Education, Women's Empowerment and Political Selection. November 2015 Preliminary. Duha T. Altindag Auburn University

Education, Women's Empowerment and Political Selection November 2015 Preliminary Duha T. Altindag Auburn University altindag@auburn.edu Naci Mocan Louisiana State University, NBER, IZA mocan@lsu.edu Abstract: Using data on candidates in Turkish elections of 2002, 2007 and 2011, we investigate the impact of voters preferences over the sex of politicians on the selection of male versus female candidates for political office. Our identification strategy, which is implemented in two steps using two different data sets similar to the two sample IV method, utilizes the change in sex preferences of voters due to an increase in their education. In the first stage, using individual level voter data, we estimate the effect of voter s education on their preferences over the sex of politicians. We find that an increase in female (male) voters educational attainment reduces (increases) their tendency to prefer male politicians. In the second step (reduced form), using city level data, we estimate the effect of average education in the city on the share of male candidates nominated. Results show that increasing average female (male) voter s education in a city reduces (increases) the share of male candidates nominated in that city. We recover the effect of voters preferences on politician sex by computing the ratio of the coefficients of education in the reduced form to that in the first stage. We find that the elasticity of the share of male candidates nominated by political parties with respect to voters preference for male politicians is about 1. Parties respond to preferences of male and female voters equally. We also find evidence that a decrease in preference for male politicians in a city increases ex-ante election chances of female candidates. 1

1. Introduction Recent research highlights the significantly different economic and social implications of female versus male presence in governing bodies. For example, Chattopadhyay and Duflo (2004) show that compared to their male counterparts, female political leaders are more likely to provide public goods that reflect women s preferences. Evidence presented by Clots-Figueras (2012) and by Bhalotra and Clots-Figueras (2014) suggests that an increase in number of female officials improves educational attainment and health outcomes of infants at birth, respectively. Swamy et al. (2001) show that the number of female parliamentarians is negatively correlated with corruption in a country. Despite these favorable consequences associated with female political leaders, underrepresentation of females in political positions is common world-wide. For example, between 2000 and 2013, the world average of share of female politicians who held a seat in the parliament was only 16%. 1 Such low number of female officials around the world suggests that there is room for improvement through increasing female representation in political positions. A necessary condition for election is candidacy. In countries, such as Turkey, where a party-list election system is used, the political parties nominate individuals for parliamentary seats. 2 That is, parties have the sole power to determine the sex-mix of their candidates. In this paper, using Turkish data that span the period between 2002 and 2011, we investigate the effect of voters education on voters preferences over the sex of the politicians and the sex-mix of the candidates nominated by political parties. Although education s effect on economic outcomes such as earnings (Brunello Fort and Weber 2009, Oreopoulos 2006, Duflo 2001), health 1 Source of these statistics is the World Bank s World Development Indicators database. Underrepresentation of females in other administrative positions is also common. For example, in an average country, the share of firms with a female top executive was 17%. 2 In Turkey, proportional representation is used. Seats are allocated to parties through the d Hondt method. A more detailed discussion of the election system is presented in Section 2. 2

(Grossman 2015, 2006), crime (Lochner and Moretti 2004), and economic growth (Hall and Jones 1999) has been studied extensively, no paper investigated the possibility that educational attainment of voters changes the sex-mix of the politicians. Ours is the first paper that tackles this question. Educational attainment of voters can affect the sex-mix of politicians based on two premises. First, political parties strategically nominate candidates that reflect voters preferences over politician sex. This is because, sex of the candidates is likely to be a determinant of votes, as the sex of the elected officials affects the implemented policies. 3 For example, women voters may be more likely to prefer female politicians, as female politicians are more likely to implement woman-friendly policies (Chattopadhyay and Duflo 2004). 4 As political parties objective is to maximize their votes, they have incentives to nominate candidates with characteristics consistent with voters preferences. Specifically, if the marginal voter, who is otherwise indifferent between ideologies of two parties, is more likely to vote for a certain party because that particular party nominated candidates of a certain sex that are more favorable to the voter, then all parties would have tendency to nominate candidates of that sex. Second part of the link that combines voter education to the sex-mix of politicians calls for a change in voters preferences due to their educational attainment. Education may influence voters preferences over the sex of politicians through its impact on dependence of one on their partner or the other sex. If education increases one s independence by empowering them (for example by improving their labor market prospects), then individuals who are more educated are 3 Median Voter Theorem, first proposed by Downs (1957) who predicts that identity (i.e. sex) of the elected person should have no effect on the policies of the government. Instead, the policy preference of the median voter will be implemented. This is in contrast to the citizen-candidate model, a framework developed by Besley and Coate (1997) and Osborne and Slivinski (1996). They suggest that implemented policies will diverge from the median voter s preferences towards the preferences of the elected person. Consequently, the characteristics of the elected official are among the determinants of the policies implemented. 4 Chattopadhyay and Duflo (2004) show that reservation of political leadership positions for women in Indian villages lead to greater public spending that reflect women s preferences. 3

