You seem like a great candidate, but…: race and gender attitudes and the 2020 democratic primary

Abstract

The 2020 Democratic presidential primary unfolded in a context with significant attention to issues of racial and gender inequality and identity. The field began as an historically diverse one but a white male candidate received the party's endorsement. Did the race and gender attitudes of Democratic primary and caucus participants play a role in shaping the pool of candidates? Using a survey of self-identified Democrats, this study provides evidence that racial resentment, hostile sexism, and modern sexism enhanced the assessments on several evaluative criteria of the white male candidate, while depressing the assessment of the Black woman candidate. These relationships are driven primarily by white respondents. These findings add to our understanding of how race and gender attitudes affect the electoral process well before the general election, particularly by shaping the ultimate choice of candidates in that contest.

The JREP special issue on “Race and the American Presidency” asks authors, among other things, to explore “the ways that racial and ethnic politics have shaped the institution of the Presidency.” The focus of this contribution is to train that question on one of the earliest stages of the electoral process: how do race and gender attitudes affect the decisions of voters who participate in the primary or caucus process? These early decisions shape the ultimate choice in the general election—in particular, if attitudes at the primary stage make a harder path for racial and gender minorities, it will be rare for voters in the general election to have a choice between anything but white male candidates.

The 2020 Democratic presidential primary contest presents an opportunity to explore this question. The contest unfolded in the shadow of the 2016 election; the 2018 midterm elections; and Donald Trump's term in office more generally. This backdrop was dominated by heightened conversation and conflict over the enduring impacts of racism and sexism at a societal level, in our daily lives, and on our electoral decisions. In this environment, 28 candidates for the Democratic nomination emerged. The sheer number was notable, but the field was also noted for its unprecedented level of diversity. This diversity emerged with regard to race (with a field that included “two [B]lack senators, a Latino former cabinet secretary, an Asian-American businessman and the first American Samoan elected to Congress” (Herndon and Martin, 2019)) and gender (with a field that included six women). But the field also included the “first openly gay major presidential candidate” in South Bend, IN Mayor Pete Buttigieg (Epstein and Gabriel, 2020), and encompassed candidates with a wide range of ages—40 years separated the youngest and oldest candidates (Wilson, 2019). The field also comprised of a range of experience, from a former Vice-President to several sitting members of Congress to one-term mayors and business people (Klahr et al., 2020) and the candidates often distinguished themselves in terms of ideology, emphasizing either their moderation or their support for liberal policies.¹

However, although the field was initially touted for its diversity, by the end of 2019, observers began to note that the endorsement winnowing process was favoring historically typical presidential candidates. Senator Kamala Harris was the only non-white candidate to qualify for the December 19th debate; when she withdrew from the race, the debate included all white candidates (Khalid, 2019). Senators Elizabeth Warren and Amy Klobuchar remained relatively viable candidates, but withdrew from the race in March, after disappointing showings in several states (Corasaniti and Burns, 2020; Goldmacher and Herndon, 2020). Ultimately, the contest came down to Senator Bernie Sanders and Vice-President Joe Biden—two white men in their 70s with extensive political resumes—whose distinguishing appeals primarily focused on ideological differences between the two of them and which candidate stood the best chance of beating President Trump (e.g., Mack and Barrett, 2020; Oliphant, 2020).

What factors drove the transformation of the Democratic field from a historically diverse one to a white male nominee? In particular, did Democratic primary and caucus participants' attitudes about race and gender play a role in shaping the remaining pool of candidates? Using a survey of self-identified Democrats conducted in October 2019, I investigate the factors that influenced voters' evaluations of four competitive candidates who varied in terms of race and gender: Vice President Joe Biden; Senator Cory Booker; Senator Kamala Harris; and Senator Elizabeth Warren. In the analyses that follow, I find evidence that among white self-identified Democrats, racial resentment consistently works to Biden's benefit and Harris's disadvantage. In addition, in certain circumstances, gender attitudes work to women candidates' disadvantage, primarily among white respondents. These findings suggest that race and gender attitudes are shaping the Presidency very early in the electoral process.

1. The role of race and gender attitudes in electoral vote choice in 2016 and beyond

The 2016 election provided ample data to better understand how race and gender attitudes might affect vote choice in a contemporary context. Of course, the study of the impact of race and gender attitudes on electoral decisions has a long history, and was already increasingly salient in the Obama era. In this issue, Jacobsmeier (2020) shows that whites, and particularly racially resentful whites, perceived Obama as more liberal. House races during the Obama era also provide evidence of the importance of racial attitudes, as whites were less likely to vote for black Democratic candidates (Hale, 2020). However, the overtly racist and sexist rhetoric of the 2016 election—and the victory of the candidate behind that rhetoric—sustained the urgency to understand these dynamics in the context of the 2016 election and beyond. Consequently, recent studies have largely sought to illuminate the relationship of gender and race attitudes—as well as policy preferences and political attitudes more generally—to the propensity to vote for Donald Trump over other candidates in the primary and general elections. This research identified several measures of race attitudes that predicted support for Trump's candidacy. For example, Sides et al. (2019) found that the traditional measure of racial resentment—which measures whites' feelings toward Blacks as an outgroup—predicted support for Trump in the 2016 primary and general elections. Similarly, in this issue, Buyuker et al. (2020) found that xenophobia—another measure of whites' attitudes toward outgroups—also predicted support for Trump, even more strongly than racial resentment. Elsewhere, Lajevardi and Oskooii (2018) find effects of resentment toward Muslim Americans on support for Trump. And, in another strand of research, scholars have developed and tested measures of whites' feelings toward the ingroup. As Jardina (2019) argues, these feelings of group solidarity operate independently of measures of discrimination against an outgroup, but provide another mechanism by which whites' racial attitudes affect vote choice. Here, too, whites' racial attitudes affected their support for Trump, in the primary and the general elections (see, also, Lopez Bunyasi, 2019).

