The Contingent Effects of Candidate Sex on Voter Choice

The Contingent Effects of Candidate Sex on Voter Choice The Contingent Effects of Candidate Sex on Voter Choice Abstract A prominent explanation for why women are significantly underrepresented in public office in the U.S. is that stereotypes lead voters to favor male candidates over female candidates. Yet whether voters actually use a candidate’s sex as a voting heuristic in the presence of other common information about candidates remains a surprisingly unsettled question. Using a conjoint experiment that controls for stereotypes, we show that voters are biased against female candidates but in some unexpected ways. The average effect of a candidate’s sex on voter decisions is small in magnitude, is limited to presidential rather than congressional elections, and appears only among male voters. More importantly, independent voters have the greatest negative bias against female candidates. The results suggest that partisanship works as a kind of “insurance” for voters who can be sure that the party affiliation of the candidate will represent their views in office regardless of the sex of the candidate. The number of female candidates running for elective office in the United States has been increasing for decades, but women are still underrepresented in politics compared to men. Women hold fewer than one in five congressional seats despite the fact that women compose a majority of the U.S. population. Why is there such an immense disparity in descriptive representation? One prominent answer is that voters are biased against female candidates. When asked directly, around 20 percent of respondents in the United States answer that they believe men make better political leaders than do women.1 Even attempts to elicit sensitive attitudes in a more honest manner find a segment of the population hostile to seeing a woman as president Burden, Ono, and Yamada 2017; Streb et al. 2008). Yet the reasons for this bias remain unclear. Theory suggests that voters might use candidate sex as a proxy for other information that actually affects how they vote. Voters frequently make decisions based on heuristics or information cues (Lupia and McCubbins 1998; Popkin 1991), so those who are biased against (or toward) female candidates might readily use candidate sex as a shortcut that implies other information about candidates. This is often done through stereotyping. Although voters do not necessarily assign feminine attributes to female candidates (Bauer 2017; Brooks 2013; Dolan 2014; Schneider and Bos 2014), voters often stereotype candidates based on their gender and associate certain personality traits and policy positions with men and women (Alexander and Andersen 1993; Huddy and Terkildsen 1993; Kahn 1994; Koch 2002; Lawless 2004; McDermott 1997; Sanbonmatsu and Dolan 2009; Sapiro 1981). For example, lacking direct knowledge of such things, a voter might assume that a female candidate is more focused on domestic policy 1 This result was drawn from the World Values Survey Wave 6, conducted in the United States in 2011 (http://www.worldvaluessurvey.org/WVSDocumentationWV6.jsp). 1 issues such as education or that a male candidate is more experienced. At the same time, candidate sex is only one of many information cues that are available to voters. It is likely that other important cues overwhelm the effect of candidate sex on voting decisions. Political party is the most likely of these cues. Scholars have documented that politics is becoming highly polarized along party lines not only at the elite level but also at the mass level, a trend that makes party a more useful diagnostic for how a candidate would act in office. The polarization of partisans spills over beyond political choices and even exerts a significant influence on social judgments and economic behavior (Iyengar, Sood, and Lelkes 2012; Iyengar and Westwood 2015; Mason 2015; McConnell et al. 2017). For partisan voters who have hostile feelings toward the opposing party, a candidate’s party affiliation might be a determinative cue for their choice of candidates that overwhelms other information about candidates. Even if stereotypes are influential amidst strong party cues, it remains unknown whether voters also discriminate against female candidates based on sex per se. Our contribution in this manuscript is threefold. First, and most importantly, using an original conjoint experiment, we uncover the direct effect of candidate sex on voter decisions after accounting for various other information about candidates, especially their partisanship and gender stereotyped attributes. Our research design, which is relatively new in political science, enables us to not only to more closely reflect how candidates are presented to voters than do traditional survey experiments by jointly varying numerous candidate attributes at a time, but also to identify the extent to which candidate sex matters in voter evaluation relative to other crucial cues about candidates. Second, we manipulate the level of office that candidates in our experiment are seeking, to assess whether the bias toward female candidate varies between congressional and presidential elections. No experiment has tested whether voters’ treatment of 2 male and female candidates depend on the political office being sought. Our experiment provides a unique opportunity to test the theory of gender-office congruency. Third, we demonstrate how the marginal effect of candidate sex varies across subgroups within the population by carefully scrutinizing the interactions of candidate sex with voters’ background characteristics such as gender and partisanship. While the gender-affinity effect has been tested extensively in the literature, only limited attention has been paid to whether same-gender voting still appears when