more likely to prefer politicians of their own sex. As an example, consider the marital unions which generally involve an income transfer from the higher earning partner to the low earning one. In Turkey, where men have better labor market prospects than women do 5, the direction of the transfer is from men to the women. As a result, married women who are dependent on their partners benefit indirectly from policies that improve men s earnings. Such women are more likely to prefer politicians who implement policies that benefit men, i.e. male politicians. Obtaining more education increases women s independence by improving their labor market opportunities and therefore they benefit more from policies that favor women, female politicians. In the Appendix, we demonstrate an example where women s dependence on men is reduced due to an increase in their education (market wages), and this consequently leads female voters to prefer female candidates over male candidates. Consistent with this idea, Edlund and Pande (2002) show that reduction in marriage rate leads women to prefer more leftist (redistributive) policies. Turkish education reform is implemented in 1997 with a law that increased the mandatory years of schooling from 5 to 8. Specifically, the law required individuals born in 1986 or later to stay in school until they complete the 8 th grade (Cesur and Mocan 2015). Older individuals are not affected by the reform. Several accounts consider the law as a surprise that created an exogenous shock in educational attainment of the young Turks. This reform has been used to investigate the effect of education on wages (Mocan 2015); fertility and birth outcomes (Dinçer, Kaushal and Grossman 2014); mortality and birth rates (Cesur et al. 2015); and religious beliefs (Cesur and Mocan 2013) in Turkey. 5 For example, according to Turkish Statistical Institute s Household Labor Survey in 2010, the average full time male worker earns 10-20% more than his female counterpart does. 4

Our empirical analysis is conducted in two steps using two different samples. This is similar to the two sample instrumental variables method suggested by Angrist and Krueger (1992). Specifically, in the first stage, using a nationally representative individual level data set (the World Values Survey), we estimate the effect of education on voters sex preferences. In estimations, we recognize that voters education could be endogenous. To avoid the potential bias, we use the education reform that took place in 1997 in Turkey as an instrumental variable. This is our first stage regression. In the second step, using a district level data set, we estimate the effect of the educational attainment of the average voter in the district on the sex-mix of candidates nominated by the parties. This is the reduced form regression. Finally, computing the ratio of the coefficient of education in reduced form to that in the first stage allows us to recover the effect of voters preferences over politician sex on the sex-mix of candidates nominated by the political parties. Our results show that education increases voters bias against the opposite sex politicians. That is, men (women) who obtain more schooling have greater preference for male (female) politicians compared to men (women) who are less educated. Specifically, for every year of completed schooling, a male voter s preference for male politicians increases by 7%. On the other hand, when a female voter acquires one additional year of schooling, her preference for a male politician decreases by about the same amount. In addition, results obtained from the city level regressions suggest that an increase in female (male) voters educational attainment reduces (increases) the share of male candidates nominated by political parties. We find that parties strategically nominate a mix of female-male candidates that reflects voters preferences over sex of politicians. A one percent increase in preference for male politicians of the average voter in a city increases the share of male candidates in a party s ticket by about 1%. Parties are equally 5

responsive to the sex preferences of male and female voters. We also show that parties alter the sex composition of candidates in ranks and/or districts where election probability is higher. That is, an increase in the demand for politicians of a certain sex increases that sex s chances of election. Our paper has numerous contributions. First, this is the first paper to study the effect of education on individuals preferences over politician sex. Secondly, we add to the research that investigates the determinants of female political representation. Previous papers have mainly focused on the effect of gender quotas (De Paola, Scoppa and Lombardo 2010; Beaman et al. 2009; Bhavani 2009; Frechette, Maniquet and Morelli 2008), co-worker discrimination (Gagliarducci and Paserman 2012). Ours is the first paper that directly estimates the effect of voters preferences over politician sex on the selection of females versus males for the political office. A close paper to ours is the one by Esteve-Volart and Bagues (2012) who study Spanish elections. They argue that voters have a taste for female politicians, and parties nominate more female candidates to attract more votes. However, parties are likely to nominate women candidates in positions where their election is unlikely. This strategy of parties improves election chances of their male candidates. Their findings suggest that party bias rather than voter bias drives this outcome. In our paper, we use a direct metric of voters preferences over politician sex results unlike Esteve-Volart and Bagues (2012) who use parties vote shares as the measure of voters sex preferences. Our findings imply that an increase in voters preferences for female politicians increase female candidates election chances. In the remainder of the paper, we provide background information about Turkish elections in section 2. Section 3 describes our data sets. In section 4 and 5, we explain our 6

