal bandwidth (column 6), and when restricted to a very narrow bandwidth of 0.05 around the discontinuity and we compare the mean of the dependent variable at both sides (column 7). Finally, the estimated coefficients also retain their size and significance when we control for stateyear fixed effects and higher order polynomials (columns 9, 10). As described earlier, the state*year fixed effects specification with quadratic controls (column 8) is our preferred specification for the RD estimation. 13 The optimal bandwidth is

12 5.2.2 Continuity of the vote margin Even if the manipulation of the vote margin is unlikely in this setting, if we observed a jump in the distribution of vote margins at zero, the use of regression discontinuity could be questioned. Figure 3A shows that the density of the vote margin does not show a discontinuity around zero. Following McCrary (2008), we estimate the density of the vote margin on either side of the zero point (Figure 4B), and test explicitly for any discontinuities. As we can see, the densities on either side share a substantial overlap. The estimated discontinuity is with a standard error of , which is statistically insignificant Are there jumps at other points? A different way to verify the validity of this strategy is to verify that the outcome does not vary discontinuously with the vote margin at points other than the zero point, which separates a woman winning from a woman losing. We do this by re-computing the RD estimates, assuming that the discontinuity (i.e. woman winning) happens at points other than zero vote margin. Following Imbens and Lemieux (2007), we choose the median of the two subsamples on either side of the cutoff value as our fake discontinuity points. We find that the RD estimates using these points results in coefficients which are much smaller in magnitude that the actual coefficient for female share of major party candidates, and these estimates are statistically insignificant (table 4, panel A) Do covariates and pre-determined variables show discontinuities? As a further check on the RD design, we verify that none of the demographic covariates (population gender ratios, literacy rates, urbanization rates) vary discontinuously at the vote margin of zero (table 3, panel B). We have also verified that pre-determined electoral variables, such as the total number of female voters, the number of candidates, voter turnout and the number of female candidates from major parties in the previous election, do not show any significant differences at the discontinuity. By definition, RD estimates are local treatment effects that hold for those constituencies with close races in a particular election, but the identity of these constituencies will tend to vary with election year. We investigated this and we find that there is some churning in the sample of constituencies with close elections. Since data pooled across elections may not recover the average of the year by year election effects (Sekhon and Titiunik, 2012), we re-estimated the RD equations by election year. The results, available on request, are very similar across the years. 12

13 5.2.5 Do characteristics other than gender vary discontinuously? The original study of Lee (2008) investigated party incumbency advantage in the US, looking at narrow victories of Democrats against Republicans or Republicans against Democrats. Recent critiques of RD on close elections (Caughey and Sekhon 2011, Grimmer 2011) refer to this context, showing evidence that the incumbent party tends to have systematically greater chances of winning even when elections are close, likely, because of better access to financial and other resources that are used to manipulate votes. This critique is evidently less relevant in our specification as we compare incumbent women with incumbent men. Also the main outcomes of interest are not the chances of re-election of the incumbent, but rather the entry of new female candidates and electoral turnout. We nevertheless address this critique below. We find that winners and losers in close elections between women and men are not significantly different along a range of important characteristics (Figure 4). If this were not the case, we may then mistakenly attribute to gender the characteristics of winners. Figure 4 graphs several different characteristics of winners and runners-up in close elections. The x-axis in all of these figures is the vote margin between the winner and the runner-up, regardless of the gender of the winner. Figure 4A shows that women and men are equally likely to win close elections. Figure 4B shows that winners in close elections are equally likely to have criminal charges filed against them, Figures 4C and 4D verify that there are no differences in education categories or net worth across winners and losers of close elections, finally, Figures 4E and 4F show that winners and losers in close elections are equally likely to belong to major parties or to be the incumbent. This makes us more confident that we are capturing the effects of the gender of the winner and not some other characteristics of close election winners. 6. A Simple Model of Candidate Choice In this section, we discuss and analyze a simple model of candidacy, based closely on the model of Casas-Arce & Saiz (2011). The model incorporates three main barriers to entry of women: bias against women candidates by party officials, bias against women on the part of voters and selfselection by women into politics. We should note that changes in any of these factors would result in an increase in female candidacy following a woman s electoral victory i.e. all of these mechanisms are consistent with the observed results so far. The purpose of writing down a stylized model is to 13