party cues are present. To mimic the different role that party plays in primary and general elections, our conjoint experiment also randomly creates the context of electoral competition where a pair of candidates shown to voters either shares or does not share the same party label. This enables us to identify whether and to what extent a candidate’s party label helps partisan voters to overcome bias against male or female candidates. To preview the results, we find that voters use candidate sex as a heuristic even when provided with information on other attributes of candidates such as party affiliations and policy positions. Voters are generally biased against female candidates and punish them compared to an identical male candidate. Although the average effect of a candidate’s sex on vote choice is relatively small in magnitude compared to the effects of information about party affiliations and policy positions, its impact on election outcomes could be decisive in tight races such as the 2016 presidential elections where small margins of votes separate the candidates. This reality might deter women away from running for elected office in the first place. At the same time, the bias against female candidates appears only among male voters and is limited to presidential rather than congressional elections. The results of our study further reveal that independent voters, who do not rely on a candidate’s party affiliation as a cue, have the greatest negative bias against female candidates among the public. Moreover, Republican 3 voters lose their hostility toward female candidates when candidates can be differentiated by a party label, while Democratic voters do not use candidate sex as a voting heuristic in any context. This finding suggests that partisanship works as a kind of “insurance” for Republican voters who can be confident that the party affiliation of the candidate will represent their views in office regardless of the sex of the candidate. This further implies that the particularly acute underrepresentation of women in the Republican Party may be partly driven by the biases that female candidates are likely to face in primary elections, where the party label is not a point of differentiation. Candidate Sex and Voters’ Evaluation of Candidates Previous studies suggest that voters make inferences based on a candidate’s sex (McDermott 1997). For instance, voters often presume that female candidates lack masculine traits such as competence and strong leadership, traits that are often considered to be significant for elected officials to achieve success in politics (Alexander and Andersen 1993; Huddy and Terkildsen 1993; Lawless 2004; Schneider and Bos 2014). Some studies argue that such gender stereotypes lead voters to favor male candidates over female candidates (Dolan, Deckman, and Swers 2015; Lawless 2004). In contrast, others claim that gender stereotypes exert almost no influence on the evaluation of female candidates among voters (Brooks 2013) and that party and issue cues are weightier than gender stereotypes (Anderson, Lewis, and Baird 2011; Dolan 2014a, 2014b; Hayes 2011; Matland and King 2002; Thompson and Steckenrider 1997). Female candidates are also assumed to be able to deal more effectively with “women’s issues,” such as those concerned with the environment, education, and healthcare, while male candidates are viewed to be well suited to deal with issues such as defense, crime, and the economy (Dolan 2010; Huddy and Terkildsen 1993; Kahn 1994; Sanbonmatsu and Dolan 2009). 4 Not only are male and female candidates thought to be interested in different policy areas, but they are also considered to have different stances on those issues. Voters in particular tend to think female candidates are more liberal and progressive than their male counterparts on various issues (Koch 2000; Sapiro 1981). As important as these perceptions are, they are distinct from the direct effects of candidate sex that result from a voter’s “baseline gender preference” (Sanbonmatsu 2002). While scholars generally agree that voters have certain gender stereotypes toward men and women running for public office, it remains an unsettled question whether candidate sex has an independent effect on voter evaluation. Failure to account for stereotypes and other candidate characteristics makes it ambiguous as to whether female candidates are disadvantaged compared to their male counterparts. We hope to resolve three key sources for this ambiguity. First of all, if voters display a bias against female candidates, it might be because they dislike the idea of women in office per se – a “baseline gender preference” for male politicians – or because they associate female politicians with stereotypes such as passivity and a focus on issues such as health care that they value less than male-linked stereotypes such as strong leadership style and a focus on economic policy. In addition, during electoral campaigns, voters receive other significant information about candidates, including characteristics such as party affiliation, policy positions, personal background, and polling results (Lau and Redlawsk 2006). Because voters often enter the campaign with ideological orientations, issue preferences, and attachments to particular parties, information about these attributes of candidates may play a more significant role than candidate sex in deciding their vote choice even if voters have certain stereotyped views toward men and women running for electoral office. This leads to a testable hypothesis that voter do not use candidate sex as a voting heuristic after accounting for 5 stereotypes and other candidate characteristics. Any “baseline gender preference” that exists within the population might vary across subgroups. The literature has paid particular attention to the difference between