empirical framework and discuss the empirical results, respectively. In section 6, we overview a summary of the paper and conclude. 2. Background Information on Turkish Elections Turkey uses a party-list proportional representation system with a 10% national threshold for representation. 6 In this system, each party determines and officially announces a list of N candidates (ordered from 1 to N) in a given district prior to the election, where N is the number of contestable seats in that district. Party leaders choose the candidates freely. Voters cast ballots for a particular party, not for any given candidate in 85 electoral districts. Using the d Hondt method, a seat-allocation formula based on vote-shares of parties, the number of seats won by each party is calculated. Altindag and Mocan (2015) utilize this seat-allocation formula to study the party switching behavior of elected politicians. Once the number of seats won by each party is determined in each district, parties send their winning candidates from their list to the Parliament in descending order. As an example, consider the Table 1 which depicts the competition of 3 parties (A, B and C) in a district with 5 seats. This is the median district in Turkey. Each party nominates 5 candidates as there are 5 contested seats in the district. Suppose the vote shares of the parties A, B and C are 50, 30 and 20, respectively. d Hondt system divides each party s vote share by consecutive numbers 1 through 5. The parties with the largest 5 quotients win the seats in the district. In this example, the largest 5 quotients are 50, 30, 25, 20 and 16.7. These quotients are underlined in Table 1. Since party A has 3 of the 5 largest quotients (50, 25 and 16.7), they obtain 3 seats. As a result, party A s top 3 candidates are elected. Parties B and C each has one of 6 That is, only those parties that have obtained at least 10% of the nation-wide vote are represented in the parliament, regardless of the outcome in any particular district. 7

the largest 5 quotients (30 and 20, respectively). Consequently, they acquire one seat each, and their first ranked candidates are elected. In this election system, candidates rank and the number of contested seats in the district and the number of competing parties determine candidates ex-ante election chances. Figure 1a depicts a first ranked candidate s probability of election as a function of their party s vote share in the hypothetical district with 5 seats mentioned above. All first rank candidates in that district are elected with certainty if their party s vote share is greater than 16.7%. That is, the share of votes sufficient for election of a first ranked candidate is 16.7%. Figure 1b depicts the election probabilities of all candidates. The greater the rank of the candidate is, the higher the sufficient share of votes is and the worse chances a candidate has for election ex-ante. In Figure 1c, we present the ex-ante election chances of candidates in a district with 4 contested seats. Notice that sufficient vote share for all candidates is greater in this district. Following the formulation provided by Palomares and Ramirez (2003), we compute vote share sufficient for an candidate s election (the minimum proportion of votes required for guaranteed election). Specifically, it is calculated by: (1) Sufficient % vote = 100 Rank Seats+1 where Rank is the candidate s rank and Seats is the number of contested seats in the district. Greater the Sufficient % Vote is, the worse election chances a candidate has. 3. Data 8

We use several data sets in our empirical analysis. The source of the individual level data is the 2001, 2007 and 2011 waves of the World Values Survey 7 which contains information of nationally representative samples of 1,500 to 3,000 individuals in Turkey. The surveys include information about the key variables that we employ in our estimation. For example, besides personal characteristics such as sex, income, marital status and so on, these data include variables about the educational attainment of the individual, as well as their year of birth and whether they agree with the statement: On the whole, men make better political leaders than women do. We use this question to construct voters preferences for male politicians. Table 2 provides the summary statistics of the key variables in this individual level data set. Specifically, Panels A and B of Table 2 show the average educational attainment of female and male voters, respectively. The first (second) large column in each panel reports average years of completed schooling of the voting age individuals who are unaffected (affected) by the reform. There is no statistics for the Affected column in 2001, since, as of then, no individual born in 1986 or later was old enough to vote. 8 Consistent with the findings of previous papers (for example, Cesur and Mocan 2014), the statistics in Table 2 points to a positive correlation between the reform and individual s educational attainment. Specifically, the education reform increased individuals average years of schooling by about 3 years. 9 Table 2 also demonstrates that, between years 2007 and 2001, compared to their unaffected counterparts, female voters who were affected by the education reform became 0.11 less likely to agree with the statement: On the whole, men make better political leaders than women do. On the contrary, treatment by the 7 http://www.worldvaluessurvey.org/wvs.jsp 8 Voting age is 18 in Turkey. The oldest cohort that is affected by the reform was 15 years old in 2001 and they could not vote in the 2002 elections. 9 For example, in 2007, education of females who were exposed to the reform increased by 4.6 years (from the baseline of 6.6 in 2001), while education of females who were not subject to the reform increased by 1.3 years. That is, female voters who were treated by the education reform completed 3.3 more years of schooling. Similarly, education of male voters who were treated by the reform increased by 3.7 years compared to male voters who were not treated. 9

education reform increased male voters preference for male politicians by 0.04 in 2007. That is, more educated voters were more likely to prefer politicians of their own sex. In each election, political parties determine and nominate candidates prior to the elections. For example, if a city is apportioned N seats in the Parliament, then each party that competes in that particular city nominates N candidates on their tickets. These candidates are published in Resmi Gazete, which is the official gazette of the state. From this journal, we obtained the names of the candidates on parties tickets in each city. This list of candidates does not contain any information about their sex. In order to identify whether a candidate is male or female, we use the Turkish Language Association s (Turk Dil Kurumu) Names Dictionary (Kisi Adlari Sozlugu). 10 Besides describing the meaning of the name, this dictionary identifies each name with a sex. Some names are unisex, while some are exclusively male or female. We identified a candidate to be male, if their first name includes a name associated with males in the Names Dictionary. The source of all city level data is Turkish Statistical Institute which is the official institution for collecting and distributing economic and social data in Turkey. The average years of education data are obtained from the 2000, 2008 and 2011 censuses. These censuses contain information about the number of individuals with certain degrees by age group and sex in each city. 11 Using these data, we computed the average years of schooling of male and female voters in each city. Table 3 provides summary statistics of the key variables in aggregate level analysis. For example, in 2007, education of the female voters who were subject to the education reform (the affected) is 3.1 years greater than those who were not exposed to the reform (unaffected). 10 Turkish Language Association is the official authority in Turkish language. They publish the official Turkish Dictionary and do research in Turkic languages. http://tdk.gov.tr/ 11 For example, we know that in Istanbul, among 120,000 females aged 20-24, 12,000 had a college degree or more, while in Ankara among 55,000 males aged 50-54, 8,000 had a middle school diploma. 10

