ESTIMATING THE INCUMBENCY BENEFITS OF WOMEN IN CLOSE CONGRESSIONAL ELECTIONS

Transcription

1 ESTIMATING THE INCUMBENCY BENEFITS OF WOMEN IN CLOSE CONGRESSIONAL ELECTIONS A Thesis. submitted to the Faculty of the Graduate School of Arts and Sciences of Georgetown University in partial fulfillment of the requirements for the degree of Master of Public Policy By Benjamin Struhl, B.A. Washington, DC April 15th, 2011

2 Copyright 2011 by Ben Struhl All Rights Reserved ii

3 ESTIMATING THE INCUMBENCY BENEFITS OF WOMEN IN CLOSE CONGRESSIONAL ELECTIONS Benjamin K Struhl, B.A. Thesis Advisor: Peter L Hinrichs, Ph.D. ABSTRACT This paper suggests that women are at a significant disadvantage in close U.S. Congressional races compared to men. Previous political science literature has argued that women do not have a disadvantage in U.S. politics. Much of this previous literature argues that the apparent disadvantage women face in elections comes from the fact that there are fewer incumbent women in Congress and that men are continuing to win seats in Congress due to incumbency. This paper uses a Regression Discontinuity Design to test how repeat candidates perform in elections, looking at the differences between women and men and the differences between incumbents and challengers. Studying candidates in the U.S. House from 1968 to 2008, this paper finds that when examining candidates in the most competitive races and controlling for the candidate s success in the previous election, women do not compete as well as their male counterparts. The results of this paper suggest that women challengers (women not running as incumbents) are especially disadvantaged compared to male challengers and that women incumbents in close election races are also at a disadvantage compared to similar male candidates. The paper also found that these disadvantages were strongest in close elections and not apparent when including candidates who did not run close races. Overall, these findings suggest that unless they are entrenched incumbents, women do not perform as well as men do in elections. This finding supports the fact that America lags behind the rest of the world in the percentage of women serving in its national legislature. iii

4 Table of Contents Introduction... 1 Review of Existing Literature... 6 Data... 9 Use of Regression Discontinuity Design Models Results Conclusion Tables Bibliography iv

5 I. Introduction In the 2010 elections there were high expectations for women in Congress. Fifteen women ran for the U.S. Senate and 138 women ran for the U.S. House, record numbers in both cases. High profile journalists and politicians speculated that likely Republican gains and several prominent female Republican candidates would lead to The Republican Year of the Woman. Despite this, the 2010 results contained a surprising twist for the first time since 1978 the number of women serving in Congress decreased. State races also offered no signs of progress for women not only were there no changes in governorships, but the number of women state legislators decreased by over one hundred. Many of these losses were due to the fact that most of the women serving in Congress in 2010 were Democrats, and Democrats performed particularly poorly overall. Still, the results called into question the progress that women have been making in attaining political office. In 2008, when representation of women in Congress was at an all-time high, women still made up only 17% of the U.S. House. Even this lopsided margin of men to women in Congress is an improvement over the historical state of the U.S. House. As recently as 1978, in fact, women held about 4% of House seats. Women have been increasing their representation in a number of fields, in fact. The share of highly placed women executives in major corporations has made similar increases and now hovers near ten percent (Palmer and Kimball 2010, ION Network 2011). As an interesting contrast, 2010 was also notable in that it was the first year where more women received PhDs than men (Washington Post, 2010). 1

6 Yet, Congress is an especially interesting place to monitor the gains made by women in the workplace while other jobs might employ selection criteria that are highly influenced by an elite few at the top of the organization, representatives are selected by the popular vote. Curiously, women candidates are not using that vote to increase their representative share in a country where most of the voters are women. The U.S. census says that 50.7% of the country is female, and according to CNN women were 52% of voters in the 2010 Congressional elections and 53% in Many scholars have attempted to explain why women haven t used their powerful franchise to elect women to Congress, and several of these scholars have written entire books on the subject (Darcy, Welch and Seltzer, 1994, Carroll, 1994, Newman and Leighton, 1997, and Lawless and Fox, 2005). On this subject there is some disagreement between political scientists and political practitioners. Many journalists and consultants believe that women are genuinely disadvantaged in the political process judged unfairly by voters, covered unfairly by the media, and generally held to a different standard than male candidates (Milligan 2010). A few scholars agree with this analysis (Kahn, 1992, 1994). Many political scientists say that this is not the case, however. Political scientists have published several regression analyses suggesting that, controlling for incumbency, women do not perform worse than men as candidates (Seltzer, Newman and Leighton, 1997, and Darcy and Schramm, 1977). In addition, controlling for incumbency even suggests that women do not receive fewer campaign contributions than their male counterparts (Uhlaner and Schlozman, 1986). These political scientists say that incumbency explains why women are not being elected (Darcy and Schramm, 1977). 2

