Race and gender discrimination

Opportunities for advancement in the United States reflect persistent racial inequality, with discrimination based on race/ethnicity prevalent within a wide range of domains, including employment, housing, credit, and consumer markets [5]. Multiple audit studies, or correspondence experiments, document high rates of discrimination against minority applicants over a period of 50 years. For example, one study matched resumes for both black and white candidate pairs, which were then used to apply for 1,008 national jobs. The findings revealed that black candidates received significantly fewer employer responses than their white counterparts, and that even an elite university degree did not provide protection from discrimination [6]. According to a meta-analysis of 43 studies and 738 tests, furthermore, members of racial and ethnic minority groups encountered substantial discrimination in hiring in 18 countries; they faced approximately half the odds (49%) of obtaining a job interview as compared to majority group applicants [7].

Job discrimination applies to gender, as well. An audit study of 1,372 job offers found evidence of biases in favor of male over female applicants in Spain [8]. Another audit study reported that job applicants who were mothers, as compared to women who were not mothers, and as compared to men, tend to be discriminated against [9]. In addition, working women are much more likely than men to recount experiences of job discrimination [10], and, of direct relevance to this study, women are vastly underrepresented in leadership positions within United States politics and business [11].

The publicly televised debate between two candidates for the position of Vice-President represented a national job interview. In the first and only Vice-Presidential debate during the 2020 presidential election, the incumbent, Republican Vice-President, Mike Pence, and then-Senator Kamal Harris (Democrat-CA) appeared together on the stage in Salt Lake City at the University of Utah on Oct. 7. The moderator, Susan Page of USA Today, posed questions to each candidate with the intention of gaining information about their knowledge and qualifications for the position, with the entire interchange accessible on public television. We examine responses on Twitter to the Vice-Presidential Debate to ascertain candid reactions, in real time, on the part of a segment of everyday citizens who are active online. The accessibility of content allows us to investigate the prevalence of online harassment and the use of gender and racial stereotypes, as reflected in slurs, regarding the performances of the two candidates during this job interview.

Online harassment

Most Americans use social media platforms. According to the Pew Research Center [12], approximately 22% of adults in the U.S. use Twitter, which makes it a popular social media site today, similar in following to Snapchat and WhatsApp, but less used than Facebook and Instagram. On average, Twitter users are younger, with a median age among U.S. users of 50, more likely to have a college degree than the general public, and identify more with the Democrat, rather than Republican, political party [12].

Harassment, hate speech, cyberbullying, and other forms of problematic messages are common occurrences on Twitter and other forms of social media. Close to two-thirds of Americans report experiences with online harassment personally [13], and over 150,000 bullying messages occur on Twitter daily [14]. Racist, sexist and homophobic content appears in numerous tweets [15,16], approaching half a million tweets per day that include sexist slurs [17]. These forms of demeaning online interactions can reap psychological and emotional harm [18,19] and cause victims to stifle their online activity [13].

The question we entertain is to what extent this type of harassing language manifests itself in messages regarding the vice-presidential candidates. We focus on “paradigmatic slurs,” which are expressions of a derisive attitude towards a group of people based on factors such as race and gender [20]. These slurs invoke negative stereotypes, typically applied in ways meant to derogate and insult people [20].

Online harassment based on gender and race

In a nationally representative survey [21], approximately 41% of Americans report having personally experienced some form of online harassment, with 18% describing severe behaviors such as physical threats and sexual harassment. Blacks (25%) and Hispanics (10%) were more likely than Whites (3%) to report victimization because of their race/ethnicity. In a recent study [22], online harassment based on race increased dramatically for Black adults from 27% in 2018 to 42% in 2020, and in just one year, to 59% in 2021. Asian-Americans experienced the largest yearly growth in severe forms of hate and harassment, from 10% in 2020 to 17% in 2021.

Additionally, women are frequent victims of negative, online content, with an abusive or problematic message sent every 30 seconds [23]. Several studies examine the way digital forms of harassment target, and affect, women [15,24–26]. These troublesome messages often reinforce traditional feminine stereotypes, such as expectations of physical beauty, sexual “purity,” and a temperament that is pleasant, kind, and soft [27]. Online content also can aim to degrade women sexually and contain threats of sexual violence [26,28]. These negative messages align with the sociocultural ethos foundational to the social web, which derives from widespread cultural forms of misogyny [24]. Furthermore, many men view these digital spaces as primarily masculine, and for men [29]. As women join these venues, thus, there has been a misogynist backlash that attempts to “reclaim” these spaces for men alone, which contributes to the rise of online abuse and vitriol directed at women [24,30,31].