Table 3 also shows that the share of male candidates in political parties tickets is decreasing between years 2002 and 2011. This decrease accompanies an increase in education level of both male and female voters. However, note that the increase in female voters education is greater than the increase in male voters education. In our regressions, we control for characteristics of individuals/districts that could be related to their sex preferences/sex-mix candidates. For example, the individual level controls include individual s age, income, their employment status and their self-positioning on a political spectrum that ranges between most left (1) to most right (10). In district level analysis, we control for share of votes obtained by center-right parties in the most recent local government elections (% Votes Right Partiest-1), the share of seats in the city won by female politicians in the last 25 years (Past Female Representation), number of individuals incoming from abroad per capita (Incoming International Travelers), ratio of salaries of the members of the Parliament to the average income in the city where the candidates are nominated (Relative Salary), average age of the voters (Avg. Voter Age), male and female employment rate and the share of the candidates who has a college degree or more. The descriptions and summary statistics of the full set of control variables that are used in the individual level and the aggregate level regressions are presented in Panels A and B of Table 4. 4. Empirical Framework Effect of Education on Voters Preference for Male Politicians In order to investigate the effect of individual s educational attainment on their preferences for male politicians, we estimate the equation depicted below: (2) D i M = β M Edu i + X i γ M + m i 11

(3) D i F = β F Edu i + X i γ F + f i where D i M and D i F are preference (demand) for male politicians of individual i for Male and Female voters, respectively. These variables are equal to 1 if the individual agrees or strongly agrees with the following statement: On the whole, men make better political leaders than women do. Edu i is individual i s completed schooling in years, and X i is a vector observable characteristics of the individual. We estimate these equations separately for male and female voters, so that the effect of education on preferences for male politicians differs by sex (β M and β F ). In (2) and (3), Edu i could be endogenous. Specifically, unobserved determinants of preference for male politicians could be correlated with schooling. For example, a conservative individual could value education less, and at the same time they may have greater discriminatory feelings towards women. Alternatively, more patriarchal families may invest in their sons education more than their daughters education, and at the same time these sons and daughters of patriarchal families may have greater preference for male politicians. To guard against the possible bias due to the endogeneity of education in equations (2) and (3), we use the variation in individuals educational attainment due to the Turkish education reform. The reform is implemented by a law that increased the mandatory years of schooling in 1997. According to the law, individuals born in year 1986 or later had to complete 8 years of compulsory education rather than 5 years. Numerous papers have used this law as an instrumental variable for individuals education (Mocan 2015; Dinçer, Kaushal and Grossman 2014; Cesur et al. 2015; Cesur and Mocan 2013). 12 Because the voting age in Turkey is 18, 13 the law did not affect the educational composition of the voters in the 2002 elections. This is 12 A detailed discussion of the exogeneity of this law is discussed in Cesur and Mocan (2015). 13 That is, only the individuals who are 18 years of age or older are eligible for voting. 12

because, the oldest cohort who were subject to the law in 2002 was only 16 years old. However, some of the voters in 2007 and 2011 elections were subject to the reform. To operationalize this instrumental variables strategy, we employ a dummy variable, Treated i, in our regressions as instrument. Treated i takes the value of one if individual i was subject to the Turkish education reform. The exclusion restriction is that the Turkish education reform influences individuals preference for male politicians only through changing their educational attainment. In the results section, we show that this instrument is highly relevant for individual s education. Effect of Voters Sex Preferences on Nomination of Male Candidates The determinants of the proportion of male candidates nominated by the political parties can be represented by the equation below: (4) Male Candidates ct = ρ F D ct F + ρ M D ct M + W ct δ + ε ct where Male Candidates ct denotes the share of male candidates nominated by parties in city c at election t. D ct F and D ct M stand for the preference for male politician of the average female and male voters, respectively. Their effects on the sex-mix of the candidates are denoted by ρ F and ρ M, respectively.w ct is a vector of other determinants of the candidates sex-mix. We do not have a metric for D ct F or D ct M (the average preference for male politicians of male and female voters) in equation (4), therefore, it is not directly estimable. Note that D ct F and D ct M F M can be written as function of D i and D i (whether female and male voter i prefers male politicians): 13