7 Yet, critics say this type of simple regression could mask real problems that are occurring in the data (Carroll 1994). For instance, one could argue that male candidates are more likely to be either strong candidates who would defeat women in close races, or to be sacrificial candidates who would be put up by a party without support or any real chance of winning. The average of these two groups could be statistically similar to the average performance of female candidates, but misleadingly so. If this were the true state of the world it would be consistent with observational data suggesting that incumbency was much of the reason women were failing to make inroads more quickly. If men are recruited more often as non-serious candidates, then it would bring down the average performance of male candidates, but it would not reveal whether or not men have an advantage when the quality of the candidates is equal. Candidate recruitment has been identified by many scholars as a significant factor in the electoral fortunes of women, and it is the subject of multiple books (Carroll, 1994, Niven, 1998, and Lawless and Fox, 2005). Setting incumbency as the explanation for low levels of female representation is complicated by the fact that incumbency itself is a very difficult subject to study. Over the past 50 years, many political scientists have made stabs at isolating incumbent benefits to re-election. Many of them have overlooked important factors about candidates that have made the findings of their studies questionable. The fact that over 90% of incumbent Congressmen are re-elected most years, for instance, does not in itself say how much advantage incumbency confers. After all, those candidates were elected in the district originally for a number of factors that continue to give those candidates advantages in the next election. For one, most Congressmen attain their positions only after besting multiple political opponents through primary and general elections. Furthermore, Congressmen are often elected because they are perceived to fit their district s 3

8 ideological tilt well, and candidates of other ideologies would not be favored to prevail over them. Given these factors it isn t so surprising that incumbents are re-elected at such high rates. Attempts over many years to establish the rate of advantage for simply having the incumbent I next to a name (and the mailing and other privileges that come with the office), and controlling for other factors have produced mixed results. Endogeneity is a major problem that is difficult to correct for in these regression analyses. There are also problems of overcontrolling if incumbents have an advantage in raising money, the results will be skewed when controlling for fundraising as any incumbent who raises less than their challenger is in many cases already perceived as vulnerable or likely to lose. In 2007, however, David Lee attempted to approach the incumbency problem through a new approach. Studying Congressional races over 30 years Lee used a regression discontinuity design (RDD) that only looked at the jump in outcomes in close races. Regression discontinuity allowed Lee to look at results barely on either side of the re-elect line that is to say, the difference in subsequent election outcomes between candidates who lost by a small margin (with the limit approaching zero) and candidates who won by a small margin. This difference is similar to random selection, and therefore problems with endogeneity and omitted variable bias are avoided. As expected, Lee found that far fewer of the candidates who lost by the narrow margin chose to run again the next year, and of those who did run, significantly fewer won their races. Lee estimated the incumbency benefit as about 5 percentage points, which matched what some political consultants had assumed for years based on first-hand observation of elections. The RDD also has potential to shine some light on the outcomes of women candidates on either side of the re-election margin. If female candidacies played out differently on either side of 4

9 the dividing line than male candidacies did, then it might help explain why women haven t expanded their representation more. Perhaps women receive less of an incumbency benefit, for instance that would make it harder for women to achieve parity with men in the U.S. House. Unfortunately, Lee did not study the effects of gender in his model. Also, when Lee did his study he used data only up to 1998 since there have been more women candidates and representatives in recent years, adding will add a greater amount of data concerning women candidates. Overall, there were three interesting dimensions on which to analyze this question. The first was determining whether female candidates are at a disadvantage when looking at narrow ranges around the treatment threshold (close to 50% of the vote). The second question was whether or not the gap between women challengers and incumbents is as large, larger, or smaller than the gap between male challengers and incumbents. The third question was in what other ways the female candidate model for the regression discontinuity design differs from the male candidate model. Taking data from the Inter-University Consortium of Political Science Research and the Federal Election Commission, I built a database of all the candidates in U.S. House races from 1968 to 2008 with a few states missing data in certain years. Table 1 shows which states are missing data. Loosely following the methodology used by Lee, I regressed election results in year t on election results in year t+1 with dummy variables for candidate gender and incumbent status, as well as the interaction of gender and incumbency status. In the regression discontinuity design, I examined four ranges of candidates. First, I looked at all candidates, then confined my 5