Women politicians are targeted frequently by violent online threats and misogyny, presumably aimed at dissuading them from public activity [32]. During the US presidential campaign of 2020 [33], for example, women Congressional candidates received many more abusive messages on Twitter than men, and women of ethnic minority backgrounds were targeted disproportionately. US and Canadian women politicians who were highly visible also were subjected to greater incivility in social media than men [34], although equivalent gender biases do not always surface in print venues [35,36]. A report on 300,000 posts regarding three female politicians found extensive gendered and sexualized abuse and disinformation across six social media platforms; then-Senator Kamala Harris, a mixed race, influential, and visible politician, was by far the most frequent target, representing 78% of such cases [28].

Multiple oppressions ensconced in gender and race systems can influence individuals simultaneously [37–39], and scholars emphasize the need to examine race, gender, and class as intersectional systems of societal oppression. The concept of “misogynoir” [40], for example, highlights the distinctive, interlocked blend of hostility experienced by Black women. Moreover, online abuse oriented towards women of color can be particularly pernicious. Problematic messages on Twitter disproportionately target Black women, and women of mixed-race backgrounds encounter abuse in multiple forms, including sexism, racism, and physical and sexual threats [23]. Thus, intersectional theories and research suggest that the joint experience of being a Black and Asian woman is apt to shape Harris’ online and offline experiences and societal reactions to her.

It is important to note that Harris’s racial identity influences both how Kamala perceives her own identity as well as how voters view her as a political candidate. Research on Black women and their political ambitions finds that Black women may question their ability to be successful in politics, but they often overcome these doubts as a result of support and encouragement from their peers and community [41]. For this group of women, political engagement tends to represent a form of resistance to societal marginalization and an approach to achieving greater equality [42]. Kamala Harris’s experience as a minority candidate is not only as a Black woman, however, but also as an Indian woman. Harris’s multiracial identity casts light on how multiracial identities are complicated and complex. Multiracial individuals, in particular, must manage how others perceive them [43,44]. Within political campaigns, multiracial candidates have the advantage of building partnerships with voters from multiple groups, but they are disadvantaged when appealing to those of the same race with strong racial identities [45]. For instance, some South Asian groups question Harris’s representation of them and her motivations for highlighting her Indian ancestry [46].

Theories of online aggression

According to theories of online aggression, two basic, social psychological processes contribute to the proliferation of harmful, abusive messages [47]. These processes are fundamental to everyday, social interaction, and include the development of social norms and the evolution of informal status hierarchies, drawing from formative work on group dynamics [48]. When people interact, they engage regularly in the enforcement of social norms, for example, where norms refer to group-level evaluations of behavior that are backed by social sanctions [49]. Individuals who deviate from standards set by social norms are likely to be subject to negative consequences. In the case of online aggression or harassment, people routinely reinforce traditional stereotypes based on race, gender, age, and other characteristics [17]. Perpetrators direct their derogatory attacks at individuals believed to depart from common, stereotypical, normative expectations for behavior and deportment.

Informal, status hierarchies also emerge in social interaction, and competition for status represents a second social process that governs bullying and harassment, and one that is common both online [47] and offline [50,51]. When people interact, they vie regularly for recognition, attention, and esteem from others, which can generate conflict, especially in situations where rank remains ambiguous [52]. Within social media, one version of such competition involves the motivation to get posts noticed by others and to gain followers. Since negative Twitter comments tend to receive more retweets than their positive counterparts [27], posting derogatory messages can speed the attainment of these attention-seeking goals.

Note, too, that communication on social media sites often remains anonymous, or confers the perception of pseudo-anonymity, with posts typically originating without accompanying, formal identifiers and labels, other than a Twitter handle. As such, electronic communication has been welcomed as a tool to provide “voice to the voiceless.” At the same time, the ability to remain anonymous fuels the likelihood of aggressive responses, and seeing others model such actions amplifies aggression [53]. Online communication also facilitates targeting someone of elevated social standing, such as a politician, with minimal, if any adverse consequences such as censorship or physical retaliation.