(5) D ct F = ( FF i=1 D i F ) ct FF ct and D ct M = ( MM i=1 D i M ) ct MM ct M where FF ct and MM ct represents number of female and male voters in city c at time t, and D i F and D i are equal to one if voter i agrees with the statement: On the whole, men make better political leaders than women do. M F We substitute the expressions of D i and D i in equations (2) and (3) into the expressions in (5). This substitution results in equations (6a) and (6b) below: (6a) D ct F = β FF F ( i=1 Edu i) ct + ( FF i=1 XiγF ) ct FF ct FF ct + error ct (6b) D ct M = β MM M ( i=1 Edu i) ct + ( MM i=1 XiγM ) ct MM ct MM ct + error ct where the average preference for male politicians of female and male voters in a city (D ct F and D ct) M is written as a function of the average education level of male and female voters and other characteristics of male and female voters, X i. Some of variables in X i are not available for males and females separately at the aggregate level. Therefore, to operationalize equations (6a) and (6b) in our empirical analysis, we use modified versions of the equations (6a) and (6b) as depicted in equations (6a ) and (6b ): (6a ) D ct F = β FF F ( i=1 Edu i) ct FF ct + X ct γ + error ct (6b ) D ct M = β MM M ( i=1 Edu i) ct MM ct + X ct γ + error ct where X ct is the average voter characteristics in in city c at election t. 14

Finally, inserting these expressions of D ct F and D ct M (average voter s preference for male politician) into equation (4) yields our estimation equation: (7) Male Candidates ct = ρ M β M ( i=1 Edu i) ct + ρ F β F ( Edu i) MM ct MM FF i=1 FF ct ct + W ct δ + ε ct where the first and second variables on the right hand side are average educational attainment of male and female voters. W is a vector containing W, the determinants of the candidates sexmix (in equation 4), and determinants of individuals politician sex preferences, X ct (in equations 6a and 6b ). Unlike equation (4), equation (7) is estimable, as we can observe all variables on the right-hand side. Running this regression provides an estimate of ρ M β M and ρ F β F. In order to recover ρ M and ρ F in equation (4) (the effect of male and female voters average preference for male politicians on the share of male candidates nominated by parties), we compute the ratio of the estimates of ρ M β M and ρ F β F in equation (7) to the estimates of β M and β F in individual level regressions (2) and (3). (8) ρ F = ρ F β F βf, and ρ M = ρ M β M β M Assuming zero covariance between ρβ s and β s, we compute the variances of ρ s, we use the delta method to calculate the standard error of the estimates of ρ. This is the standard procedure used in previous research (for example, see Mocan and Raschke 2013). Specifically, the variances of ρ s are obtained by: (9) V(ρ ) = ( β )2 ρ V(β ) + ( ρ ρβ )2 V(ρβ ) = ( 1 ) β 2 [( ρβ 2 ) V(β ) + V(ρβ )] β 15

5. Results Effect of Education on Voters Preference for Male Politicians Panels A and B of Table 5 present the estimates obtained from equations (2) and (3), respectively. Preference for Male Politicians is measured by whether the individual agrees with the statement On the whole, men make better political leaders than women do. The variable of interest is Education, and it stands for the years of individual s completed schooling. Election year fixed effects are included in the regressions. Sampling weights are used, and robust standard errors are reported. Column (1) in each panel shows the marginal effects obtained from estimating equations (2) and (3) by probit. Results suggest that education reduces the preference for male politicians for both male and female voters. However, this regression ignores the possibility that education of an individual could be endogenous. For example, individuals who are members of conservative/patriarchal families are likely to obtain less education, and at the same time, such individuals might have greater taste for male politicians. To guard against this type of selection, we use the variation in individual s education induced by the education reform to identify the effect of education on preference for male politicians. Our instrumental variable, Treated, indicates whether the individual was subject to the Turkish education reform, and consequently had to complete 8 years of education instead of 5. Individuals who are born in 1986 or later are subject to this law. The result of the first-stage regressions where individual s education is regressed on whether the individual was subject to the education reform is presented in columns 2 of Table 5. Results show that both treated males and females complete more years of schooling compared to 16

their untreated counterparts. The increase in females educational attainment due to the reform, 1.8 additional years, is greater than the increase in males education, 1 year. This is probably because, prior to the enactment of the law, education of women was much smaller compared to that of men. Males were more likely to complete 8 years of education regardless of whether they were subject to the reform. As a result, more females than males had to comply with the law, and females education is increased more. Columns 3 in panels of Table 5 show the result of the reduced form regression, where individual s preference for male politician is regressed on whether they were subject to the education reform. Being treated by the education reform reduces female voters preference for male politicians, while it increases males. However the coefficient of Treated in male regressions (Panel B) is not statistically significant. In columns (4) of Table 5, we present the IV estimates of the effect of individual s education on their preference for male politicians. Results suggest that education increases one s bias against politicians of the opposite sex. Specifically, when a female voter acquires one more year of schooling, her preference for male politician decreases by 3.6 percentage points (6% of the mean 60 percentage points). On the other hand, a year increase a male voter s educational attainment raises his preference for male politicians by 4.6 percentage points (7% of the sample mean 69 percentage points). These are the estimates of β F = 0.036 and β M = 0.046, which we will use to recover the effect of the average voter s preference for male politicians on parties choice of sex-mix of their candidates (ρ). Effect of Voters Preference for Male Politicians on the Share and Number of Male Candidates 17

