1 Introduction
Before the 2020 election, only a handful of women had run in a major party presidential primary in the United States, most of them within the past two decades, and only five women had made it to a major party primary debate stage
1 (Zhou
2019). This history made the first couple of Democratic debates in the summer of 2019 striking for their gender diversity, as six female candidates qualified: Senator Elizabeth Warren, Senator Amy Klobuchar, then-Senator Kamala Harris, Senator Kirsten Gillibrand, Representative Tulsi Gabbard, and author Marianne Williamson. These initial debates were the first time in U.S. history that more than one female candidate was onstage (Zhou
2019). The 2020 Democratic primary also featured the first openly gay major presidential candidate, Mayor Pete Buttigieg, and multiple candidates of color.
Despite this recent rise in the number of presidential candidates from politically under-represented groups, there has yet to be a female or openly gay President of the United States. However, gender representation in U.S. politics has continued to improve, slowly approaching 25% of Congress (Women in the U.S. Congress
2020). Studies show that once women do decide to run, they are just as likely to win as men (Fulton
2013). However, this parity does not address potential differences in candidate quality. When female and male candidates have equal qualifications, a gender penalty of approximately 3% has been observed in prior studies (Fulton
2013). These results indicate that the observed gender parity in winning elections is due to overall higher candidate quality among the female candidates overcoming an otherwise systemic gender penalty. Additionally, the observed gender parity in winning elections has not yet been seen at the presidential level, where gender stereotypes may play a larger role in the mind of voters (Schneider and Bos
2019).
There are likely multiple contributing factors to why female candidates may be penalized at the ballot box. One major reason may be perceived gender roles and implicit bias (Schneider and Bos
2019; Conroy et al.
2020). Another may be media coverage (Oates et al.
2019). Previous studies have shown that female candidates get less traditional media coverage than their male counterparts, and new evidence is emerging that social media treatment of female candidates may be similar to traditional media coverage (Oates et al.
2019).
A recent study of the 2020 Democratic candidates analyzed the Twitter conversations surrounding the launch of their presidential campaigns. The study found that the female candidates’ (Warren, Klobuchar, and Harris) top narratives were mostly negative and about their character or identity, while those for the male candidates (Sanders, Buttigieg, and Biden) were all about their electability or lack thereof. The female candidates also received less mainstream coverage and were more likely to be attacked by right-wing users and fake accounts (Oates et al.
2019; Haynes
2019; Bowden
2019). Fake accounts, including bots, have been used widely in the spread of election misinformation on social media (Ghanem et al.
2019).
In this paper, we investigate the role social media plays in female presidential candidates’ campaigns. This work aims to further explore the social media treatment of the Democratic presidential candidates to determine whether there are any impacts of gender and sexuality on Twitter conversations throughout the presidential campaign. If there are differences, we plan to investigate if this differential treatment is coming from regular people, bots, or both, as that may inform how campaigns address their social media presences in future.
From December 2019 through April 2020, we collected Twitter data on the conversations surrounding the top five Democratic presidential candidates: Joe Biden, Bernie Sanders, Elizabeth Warren, Pete Buttigieg, and Amy Klobuchar. The conversations were found by collecting tweets, retweets and replies that used election-related hashtags or a candidate’s handle. We used NetMapper software to get linguistic cues associated with all the tweets in the data set, such as the number of abusive words in each tweet (Carley and Malloy
2020). We use this data set to address the following research questions:
1.
How did the volume of Twitter conversations surrounding the presidential candidates change over time? How do the candidates compare with each other?
2.
Was there differential treatment of the Democratic primary candidates on Twitter in terms of general abusive language and gendered abusive language?
3.
If there are differences between the candidates in the above RQs, were they due to bots or regular users?
We build on prior research in the social cybersecurity field by using network analysis to characterize behavioral and societal changes in a cyber-mediated information environment such as Twitter. We conduct network analysis in ORA (Carley
2017) and statistical analysis in R to help answer these research questions. Analyzing how different presidential candidates are discussed on social media can help us understand why gender parity has still not been reached in politics. This research may also help female candidates in future better prepare to counter false narratives and bot accounts.
This work draws on previous research on gender in politics as well as studies that have analyzed the spread of misinformation and hate speech on social media.