14 generate some empirical implications which might help us in distinguishing the mechanism behind the observed increase in candidacy. Each of the mechanism we model party leaders bias against women, voter bias against women, and women candidates being less likely to come forward-- has been documented to have some empirical relevance. For instance, Lawless and Fox (2010) find that women more likely to perceive politics as competitive and less likely to perceive themselves qualified for office; they are also less likely to be encouraged by party insiders and others to run for office. Casas-Arce & Saiz (2011) use the implications of their model for gender quotas to find that parties in Spain sacrifice vote share by not having female candidates, indicating a substantial role for gender-biased party machines. Beaman et al (2009) show that repeated exposure to women leaders (by means of a quota) significantly changes voter attitudes on the desirability of female leaders, and increased female candidacy in future elections. But their setting, of village level elections in India, does not involve parties as gate-keepers into politics. 6.1 Timing of events in our model In our model, we make the assumption is that party positions are decided before any efforts to find candidates (A1). This is reasonable in the Indian setting, where policy positions are decided by the party leadership at the state level and candidates are then chosen at the constituency level. Potential candidates will take this into account before being in contention for the candidacy. If we allow policy positions to be endogenously determined, we are unlikely to isolate the determinants of female candidacy; such models have been used fruitfully to discuss the policy consequences of women s political representation (Chattopadhyay and Duflo, 2004) but not to discuss whether there will be a woman candidate or note. There is a three stage process to candidate selection. In the first stage, the candidates decide whether to be in contention or not. This creates a pool of potential candidates for the parties to choose from. Here there can be differences between men and women if the cost of entering politics is different for women, perhaps because of social norms regarding the public roles of women, or because women believe the payoff to running is lower or because women perceive a career in politics to be incompatible with work-life balance or family responsibilities. All of these perceptions have been documented for US women in Lawless and Fox (2010). In the second stage, the parties choose a candidate for each constituency, in a process described in more detail below. Finally, the voters vote, taking into account party positions, candidate genders and candidate abilities. 14

15 6.2 Voters There is a continuum of voters with mass 1 and their most preferred policy outcome distributed uniformly along a policy continuum between 0 and 1. There are two parties in our model, with their policy positions exogenously given at 0 and 1. Voters care about policies, candidate ability and gender. A voter with preferred policy x ε [0,1] obtains the following utility by voting for party p ε {0,1}: U(x,p) = C (1/2)*λ* p-x + A p d v *W p where C is a constant and λ measures the extent to which voters penalize deviations of party ideology from their own policy preference. If λ is low, voters do not care a lot about ideology and are more likely to switch based on candidate characteristics; in this sense, elections are more competitive when λ is low. A p is the ability of party p s candidate, W p equals one if party p s candidate is a woman and d v 0 measures the extent of voter bias against women. We should note that this bias is a taste discrimination parameter: candidates ability is known and therefore there is no statistical discrimination (Assumption A2). For simplicity, the voter bias against women is assumed independent of the policy preferences of the voter. In future work, we will work through the implications of a model of statistical discrimination where candidate abilities are observed with noise, and there is greater variance in the observability for women. Voters turn out to vote as long as the net utility of voting is higher than a reservation utility u 0 of not voting at all. For now, we assume that this opportunity cost of voting is low enough that all voters turn out to vote (assumption A3). In our setting, this is without loss of generality, since we do not find any changes in voter turnout following a woman s electoral victory. Given the full turnout assumption, the equilibrium outcome is such that all voters in [0, x m ] vote for party 0, and all voters with ideal points in [x m, 1] vote for party 1, where the marginal voter x m satisfies (3) x m = (1/2) + (1/ λ)* [(A 0 -A 1 ) d v (W 0 -W 1 )] Party 0 gets the vote share x m and Party 1 has vote share (1- x m ). 15