10 analysis to candidates within ten percentage points of the treatment threshold, then to candidates within five points of the threshold, and finally to candidates within two points. Although the results examining all candidates found a statistically insignificant coefficient for female candidates, all attempts to narrow the range of the analysis to the candidates who were close to the treatment threshold found statistically significant coefficients for females. As the range narrowed, the coefficient on the gender dummy became more and more statistically significant. The narrowest range around the treatment threshold (48-52%) found a coefficient on the gender dummy that was highly statistically significant. However, in none of the regressions was I able to identify a different result between male and female incumbents. The difference between the performance of male and female incumbents seems to be the same as the difference between male and female challengers. Finally, a separate analysis of likelihood to run again after an election found that not only do female candidates differ from males in likelihood to run again, but female incumbents also differ from male incumbents in repeat candidacies. Females who were unsuccessful challengers were found to be ten percentage points less likely to run again compared to their unsuccessful male counterparts. Female incumbents, on the other hand, were found to be four percentage points more likely to run again than their male incumbent counterparts. II. Review of Existing Literature: There is a large body of research that is relevant to examining the intersection of gender and incumbency. There are at least three significant research areas that are informative to this 6

11 paper. One key area concerns the possible electoral disadvantages for women candidates in general. Much of this literature follows what has now become almost conventional wisdom on the subject of why women aren t elected to political offices: there is no effect of gender on the success of candidates controlling for incumbency, and therefore incumbency is almost the entire explanation that women continue to be underrepresented in politics (Darcy and Schramm 1977; Seltzer, Newman and Leighton 1997). Schwindt-Bayer (2005) confirms this view with the finding that as incumbent retention increases in a country there is a negative effect on female representation in national legislatures. Other studies have looked at election results and seen gender effects that suggest female candidates have advantages in support from women voters (Smith and Fox, 2001) and that people treat female candidates differently according to their party (Dolan 2004). Kahn (1992) addresses the issue differently, finding that the media produces different perceptions of candidates depending on gender, with a possible disadvantage for women. Specific incumbent effects are of great interest, although many of them deal only with incumbency in general and do not differentiate between male and female incumbents. There has been a wide range of disagreement over the past forty years about whether an incumbency advantage exists and how it should be measured. Early papers found an incumbency effect and suggested it was growing over time (Erikson 1971; Cover 1977). Alford and Brady (1988) found that there was no incumbency effect and possibly even evidence of a negative incumbency effect in some races their methodology was based on the assumption that they could average one overestimated measure of incumbency advantage and one underestimated measure of this advantage to produce an overall unbiased measure. Gelman and King (1990) however, disagreed with this finding and specifically found that the measures used by Alford and Brady were not 7

12 balanced or unbiased as the authors claimed, and they suggested that previous studies had been measuring incumbency effects with a large degree of bias. They concluded that there was in fact a significant but relatively small two percentage point incumbent advantage on average. Despite the difficulty of measuring the actual incumbency advantage, many more studies have been attempted. Cox and Katz (1996) tried to avoid the question of what the incumbency advantage was and looked at factors that might affect the incumbent retention rate. Their study found an association between increasing benefits from experience in Congress and increased incumbency retention. More recently, David Lee (2007) proposed a new way of measuring the incumbency effect with a regression discontinuity design. Building off of an earlier application of RDDs (Lee, Moretti, and Butler 2004) this study found a large and significant advantage for U.S. Congressional incumbents, the equivalent of about five percentage points in voting. Uppal (2009) confirmed this finding with a similar RDD, this time looking at state legislatures. There are comparatively fewer studies, however, on the subject of incumbency benefits for women. Darcy and Choike (1996) used a model based on the Markov process to show that the number of women legislators should approach the number of women candidates eventually if voters are not biased. The larger the effect of incumbency on re-election, however, the longer it would take for this to occur. Significantly, this paper also found that increasing the re-election rate of female incumbents would increase the share of women in a state legislature by more than simply having more women candidates running. Gagliarducci and Paserman (2009) looked at reelection of Italian mayors and found that women were re-elected less often than men. The study suggested that male legislators had a harder time accepting and cooperating with the female mayors. Ferreira and Gyourko (2010) also studied the re-election of female mayors, this time in the U.S. Their study found that women were more likely to be re-elected as mayors, but the study 8