Twitter, politics, and online harassment

Forms of social media have revolutionized today’s political landscape [54], and widely facilitated the rise of populism [55]. An experiment with 61 million Facebook users found that political mobilization messages delivered during the 2010 presidential election directly shaped actual, voting behavior, as well as increased political self-expression and information-seeking [2]. Messages shared in the experiment affected not only the behavior of the individuals themselves, but the effects spread to their friends and to their friends of friends. Another study [3] documented that Twitter influenced the 2016 Presidential election, demonstrating the potential of this form of social media as a political platform. Twitter activity during the 2020, Vice Presidential debate is noteworthy not only because of its potential for abuse, thus, but also because such commentary can shape political outcomes.

Furthermore, although Twitter facilitates communication and information dissemination, it also can present a biased impression of public opinion. According to an investigation of public debate on Twitter [56], users who retweet far-right politics on Twitter are significantly more active, and with more visible platforms, than their more moderate or left-leaning counterparts. Reactions from relatively small minorities of individuals can appear more common than those of the majority, therefore, and give the inaccurate impression that such responses are widely acceptable. One reason we believe our study is important, therefore, is the potential for discriminatory and biased tweets to influence public opinion and political outcomes.

Twitter represents a unique source of data for this topic. Although it does not offer a representative sample from the broader population, Twitter serves as a public forum for information dissemination and debate, and one that influences news reporting [1]. Moreover, when topics or events are contentious, individuals may feel less pressure to self-censor than when confronted with traditional measures such as surveys [57,58]. Individuals’ interactions on Twitter are unprompted by any researcher, instead reflecting precisely what, how, and to whom the individuals communicate.

Despite gaps in representativeness, Twitter sentiment toward feminism also highly correlates with individual gender attitudes taken from the General Social Survey [59], a finding that is relevant to our study of gender stereotypes. In these ways, gathering data from Twitter presents unique advantages for the study of immediate reactions to the Vice-Presidential debate.

Data and search procedure

We used the Twitter API, via academic research developer accounts, to gather public tweets directly from Twitter and collected messages containing key terms and hashtags related to Harris and Pence, as shown in Table 1. We restricted our data collection to hashtags that were neutral in description (e.g., #vpdebate2020; #Harris; #Pence), in addition to one popular, comparable, negative hashtag (#kamalalies and #pencelies), to increase the standardization of our comparison of message content. Hashtags represent a common way on Twitter that individuals contribute to specific conversations; searching for a given hashtag within the platform allows the individual to see all recent tweets that have included the hashtag. Neutral Twitter hashtags frequently serve as indices, marking the individual’s tweet as joining a growing conversation [60]. In contrast, negative hashtags often do not serve only as indices. They also can represent an instance wherein individuals both join a conversation and express an opinion that aligns themselves with a particular “side”. Finally, observing the frequencies of the hashtags searched (see Table 1) and the descriptive statistics of the tweets in our dataset (see Table 3), we note that the hashtags we used to scrape tweets capture both highly visible, popular content as well as low-visibility, “ordinary” content [60]. Therefore, while employing hashtags to collect data does not yield the entire corpus of tweets that pertain to the debate, we can be reasonably sure that we have collected tweets that were intentionally entering into the “vice-presidential debate” conversation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Search keywords and frequencies in the dataset.

https://doi.org/10.1371/journal.pone.0280828.t001

We utilized the R package, “rtweet”, to collect our data from the Twitter API [61]. Our main dataset is comprised of nearly 250,000 interactions, including tweets, retweets, replies, and mentions. We collected data from October 2020 to November 2020 to capture tweets most relevant to the vice-presidential debate. This research was approved by the Penn State University, Office for Research Protection, Institutional Review Board (STUDY00004666). Informed consent was not required for our use of publicly available, Twitter data. We also removed personal identifiers from specific illustrations of tweets, and paraphrased the content, to provide greater anonymity.

The key terms we chose as gender and race slurs derive from the top curse words that appear on Twitter targeting men and women [62]. We include four of the most common feminine curse words, or slurs (e.g., “b*tch”), and four of the most frequent masculine curse words (e.g., “d*ck”) [62]. Note that both types of slurs, “feminine” or “masculine,” can be used to attack a woman or a man. The most common words used for race in our dataset included Black (3,846) followed by White (2,085), and the most frequent words used as racial slurs included “cracker” (36) for White race, and Hindu (99) for minority races. Anti-Hindu sentiment, or “Hinduphobia,” contributed to the common, negative use of the term, Hindu, in the data, which invoked negative stereotypes such as “Cows. crapping on the streets of Hindu USA.” The most frequent feminine slur was “b*tch” and that for masculine slurs was “b*stard” (see Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Race and gender terms, with frequencies in the dataset.