male and female voters to examine whether female voters support female candidates at higher rates than do male voters due to “gender affinity” (Dolan 2008; Rosenthal 1995; Sanbonmatsu 2002). The gender affinity effect has received mixed support in the literature. Some studies show that women do not necessarily vote for female candidates more than they do for male candidates (Ekstrand and Eckert 1981; Higgle et al. 1997; Lynch and Dolan 2014; Sapiro 1981) and that this effect is limited under the presence of policy issue cues (Anderson, Lewis, and Baird 2011). Our study is able to resolve whether there is a baseline preference for candidates of one sex or the other for both male and female voters. Second, female candidates appear to face a greater challenge when they run for executive office than when they run for legislative office (Huddy and Terkildsen 1993; Lawrence and Rose 2014; Rose 2013). The common explanation for the difference across offices also hinges on stereotypes in that voters may perceive male candidates are more likely to have characteristics such as strong leadership and to emphasize issues such as foreign policy that align well with expectations of presidents, while female candidates are more likely to be seen as compassionate and emphasizing domestic issues such as health care and education that are wellsuited to being a legislator (Eagly and Karau 2002; Kahn 1996; Koch 2002; Sapiro 1981). This “gender-office congruency theory” thus generates a hypothesis that the use of candidate sex as a voting heuristic among voters does not vary across political offices being sought when we isolate the effect of candidate sex from stereotypes that voters associate with male and female politicians. 6 The third question we address is the degree to which partisanship dominates other factors. Partisanship has been considered as the most important factor behind most voting decisions; in the contemporary polarized environment, party labels convey a lot of information and might leave little room for partisan voters to rely on candidate sex. We analogize that voting for a co-partisan provides a kind of “insurance” that essentially guarantees how a politician will act in office. This leads to a hypothesis that independents, who do not have any attachments to particular parties, are more likely to rely on candidate sex when they evaluate candidates and to exhibit the greatest bias against female candidates among voters. But this effect may be limited to general elections where party labels differentiate candidates. Candidate sex is likely to operate differently in a primary where voters make choices between two competing candidates of the same party. For Democrats, candidate sex does not convey as a clear ideological signal; thus, the effect of candidate sex is expected to be marginal regardless of a candidate’s party label. In contrast, Republicans tend to view female candidates more liberal than comparable male candidates, which might lead them to withdraw their support to female candidates unless a party signal differentiates the two competing candidates (see King and Matland 2003). This implies that both Democrats and Republicans do not rely on candidate sex when the party label is a point of differentiation, and that candidate sex has different effects between them when they choose a candidate from those running from the same party. Research Design It is challenging to isolate gender effects in observational data where female candidates might have been selected differentially during the recruitment process (Dolan and Sanbonmatsu 2011). The quality of emerging female candidates indeed differs significantly from their male counterparts (Anzia and Berry 2011; Fox and Lawless 2010; Lawless and Pearson 2008). 7 Candidates are also strategic actors whose anticipation of the electorate’s response can shape their decisions. Thus, an extensive amount of research has conducted survey experiments to understand the effect of candidate sex on voter decisions (e.g., Bauer 2016; Brooks 2013; Fridkin, Kenney, and Woodall 2009; Iyengar et al. 1996; Kahn 1994; Sanbonmatsu 2002; Sapiro 1981). These experimental studies contribute much to our understanding by intentionally manipulating candidate profiles and behavior. We build on these studies by varying more candidate attributes simultaneously and randomizing the order of those attributes. We test the hypotheses laid out above using a conjoint survey experiment. Recently introduced to political science, conjoint experiments were widely used in marketing to assess the impact of many product characteristics simultaneously. Unlike more familiar factorial designs, conjoint experiments vary all treatments simultaneously to allow for assessment of the impact of the independent and interactive effects of multiple variables on a common outcome metric without sacrificing much statistical power or imposing assumptions about functional form (Hainmueller, Hopkins, and Yamamoto 2014). This degree of power and flexibility is appealing in our application because it allows us to randomly vary many more candidate traits than have previous studies. This permits us not only to separate the baseline effect of candidate sex from the other characteristics such as gender stereotypes or partisanship that voters might infer, but also to assess simultaneously the relative importance of these factors on voter evaluation. In our experiment, we present each subject with a pair of opposing candidates whose profiles are randomly generated from the set of characteristics, and then ask him/her to choose between the two candidates. Although some studies give each subject only one candidate at a time to evaluate, presenting two opposing candidates is more realistic as it mimics the decision that voters must make on real ballots. We asked respondents which