13 also found this was despite bias against women. The authors found that because there are fewer female mayors, those who get elected are better candidates in general. Schwarz (2010) studied election and re-election of women and men in local California races. Schwarz found that there was some degree of gender balancing occurring from election to election. According to Schwarz, the fewer females that sit on a school board, the greater the likelihood a female candidate would win a future race. Conversely, the more women on a board, the greater the likelihood a man would win a future race. III. Data Election data is not easy to find beyond the aggregate level, so this study uses a data set assembled from several sources. Four separate sources were used to build the final dataset, many of them providing the same type of data but in different years. The first source of data is the Inter-University Consortium of Political Science Research (ICPSR). This includes information on all the candidates from the U.S. House from the 1700s to The data includes candidates from 47 states and occasional observations from the District of Columbia. Table 1 contains information on which states are missing data. Since DC does not have data on all years, nor does it have a proper voting representative, I did not include DC results in the data set. The other fields in the data set were year, district, vote total, and party. I did not change any of these except for political party, where I changed Minnesota s Democrat- Farmer-Labor to simply Democrat since they are functionally the same party. This helped with dataset cohesion. Similarly, North Dakota Democrats often go by the designation DNL for Democratic Nonpartisan League. These names were also all changed to Democrat. 9

14 The ICPSR data does suffer from some missing data, which concerns a small number of states and years. As Table 1 shows, Louisiana and Pennsylvania are both missing data from 1978 and The missing ICPSR data is likely random, meaning it should not bias the results, and it also has very limited power to change any of the results. The second and third sources of data I used were further election and candidate data from the FEC and the U.S. House Clerk. This is similar data with similar fields, but the years range from for the House Clerk and for the FEC. Table 2 shows that all of the datasets together combined to contain 16,533 candidates 8,488 Democrats and 8,045 Republicans. The fourth source of data was the Rutgers Center for Women in Politics. It was a list of all the women candidates who had ever run for the U.S. House or Senate. This data went from 1968 to 2008 and only contained women candidates. In the process of merging the data sets I decided to only include candidates from 1968 to This was not only because the list of female candidates only went back to 1968 but also because there were relatively few women candidates before that election, so including more elections would not add very much data about women. Also, it may not be appropriate to pool the results about women candidates from over forty years ago with modern women candidates. Some observations may not have merged correctly because names were not always recorded exactly the same from year to year. To maximize success in merging observations, I created a code involving the first five letters of the last name, the first initial, the state, and the district. I matched this code from the Rutgers data set to the base data set. Of the general election women 10

15 in the time frame of interest, over 97% of the women in the Rutgers data set were matched to candidates in election results. Most of these candidates were matched with the algorithm, but some had to be corrected manually to account for changed names, nicknames, and data entry problems. Of the remaining 3% that could not be matched, most of them were candidates from years when the candidate s state was missing from the data. The matching of repeat candidates from year to year no doubt encountered similar problems, but there was no good way to go back and correct these results. However, these results also would have been more accurate than the data matched between the two data sources, as inter-source inconsistencies and missing data accounted for most of the matching problems, and the candidate data was usually consistent, even when wrong. Ultimately, my dataset contained 1,625 female candidates 1,013 of those were Democrats and 612 were Republicans. Table 3 shows further summary statistics. Additionally, because Congressional districts often changed between years ending with 0 and 2 due to redistricting, it is not appropriate to make incumbency judgments between those years if a candidate who represented district X in 2000 loses in 2002 it might be because they actually represented a wholly different part of the map earlier and were, therefore, not true incumbents. Therefore, I carefully made indicators for anyone running in a 0 year. Those people would be taken out of comparison to the next election s results, although they would still be useful in looking at their relationship to the previous 08 election. In order to capture all the major candidates and their traits, the unit of observation I decided upon was at the candidate level. Each election contributes one or two observations, depending on whether both of the two major parties advanced candidates. The next step was to 11