Similar scrutiny has been put to understanding how gender attitudes shaped candidate preferences in 2016 and beyond. Based on Donald Trump's treatment of and attitudes toward women, many in the popular press predicted the emergence of a significant gender gap in preference for Trump compared to Clinton (e.g., Gabriel, 2016). Initial analysis of vote totals reinforced the narrative of the gender gap, but this gender gap did not hold for white women and many attempted to understand how a majority of white women voted for Trump (e.g., Chira, 2016). Scholarly treatments of voting behavior in 2016 incorporated measures of gender and race attitudes in order to understand how race, gender, party identification, economic attitudes, and education came together to predict whites' votes for Trump. In particular, Schaffner et al. (2018) found that gender and race attitudes were far stronger predictors of a Trump vote than were economic attitudes. Frasure-Yokley (2018) found that ambivalent sexism—a measure that includes measures of hostile and benevolent sexism (Glick and Fiske, 1996)—strongly predicted support for Trump in the 2016 election among white women; ambivalent sexism was not a predictor of Trump support among Black women (see also Bracic et al. (2019) and Phillips (2018) for additional evidence of the impact of gender attitudes for white women's support of Trump). Similarly, Cassese and Barnes (2019) look specifically at white women's voting behavior in 2012 and 2016. Hostile sexism and modern sexism were particularly strong predictors of voting for the Republican candidate in the 2016 election.

These findings illustrate two important themes regarding race and gender attitudes. First, measures of specific attitudes about race and gender provide more insight into voting behavior than do analyses that simply account for whether respondents are men versus women or whites versus people of color. Second, while implementing these attitude measures, it is clear that analyzing aggregated samples obscures important differences between those who identify as white and those who do not. The recent study illuminating how gender attitudes operate differently for white women and Black women (Frasure-Yokley, 2018) builds on well-established findings of the interplay of race and gender identification for Black women (Gay and Tate, 1998; Simien, 2005; Philpot and Walton, 2007). As a result it is vital to disaggregate samples by politically-meaningful identities so that the relationships of majority groups in the data are not extrapolated to smaller, but politically consequential, groups.

Party identity also presents itself as a key additional identity when it comes to assessing the contextual importance of the manifestation of race and gender attitudes. Descriptive statistics alone suggest differences in attitudes between identifiers of the two major parties. On average, Democrats report lower levels of hostile sexism than the general public (Luks and Schaffner, 2019). There are consistent partisan differences in attitudes on the enduring impact of race. For example, Democrats are more likely than Republicans to attribute disparities to systemic sources as opposed to individual behavior (Horowitz et al., 2019). From the perspective of the Democratic Party, these differences are probably causes and effects of the way the party has worked to position itself in relation to issues of gender and race. At least in the past few decades, the Democratic Party has moved toward representing liberal positions on women's issues such as abortion rights (Wolbrecht, 2000; Sanbonmatsu, 2004) and positioning itself as the alternative to the Republican Party, which, in recent popular parlance, has been accused of waging a “war on women.” Democratic voters appear to bear out these women friendly-attitudes, given that women candidates prevail at rates equal to men candidates in Democratic congressional primaries (Lawless and Pearson, 2008), and that Democratic identifiers tend to have a lower baseline preference for male candidates (Sanbonmatsu, 2002). Similarly, since the 1960s, the party has moved to position itself as the party of civil rights, appealing to racial minorities (e.g., Carmines and Stimson, 1989). This emphasis is reflected in the Democratic Party's identifiers, who are significantly more racially and ethnically diverse than the Republican Party's (e.g., Horowitz et al., 2019).

However, this idealistic view of the party—and, perhaps more specifically, of party members—obscures heterogeneity within the Democratic Party (Casesse et al., 2015; Luks and Schaffner, 2019). Indeed, hostile sexism, which measures gender attitudes in a way that approximates the classic understanding of prejudice (Glick and Fiske, 1996), depressed turnout and campaign engagement among Democrats in 2016 (Banda and Casesse, nd), while racial resentment depressed turnout among white Democrats in the 2010 midterm elections (Luttig, 2017). Moreover, evidence from the 2008 Democratic primaries suggests that racial attitudes were a significant predictor of preferences between Obama and Clinton—those with higher levels of symbolic racism were more likely to support Clinton as compared to Obama (Jackman and Vavreck, 2010), and a small but notable proportion of whites voted against Obama because of his race (Huddy and Carey, 2009, 92). Huddy and Carey (2009) also found that Black racial identity was a stronger predictor of preference for Obama than women's gender identity was for Clinton, although both group identities showed an effect. Similarly, in the 2016 primary, Democrats with high levels of sexist attitudes were less likely to support Clinton, especially among men (Sides et al., 2019, 121). In short, then, race and gender attitudes are at work in the decisions of Democrats. Given the enduring salience of issues associated with race and gender, it is plausible that these attitudes would exert influence in the 2020 Democratic primary/caucus process.

2. Expectations

From a decision-making perspective, the Democratic primary contest presented a potentially overwhelming task—to determine one's preferred candidate, in a field that started at 28, without the benefit of a differentiating party heuristic, the most relied-upon signal for most political decision-making. Voters did not have the benefit of an incumbent or quasi-incumbent. Nor was there a presumed consistent front-runner, as the volatility of the race proved,² and, early in the race, no sense of momentum for one candidate over another (e.g., Bartels, 1988). Lau and Redlawsk (2006) offer some guidance as to what voters in such a primary situation might do. They found that “heuristic search”—that is, searching for candidate-related information that facilitates using “cognitive shortcuts” in making political decisions—is higher during primary contests than general elections (235). One significant factor that predicts this heuristic search is the number of candidates—large candidate fields encourage heuristic search. The 2020 Democratic primary field, then, would encourage heuristic search.

