Sexism is Probably One Reason Why Elizabeth Warren Didn't Do Better
By Brian Schaffner and Jon Green
At one point last year, Elizabeth Warren looked poised to become a frontrunner for the Democratic nomination. She was second in the national polling during much of the fall and was even leading polls in key states like Iowa and New Hampshire. But, by December, she had faded in the polls. She didn’t finish higher than third in any state through Super Tuesday, including her home state of Massachusetts. Today, she is dropping out of the race.
Why didn’t she catch on?
Primaries are idiosyncratic contests, where candidates with broadly similar policy platforms compete for voters with broadly similar policy preferences. Without party labels to rely on and with few ideological differences between the candidates, voters often base their decisions in primary elections on other things -- such as social identity or strategic considerations -- and differences in early results on the margins can have dramatic effects on whether the media treats a candidate as a contender later on. A variety of compounding factors likely contributed to the gap between Warren’s early polling strength and her low delegate count.
One such factor -- again, among many -- appears to be sexism. The idea that sexism has hampered the campaigns of the women running for the Democratic nomination is not new. Last summer, there was a clear relationship between Democratic voters’ sexism and their willingness to support women candidates like Warren, Kamala Harris, Amy Klobuchar, or Kirsten Gillibrand. Here, we show that this effect has persisted into 2020.
Data for Progress surveyed 2,953 likely Democratic primary voters in August, 2019 and then re-interviewed as many of them as possible (n = 1,619) at the end of January, 2020 -- just before the Iowa caucuses. In the first wave of the survey, respondents reported how much they agreed or disagreed with four statements that are meant to gauge one’s level of “hostile sexism”:
Most women interpret innocent remarks or acts as being sexist.
Women are too easily offended.
Most women fail to appreciate fully all that men do for them.
Women seek to gain power by getting control over men.
As you might imagine, many Democratic primary voters tend to strongly disagree with all of these statements, but this is not true for everyone. Roughly one-third of likely Democratic voters do not, on average, disagree with these statements.
If sexism mattered for Warren’s failure to gain traction in the early primaries and caucuses, then we should find that how people answered these questions last summer strongly predicted whether they planned to vote for Warren when we interviewed them in January. In purely descriptive terms, this certainly seems to be the case. As our first graph shows, Warren’s support was concentrated among those who reject sexist sentiments. To be clear, this does not show that voters who did not support Warren are necessarily sexist -- there are plenty of likely Democratic primary voters who strongly disagree with the items in the hostile sexism battery and supported other candidates. Instead, what this shows is that Warren received little-to-no support from the roughly one third of the Democratic primary electorate that does not reject these sentiments. The current front-runners, Joe Biden and Bernie Sanders, have support from voters with a variety of views on these items.
To test the importance of sexism as a predictor of candidate support, we estimate its variable importance in random forest classification trees for each of the top three candidates in the second wave of the survey: Elizabeth Warren, Joe Biden, and Bernie Sanders. A random forest is a machine learning algorithm that sequentially partitions the outcome of interest at different levels of random subsets of variables, repeating the process hundreds of times to identify generalizable patterns that are useful for prediction. Random forests can be used to estimate how important each independent variable is for successfully predicting the outcome of interest based on how well the model performs when they are or are not included in the subsets of variables used for prediction. As in, if the model is less accurate when a variable is left out, that variable is important for predicting the outcome.
The plot below shows the results of these tests for variable importance on whether respondents supported Warren, Biden, and Sanders in the January wave of the survey. We have highlighted the importance of responses to the hostile sexism battery in each plot. In order to keep the plot from becoming more unwieldy than it already is, we show the top ten most important features in each model.
As one might expect, first-wave support is the most important predictor in each model. If you want to know whether someone supported Warren, Biden, or Sanders in the second wave of the survey, the most important thing to know about them is whether they supported them in the first wave. The second-most-important thing to know is if they supported a different top-tier candidate in the first wave (as opposed to a candidate who dropped out). However, when it comes to generalizable traits, we found that how people responded to the hostile sexism battery in the first wave of the survey was the most important for predicting whether they supported Warren in the second wave, but these responses were less important for predicting support for Biden or Sanders relative to other variables. For Biden and Sanders, factors like age, race, and political identity were more important for predicting their support.