16 convert vote totals to this single vote share. Because many candidates ran with write-ins or third party candidates on the ballot, a winning percentage might be less than 50%. To correct for this, vote share was calculated as among the top two candidates and expresses the percent of that share that one of the major political parties received. This preserved the effect of a candidate over 50% being re-elected, and a candidate under 50% just missing re-election. I created two indicator variables that would tell if a candidate ran again in the next year and whether they won election in the next year. I used the same method for matching names as I had earlier with the Rutgers data set. David Lee used a similar matching method and found it to be the most effective method in matching candidates. My own experience also found this to be the most effective way of matching available. IV. Use of Regression Discontinuity Design Use of a regression discontinuity design (RDD) is appropriate and useful in the context of political elections, as elections have the treatment threshold required to perform an RDD analysis. RDD attempts to replicate the results of an experiment by using a treatment threshold a specific point in a continuous variable where you believe there is a meaningful difference in outcomes. For instance, at over 50% of the vote, candidates are elected to office; at under 50%, they are not. Therefore, 50% is the treatment threshold. In an ideal experiment we could assign people to incumbency randomly and then see if those assigned to incumbency performed better in an election. In reality, this would require an entire election cycle where representatives were all chosen randomly, and although this alternative might place a greater number of talented individuals in Congress than usual, it would 12

17 of course never occur. Approaching the limit of the treatment threshold (getting very close to 50%), however, the result of the election is in fact influenced by truly random events, such as the weather. Therefore, we can assume that candidates who each got very close to 50% of the vote were assigned their incumbent status close to randomly. RDDs attempt to quantify the difference in the electoral prospects of candidates who barely won on the previous election and candidates who barely lost before. If a candidate who approaches 50% in 2002 but loses and then receives 45% of the vote in 2004, and a winning candidate who received little more than 50% of the vote in 2002 then receives 55% of the vote in 2004, then the RDD suggests that winning in 2002 causes a candidate to perform ten percentage points better than he or she would have if he or she had lost in The causal inference that an RDD brings to the table is analytically powerful, but not flawless. One key assumption behind the RDD is that chance has an equal likelihood of pushing either candidate over the line. If one party had a better ability to contest close elections, then they might have a greater chance of winning at the 50% threshold, at which point treatment to incumbency would no longer be random. The fact that treatment is random is especially important to RDDs, since random treatment allows this analytical method to avoid omitted variable bias problems the same way a conventional randomized experiment does. V. Models I used three models to examine the relation between gender and incumbency in politics and also performed a simple regression to compare to previous standard regression-based 13

18 analyses. The first model is based on the standard format for RDD regressions the dependent variable is the vote share in the next election (t+1), and the independent variables are the continuous variable on which the treatment threshold is based (VoteShareElection t ), the dummy variable for the treatment threshold (WonElection t ), a dummy variable for female, and an interaction of the female dummy variable and the treatment threshold. It would also be possible to include an interaction of the VoteShare variable and the treatment threshold. But given the narrow range that many of the regressions looked at, adding the additional interaction term does not decrease bias by very much and often increased the standard errors by a large amount. My primary RDD model ended up being similar to other RDD approaches to incumbency such as those done by David Lee. One difference between my model and others is that I included variables considering candidate gender, while previous studies have not. I also changed the unit of analysis to focus on candidate rather than election. This allowed me examine female candidates of both parties simultaneously. Lee observes candidates on the election level, given the assumption that a Democrat getting 40% of the vote will correspond to a Republican getting approximately 60% of the vote. Although this assumption is valid, it was important to distinguish candidates in elections to determine gender-based effects. Also, focusing on the candidate rather than the election is arguably a better way of considering the question of incumbency, which seems based on individuality as much as it is on political party. This last point is particularly supported by the number of candidates, especially southern Democrats, who were successful after changing parties, although there is also some contrary evidence of candidates who do not find success in party switches. 14

19 I first estimate the model VoteShareElection it+1 = β 0 + β 1 VoteShareElection it + β 2 WonElection it + β 3 Female it + β 4 Female it *WonElection it + ε it. (1) This model examines the relationship between one election and the following election, with a dummy variable indicating whether the candidate won or lost the previous election. This model looks at candidates who got between 40 and 60% of the vote, 45% to 55% of the vote, and 48% to 52% of the vote, narrowing in on the limit of the treatment threshold (although necessarily constrained by sample size). My second model was another look at the same question, but this time I had a simple dummy variable for winning the election in question as the dependent variable and the model is a linear probability model. This second equation I estimate is WonElection it+1 = β 0 + β 1 VoteShareElection it + β 2 WonElection it + β 3 Female it + β 4 Female it *WonElection it + ε it. (2) Another side of the incumbency question I thought would be important to consider was the deterrent effect on candidates running again for office after losing a bid. This is a significant factor in studies of incumbency, as many very capable candidates who closely lose their elections do not go on to contest the seat again. On the other hand, many who are elected to Congress only do so after repeated attempts to run for the seat. Therefore, whether or not a candidate chooses to run again makes a difference in how competitive elections can be if seasoned candidates don t run again then incumbent politicians are the likely beneficiaries. The third model tries to examine this question. The model is 15