Particularly relevant to our purposes is the specific content of those heuristics. They found that one of the kinds of information that primary voters sought out was information related to “person stereotypes,” measured by respondents' propensity to seek out images of the candidates, which helps respondents assess key characteristics of the candidate, such as the candidate's race and gender. Certainly, given the image-rich information environment of the Democratic primary race, voters would not have to work hard to know this information about each of the primary candidates. First, then, it is reasonable to expect that most Democratic candidates would know this heuristic information about each of the candidates; it would be easily accessible knowledge in primary and caucus participants' minds.

It is the campaign context, then, that provides reason to believe that this information would be not just accessible, but actually used to evaluate candidates. Issues of race and gender remained salient in the primary electoral environment, given the heightened awareness of race and gender issues stemming from the 2016 election and further reinforced by the 2018 midterm elections and discussions of the 2020 primary field as it relates to its diversity. Campaigns' emphases on gender (Bauer, 2015) and race issues (Karl and Ryan, 2016) can influence the extent to which voters rely on stereotypes. Similarly, the salience of race and gender-related issues in the broader context may make primary and caucus participants more likely to incorporate these “person stereotypes” into their evaluations.

Most simply, then, we would expect gender and race attitudes to influence voters' evaluations of candidates. As a respondent's sexist attitudes increase, their evaluations of male candidates should increase, and as a respondent's level of racial resentment increases, their evaluations of a white candidate should increase. The benefit of this particular candidate pool, however, is that we can also observe how voters respond at the intersection of candidates' identities. In the case of the 2020 Democratic primary contest, we can observe voters' evaluations of a range of candidates: men and women of color and white women and men. These intersections defy a simple ranking of advantage—while it may seem straight forward to predict that white male candidates will be doubly advantaged while Black women candidates will be doubly disadvantaged, the latter, at least, is not a forgone conclusion. Tate (2003) argues that Black women candidates may be at an advantage in terms of mobilizing voters along race and gender identities. Philpot and Walton (2007) provide evidence for this argument, as they find particularly strong support for some Black women candidates from Black women voters and, to a somewhat lesser degree, Black men voters.

These findings make clear that, while the race and gender identities of the candidates are of central importance in these calculations, the race and gender identities of the respondents will also be consequential factors in whether and how respondents weigh race and gender attitudes in their candidate assessments. This is a particularly important consideration for a study of the Democratic electorate, given that this group is more racially diverse and includes a higher percentage of women than the general electorate. Beyond group membership, however, previous research provides evidence that race and gender attitudes operate differently for white men, white women, women of color, and men of color (e.g., Bracic et al., 2019; Yokley-Frasure, 2018; Cassese and Barnes, 2019). This evidence leads me to believe that it will be white respondents who will most clearly demonstrate a relationship between their gender and race attitudes and candidate evaluations, contingent on the race and gender of the candidate.

3. Methods

3.1 The sample

In order to measure the attitudes of self-identified Democrats, I undertook a survey in the fall of 2019 through Dynata. Dynata is a web-based service that uses multiple methods to contact potential survey takers and screen them into appropriate surveys. The survey was in the field from September 28, 2019 to October 2, 2019. Dynata only recruited respondents who pre-identified as Democrats; in total, 16,830 Dynata participants received an email inviting them to participate in the survey; 1,475 of those individuals clicked in to take the survey, for a response rate of 8.8%. The survey begins by asking respondents for their party identification; despite the party-based pre-screen, some non-Democrats were recruited. Once Republican and non-leaning independent respondents were dropped, as well as some respondents who sped through the survey, 1,286 respondents completed the survey.³ In total, 87% of respondents who began the survey are included in the analyzed sample. The median time for completion was 9.4 min. Table 1 provides an overview of the sample on some key variables.

Table 1. Descriptive statistics of the sample on key variables

Unless otherwise indicated, first line is percentage (frequency).

Although the use of “non-probability” web-based samples are becoming more common—partially out of sheer necessity, given the increasing difficulty of contacting a representative sample using traditional methods—it is not specifically established as to how accurate these sampling methods are at capturing respondents that mirror the broader population (e.g., Baker et al., 2010), particularly as they compare to traditional RDD telephone and face-to-face probability sample surveys (Fahimi et al., 2015).

Although imperfect, one method for assessing the representativeness of a non-probability sample is to compare metrics to a high-quality benchmark (Yeager et al., 2011), particularly on variables that are likely to correlate with primary independent variables of interest (Lopez Bunyasi, 2019). The American National Election Study (NES) provides such a high-quality benchmark. Table A1 in the Appendix provides a comparison of key variables in the current study with Democrats surveyed in the 2019 NES Pilot Study (conducted in late December 2018) and the 2016 NES study.⁴ By including two previous NES studies, we can compare the present data to a relatively recent sample (2019 NES Pilot) and the most recent presidential election year (2016 NES). Moreover, including two NES surveys provides a wide range of variables to compare. All three surveys ask basic demographic questions (i.e., race, gender, age, and education) and measure ideology and racial resentment. The 2016 NES provides a benchmark for sexism measures (which were not included in the 2019 NES) and the 2019 NES provides a comparison for candidate choice.

Across race, gender, age, and education of the respondents, the summary statistics in Table A1 are primarily characterized by a lack of difference across the samples. The present sample may somewhat overrepresent women respondents (63% of respondents are women, as opposed to 57 and 54% in the two NES samples) and may somewhat underrepresent Black respondents (17% of the respondents in the current study are Black, as compared to 19 and 20% in the two NES samples). Respondents in the current study may be slightly older and more educated compared to the NES samples, but these differences fall within one standard deviation of the present sample.