To investigate the effect of voters education on party behavior, we estimate equation (7). The results are presented in Table 6 where the unit of observation is a district-election. Outcome variable in columns 1 (column 2) of Table 6 is the share (number) of male candidates nominated by all parties in a district in one election. In addition to the variables listed, all regressions include district fixed effects and election year dummies. Standard errors that are clustered at the district level are presented in parentheses. In column 1 of Table 6, the coefficients of Average Edu. of Fem. Voters and Average Edu. of Male Voters are -0.032 and 0.037, respectively, and they are both are statistically significant. That is, for each additional year of schooling female (male) voters in a district obtain, parties nominate 3.2 percentage points fewer (3.7 percentage points more) male candidates. The sample mean of the outcome is about 0.88. These are the point estimates of ρ F β F and ρ M β M. Using the estimates of β F and β M (-0.036 and 0.046) from equations (2) and (3) which are presented in columns 4 of Panels A and B of Table 5, we recover the point estimates and standard errors of ρ F and ρ M, which are reported at the bottom of column 1 in Table 6. Specifically, ρ M = ρm β β M 0.037 = = 0.82 and ρf M 0.046 = ρf β β F = 0.032 F 0.036 = 0.87 This point estimates of ρ s (which are statistically significant at 10% level) suggests that the elasticity of share of male candidates in tickets with respect to preference for male politicians is about 0.6-0.7. 14 When we use the Number of Male Candidates as the outcome variable (column 2 14 For example, from equation (4), Male Candidates ct D ct F Male Candidatesct D ct F following: Male Candidates ct = ρ F D ct F D ct F with respect to average preference of male voters is computed to be 0.65 = ρ. F Therefore, the elasticity can be computed using the Male Candidates ct 18 = 0.87 0.60 = 0.6. Similarly, elasticity 0.88

of Table 6), the estimate of ρ F and ρ M are 88 and 150, respectively (s.e: 54 and 66). These estimates imply that number of male candidates in a district increases by 0.7 and 1.3 percent when the average preference for male politicians of female and male voters increase by one percent in that district, respectively. All elasticities are in 0.6-1.3 range. In addition, the estimates of ρ s are not statistically different from each other (that is, they are about 1 standard deviation from one another). This implies that parties responsiveness to male and female voters preference for male politicians is equal. In a separate regression, we estimated a model that restricts parties behavior such that they respond to the average preference of all voters (not separately to average of male and female voters). The unreported results suggest elasticity of proportion of male candidates nominated by parties with respect to preference for male politicians is about 1. Coefficients of the control variables have expected signs. Share of male candidates in a city is increasing at a decreasing rate with respect to the vote share of right parties in the most recent local election. Per capita number of individuals coming to Turkey from abroad reduces male candidates in that city. This could be due to multiple reasons. For example, individuals who have been outside of Turkey could be exposed to ideas of equality of men and women more than individuals who have not been abroad. Alternatively, the effect could be due to the effect of high income. Individuals arriving from abroad are either tourists visiting Turkey or Turkish citizens traveling abroad. Both groups should have higher incomes compared to those who never travel abroad. We also find that previous female representation in a city (measured by the share of the seats in the district occupied by female politicians in the last 25 years) decreases the number of male candidates. Though, this effect is not statistically significant. 19

An interesting result that emerges in Table 6 is the effect of Relative MP (Member of Parliament) Salary. The ratio of MP salary to the average income in the city reduces nominations of male candidates. Specifically, our results indicate that if MP salary relative to the average city income doubles, the share of male candidates nominated in that city decreases by 2 percentage points. Notice that none of the candidates are yet to be members of the Parliament. However, the prospect of a higher earning relative to the earnings of the average person in their home city induces more female nominations. That is, the coefficient of Relative Salary could be interpreted as the responsiveness of political labor supply of males versus that of females. Specifically, females could be more likely to run for a seat in the parliament compared to the males when parliamentarian salaries increase relative to the average income in their city. The employment rate (ratio of the employed to the working age population) of males increases share of male candidates, but the employment rate of females reduces it (these effects are not statistically significant). In cities with older voters, more male candidates are nominated. Effect of Voters Preference for Male Politicians on the Male Candidates Ex-Ante Election Chances Results in the previous section shows that parties nominate more male (female) candidates in districts where preference for male politicians is higher (lower). However, the change in sex-mix of the candidates has little or large influence on who gets elected depending on whether the change takes place at the top or bottom of parties tickets. Specifically, as explained in Section 2, the lower ranked candidates and those who are nominated in district with fewer contested seats have smaller chances of election. If parties nominate more females towards the bottom of their tickets or in districts with small number of available seats, then nomination of 20