20 RanInElection it+1 =β 0 + β 1 VoteShareElection it + β 2 WonElection it + β 3 Female it + β 4 Female it *WonElection it + ε it. (3) All three of these models are looking at possible differences between men and women in whether they chose to run for election subsequently, as a function of their previous vote share and whether they won the previous election or not. Since there is expected to be a large jump around 50% of the vote, a specific dummy variable is used for candidates who scored over or under that percentage of the vote. The gender dummy is included in all three models, as it explores my primary research question of how men and women differ in their performance as repeat candidates. Also included is an interaction term, to test if there is a difference between the likelihood of female candidates who won to run again, and female candidates who lost to run again, versus the same categories of male candidates. I report clustered standard errors for all three models, with clusters at the state level. VI. Results Although I was primarily concerned with more sophisticated analyses, I initially ran two regressions based on some of the simpler models used to describe election results. These models, with results given in Table 4, looked only at the association between vote share in a given election and the candidate s party, gender, and incumbency status. The first regression looked at only gender and party as explanatory variables for vote share and found significant results for both coefficients, with women appearing to be at a disadvantage in terms of vote share. The second regression added incumbency and found that the coefficient for female candidates was now no longer statistically significant, while both incumbency and party were highly statistically 16

21 significant. The relationship between these two regressions is consistent with all of the previous literature that I reviewed, which finds a coefficient for candidate gender that is not statistically significant from zero once incumbency is included in the models (Darcy and Schramm 1977; Seltzer, Newman and Leighton 1997). This second regression still suffers from omitted variable bias by leaving out items such as fundraising and personability, but these aforementioned studies suggest that if this data were included it would not change the results on the significance of the gender coefficient. However, the validity of any of these models is, in my mind, questionable. I believe all of these models miss relevant factors such as female candidates being dissuaded from running against incumbents and female challengers struggling more than males in raising money. Nevertheless, the experimental aspect of the RDD allows the analyst to sidestep these questions. Table 5 illustrates the four separate models that I ran that look at the relationship between vote shares in subsequent elections. By looking at candidates in successively smaller windows around the election point, the analysis approaches the limit of the treatment threshold (50%) defined by the Regression Discontinuity Design. Again, elections with different candidates running in separate years were not examined. The analysis leaves out comparisons between candidates running in a year ending in 2 and also running in the previous 0 year election. Although these elections are sometimes in identical districts, often the lines and even the candidates have been altered through redistricting. As a given district numbered eight may roughly be turned into a district numbered seven after redistricting, care must be taken not to compare election districts across redistricting. In examination of election vote shares, the first model ( All Elections ) is the most general, encompassing all valid data points of candidates who both ran in election t+1 and in the 17

22 same district in the previous election cycle. This analysis reveals highly statistically significant coefficients on the vote share from the previous election and whether or not the candidate was an incumbent. However, the coefficient on whether or not a candidate is a woman is statistically insignificant from zero. This seems to support the finding of earlier researchers who could not identify a separate effect for female candidates beyond incumbency however, as the range of analysis is narrowed this conclusion changes. The next three models examine increasingly narrow ranges of candidates around the treatment threshold first, using a generous range of ten points on either side, then narrowing to five points, and finally to two points. Across the more specific models, the effect of incumbency remains fairly constant and highly statistically significant, and positive. The effect of the previous election s vote share becomes stronger the smaller the range of analysis, but the standard errors also quickly increase in size. Most interesting is the coefficient on the female variable, which becomes both greater in magnitude and more statistically significant as the analysis narrows. In closer elections, there is a pronounced negative effect of being a female on the election results which is not visible in a snapshot of all the elections. Using the most specific model, it is informative to generate fitted values for different hypothetical candidates who got approximately 50% of the vote but fell to just one side (49.99% or 50.01%). For instance, a male candidate who got 50.01% of the vote in election t-1 is predicted to gain about six points in the next election, or 56.2%. 18