16 6.3 Choosing a Candidate Potential candidates vary in their ability, with ability distributions f pm and f pw for men and women within party p. Candidate ability is initially unknown. We assume that parties draw a random sample from the male and female candidate pool within their parties, after which the candidate abilities are revealed. 14 Once the ability of male and female candidates is revealed, parties choose the one to maximize their utility. We also assume that parties choose candidates simultaneously (assumption A4). Again, this is relevant to the Indian political system where there are no official primaries to enable parties to predict the other party s candidate choice perfectly and all parties have a common deadline to declare an official candidate, typically one month before the election. Parties are interested in winning elections. We therefore assume that they are interested in maximizing their vote share, 15 but that they also discriminate against women in a taste-based way i.e. that party p experiences a disutility d p from having women candidates. After Party 0 obtains candidate abilities a w and a m, it has to decide whether to choose the man or the woman to maximize utility U 0 = x m d 0 W 0. Taking the choices of Party 1 to be given (i.e. W 1 and A 1 are fixed), this works out to: Max (1/t)*A 0 (1/ λ)* d v W 0 d 0 W 0 Party 0 therefore chooses the woman candidate over the man if (4) a w > a m + d v + λ d 0 A similar equation applies to party 1 with d 0 replaced by d 1. This equation suggests three potential mechanisms why we might observe a rise in female candidacy in party 0 after a woman from that party wins an election: (A) a decline in the party s disutility of having a woman candidate (d 0 ), perhaps because party leaders are now favorably impressed by the woman winner, (B) a decline in the voter bias against women (d v ), as voters get 14 This is a simplified representation of the search process within actual Indian political parties. Politicians from several different Indian political parties described the process of candidate selection as follows: parties draw up an initial short list of 2-5 candidates from each constituency, and then embark on serious information gathering exercises, including third-party voter surveys, to assess candidate quality. The main candidate quality emphasized by all parties was the ability to win the election. To this end, several metrics including the candidate s name recognition within the constituency, service to the party, financial resources, caste identity and internal party support were considered relevant. 15 This is without loss of generality, since winning is an increasing function of vote share with a discontinuity when vote share exceeds half in a two-party setting, and a different (endogenous) threshold in a multi-party setting. 16

17 accustomed to having women in leadership positions, and (C) a rightward shift in the ability distribution of potential women candidates, due to the entry of more qualified women candidates (due to the efforts of the woman winner herself to recruit more women, or a more welcoming environment in the party, or a rise in ascendance of woman-friendly factions within the party or changed perceptions about politics among potential women candidates). We now conduct some comparative static exercises with a view to generating empirically testable hypotheses that can separate among these possible mechanisms. By definition, the comparative statics assume that only one of these mechanisms is in operation at any given time. In reality, of course, some of these effects might work simultaneously e.g. a decline in party bias against women might also generate a right-ward shift in the ability distribution of potential female candidates. But this is a modeling exercise to generate separate predictions of each specific channel, and to see if some of these can be tested in the data. We will emphasize that all of these channels will lead to the implication that the probability of choosing a woman candidate is now higher than it was before; the random draw assumption also gives us the implication that the woman candidate in the subsequent election need not necessarily be the incumbent woman herself. In other words, all of these mechanisms are consistent with our candidacy results. We now want to generate predictions beyond this robustly documented fact. 6.4 Implications of reduced voter bias A key implication of reduced voter bias (a reduction of dv) is that the ability threshold for choosing a female candidate is lowered for both parties, so that we expect to see greater female candidacy across all parties. On the other hand, if a woman s victory reduces party bias against women or leads to greater entry of potential female candidates only within the woman winner s party, then we will see an increase in female candidacy only within the party where the woman won and not within other parties. This enables us to distinguish the voter bias reduction hypothesis from the other mechanisms by running the following regression: (H1) Femcand pit = a + b 10 *WomanWon i, t-1 + b 11 *WomanWon i, t-1 * IncumbentParty pit + u pit where Femcand pit is the share of female candidates from party p in constituency i and time t, WomanWon i, t-1 is a dummy which equals one if a woman won the last election in that constituency, and IncumbentParty pit is a dummy which equals one for the party from which the woman won the 17