2.1 Gender and sexuality in politics
The U.S. has seen a growing number of women in politics since the 1990s, with women occupying approximately 25% of the seats in the 2021–2022 U.S. Congress (Women in the U.S. Congress
2020). However, the U.S. has still yet to see a female president. Voter perceptions of female candidates likely contribute to this issue. A 2010 study by Okimoto and Brescoll found that the perceived ambition of a political candidate leads to negative perceptions of female candidates but has no effect on perceptions of male candidates or the likelihood of voting for a male candidate (Okimoto and Brescoll
2010). This difference in perception is most likely due to the perceived lack of stereotypically female personality traits like warmth and compassion (Schneider and Bos
2019; Okimoto and Brescoll
2010). Additionally, previous research by Valentino et al. found through survey analysis that sexism was underestimated as a factor that contributed to Clinton’s loss in 2016. Even when controlling for partisanship, authoritarian preferences, and ethnocentric beliefs among whites, hostile sexism was highly correlated with support for Trump. Only party identification was more strongly related to his support (Valentino et al.
2018).
While many news articles during the 2020 primary focused on potential sexism regarding Senators Warren and Klobuchar (Schneider and Thompson
2020), additional reporting has shown that the United States may not be ready for a gay president either (Cummings
2019). Polls show that roughly 94 and 76% of Americans would support a female candidate and a gay candidate for president, respectively (The Economist
2020; Mercier et al.
2022). While 94% seems high, surveys measure explicit prejudice and the potential presence of social desirability bias in this survey may mean these self-reported numbers could be slightly inflated. Recent presidential elections in the United States have been incredibly close, with only one race since 2000 having a popular vote margin of over 5% (2008). Even a few percentage points or fractions of a percentage point can make all the difference.
On the other hand, both Democrats and Republicans but especially Democrats tend to underestimate the electability of individuals from politically under-represented groups. For example, Democrats in a 2020 survey estimated that only 61% of Americans were ready to vote for a female candidate, while 94% of Gallup survey respondents said they were ready (Mercier et al.
2022). This may lead Democrats to excessively fear the potential unelectability of female candidates, especially after Clinton’s loss in 2016.
However, a recent study showed that the 2020 Democratic female presidential candidates received more negative interactions from both less-credible and more right-leaning accounts when compared to their male counterparts (Oates et al.
2019). Given that previous work has shown the importance of social media as a source of election news for American voters (Allcott and Gentzkow
2017), this could impact elections and influence voter choices. This previous research on social media engagement and news coverage of various candidates motivates our first research question:
RQ1 During the 2020 Democratic presidential primary, how did the volume of Twitter conversations surrounding the presidential candidates change over time? How do the candidates compare with each other?
The study analyzing the 2020 Democratic female candidates used data from the first half of 2019 surrounding the candidates’ campaign launches and the first debates (Oates et al.
2019). Our work analyzes similar research questions as this previous study and builds on it by analyzing data collected later in the election cycle when the primary was ongoing.
2.2 Gender and sexuality on social media
In addition to the research showing that female candidates may get less media coverage, previous studies have shown that sexist language is prevalent on Twitter, furthering the differential treatment of the female candidates by the general public and potentially reinforcing gender stereotypes (Jha and Mamidi
2017; Hardaker and McGlashan
2016; Felmlee et al.
2019). A study analyzing the Twitter and Facebook conversations surrounding the 2020 U.S. Congressional elections found that female candidates, especially those from a minority background, were substantially more likely to face online abuse, and that abuse was more likely to be related to their gender when compared with male candidates (Guerin and Maharasingam-Shah
2020). These attacks often focused on supposed incompetence and the candidate’s physical appearance, while male candidates were more likely to be attacked on their political ideas (Guerin and Maharasingam-Shah
2020).
Another study investigating sexist slurs collected Twitter data on the four most commonly used terms: “bitch,” “cunt,” “slut,” and “whore,” with the “bitch” data stream accounting for 87% of their data. All four of these words show up in the top 20 curse words used on Twitter, with “bitch” at 4th, “whore” at 7th, and the most used male-based slur, “dick” at 8th (Wang et al.
2014). The authors found that these sexist slurs are often used to reinforce gender stereotypes about traditional feminine norms by insulting a woman’s appearance, age, competence, and sexual experience (Felmlee et al.