candidate they would vote for 8 if it was an actual election, a familiar task for most of the electorate. The profiles of candidates are created in line with the existing literature on voter decisions as well as gender stereotypes more specifically (Lau and Redlawsk 2006; Lynch and Dolan 2014). We manipulate candidate attributes within four broad categories of information that a voter might encounter in a salient campaign: personal information, party information, issue information, and polling information. First, to vary personal information that describes a candidate’s backgrounds and personality in the candidate profiles, we include a candidate’s sex, race/ethnicity, age, marital status, experience in public office, and salient personality trait. These attributes have been considered to play an important role when voters make decisions. For instance, scholars have paid extensive attention to the effect of a candidate’s race on voter decisions in expectation that black candidates are disadvantaged compared to white candidates. We are especially interested in personality traits because, in the absence of direct information, these are the stereotypes that voters are most likely to ascribe to male and female candidates. These perceptions rather than sex itself might drive voting choices. For example, the perceived competence and personality traits of candidates have been discussed as possible sources of disadvantage to women running for public office (Fridkin and Kenney 2011). Most importantly, as noted above, male candidates are often viewed to be more decisive and stronger leaders than female candidates. Hence, having experience in public office may help female candidates fend off criticism over the lack of competence toward them; and having personality traits that contradict their gender stereotypes may alleviate bias against female candidates. The personality traits of candidates are adopted from survey questions conducted by the American National Election Studies that routinely ask whether candidates provide strong leadership, are compassionate, are honest, are intelligent, are knowledgeable, and really care about people like you. 9 Second, many American voters have psychological attachments to one of the major political parties, making it natural to rely on party information about candidates. Some recent studies suggest that, even if voters hold gendered attitudes, they are still heavily influenced by a candidate’s party label rather than a candidate’s sex when they decide for whom to vote (Anderson, Lewis, and Baird 2011; Dolan 2014a; Falk and Kenski 2006; Hayes 2011; Matland and King 2002). Hence, we include a candidate’s party label as an essential attribute to vary in the candidate profiles. Importantly, by varying the party labels for both candidates, we are able to investigate the effects of candidate sex when the candidates are from opposing parties (as in a general election) and when they are from the same party (as in a primary election). Third, the effect of candidate sex on voter decisions may be dampened when issue information such as a candidate’s policy positions and expertise is given to voters. Female candidates are often seen to have different policy priorities and preferences from their male counterparts (Swers 2002). The issue information attributes in the candidate profiles include a candidate’s positions on abortion, immigration, the federal budget deficit, and national defense. In addition, to distinguish positions from emphasis, we include the following six policy areas as varying attributes to describe the policy specialization of candidates: economic policy, foreign policy, public safety (crime), education, health care, and the environment. Fourth, and finally, we vary polling information that describes a candidate’s popular support in the public. Voters have been known to rely on polling data as a sign of the relative desirability of candidates. Therefore, they may engage in strategic voting behavior by jumping on the bandwagon when one candidate is performing well in the polls. Voters with such incentives may pay attention to the public opinion when they evaluate candidates. The reported favorability rating of each candidate is randomly varied among five levels from relatively 10 unpopular (34%) to highly popular (70%). Table 1 summarizes all the attributes of candidate profiles used in our experiment. There is a total of 13 varying attributes. Some attributes take only two values, as in the case of partisanship (Democrat or Republican), whereas others take on several values, as in the case of polling favorability (five values ranging from 34% to 70%). For each profile, we randomly assign a value of each attribute. This research design yields 9,953,280 possible combinations of candidate profiles. The use of a conjoint experiment rather than a factorial design makes it possible to allow all of these combinations to be realized and to estimate the effect of each attribute with a modest number of observations. [Table 1 about here] Our experiment asks respondents to review the profiles of two candidates that are randomly created from the set of attributes and then to choose between them. This evaluation task is repeated ten times, with each pair of candidates displayed on a new screen. The categories of attributes of candidates such as age and experience in office are shown in randomized order across respondents so that the exercise does not inadvertently focus respondent attention on specific attributes.2 Because so many attributes are varied, we think it is unlikely that respondents would be able to surmise the purpose of the experiment and behave strategically rather than simply choosing the candidate that seems most appealing. Figure 1 presents an example of one set of congressional candidate profiles that was shown to a respondent in our experiment. The visual presentation mimics the one used by Hainmueller, Hopkins, and Yamamoto (2014) and these examples illustrate how types of 2 The order is fixed across ten pairs for each respondent to minimize his or her cognitive burden. 