18 previous year. If voter bias reduction is the only mechanism at work, we do not expect any additional impact for the previous winner s party i.e. b 11 =0. On the other hand, if only an intra-party mechanism is at work, then we expect b 11 >0 and b 11 >b 10. We should note that our lack of any effect on gender-specific voter turnout already suggests that changing voter behavior is unlikely to be the main mechanism at work. But equation (H1) provides a more comprehensive test derived from the model. A second implication of the voter bias reduction hypothesis is that keeping ability constant, women candidates can expect a bigger vote share in the current election. This can be tested using the vote share of non-new female candidates in the current election, since our model assumes that the ability of these candidates is already known to voters from the past election. If these candidates do not command a higher vote share this time around, this would provide further confirmation that voter bias reduction is not the main mechanism at work. We test this by running the following regression (in the RD specification): (H2) Voteshare of non-new female candidates it = a + b 20 WomanWon i, t u it If there is a reduction in voter bias, we expect b 20 >0. On the other hand, if there is no reduction in voter bias, then even if the other two mechanisms are in operation, we do not expect a change in the vote share of candidates whose abilities are already known to the voters. 6.5 Implications of reduced party bias In this section, we focus on generating hypotheses which will enable us to distinguish between a reduction in party bias and a right-ward shift in the ability distribution of potential women candidates. Equation (4) generates the implication that if there is a reduction in d 0, the effect on female candidacy will be higher in places where λ is high i.e. in less competitive races. The intuition for this is that in highly competitive places, indulging anti-women preferences likely comes at an electoral cost and hence is unlikely to be happening in the first place. Given that the RD specification by definition focuses on the highly competitive races, we test this implication using an OLS specification and the following regression: (H3) Femcand it = a + b 30 *WomanWon i, t-1 + b 31 *WomanWon i, t-1 * Competition i,t-1 + u pit 18

19 where we proxy electoral competition in the previous period by the vote margin in the previous election (higher vote margin means less competition). A reduction in party bias would imply that b 31 <0. Note that we would not see this necessarily happening under the increased candidate supply or the reduced voter bias hypothesis. In general, the change in vote share for party 0 between period t-1 and t is given by: V 0,t - V 0,t-1 = (1/ λ)* [(A 0,t -A 0,t-1 ) - (A 1,t -A 1,t-1 ) d v,t-1 (W 0,t -W 1,t ) + d v,t-1 (W 0,t-1 -W 1,t-1 )] Under the candidate supply hypothesis i.e. that more qualified women now come forward within party 0 after that party has a female winner, we have E(A 0,t - A 0,t-1 ) >0 provided party 0 fields a woman candidate both times. If there is no change in voter bias (d v,t = d v,t-1 )and no change in the ability distribution in party 1 (i.e. E(A 1,t - A 1,t-1 ) =0), then we can say that expected change in vote share for female candidates is positive i.e. E [V 0,t - V 0,t-1 ] > 0. Since this will not happen if the same woman runs again (A 0,t = A 0,t-1 ), we get the implication that vote share of new female candidates from the incumbent s party should go up under the candidate supply hypothesis. On the other hand, if party bias reduces, then the ability threshold has been lowered for women potential candidates and new women candidates are on expectation of lower ability than before i.e. E(A 0,t - A 0,t-1 ) <0. We can thus use the following regression to separate between these hypotheses: (H4) Voteshare of new female candidates pit = a + b 40 WomanWon i, t-1 + b 41 *WomanWon i, t-1 * IncumbentParty pit + u pit The increased candidate supply hypothesis implies that b 41 >0, while the reduction in party bias implies b 41 <0. 7. Testing Model Implications We find empirical support for increased within-party candidate supply as the main driving mechanism behind the observed increase in female candidacy. We begin by testing equation (H1), and find that the increase in female candidacy is restricted to the party from which the woman won the previous election (table 5, column 1). The result also holds in the OLS specification, as a confirmatory check (column 2). In our view, this provides strong evidence against the reduction in voter bias channel. Further evidence against voter bias reduction is provided by the fact that the vote 19