more females will not translate into more elections of female candidates. In this section, we investigate whether voters preference for male politicians influences ex-ante election probabilities of male versus female candidates. In order to measure election probabilities of male versus female candidates, we first compute the minimum share of votes a candidate s party must obtain in order for the candidate to be elected for sure (Sufficient Vote Share). This threshold is computed using equation (1). Then, we separate candidates into three groups: those with High Election Chances (Sufficient vote share < 33%), Medium Election Chances (Sufficient Vote share is between 33% and 67%), and Low Election Chances (Sufficient Vote share is > 67 %). Within each group we calculated share of male candidates. For example, in the hypothetical district presented in Table 1 where there are 3 contested seats, the sufficient vote shares for the 1 st, 2 nd and 3 rd ranked candidates are 25%, 50% and 75%, respectively. For this specific district, the share of males in candidates with high election chances is the share of male candidates among all first ranked candidates. In another district with 5 contested seats, the sufficient vote shares of 1 st and 2 nd ranked candidates is 17% and 33%, respectively. For this district, the share of male candidates with high election chances is the share of male candidates among all first and second ranked candidates. We estimated equation (7) using the share of male candidates with High, Medium and Low election chances as the outcome variables. Results are presented in Table 7. In the first (last) 3 columns, the share (number) of male candidates is the outcome variable. In addition to the variables listed, regressions include district fixed effects and election year dummies. Results in column 1 of Table 7 suggest that parties respond to increases in average education of male and female voters by changing the sex composition of their candidates at the top of the tickets and/or in districts with better election prospects. For example, the coefficients 21

in column (1) of Table 7 shows that a one year increase in the average female (male) voter s education reduces (increases) share of males in candidates with high election chances by 3.7 (3.8) percentage points (mean: 92 percentage points). These estimates imply that ρ F and ρ M are 1 and 0.85, and they are statistically not different from one another. A one percent increase in the male voters preference for male politicians in a district increases the share of male candidates who are nominated in ranks/districts with high ex-ante election chances by 0.7 percent. Similar results are obtained when the number of male candidates is used as the outcome variable instead of share of male candidates (column 4). An additional year increase in male or female voters education changes the number of male candidates with high election chances by 1. 6. Summary and Conclusion Despite its favorable consequences on economic outcomes such as improved public health and increased education, females representation in political office is low. In an average country in the world, only 16% of the seats in the national parliament are occupied by female politicians. As a solution, previous research proposed gender quotas which reserve a certain number of seats in the parliament for women. Although these quotas were successful in increasing the share of female politicians, they have failed to attain equal male and female representation in politics. In this paper we investigate possibility of another tool to eliminate gender inequality in politics, education of voters. Since the necessary condition for election is candidacy, we focus on the nomination of female versus male candidates by their parties. We use data obtained from Turkish elections 2002, 2007 and 2011 where a party-list proportional representation system is used. In this system, parties have the sole power to determine their candidates. 22

We show that voters education affects the sex mix of the candidates, through its impact on voters preferences over the sex of politicians. We find that more educated voters are more likely to prefer politicians of their own sex. Specifically, a one year increase in completed schooling of a female (male) candidate increases (decreases) their preference for a male political leader by about 6%. In addition, we find that political parties increase (decrease) the share of their male candidates by about 4 percentage points (from a baseline of 87 percentage points) when the educational attainment of the average male (female) voter in a city increases by one year. Using an empirical method similar to the two sample IV, we recover the impact of voters preferences of politician sex on nomination of male versus female candidates. Our findings imply that for every one percent decrease in preference for male politicians leads parties to nominate 1-2% more female candidates. We show that parties are equally responsive to the preferences of male versus female voters. In addition, our results suggest that smaller preference for male politicians improve ex-ante election probabilities of female candidates. Our results are consistent with the recent developments in Turkey. Particularly, because of the law which increased the years of mandatory schooling from 5 to 8 in 1995 in Turkey, the educational composition of voters have changed significantly starring 2007 elections. First, the share of voters who are more educated has risen since the 2007 election compared to the 2002 election where no individual who was subject to the mandatory schooling law was old enough to vote. In addition, this law increased the education of females more than it increased males education, as females were less likely to complete 8 years of education compared to males before the law was enacted. That is, due to the change in the mandatory schooling law, the average female voter s education increased more than the average male voter s education. In the same time period, starting with elections in 2007, Turkey experienced a dramatic increase in the 23

number of female candidates and the number of female politicians. For example, share of female candidates has increased from 10% to 16% between 2002 and 2011, and the number of female politicians in the parliament increased from 24 to 79 in the same time period. Our results suggest that greater female representation is mainly driven by the increase in educational attainment of female voters versus male voters. 24

Figure 1a: Ex-Ante Election Probability of a First Ranked Candidate (3 Parties Competing for 5 Seats in a District) First Ranked Candidate 0 Prob(Election).2.4.6.8 1 0 20 40 60 Vote Share of 80 100 the Candidate's Party Notes: Probability of election of a 1 st ranked candidate who is nominated in a district with 5 seats as a function of the share of votes their party obtains. 25