In terms of race attitudes, the means across all three samples fall within 1.29-units of each other (on a 16-point scale), with the present sample's mean falling between the two NES means. In comparing gender attitudes, the present sample's mean and the 2016 NES mean fall within .05-units of one another (on a 16-point scale for hostile sexism) and within .11-units of one another on the modern sexism measure (on a 5-point scale). Overall, ideology remains stable across all three samples as well, with the mean respondent in the present sample and the 2016 sample indicating they are “slightly liberal.”⁵

Finally, Online appendix Table A1 compares the current study and 2019 samples based on their most preferred candidate. This is a more difficult comparison to make given the quickly changing nature of the Democratic candidate field—for example, when the NES fielded its survey in December 2018, few candidates had officially announced their candidacy. In the question that asked who the respondent would be most likely to vote for in the primary/caucus, the response options included Biden, Booker, and Warren by name, but not Harris. Taking into account this difficulty in comparison, the present sample may be somewhat more supportive of Vice President Biden (37.8 versus 27.5% in the 2019 NES), while support for Senators Booker and Warren is relatively stable across the three samples. In comparing respondents' reported basis for their candidate decision, similar percentages reported that electability was the most important concern (56% in the present survey sample compared to 54% in the 2019 NES sample). (See Appendix Table A1 for more details on these comparisons.)

3.2 Measurement

3.2.1 Evaluating the candidates

Once the respondent opted into the survey, they first answered a screener question to ensure that the sample only included self-identified Democrats. Respondents also answered a set of questions that assessed their previous electoral participation, vote choice, and strength of identification with the Democratic Party. Respondents were then presented with text that indicated that, in the next set of questions, they would use four different criteria to choose a candidate from four randomly selected Democratic presidential candidates. In reality, all respondents chose from the same four candidates: Vice-President Joe Biden, Senator Cory Booker, Senator Kamala Harris, and Senator Elizabeth Warren. In order to get meaningful results with a reasonable number of respondents, the number of candidate options respondents considered needed to be focused. In choosing these four candidates, I attempted to balance the gender and race of the candidate with their competitiveness. First, with this set of four candidates, I had one white man (Biden), one Black man (Booker), one Black woman (Harris), and one white woman (Warren). Certainly, given that this survey took place in real time with real candidates, respondents held opinions about these candidates based on factors beyond the candidate's race and gender. However, the intention was to give respondents a member of each race/gender combination in order to begin to assess whether the respondents' race and gender attitudes influenced their choice of candidate on each criterion.

With the candidate's race and gender in mind, the candidates also needed to be relatively similar in their competitiveness. Name recognition, for example, is a factor in primary candidate choice (e.g., Norrander 1980), so it was vital to hold this factor relatively constant. Two of these chosen candidates—Biden and Warren—were consistently polling among the top three candidates at the time the survey was fielded (RealClearPolitics, fivethirtyeight.com). Although Booker and Harris were not polling as well, they were performing well enough to be included in the debates through November. In order to qualify for these debates, candidates needed to meet fundraising and polling targets—all four of these included candidates met both those targets until December 2019.⁶

Respondents were asked to identify which of these four candidates they preferred based on four criteria. First, respondents were asked, “If you chose your candidate strictly based on who best represents you, who would you choose of the following 4 candidates?” Respondents were then asked to make the same type of decision while considering the candidate's policy positions, the candidate the respondent thinks has the best chance of beating Donald Trump (a.k.a., electability), and the candidate the respondent likes the most. Respondents considered these evaluative criteria in the same order. The order of the candidate response options, however, was randomized for each respondent, although that randomization was held constant throughout the four criteria for each respondent.⁷

Importantly, these four questions asked respondents to evaluate the candidates rather than to indicate which candidate they intended to vote for. Given that this survey was conducted early in the primary cycle, evaluations of candidates seemed like a measure that would be more stable as candidates dropped out, voters needed to move to their second or third choices, and the context around those candidates continued to change (perhaps, most saliently, in terms of who the specific opponents were). These evaluations allow for a range of possible effects on vote choice. For example, Norrander (1986) found effects of “candidate qualities” on vote choice, which is reflected in the evaluative questions about likability and which candidate represents you best.

The question about congruent policy beliefs flows from Aldrich and Alvarez (1994), who found that a candidate's policy emphases predict voter support. Finally, the fourth criterion, electability, asks the respondent to consider not just their personal proximity to each of the candidates, but also who they think other people will and will not support. This criterion received considerable attention in the media, as it was a central concern among a Democratic electorate that was particularly motivated to beat the incumbent president (see, e.g., YouGov/HuffPost, 2019). Moreover, this has been a consideration of previous primary electorates in certain circumstances (Norrander, 1986).

In the next section of the survey, respondents were asked to indicate their first choice of candidate overall. At the time the survey was conducted, there were 19 declared candidates for the Democratic nomination. Each of the options was listed, as well as an “other” option. Respondents also indicated their basis for choosing their most-preferred candidate from the following options: electability, likeability, stances on policy positions, and feeling represented.

3.2.2 Attitudinal variables

The final section of interest included several questions that measured attitudinal variables. In this section, respondents first considered six policy questions; combined, these items represent how liberal the respondent is on policy specifically.⁸

Also key to this section were measures of hostile sexism, modern sexism, and racial resentment. Respondents answered the following:

• • Six items from the hostile sexism scale. Hostile sexism most closely resembles traditional ideas of sexist prejudice (Glick and Fiske, 1996). Each of these items is measured on a 5-point scale, and asks the respondent to agree or disagree with statements such as, “women put men on a tight leash” and “many women interpret innocent remarks or acts as being sexist.”⁹ These were rescaled so that, for each, higher numbers indicate higher levels of hostile sexism. The six questions are aggregated, resulting in a scale that runs from 6 (least hostile sexism) to 30 (most hostile sexism). Cronbach's alpha for these six items was high, indicating they could be combined (α = .89).