20 share of non-new female candidates (i.e. those who were in the electoral race the previous time as well) does not change significantly after a woman s electoral victory (table 5, column 3). Our evidence is consistent with the survey evidence in Beaman et al (2009), which documents a change in voter views about women in leadership roles only after repeated exposure to women leaders and no change over a single electoral cycle. We also do not find evidence to support the reduction in party bias hypothesis. The key implication that female candidacy increases will be higher in less competitive places is not borne out by the data, which shows no relationship between female candidacy in the wake of electoral victory and the extent of competition in the constituency, proxied by the vote margin in the previous election (table 5, column 4). 16 We do, however, find strong evidence for hypothesis H4. The vote share of new female candidates shows a significant increase within the party from which the woman won last time (table 5, column 5). Note that this is not simply an incumbency advantage, since by definition the incumbent is not running this time. We provide an important caveat to this candidate supply hypothesis, namely that the hypothesized increase is purely within-party rather than across the board. We see that there is no increase in the number of independent women candidates following a woman s electoral victory (table 6, column 1). A further test compares the increase in candidacy before and after the implementation of gender quotas in local government elections. We do not find a bigger increase in female candidacy after the generation of this increased potential candidate pool via the quotas at lower levels of government (table 6, column 3). 8. Conclusions We document that a victory by a female candidate in Indian state elections has a large, robust and statistically significant effect on the political candidacy of women in future elections. These effects are particularly large for major parties, and increased candidacy does not come at the cost of political viability. About half of the increase in major party female candidacy is attributable to the entry of new women who had not contested the previous election. Our simple theoretical framework generates testable implications to distinguish between several possible mechanisms which might be driving increased female candidacy. Our empirical evidence supports the hypothesis 16 Other measures of competition, such as the Herfindahl index of party vote shares, also do not show a robust relationship with candidacy increases. 20

21 of an increased supply of female candidates within the party after a woman s electoral victory, and convincingly rejects the implications of a reduction in voter bias. Our results are important in showing that barriers to women s candidacy in politics can be significantly lowered by the power of a good example. Whether this can act as a justification for promoting electoral quotas for women depends whether quotas would also result in an increased supply of qualified candidates. On the one hand, if women perceive electoral politics as being unduly competitive, the reduction in competition due to quotas might have a similar effect. On the other hand, if women do not come forward due to other factors which make a political career unattractive to them, then quotas might need to be complemented with other changes to effect a sustained increase in women s political candidacy. 21

22 References Angrist, Joshua and Victor Lavy, 1999, Using Maimonides' Rule to Estimate the Effect of Class Size on Scholastic Achievement", Quarterly Journal of Economics, 114 (2). Bagues, Manuel and Berta Esteve-Volart Are Women Pawns in the Political Game? Evidence from Elections to the Spanish Senate, Journal of Public Economics 96, pp Bagues, Manuel and Berta Esteve-Volart Can Gender Parity Break the Glass Ceiling? Evidence from a Repeated Randomized Experiment, Review of Economic Studies 77(4), pp Beaman, Lori, Esther Duflo, Rohini Pande and Petia Topalova Political Reservation and Substantive Representation: Evidence from Indian Village Councils. India Policy Forum 7, Beaman, Lori, Esther Duflo, Rohini Pande and Petia Topalova Female Leadership Raises Aspirations and Educational Attainment for Girls: A Policy Experiment in India. Science: January 12, Beaman, Lori, Raghabendra Chattopadhyay, Esther Duflo, Rohini Pande and Petia Topalova Powerful Women: Does Exposure Reduce Bias? Quarterly Journal of Economics 124(4): Bhavnani, Rikhil Do Electoral Quotas Work After They Are Withdrawn? Evidence from a Natural Experiment in India. American Political Science Review 103 (1): Broockman, David E. (2012). Do Female Politicians Empower Women To Vote or Run for Office? A Regression Discontinuity Approach. Mimeograph, Department of Political Science University of California, Berkeley Brollo, Fernando and Ugo Troiano What Happens When a Woman Wins a Close Election? Evidence from Brazil, Working Paper, Harvard University. Campa, Pamela Gender Quotas, Female Politicians and Public Expenditures: Quasi- Experimental Evidence. Working Paper, Stockholm University. Casas-Arce, Pablo and Albert Saiz Women and Power: Unwilling, Ineffective, or Held Back? IZA Discussion Paper No Caughey, Devin and Jasjeet Sekhon (2011). "Elections and the Regression-Discontinuity Design: Lessons from Close U.S. House Races, " Political Analysis, 19 (4): Chari, A.V Gender, Social Norms and Voting: Female Turnout in Indian State Elections. RAND Working Paper. 22