2019).
Previous research has also shown that social media users use female gender-based slurs more often than male gender-based slurs, and in general, they use them more often against women (Gauthier
2021). A previous U.K. study conducted using English language Twitter data from 2015 found that while men swore significantly more often than women in their data set, they used similar language (Gauthier
2021). Men and women used “bitch” and “cunt” as their two most frequently used gender-based swear words, with these words most often being used to describe a woman (Gauthier
2021). While there are some swear words predominantly used to insult men (“bastard,” “prick,” and “dick”), this study showed that both men and women use those insults less frequently. Combined, those three male slurs were used less frequently than both “cunt” and “bitch” by both women and men (Gauthier
2021).
Not all sexist language on social media comes in the form of gender-based slurs. Sexist language, mostly targeted at women, can take both a benevolent and hostile form (Jha and Mamidi
2017). Benevolent sexism typically uses seemingly positive language or back-handed compliments. Common phrases include “as good as a man” or “smart for a girl,” as well as referring to successful women as “the wife of [successful man].” Hostile sexism typically comes from three sources: paternalism (“women should stay at home”), gender differentiation (“women are unqualified”), and aggressive heterosexuality, including (“I’d like to fuck that slut”) (Jha and Mamidi
2017). Sexist language is not only used to describe women. A Twitter study on a female-named storm in the U.K. in 2018 found that the storm was personified in one of three ways: promiscuity (“slut,” “slag”), an animal (“bitch,” “cow”), and genitalia (“cunt,” “twat”) (Ablett
2018).
These previous studies give some background information that may lead us to suspect differential treatment of Buttigieg, Warren, and Klobuchar versus Biden and Sanders, who are both more traditional presidential candidates demographically. This prior work motivates our second research question:
RQ2 Was there differential treatment of the Democratic primary candidates in terms of general abusive language and gendered abusive language?
2.3 The spread of false news
In the aftermath of the highly polarizing 2016 U.S. presidential election and the 2016 Brexit vote, many researchers have focused on the potential impact of Russian bots and trolls in shaping the election narrative and how to detect these actors (Ghanem et al.
2019; Beskow and Carley
2018). The new interdisciplinary field of
social cybersecurity has emerged in response to these online threats. Social cybersecurity concentrates on characterizing and analyzing the impact of cyber-assisted maneuvers on both human behavior, and societal and political outcomes. Adversaries use information maneuvers to spread specific content, including falsehoods, conspiracy theories, and polarizing content. They also often employ network maneuvers, which include creating or breaking up groups. Misinformation campaigns use these maneuvers, often boosted by bot accounts to reach more people, to effectively spread their messages (Carley
2020; A Decadal Survey of the Social and Behavioral Sciences
2019)
Researchers in this field continue to analyze the impact of mis-/dis-information campaigns that target democratic elections (Grinberg et al.
2019). Automated accounts, or bots, during the 2016 election were shown to have had a disproportionate part in the spreading of false stories (Shao et al.
2018). Previous research suggests that these false stories may not change vote choices, but they may increase polarization or suppress some demographics from political participation (Allcott and Gentzkow
2017).
In general, false news has been shown to spread much more rapidly than true stories, perhaps due to novelty or emotional reactions incited in the recipients (Knight Foundation
2018; Vosoughi et al.
2018). During the 2016 U.S. election, Russian information campaigns were observed spreading extremist content across the political spectrum to escalate polarization and cause democratic instability (Matishak and Desiderio
2020). This polarization is frequently used around social issues such as pro/anti-women’s rights and pro/anti-LGBTQ+ (Carley
2020). Using social media to systematically impact voters’ attitudes and behaviors, escalate polarization, and spread disinformation about candidates draws on research concerning social cybersecurity. Our final research question is motivated by the importance of bots in the spread of misinformation and their possible impact on politics:
RQ3 If there are differences observed between the candidates from the previous research questions, are they due to bots or regular users?
It is not known if the polarization process, which may involve the use of bots, has any impact on the portrayal or perception of female candidates. This paper begins to shed light on this.
5 Discussion
The diversity of the 2020 Democratic presidential primary allowed for a direct comparison of the social media treatment of female candidates versus male candidates vying for the highest political office in the United States. This case study provided ample data to investigate abusive and gendered language, bot levels, and media coverage of female presidential candidates.