• • One modern sexism item. Modern sexism assesses how willing respondents are to believe that gender inequality is a result of discrimination (Swim et al., 1995). The survey measures modern sexism by asking respondents to indicate how much discrimination exists against women, on a 5-point scale, from none at all to a great deal (Cassese et al., 2015; Cassese and Barnes, 2019). This is the same measure used by the NES.

• • Four racial resentment questions from the NES. Racial resentment, in short, measures whites' animosity toward Blacks as an outgroup, with that animosity couched in terms of violating norms of individualism (Kinder and Sanders, 1996). Similar to the hostile sexism scale, respondents are asked to agree or disagree with four items on a 5-point scale. These items include questions such as, “Irish, Italians, Jewish, and many other minorities overcame prejudice and worked their way up. Blacks should do the same without any special favors,” and “Over the past few years, Blacks have gotten less than they deserve.” Items are rescaled so that higher numbers indicate higher levels of racial resentment. The four items are aggregated, resulting in a scale that runs from 4 (lowest racial resentment) to 20 (highest racial resentment). The Cronbach's alpha for this scale was high enough to support aggregating the items (α = .77).

Respondents finished with several demographic questions, including respondent ideology, measured on a 7-point scale (from extremely conservative to extremely liberal). Lau and Redlawsk (2006) found ideology heuristics to be another important source of information for primary voters. Particularly given the ideological diversity of the candidates included for evaluation, this is an important control variable.¹⁰ Respondents also answered demographic questions regarding race, gender, education, and age.

4. Results

In the face of 19 potential candidates, at least one respondent chose each candidate as the candidate they intended to vote for in the primary/caucus (except for Wayne Messam). As Table A1 indicated, Biden is the most preferred candidate with almost 38% of respondents choosing him. Warren follows with 22%. Sanders is third, with almost 11%. In descending order after Sanders is Harris, Buttigieg, and Booker. Table 2 shows respondents' priorities when it comes to choosing a candidate. Reflecting general assessments at the time, a majority of respondents indicated that electability was their most important consideration when they chose their most-preferred candidate (56%). Policy positions was the criterion that was top-ranked second most often, with almost 25% of respondents choosing this criterion as most important. Feeling represented and likeability came in third and fourth places, respectively.

Table 2. Most important consideration in choosing a candidate; candidate who best achieves that criterion

^a These represent the respondents' most preferred candidate for each criterion when asked to choose from these four candidate options only.

Bringing these two considerations—candidate preference and evaluative criteria—together, Table 2 also displays the percentage of respondents who preferred each of the four candidates for each evaluative criterion.¹¹ Across the board, Biden has a majority or strong plurality—he was chosen by the most respondents when asked to choose the candidate who is most electable (61%). His plurality is the smallest when respondents chose the candidate who is most likable (43%). On the contrary, Booker fares the worst across all four categories—he is least competitive when respondents chose the most electable candidate (5%) and does best when respondents chose the candidate who is the most likable (10%). In fact, in all four evaluative criteria, the candidates finish in the same order: Biden is first, followed (sometimes within 10 percentage points) by Warren, then Harris and Booker. These raw percentages might speak to some advantage for white candidates, as the top two choices across all categories are the white candidates.

Turning now to the gender and race attitudes of the sample, Table 3 displays the distribution of hostile sexism, modern sexism, and racial resentment across low, medium, and high categories. These distributions reveal the sizable variation for each of these attitudes. More than 10% of respondents fall into the highest third of the scale for each measure—specifically, 11% of the sample fall into the top third of the hostile sexism scale, 21% on the modern sexism scale, and 13% on the racial resentment measure. At the same time, a plurality of the sample is in the lowest third of the scale for each of the measures. For hostile sexism, 53% of the sample falls into the lowest third of the scale. For modern sexism, almost 43% of the sample falls into the lowest third, while almost 44% of the sample does so for racial resentment. Although each of the scales leans toward low levels, the distribution illustrates variation in the sample.

Table 3. Distribution of hostile sexism, modern sexism, and racial resentment

Table 4, then, pulls these considerations together, and compares mean levels of hostile sexism, modern sexism, and racial resentment for those who choose each candidate on each criterion. Across all these means, the most consistent statistically significant difference is between those choosing Biden and those choosing Warren. Those who choose Biden have a higher mean level of hostile sexism than those who choose Warren on three out of four criteria (i.e., policy, likability, and represent). These patterns persist through measures of modern sexism and racial resentment—on each of these criteria for both measures, those who choose Biden have a statistically significantly higher level of modern sexism and racial resentment than those who choose Warren. Moreover, Biden supporters have statistically significantly higher levels of modern sexism and racial resentment than those who choose Harris on the policy and likability criteria, and Biden supporters maintain that difference over Harris supporters in terms of racial resentment for the representation criterion. In short, the bivariate results suggest there are some differences in gender and race attitudes opening up between Biden supporters and those who support Warren and Harris. This serves as a piece of evidence that gender and race attitudes may manifest as a benefit to the white male candidate as compared to the women candidates. (Table A2 in the Online appendix provides additional bivariate comparisons on key variables.)