23 Chattopadhyay, Raghabendra and Esther Duflo Women as Policy Makers: Evidence from a Randomized Policy Experiment in India. Econometrica 72(5): Clots-Figueras, Irma Women in Politics: Evidence from the Indian States. Journal of Public Economics 95: Deininger, Klaus and Hari Nagarajan Does Female Reservation Affect Long-Term Political Outcomes? Evidence from Rural India. World Bank Policy Research Working Paper Grimmer, Justin., Eitan Hersh, Brian Feinstein and Dan Carpenter (2011), Are Close Elections Random? Mimeograph. Harvard School of Government. Hahn, J., Todd, P., Van Der Klaauw, W., Identification and estimation of treatment effects with a regression discontinuity design. Econometrica, Vol. 69. Hughes, Melanie M. "Intersectionality, Quotas and Minority Women's Political Representation Worldwide," American Political Science Review 105 (3): Imbens, Guido and Thomas Lemiuex, 2007, Regression discontinuity designs: A guide to practice", Journal of Econometrics, doi: /j.jeconom Imbens, Guido, and Karthik Kalyanaraman Optimal Bandwidth Choice for the Regression Discontinuity Estimator. Review of Economic Studies. Iyer, Lakshmi, Anandi Mani, Prachi Mishra and Petia Topalova The Power of Political Voice: Women s Political Representation and Crime in India. American Economic Journal: Applied Economics, 4(4): Lawless, Jennifer L. and Richard L. Fox. (2010) It Still Takes A Candidate: Why Women Don't Run for Office, New York: Cambridge University Press. Lee, David S., Enrico Moretti, and Matthew J. Butler Do Voters Affect or Elect Policies? Evidence from the U.S. House. Quarterly Journal of Economics, 119(3), McCrary Manipulation of the Running Variable in the Regression Discontinuity Design: A Density Test, Journal of Economic Literature 142 (2). Pande, Rohini and Deanna Ford Gender Quotas and Female Leadership: A Review, Background paper for the World Development Report on Gender. Pande, Rohini. Understanding Political Corruption in Low Income Countries. Chapter for the Handbook of Development Economics, Volume 4, April Rajaraman, Indira and Manish Gupta Public Expenditure Choices and Gender Quotas. Working Paper, Indian Statistical Institute, New Delhi. 23

24 Ravishankar, Nirmala The Cost of Ruling: Anti-Incumbency in Elections. Economic and Political Weekly 54 (10): Rehavi, Marit Sex and Politics: Do Female Legislators Affect State Spending? Working Paper, University of British Columbia. Sekhon, Jasjeet S. and Rocio Titiunik When Natural Experiments are Neither Natural Nor Experiments. American Political Science Review 106 (1): Snyder, Jason Detecting Manipulation in U.S. House Elections." Unpublished manuscript. Washington, Ebonya Female Socialization: How Daughters Affect Their Legislator Fathers' Voting on Women's Issues. American Economic Review 98(1): Washington, Ebonya How Black Candidates Affect Voter Turnout. Quarterly Journal of Economics 121(3):

25 Figure 1 Fraction of Female Candidates in Major Indian States,

26 Figure 2: Regression Discontinuity Estimates for Female Political Participation A: Fraction of female candidates B: Female share of major party candidates C: New female share of major party candidates D: Female voter turnout

27 Figure 3: Testing the Continuity of the Victory Margin between Female and Male Candidates A. Density of the victory margin B. Testing for density discontinuities at zero (McCrary Test) Density margin of victory Notes: Sample restricted to those where a man and a woman were the top two vote-getters in the election. Discontinuity estimate in Figure B (log difference in height): ( s.e. = )

28 Figure 4: Characteristics of Winners and Losers in Close Male-Female Elections A. Gender of winning candidate B. Any criminal charges filed against candidate C. Candidate has less than high school education D. Log (net worth) E. Candidate belongs to a Major Party F. Candidate was the incumbent