First, we found that the candidates most popular with the voting public, President Biden and Senator Sanders, were also the most talked-about on Twitter. We also found that President Biden had the most number of tweets in his data set from verified news agencies by far. Both Biden and Sanders dominated the narrative even before the election as well, possibly because they had higher levels of name recognition. Senator Sanders even surpassed President Biden in number of tweets in the middle section of our data (mid-January to mid-March), when the primaries were most competitive. This observation may be because Sanders was highly competitive and won two of the first four primary elections (Nevada, New Hampshire) and almost won a third (Iowa). Or this result could mean that Sanders was more controversial or interesting to talk about. In the last time period of our data set (mid-March through April), Biden’s data size skyrocketed, while the other four candidates declined. This result would be expected given Biden’s status as the presumptive nominee by the end of March.
Our second research question focused on analyzing if there was differential treatment of the candidates based on abusive language or gender slurs. We found that the most popular candidates received more abusive tweets and had a higher percent of abusive tweets in their network. This result is in line with our previous results, showing higher engagement overall with President Biden and Senator Sanders. One possible explanation for why the most popular candidates were attacked more often could be that they were more popular and therefore seen as more of a competitive threat to other primary candidates or President Trump.
However, we did find that the female gender slurs were a slightly higher fraction of the tweets in the female candidates’ networks. This result corroborates two previous studies on the social media treatment of female U.S. candidates (Guerin and Maharasingam-Shah
2020; Oates et al.
2019). The first study on the treatment of the Democratic presidential candidates’ campaign launches found that the female candidates were attacked more often on their character and identity than their male counterparts (Oates et al.
2019). The second study looked more generally at all candidates running for U.S. Congress in 2020, and they found that female candidates were more likely to be attacked in general, and more likely to be attacked on their gender (Guerin and Maharasingam-Shah
2020).
For all five candidates, the top female slur (“bitch”) was used more times than the top two male slurs (“dick” and “bastard”) combined. These results support previous research that shows female gender-based slurs are used more often than male gender-based slurs (Wang et al.
2014; Gauthier
2021). Also, using derogatory, gendered terms to describe males may not be unexpected in the Democratic party, as previous studies show that the general population views the Republican party as more “masculine” and the Democratic party more as “feminine” (Schneider and Bos
2019). This difference in perception may be due to the perceived policy focus of the two parties (Republicans as being strong leaders that are tough on crime, Democrats as being compassionate with more focus on welfare) (Schneider and Bos
2019).
For our final research question, we analyzed whether bots were driving the differences we saw between the candidates. Previous work on the 2020 Democratic presidential candidates found a higher level of fake accounts in the conversations surrounding the female candidates right after their campaign launches and the first debates (Oates et al.
2019). Our work, which looked at a later time frame than this previous study, did not find the same result. We found that bots were most prevalent in President Biden’s data set, though the percent of tweets coming from bots in the other four candidates’ data sets was not much lower (see Table
13). More interestingly, a higher fraction of tweets from normal users were abusive or used gendered slurs than from bot users (Figs.
4,
5). The higher fraction of bots in Biden’s network does not explain his higher level of abusive language in his data set. The abusive language was driven more by people than by bots in our study.
Overall, we found differential treatments of the various Democratic presidential candidates on social media. The more popular the candidate was offline, the more they were talked about (and attacked) online. The women received less news media interaction (though that may have been because they were less popular as candidates) and had a slightly higher fraction of their tweets using female-gendered slurs. Using sexist language, regardless of the targeted party’s gender, further perpetuates gender stereotypes in the political sphere and society at large.
6 Limitations and future work
6.1 Limitations
This work has multiple limitations. First, data from the Twitter API are not necessarily random (Morstatter et al.
2013). The data we collected may not be representative of all election-related conversations on Twitter, let alone the conversations on all social media platforms and of the wider American electorate. Also, the data were collected on top election-related hashtags and account handles, which tried to get as much of the election conversation as possible; however, some important hashtags could have been overlooked.