Figure 1b: Ex-Ante Election Probability of All Candidates by Rank (3 Parties Competing for 5 Seats in a District) Ex-Ante Election Chances by Rank (3 Parties Competing for 5 Seats) First Ranked Candidate Second Ranked Candidate Third Ranked Candidate 0 0 0 Prob(Election).2.4.6.8 Prob(Election).2.4.6.8 Prob(Election).2.4.6.8 1 1 1 0 20 40 60 80 100 Vote Share of the Candidate's Party 0 20 40 60 80 100 Vote Share of the Candidate's Party 0 20 40 60 80 100 Vote Share of the Candidate's Party Fourth Ranked Candidate Fifth Ranked Candidate 0 0 Prob(Election).2.4.6.8 Prob(Election).2.4.6.8 1 1 0 20 40 60 80 100 Vote Share of the Candidate's Party 0 20 40 60 80 100 Vote Share of the Candidate's Party Figure 1b: Ex-Ante Election Probability of All Candidates by Rank (3 Parties Competing for 4 Seats in a District) Ex-Ante Election Chances by Rank (3 Parties Competing for 4 Seats) First Ranked Candidate Second Ranked Candidate Prob(Election) 0.2.4.6.8 1 Prob(Election) 0.2.4.6.8 1 0 20 40 60 80 100 Vote Share of the Candidate's Party 0 20 40 60 80 100 Vote Share of the Candidate's Party Third Ranked Candidate Fourth Ranked Candidate Prob(Election) 0.2.4.6.8 1 Prob(Election) 0.2.4.6.8 1 0 20 40 60 80 100 Vote Share of the Candidate's Party 0 20 40 60 80 100 Vote Share of the Candidate's Party 26

Table 1: Hypothetical d Hondt Example Parties Votes/1 Votes/2 Votes/3 Votes/4 Votes/5 A 50 25.0 16.7 12.5 10.0 B 30 15.0 10.0 7.5 6.0 C 20 10.0 6.7 5.0 4.0 Notes: The table presents an example of how seats in a district with five seats are allocated to three parties using d Hondt method. There are 100 votes cast, so votes = vote shares. The number of votes received by each party is shown in the column (Votes/1) in each panel. d Hondt method divides each party s votes by consecutive integers up to the number of seats in the district (N). In this example, N=5. The columns Votes/1, Votes/2,, Votes/5 in the table present the resultant quotients. The parties with the largest N quotients win the seats. In the examples above, the bold and underlined numbers represent the largest five quotients. Parties win as many seats as the number of largest quotients they have. In this example, parties B and C win one seat, A wins three seats. Table 2: Voters Educational Attainment and Preferences for Male Politicians Female Voters Unaffected by the Education Reform Affected by the Education Reform Survey Year Years of Schooling Men Make Better Political Leaders Years of Schooling Men Make Better Political Leaders 2001 6.62 0.59 2007 7.92 0.57 11.22 0.46 2011 7.15 0.68 11.74 0.58 Male Voters Unaffected by the Education Reform Affected by the Education Reform Survey Year Years of Schooling Men Make Better Political Leaders Years of Schooling Men Make Better Political Leaders 2001 8.90 0.67 2007 9.58 0.66 13.30 0.70 2011 9.28 0.75 12.37 0.73 Data source: World Values Survey. Voters are at least 18 years old as of the survey. Individuals affected by the education reform are born in 1986 or later. The survey question for preference for male politicians involved asking whether the individual agreed to this statement: On the whole, men make better political leaders than women do. Election Year Table 3: Voters Educational Attainment and Sex-Mix of Candidates Share of Male Candidates in Party Tickets Average Education of the Untreated Female Voters 27 Average Education of the Treated Female Voters Average Education of the Untreated Male Voters Average Education of the Treated Male Voters 2002 89.88% 4.62 6.95 2007 85.73% 5.08 8.19 7.25 8.88 2011 84.51% 6.86 9.91 9.15 10.01 Data sources for education and sex-mix of candidates are Census and Resmi Gazete (official gazette where candidates are announced publicly before elections), respectively. Table presents national averages. Treated voters are born in 1986 or later.

Variable Preference for Male Politicians Table 4: Descriptions and Summary Statistics Panel A: Voter Data (World Values Survey) Whole Sample Description N=5,986 =1 if individual agrees with the 0.645 statement On the whole, men (0.478) make better political leaders than women do. Females N=3,015 0.601 (0.490) Males N=2,971 0.688 (0.463) Education Years of completed schooling 8.366 7.307 9.393 (4.821) (4.867) (4.548) Treated =1 if the individual is subject to the education reform 0.071 0.065 0.075 (0.256) (0.247) (0.264) Male =1 if male 0.508 0.000 1.000 (0.500) - - High Income =1 if individual s income is greater than the median income bracket in a survey year 0.363 (0.481) 0.347 (0.476) 0.379 (0.485) Employed =1 if working 0.426 0.178 0.667 (0.495) (0.382) (0.471) Age Age as of the survey date 37.744 37.390 38.086 (13.936) (13.431) (14.404) Ideology Self-positioning on a political spectrum that ranges between 1 (most left) to 10 (most right) 6.055 (2.481) 6.031 (2.381) 6.079 (2.575) Panel B: District Level Data, N=255 (Census) Variable Description Mean Std. Dev. Share of Male Candidates Share of male candidates in parties tickets 0.88 0.05 No. of Male Candidates Number of male candidates in parties' tickets 81.28 63.31 Avg. Edu. of Fem. Voters Average years of schooling of female voters 5.88 1.55 Avg. Edu of Male Voters Average years of schooling of male voters 7.93 1.16 % Votes Right Partiest-1 Share of votes obtained by the right wing parties in the most recent elections for local government 0.73 0.12 28