11 attributes are varied across respondents.3 [Figure 1 about here] In addition to varying attributes of the candidates, we also vary the office being sought to examine whether candidate sex has a different effect on voter decisions between presidential and congressional elections. Specifically, we split the ten pairs of candidates being evaluated into five sets of congressional candidates and five sets of presidential candidates, and ask respondents to evaluate candidates both for president and for Congress. The order of evaluation is randomly determined across respondents. Hence, approximately half respondents in our experiment first evaluated five pairs of candidates for president and then moved on to evaluating another five pairs of candidates for the House of Representatives; the remaining respondents evaluated the two groups in the reversed order.4 3 This design creates some combinations of candidate attributes that would be rare in real elections. For instance, voters seldom encounter a candidate who is female, black, Republican, pro-choice, and who wants to increase taxes. Such implausible combinations may introduce some biases to the results by leading our subjects to make artificial judgments without much cognitive effort (Auspurg, Hinz, and Liebig 2009). However, as we show in the appendix, the likelihood of having implausible combinations is limited, and the results remain almost the same even after excluding those combinations. We address this issue at greater length in the appendix using response timers to assess whether respondents take the experiment less seriously when candidates’ policy positions seem incongruent with their party affiliations. 4 We randomize the order of evaluation in this way to mitigate the concern that respondents may change their behavior when they evaluate the latter five pairs of candidates. 12 Data and Method of Analysis We collected data through an online survey experiment that was fielded in March 2016. The sample of voting-eligible adults in the United States was drawn by Survey Sampling International (SSI).5 In collecting the data, we stratified the sample by the region of residence (Northeast, Midwest, South, and West), sex (male and female), race/ethnicity (white, black, Hispanic, Asian, and others), and age groups, based on the latest U.S. Census data. To be more specific, we first asked demographic screening questions to 3,152 people in the SSI panel, and then invited 1,733 respondents among them to our survey according to fixed quotas so that the total sample matches the adult U.S. Census population on age, sex, geographic region, and race/ethnicity.6 In the survey, we also collected data of other personal information about respondents such as educational background, social class, partisanship, political interest, and 5 SSI samples of this kind have been used in a variety of survey studies published in top tier journals in political science (Berinsky, Margolis, and Sances 2014; Bullock 2011; Iyengar and Westwood 2015; Kam 2012; Malhotra, Margalit, and Mo 2013). 6 Some might be concerned that actual voters may differ from the overall adult population. While our survey does not directly ask respondents whether they have cast a ballot in the general election, we have a measure of their political interest. The turnout and political interest are indeed highly correlated. According to data from the 2012 ANES, 92% of people who are very interested in politics report they voted for the election; in contrast, only 37% of people with no political interest report their positive turnout. However, we find that the bias against female candidates does not vary across respondents with different levels of political interest (see the appendix for more details). 13 ideological position. A total of 1,583 respondents completed the conjoint experiment tasks in our survey (a completion rate of 91.3%). A detailed descriptive statistics on our sample are shown in the appendix. Because each of our respondents evaluated ten pairs of candidates, we have data from 31,660 profiles or 15,830 evaluated pairings. This is large enough to estimate the effect of each attribute in the candidate profiles. The outcome variable of interest in this study is which candidate was chosen by a respondent. The choices are coded as a binary variable, where a value of one indicates that a respondent supported the candidate and zero otherwise. We analyze the data following the statistical approach developed in Hainmueller, Hopkins, and Yamamoto (2014) to estimate nonparametrically what they define as the average marginal component effect (AMCE). To estimate the AMCE of each attribute on the probability that the candidate will be chosen, we employ the “cjoint” package (ver. 2.0.4) developed by Strezhnev et al. (2016). The standard errors are clustered by the respondent to account for the dependence of observations across respondents. In the following sections, we first present the average direct effect of a candidate’s sex on voter decisions, including separate examinations for presidential and congressional candidates to test for the gender-office congruency hypothesis. We then present the results that test for heterogeneous treatment effects by including the interactions between candidates’ attributes and respondents’ characteristics to reveal how bias against female candidates varies among different subpopulations within the electorate to test for the gender-affinity hypothesis. Finally, we explore differences by party to test whether party affiliation moderates any gender bias that exists in the full population. 14