Table 4. Bivariate comparison of mean levels of hostile sexism (HS), modern sexism (MS), and racial resentment (RR) by candidate choice

As a next step in analysis, I construct 16 different models to predict the likelihood of picking each of the four candidates for each of the four evaluative criteria. For each model, the dependent variable is dichotomous, coded 1 if the respondent chose a particular candidate and 0 if they did not. Each of the models also includes the same independent variables. Most centrally, the models include measures of hostile sexism, modern sexism, and racial resentment. The models also include a measure of policy liberalism to account for the possibility that respondents will prefer a candidate, across evaluative criteria, who shares their views on policy (e.g., Aldrich and Alvarez, 1994). These models also include a measure of ideology, to account for the likelihood that more moderate voters will be more likely to choose Biden across the board, as he is the moderate option and primary voters rely on ideology heuristics (Lau and Redlawsk, 2006); and party strength, to account for the possibility that strong partisan identifiers may be more supportive of Biden as a de facto standard bearer of the party, having served as Vice President for the previous Democratic president. Finally, the models include dichotomous variables for Black and Latina/o/x respondents (with the omitted category being white) and women, as well as measures of age and educational level.¹²

Figure 1 displays average marginal effects of these logit models with 95% confidence intervals for the hostile sexism, modern sexism, and racial resentment independent variables. When the effects of the independent variable are statistically significant, the confidence interval does not cross the zero reference line. We see evidence for the intersectional nature of candidate identities—being a white candidate advantages the white man candidate (i.e., Biden) in all four models, but does not advantage the white woman candidate in any of the four models (i.e., racial resentment is not statistically significant in any of the Warren evaluation models). In fact, the positive relationships between racial resentment and the likelihood of choosing Biden is one of the most consistent findings. For example, in the Biden/policy model, the likelihood that respondents with the lowest level of racial resentment will choose Biden as the candidate whose policy stances most closely match theirs is 34%.¹³ For respondents with the highest level of racial resentment, the likelihood is 62%. On the contrary, the Black woman candidate, Harris, is consistently at a disadvantage when racial resentment attitudes increase. Across all four evaluative criteria, as racial resentment increases, respondents are less likely to select Harris. The likability criterion displays the widest range of predicted probabilities. Those with the lowest level of racial resentment have a 22% likelihood of choosing Harris as the most likable candidate, while those with the highest level of racial resentment have an 8% likelihood. In contrast, racial resentment is only consequential for the other Black candidate (Booker) in the model that predicts likability. As racial resentment increases, respondents become less likely to select Booker as the most likable. (Online appendix Table A3 presents the logit coefficients for the full models.)

Figure 1. Average marginal effect of hostile sexism, modern sexism, and racial resentment on choosing each candidate, by criterion.

These results underscore the importance of analyzing responses to candidates in a way that can account for candidate's multiple identities. Racial resentment does not uniformly advantage white candidates, nor does it uniformly disadvantage Black candidates. Instead, racial resentment is uniquely detrimental to the Black woman candidate and a unique benefit to the white man candidate. To further illustrate this, Table A4 in the Online appendix displays the logit coefficients for models that combine the candidates by race and, separately, by sex.¹⁴ If we relied on these aggregated models, we would conclude that racial resentment worked to the advantage of both white candidates (i.e., Biden and Warren) and both male candidates (i.e., Biden and Booker), although, when we disaggregate the models, we observe that these effects are driven by Biden and do not accrue to either Warren or Booker.

Similarly, measures of hostile and modern sexism are not simply a matter of predicting differences in selecting men and women candidates. Instead, we again see Biden, the white male candidate, benefiting as modern sexism levels rise, though less consistently than the benefit he received from racial resentment. In particular, as a respondent's modern sexism increases, they become more likely to choose Biden as the most electable candidate. At the lowest level of modern sexism, a respondent is 56% likely to choose Biden as the most electable candidate, while a respondent with the highest level of modern sexism is 69% likely to choose him. Biden's likelihood of selection across all four criteria does not vary as a function of hostile sexism. There is some evidence, however, that measures of hostile sexism are consequential for Warren. In three out of the four models, increasing levels of hostile sexism reduce the likelihood of choosing Warren, although in two of these models, p < .1 (i.e., policy stances and likability). In terms of representation, however, hostile sexism is statistically significant at the p < .05 level. The respondents with the highest level of hostile sexism choose Warren as the most representative candidate with a likelihood of 23%, while those with the lowest level of hostile sexism have a 42% chance of selecting her as the candidate that best represents the respondent. Finally, we see one other manifestation of race and gender attitudes that disadvantage Harris—in the model that predicts selecting Harris as the most electable candidate, her likelihood of being chosen decreases as the respondent's level of modern sexism increases. Specifically, those with the highest level of modern sexism have a 4% likelihood of choosing Harris as the most electable candidate, compared to 10% for those with the lowest level of modern sexism.

The results show that, on balance, Democratic primary/caucus voters use their race and gender attitudes in their evaluations of these candidates. Importantly, though, I hypothesized that the respondent's race will be consequential for whether these race and gender attitudes matter—specifically, that I expect it is white respondents who are driving these relationships. To test this hypothesis, I reran all 16 logit models from above with some minor adjustments. First, I replaced the IVs “Black” and “Latina/o/x” with the dichotomous variable “white.” This variable is coded 1 for those respondents who identify as white and 0 otherwise. I also added interaction terms between the key variables of interest—hostile sexism, modern sexism, and racial resentment—and “white.” These interactions will help assess whether white respondents display different patterns than respondents of color. Finally, I ran the models as ordinary least squares regression models, in order to make the interaction terms properly interpretable. The results of these models are presented in Online appendix Table A5.