While these results suggest that sexism is important for predicting whether a given likely Democratic primary voter supported Elizabeth Warren immediately pre-Iowa, the above plot doesn’t tell us the direction or strength of such a relationship. For this, we turn to the graph below, which shows the level of support that Biden, Sanders, and Warren received in the January poll based on responses to the sexism items in the previous wave -- adjusting for other factors that we might expect to influence vote choice, such as gender, age, race, ideology, and racial attitudes.
The graph shows that Warren’s support was strongly associated with sexism. Among likely Democratic voters who strongly disagreed with all four sexist statements, Warren was the marginal favorite over Biden and Sanders. However, among those who did not disagree with the sexist statements, Warren was a clear third (or worse). Indeed, adjusting for other factors, a respondent with neutral attitudes on the sexism items is only half as likely to support Warren as someone who strongly disagrees with all four statements.
To be clear, the direct effects of sexism are likely not the only factor holding Warren’s candidacy back. As is typically the case in primary elections, Democratic primary voters are willing to discount their own ideological preferences in favor of more moderate candidates who are viewed as more electable in the general election. As Mark White found in separate analyses of this survey, Elizabeth Warren consistently performed better on “magic wand” support than on actual vote intention. That is, Democratic primary voters were more likely to say that they would make Elizabeth Warren president if it were entirely up to them (so they wouldn’t have to worry about the general election) than they were to say they’d vote for her if their primary or caucus were held the day they took the survey. The opposite is true for Joe Biden, who consistently benefits the most from respondents saying they’d vote for him even though he isn’t the candidate they’d make president if it were entirely up to them.
While this electability penalty on Warren is real, we find that it is likely independent of any penalty she received from voter sexism. The plot below shows marginal effects from the same regression as above, with the outcome being magic-wand support instead of reported vote intention. While the intercepts change (reflecting the differences in baseline support for the different candidates across the two measures), the slopes are essentially the same: sexism carries a strong negative association with the likelihood of selecting Warren as one’s magic wand preference, with weaker positive relationships for Biden and Sanders. That sexism and electability penalties are independent makes sense theoretically, in that sexism should be related to viewing Warren unfavorably while any electability penalty should come from voters who otherwise view her favorably.
This does not rule out the possibility that second-order sexism -- that is, acting on the belief that other voters will penalize a candidate due to her gender -- hurt Elizabeth Warren’s candidacy independent of the first-order relationships we see here. Indeed, in a previous analysis of the conjoint experiment embedded in the first wave of this survey, we found that Democratic primary voters would on average prefer to vote for a woman, all else equal, but simultaneously felt that a woman would be less likely to win against Trump in a general election. And in an analysis of word association items included in the second wave, the word “win” is associated with negative sentiments regarding Warren but not Biden or Sanders -- often taking the form of respondents indicating that they view her positively overall, but don’t think she can win. Taken together, this suggests that even voters who do not actively endorse sexist sentiments may still penalize Elizabeth Warren based on the sexist attitudes they think others hold.
Ironically, voters’ concerns about sexism-based penalties for a woman in the general election are probably misguided. In general elections, the effects of sexism on vote choice appear to be driven mostly by candidate rhetoric and party reputations rather than by candidate gender. In fact, in 2018, Republicans lost votes because of sexism, and the effect of sexism on vote choice did not significantly differ based on whether the candidates were men or women.
Primary elections are complex processes, and primary voter behavior is far more difficult to understand than that of general election voters -- for whom partisanship tends to dominate. However, none of this takes away from the basic finding that Democratic primary voters who are more accepting of sexist sentiments were less likely to support Elizabeth Warren’s campaign for president, and this has likely been an important reason why her campaign did not gain more traction in the time between the first debates and the first votes being cast.