Though many of our results align with previous work on abusive language on Twitter, because of the data limitations, these findings may be different on other Twitter data sets or other social media platforms. The percentage of tweets calculated as gendered abusive language may be dependent on which words were included (we included seven female slurs). We tempered this limitation by creating our list of female-based slurs from previous work and following up our general gendered abusive analysis by comparing the most common female slur in our datasets with the two most common male slurs.
Another limitation of our work is the assumption that abusive words and gendered slur words are always negative or used when attacking the candidates. There may be some tweets with these abusive words that are not attacking the candidate mentioned, but perhaps attacking someone else or being used in a joking manner. There are some slurs, such as “bitch,” that are sometimes used by women to positively describe other women as a way to almost reclaim the term (Felmlee et al.
2019).
Finally, these results may show an association between gender or popularity with online abuse or lack of media coverage, but these results are not causal as the data set is purely observational.
6.2 Future work
Future research could analyze sexist phrases or sentences that do not necessarily contain abusive terms. This would help analyze more “benevolent” forms of sexism, including phrases like “smart for a girl” or “women should stay home.” This type of sexism may be more insidious and harder to find, but it may be having a large impact on the conversation. Future research could also analyze the network of Twitter users that are tweeting the abuse, not just the tweets themselves. This analysis could help show if there is a relationship between the accounts or if these accounts are coordinating with one another. This analysis could also help determine if these abusive users are targeting specific candidates or if they are targeting several candidates at once. Finally, survey analysis or experiments on this topic could add further evidence to this research area of potential differential candidate treatment by social media actors.
7 Conclusion
Our work contributes to the literature in two primary ways. First, we show the success of straightforward Twitter data collection and analysis to identify abusive language and ultimately protect minority candidates. This type of analysis could be used in future campaigns to analyze if gendered abusive language continues to be higher in female candidate data sets. Twitter bot detection is also an effective way to determine if there may be coordinated bot attacks against certain candidates, or if the attacks come primarily from trolls and regular people due to underlying sexist beliefs.
Second, our results have political implications with respect to the interplay of gender, politics, and social media. We see that gender continues to play a role in political campaigns, elections, and social media coverage. Our most politically impactful findings are:
1.
Popular candidates were targeted the most. This result is in contrast with previous results that showed female and minority candidates being attacked more often (Oates et al.
2019; Guerin and Maharasingam-Shah
2020). Those previous studies were on the 2020 congressional races and the early 2020 presidential primary. Perhaps in a presidential context, especially after a presumptive nominee had been chosen, attacks are tailored to be more impactful. The fact that there are more attacks on popular candidates suggests a certain sense of economy in those conducting influence campaigns; they are spending more effort where it may matter more.
2.
Normal accounts were more likely to use abusive or gendered slurs than bot accounts While bot accounts did contribute to the abusive rhetoric on Twitter, our results show that humans were behind much of the abusive environment. Even if the bot problem is addressed, regular users may continue engaging in this type of behavior.
3.
Female candidates tended to be targeted with gender slurs. Female candidates regardless of popularity had a higher fraction of their abusive speech consisting of gender-based abusive speech. This may suggest a strategy of belittlement or dismissal of female candidates and their policy ideas. This abusive social media treatment may just be a symptom of underlying gender stereotypes in society, further showing why there continues to be a 3% gender penalty at the ballot box (Fulton
2013). Or this treatment may be continuing to spread these sexist ideas and may be contributing to the continued lack of gender parity in U.S. politics. Even in the engagement with popular male candidates, the abusive tweets there used more female slurs than male slurs. These results suggest that were a women to be a popular candidate the engagement might be highly vitriolic.
Despite all the progress we have made toward increasing representation in U.S. politics, women may still be at a disadvantage when campaigning at the presidential level. While this work focuses on initial findings from the 2020 U.S. Democratic presidential campaign, it speaks to a larger problem that female candidates likely face in other elections, in the U.S. and abroad. It is important as a society to bring awareness to disparate treatment of certain types of candidates in politics, so that news agencies and regular voters alike can be more conscious in their discussions moving forward.
Acknowledgements
This work was supported in part by the Office of Naval Research (ONR) Award N00014182106, the Knight Foundation, the Center for Computational Analysis of Social and Organizational Systems (CASOS), and the Center for Informed Democracy and Social-cybersecurity (IDeaS). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the ONR or the U.S. government. The authors would also like to thank Dr. David Beskow for collecting the data used in this study.