23 For females running a second time the results were different. If female candidates barely missed attaining 50% of the vote in election t-1 they were expected to get 45.9% of the vote in the following election. When they barely attained 50% of the vote, female candidates were expected to get just 51.9% of the vote the following election. Significantly, males who just missed winning by a fraction in the previous election (49.99% of the vote) were projected to barely win the next time (50.6%, using the losing regression). This is not so for women. If these results hold, then they reveal two remarkable reasons females may continue to be underrepresented in Congress. The first is that they don t do as well as men when making an attempt for office immediately after losing. Since many candidates take more than one attempt to win election to Congress, the fact that women are less likely to be given a second shot is significant. Another important result is that female candidates don t seem to get as much of an incumbency advantage. This could be an especially compelling explanation for why female candidates have not been increasing their voting representation in Congress. Table 6 shows that the models that used the linear probability model yielded similar results. Being a female was associated with a decreased chance of being re-elected this effect became stronger and more statistically significant the narrower the range around the treatment 19

24 threshold that was examined. Some of the linear probability models also found a separate statistically significant effect associated with being a female incumbent. This effect is positive and largely counteracts the effect of being a female in the first place, although larger standard errors in the narrowest model (the one examining 48 through 52 percent vote getters) meant that this effect was not significant closest to the treatment threshold. Still, the overall results suggest that women overall have a lesser chance of being elected when running a second time compared to men and that for female challengers this effect is especially strong. The continued success of a limited number of female candidates who attain office and continue to run and benefit from their incumbency for long stretches may be covering the misfortunes of female challengers, who these results suggest are facing an additional barrier to their electoral prospects. Table 7 shows the results of the model I estimated which examined the likelihood that a candidate would run for Congress a second time, based on the election results from their previous run. I found that, as expected, being an incumbent had a very strong effect on whether candidates chose to run again, although is it not clear whether or not women are different from men in this regard. Overall, these results did not seem to add much more insight to differences that men and women might have as candidates. VII. Conclusion Some of these results might suggest partial solutions to the ongoing debates among some political scientists and some political consultants as to the disadvantage that female politicians 20

25 have in Congressional elections. It is easy to see how political scientists have determined that observable disadvantages for women are due to incumbency. Big picture results such as the RDD model using all candidates show that incumbency has a huge effect on elections and that gender does not. However, when correctly narrowing in the treatment threshold in the RDD analysis, it is evident that at the tipping point between winning and losing an election, women are at a significant disadvantage beyond incumbency. Two facts are especially worthy of note the first is that female challengers in particular are subject to negative effects regarding incumbency. Female challengers always have predicted values that are significantly lower during their second election runs. Considering the prevalence of candidates who only win office on their second try, this could be a significant barrier to increasing female representation in Congress. In addition, female challengers are shown to make a second attempt less often than men do. The other interesting fact is that female incumbents in close races have less of a predicted advantage than men do, but this disadvantage diminishes once women candidates have become sufficiently entrenched and are no longer running close election races. When looking at elections in the 40-60% range, for example, we find that the interaction variable of gender and incumbency has a positive coefficient and almost perfectly counterbalances the disadvantage that female candidates have overall. This could go a significant distance in explaining how previous results fit in with these findings, and it also points to a reason why women have not achieved significant representation in the U.S. Congress. The fact that there might be a cadre of entrenched women incumbents who get an additional incumbency benefit could explain why 21

26 previous studies have been so quick to attribute candidate advantage solely to incumbency. This study, however, concludes that women have a very real disadvantage where close races are concerned. 22

27 VII. Tables 23

28 24

29 25

30 26

31 27

32 VIII. Bibliography Adams, James, Thomas Brunell, Bernard Grofman, and Samuel Merrill. "Why Candidate Divergence Should Be Expected to Be Just as Great (or Even Greater) in Competitive Seats as in Non-competitive Ones." IDEAS: Economics and Finance Research. 11 Nov < Alford and Brady, Partisan and Incumbent Advantage in U.S. House Elections, , Center for the Study of Institution and Values, Rice University, 1988 Black, Dan, Galdo, Jose and Smith, Jeffrey. Estimating The Selection Bias Of The Regression Discontinuity Design Using A Tie-Breaking Experiment. Syracuse University (2005). Central European University. Web. < Evaluation/Black-Galdo-Smith-2005-RegressionDiscontinuity.pdf> Carroll, Susan J. Women as Candidates in American Politics. Bloomington: Indiana UP, Cover, Albert D. One Good Term Deserves Another: The Advantage of Incumbency in Congressional Elections American Journal of Political Science Vol. 21, No. 3 (Aug., 1977), pp Cox, Gary W. and Katz, Jonathan N. Why Did the Incumbency Advantage in U.S. House Elections Grow? American Journal of Political Science Vol. 40, No. 2 (May, 1996), pp Darcy, R., and Sarah Schramm. "When Women Run Against Men %u2014 Public Opin Q." Public Opinion Quarterly 41.1 (1977): Oxford Journals Social Sciences Public Opinion Quarterly. Web. < Dolan, Kathleen. "The Impact of Candidate Sex on Evaluations of Candidates for the U.S. House of Representatives*." Social Science Quarterly 85.1 (2004):