In seven out of eight cases where racial resentment predicts the likelihood of choosing Biden or Harris, the average marginal effect for white respondents is statistically significant. In those same seven cases, the average marginal effect for respondents of color on the likelihood of choosing Biden or Harris is insignificant; in these cases, then, we are unable to reject the null hypothesis that there is no relationship.¹⁵

Similarly, in the models where hostile sexism reduces the likelihood of choosing Warren, the average marginal effect for white respondents is statistically significant, while the average marginal effect for respondents of color is not. Notably, although hostile sexism does not predict the choice of Biden on any evaluative criteria in the non-interactive models, the average marginal effect of hostile sexism is statistically significant for whites' evaluations of Biden on policy, likability, and representativeness. In the Biden/electability model, the average marginal effect for modern sexism among white respondents is statistically significant, but, again, the average marginal effect for respondents of color is not. And, of the 14 relationships where racial resentment, hostile sexism, or modern sexism predicts the choice of Biden, Harris, or Warren in the non-interactive models, in thirteen of these, the average marginal effect for respondents of color is not statistically significant.¹⁶ As to the second hypothesis, then, there is evidence that racial resentment, hostile sexism, and modern sexism affect white respondents' candidate evaluations, but little evidence of a similar effect for respondents of color.

The specific results related to the impact of ideology on candidate evaluation are included in Appendix Table A3. Increasingly liberal respondents were more likely to choose Warren across all four criteria, while the opposite relationship existed for three out of the four criteria for Biden (it did not hold for the likelihood of choosing Biden as the most electable). Ideology was not a statistically significant predictor for any criteria related to Booker or Harris. These results comport with Biden's and Warren's strategies to emphasize their ideologies.

Finally, are these evaluations consequential for the ultimate choice of candidate? In other words, do these raced and gendered candidate evaluations have any bearing on the ultimate primary or caucus vote choice? To assess this, I run four more logit models that predict the likelihood of voting for a specific candidate (as opposed to evaluating the candidates). The dependent variable is the likelihood of selecting a given candidate when asked, “If you participated in a caucus or primary today, which candidate would you support?” For each of the four models, the independent variables are indicator variables, coded 1 if the respondent chose that particular candidate for that particular evaluative criterion, and 0 otherwise. Across three models that predict respondents intending to vote for Biden, Harris, and Warren, choosing that specific candidate on each of the four evaluative criteria positively and statistically significantly predicts indicating that a respondent would vote for that candidate in the primary or caucus.¹⁷ These analyses provide evidence that the individual criteria serve as bases for the ultimate vote decision, providing the final link between three of the candidates' evaluations (i.e., Biden, Harris, and Warren and, to a lesser extent, Booker)—in which race and gender attitudes play a role—and the decision to vote for them. (The results of these four models are displayed in Appendix Table A6.)

4.1 Electability—is it a special case?

Of the four evaluative criteria analyzed above, the concern about electability was particularly relevant throughout the primary season. Journalistic articles frequently highlighted members of the Democratic base that were deeply motivated to beat the sitting president. For example, the sentiment expressed in this LA Times article was common: “‘I like Elizabeth Warren's policies, I just don't think she can get elected,’ Greg Reed, a 72-year-old retired high school principal, said…‘I believe Biden can win. That's what I'm interested in. Beating Trump’” (Halper, 2019). Data in the current study reinforce the prevalence of this attitude, with 56% of respondents indicating that the most important criterion in choosing a candidate is the ability to beat President Trump.

Although proponents of using “electability” as a basis for their vote choice tended to couch their arguments in pragmatic terms, there was discomfort with this criterion for picking candidates. Journalistic outlets published arguments that the term was being defined and used in gendered ways. As Kate Manne explained, “People always say they want substance, but when it's a woman bringing it, it seems unexciting. My worry is electability is a smokescreen for this sadly common thing, which is not wanting to support a female candidate” (Klein, 2019). Others noted the term specifically advantages centrist white men, adding a critique from the perspective of the candidate's race and ideology (Wang, 2019). If these critiques are accurate, we would expect to see a particular manifestation of race and gender attitudes in respondents' selection of the white male candidate over the other three candidates when evaluating the candidates in terms of electability.

At the same time, it is plausible that we might observe the opposite relationship. Although some conversations questioned whether electability masked undertones of racism and sexism that Democrats were not comfortable expressing directly, there were also explicit conversations about racism and sexism and how the general electorate's reliance on these attitudes in their vote choice might disadvantage the Democratic candidate in a crucial election. Perhaps because Senators Warren and Klobuchar were some of the last competitive challengers to drop out of the race, these conversations seemed to focus on the gender angle. NPR's Danielle Kurtzleben summed it up thus, “This is something that many journalists (myself included) heard over and over in interviews with voters—not sexism itself driving voters’ choices, but fears about other people's sexism” (2020). In short, the concern about the electability of women candidates—but certainly with a logic that also translated to a nominee of color—may have been coming from those with very low levels of racial resentment, hostile sexism, and modern sexism, but who believed that these attitudes are genuinely held in great enough numbers in the general electorate as to doom the chances of a candidate of color and/or a woman candidate.

There is initial evidence that electability may be a special case. In the logit models related to candidate evaluation, Biden—the only candidate to benefit from higher levels of modern sexism—only did so in the electability model. However, racial resentment does not appear to operate differently for either Biden or Harris in predicting electability as compared to the other criteria. To further explore this evaluative criterion, I ran one more set of models. Given that the critiques of the term “electability” focused on the unique benefit it may have for white male candidates, I focus the analysis here on the likelihood of choosing Biden. To differentiate among those who are solid Biden supporters, solid Biden non-supporters, and those who only support him for electability reasons, I run multinomial logit models with a dependent variable with three values. If respondents did not choose Biden for any of the four criteria, they are coded as a 0; respondents who chose Biden as the most electable, but did not choose him for any of the other three criteria are coded 1, and respondents who chose Biden for all four criteria are coded 2. The independent variables are the same as those included in the previous candidate evaluation logit models. Given the findings that it is primarily white respondents driving the results, I restrict this analysis to white respondents. If it is the case that “electability” is masking sexist and racially resentful attitudes, I would expect that, as hostile sexism, modern sexism, and/or racial resentment increase, respondents will be more likely to select Biden as the most electable candidate only. On the contrary, if those who are acutely aware of the sustained effects of racist and sexist attitudes might see Biden as a safer choice with a general electorate that will be unable to overcome their own prejudices, then I would expect the opposite relationship—that as hostile sexism, modern sexism, and racial resentment decrease, the likelihood of choosing Biden as most preferred candidate for electability only will increase.