33 Erikson, RS, The advantage of incumbency in Congressional elections. Polity Vol. 3, No. 3 (Spring, 1971), pp Ferreira, Fernando, and Joseph Gyourko. "Do Political Parties Matter? Evidence from U.S. Cities." Quarterly Journal of Economics (2009): Ferreira, Fernado, and Joseph Gyourko. "Does Gender Matter for Political Leadership? The Case of U.S. Mayors." The Wharton School (2010). University of Pennsylvania. The Wharton School. Web. < Gagliarducci, Stefano and Paserman, Daniele, Gender Interactions within Hierarchies: Evidence from the Political Arena (April 2009). NBER Working Paper Series, Vol. w14893, pp. -, Available at SSRN: Gelman, Andrew, and Gary King Estimating incumbency advantage without bias. American Journal of Political Science 34(4): Huddy, Leonie and Terkildsen, Nayda The Consequences of Gender Stereotypes for Women Candidates at Different Levels and Types of Office Political Research Quarterly September : , ION Network. How Change Happens The 7th Annual Status Report of Women Directors and Executive Officers of Public Companies in Fourteen Regions of the United States. Rep. The InterOrganization Network (ION), Mar Web. < Kahn, Kim Fridkin. "Does Being Male Help? An Investigation of the Effects of Candidate Gender and Campaign Coverage on Evaluations of U.S. Senate Candidates." The Journal of Politics 54.2 (1992): 497. Print. 29

34 Kahn, Kim Fridkin. "Gender Differences in Campaign Messages: The Political Advertisements of Men and Women Candidates for U. S. Senate." Political Research Quarterly 46.3 (1993): 481. Print. Kahn, Kim Fridkin. "The Distorted Mirror: Press Coverage of Women Candidates for Statewide Office." The Journal of Politics 56.1 (1994): 154. Print. Lawless, Jennifer L., and Richard Logan. Fox. It Still Takes a Candidate: Why Women Don't Run for Office. New York: Cambridge UP, Lee, David. "Randomized Experiments from Non-random Selection in U.S. House Elections." Journal of Econometrics (2008): 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 (2004): Print. Milligan, Susan. "Lecture on Women in Politics." Speech. Lecture on Women in Politics. Harvard Crimson. 10 Nov Web. < Niven, David. The Missing Majority: the Recruitment of Women as State Legislative Candidates. Westport, CT: Praeger, Palmer, Donald, and Amanda Kimball. UC DAVIS STUDY OF CALIFORNIA WOMEN BUSINESS LEADERS A Census of Women Directors and Highest-Paid Executives. Rep. UC DAVIS, 2 Dec Web. < Pettersson-Lidbom, Per. "Do Parties Matter for Economic Outcomes? A Regression- Discontinuity Approach." Journal of the European Economic Association 6.5 (2008): Print. 30

35 Poutvaara, Panu, Berggren, Niclas and Jordahl, Henrik, The Looks of a Winner: Beauty, Gender and Electoral Success (September 2006). IZA Discussion Paper No Available at SSRN: Schwarz, Jay A., Election Behavior with High-Female Representation: A Regression Discontinuity Analysis (October 7, 2010). Available at SSRN: Schwindt-Bayer, L. "The Incumbency Disadvantage and Women's Election to Legislative Office." Electoral Studies 24.2 (2005): Print. Seltzer, Richard, Jody Newman, and Melissa Voorhees Leighton. Sex as a Political Variable: Women as Candidates and Voters in U.S. Elections. Boulder, CO: Lynne Rienner, Print. Smith, Eric R. A. N., and Richard L. Fox. "The Electoral Fortunes of Women Candidates for Congress." Political Research Quarterly 54.1 (2001): 205. Print. Uhlaner, Carole Jean, and Kay Lehman Schlozman. "Candidate Gender and Congressional Campaign Receipts." The Journal of Politics 48.1 (1986): 30. Uppal, Y. (2010), Estimating Incumbency Effects In U.S. State Legislatures: A Quasi- Experimental Study. Economics & Politics, 22: doi: /j x 31