The average marginal effects for the three key independent variables are displayed in Table 5 (and a full table of coefficients for all independent variables is included in Online appendix Table A7). The results reinforce the findings from the logit model predicting the choice of Biden as the most electable candidate above in that modern sexism appears to give Biden a particular edge when it comes to assessing him to be the most electable candidate. Table 5 shows that, as levels of modern sexism increase, the likelihood of never picking Biden as your top choice for any of the four evaluative criteria decreases. At the same time, the likelihood of choosing Biden as the most electable candidate only makes up for this difference. As levels of modern sexism increase, the likelihood of choosing Biden for electability only also increases. The effects of hostile sexism and racial resentment do not play the same role in choosing Biden as the candidate of choice for electability only, although these variables do affect the likelihood of always voting for Biden (hostile sexism and racial resentment both positively predict choosing Biden on all four criteria). In addition, increasing levels of modern sexism and racial resentment decrease the likelihood of not choosing Biden at all.

Table 5. Predictors of choosing Biden as the most electable candidate (truncated); multinomial logits

*p ≤ .05.

In sum, then, modern sexism plays a role for those who only choose Biden as the most electable candidate in a way that is distinct from the role that modern sexism plays in candidate choice for other criteria. This relationship supports the idea that discussion of “electability” may have been a covert means of expressing modern sexism.

5. Conclusion

The 2020 Democratic presidential primaries unfolded in an environment in which matters of racial and gender equality were topics of significant concern and discussion. As part of this larger environment, the survey results presented here provide evidence that this larger context influenced white Democratic primary voters' assessments of their field of candidates. Across four evaluative criteria, Biden, the white male candidate, was more likely to be chosen as a respondent's level of racial resentment increased. This relationship disadvantaged Harris, the Black woman candidate, who experienced the opposite effect—for three out of four evaluative criteria, as a respondent's level of racial resentment increased, the white respondent became less likely to indicate Harris was the most qualified candidate on that particular criterion. Hostile sexism also consistently emerged among white respondents as consequential for assessments of Warren, harming her likelihood of being selected on three of the four criteria. What is more, these criteria ultimately predicted the likelihood of voting for each of these three candidates, which provides the link from race and gender attitudes to candidate evaluation to vote choice.

The conclusion, then, seems to be that Democratic women of color face a disadvantage when a portion of their primary electorate evaluates them on key criteria. Of course, caution is necessary before these findings are generalized. The external validity advantages of conducting a survey on these matters, in real time with real candidates, is counterbalanced by a significant limitation—each and every one of these candidates bring with them baggage that is specific to the candidate. Biden, for example, may benefit from being associated with a popular former Democratic president—who is the first and only Black president. On the other hand, Biden faced criticism for his leadership role in a 1990s crime bill that is widely condemned for its racist consequences (e.g., Lopez, 2019a). Harris, on the other hand, could have an advantage in mobilizing voters along race and gender affinities (e.g., Tate, 2003), but also faced criticism for her background as a prosecutor, playing a key role in a criminal justice system that systematically targets people of color for incarceration (Lopez, 2019b).

Another factor that should be integrated into this line of inquiry is the role of ideology. Although the models analyzed here controlled for it, its correlation with race and gender attitudes (as shown in Online appendix Table A2) suggests that it warrants additional investigation, particularly as it relates to the role it may play in primary decisions, where it is a key piece of information for voters (Lau and Redlawsk, 2006). For example, ideological differences came into sharp focus in the closing months of the Democratic primary race in both 2016 and 2020. In 2020, given that the race came down to two candidates (Biden and Sanders) who were similar in many “personal heuristics” (i.e., Lau and Redlawsk, 2006), it is reasonable to expect that ideology emerged as a significant predictor of candidate choice—and that, once the supporters of other candidates who dropped out re-sorted themselves into Biden and Sanders camps, the influence of racial resentment and modern sexism may have evaporated.

That logic may have been muddied somewhat in considering the 2016 race. Although Sanders also positioned himself as the liberal candidate in the primary contest against Hillary Clinton, it is plausible that evaluations formed on the basis of a gender difference persisted until the primary's end. Again, given the correlation between ideology and gender attitudes, in this circumstance, these attitudes may have worked at cross purposes. This point may have been particularly muddy on the question of electability between Clinton and Sanders, where Clinton's gender disadvantage (assuming that translates from the current findings) may have been offset by the perceived advantage of being the more moderate candidate. Of course, the run-up to the 2016 election seems like another era. It is equally likely that the unexpected (to most) victory of Trump fundamentally scrambled Democrats' primary decision-making calculus, particularly as they approached a second contest with him.

Perhaps even more important for future research is to assess how the process of candidate evaluations affects the next generation of potential candidates. Might the winnowing of a historically diverse candidate field to a white male with a lifetime of experience send a message to future generations of candidates? Might discussions of candidates' strengths and weaknesses, electability and likability, colored by underlying attitudes of racial resentment, hostile sexism, and modern sexism telegraph a message about who is “fit to rule” (Mansbridge, 1999)? Further investigation into how potential candidates hear and process these messages is an important direction to pursue.

With these future directions in mind, the current study adds to our understanding of how attitudes of race and gender insinuate themselves into the electoral process. Although their role in the 2016 election is well-documented, these findings provide additional evidence that these attitudes exert their force early in the electoral calendar, shaping the prospects of those who bear the Democratic party endorsement. These attitudes will continue to be salient as the Democratic Party continues to grapple with embracing the diversity of its identifiers.