https://doi.org/10.1371/journal.pone.0280828.t002

We present word clouds of the top 100 words found in tweets pertaining first to the incumbent, Pence (Fig 1) and next to the contender, Harris, in Fig 2 below. Note that several of the top 100 words are common between the Pence-related and Harris-related tweets, including several names (harris, pence, trump) and political keywords (vote, democrats). There are a few top words related to gender and race, including white, black, Indian, woman, mansplained, and “Imspeaking.” Several keywords arose in light of Pence’s behavior during the debate, referred to as explicit male dominance and thin-skinned, “white fragility,” and Harris’s response [63]. In addition, we find that some of the most frequently used words in the Pence dataset concerned the coronavirus (e.g. “covid”) and the fly that landed on his head during the debate (“fly”). Other recurrent words within Harris messages focused on words related to voting (e.g., “vote,” “polling”).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Pence top 100 words.

https://doi.org/10.1371/journal.pone.0280828.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Harris top 100 words.

https://doi.org/10.1371/journal.pone.0280828.g002

Sentiment analysis

We use a supervised method for classifying the sentiment of tweets that is specifically developed to assess emotional content in online harassment on Twitter [17,64]. The classifier represents an ensemble of three popular, lexicon-based classifiers (“VADER” (https://github.com/cjhutto/vaderSentiment), “bing” [65], and “afinn”) [66]. In this approach, the word order of a tweet is ignored, and extremely common words (i.e., “stop words”) are removed. The final sentiment score is calculated as the sum of the scores of the individual words in each tweet and ranges on a scale between -4 (most negative) to 4 (most positive). When compared with sentiment scores for tweets derived from four human coders for a sample of 400 random tweets, the performance of the ensemble classifier represented an improvement over those based on VADER’s classifier alone, or over alternative combinations of the three, original, common classifiers (i.e., Vader, bing and afinn). For more details regarding the ensemble classifier see [67].

Multivariate approach

We utilize ordinary least square (OLS) regression analysis to examine trends in the sentiment of the Twitter dataset. The primary predictors in our analyses consist of gender and race references, shown in Table 2. These variables are binary indicators that take a value of “1” if the tweet contained a gender reference or a race reference in the data (see Table 2), or “0” if the tweet did not contain such a reference. We likewise have “Pence” and “Harris” variables, which indicate if the individual’s first or last name is in the tweet. The dependent variable in our analyses is the sentiment score of a given tweet.

We also include a set of control variables in our analyses: logged retweets, logged friends, logged followers, and logged favorites. Retweets are instances where an individual shares a tweet, typically authored by another user, to their own Twitter profile. Friends are those individuals for whom the user has chosen to receive updates; activity of “friends” is shown on the user’s “timeline.” In contrast, followers are individuals who choose to receive updates about the user. Therefore, friends are those ties emanating from the user, while followers are those ties directed towards the user. Finally, favorites appear as a heart icon on tweets and, broadly, indicate acceptance of the post. We use a natural log transformation of these four variables, since their distributions are skewed, which is typical of datasets containing viral tweets.

Robustness

In analyses not shown here, we compared results from this modeling approach to others, including ordered logit and proportional odds regression. Our findings were consistent with the OLS regression, where the same coefficients achieve significance, have similar magnitude, and comparable AIC scores. For these reasons, we proceed with the least complex, multivariate approach, the OLS regression. Findings were reported to three decimals to enable the recovery of more precise p-values, if warranted [68]. There was no evidence of multicollinearity among our independent variables, and no variable had a variance inflation factor (VIF) above 2.5. Finally, using a random sample of 6,510 accounts from our dataset, we found that less than 10% were identified as bot accounts, according to Botometer (version 4), a tool to detect social bots on Twitter [69]. Their inclusion as a control variable in a narrowed dataset did not substantively change our findings.

Descriptive information

We begin by comparing patterns in the descriptive statistics for tweets associated with Pence and Harris (see Table 3). Findings from t-tests indicate that the majority of our variables of interest in the Pence and Harris datasets are significantly different from each other (Welch’s Two Sample). A significantly greater proportion of tweets in the Harris dataset, compared to the Pence dataset, contain common racial words and slurs for minority individuals only, as well as more masculine slurs (minority race words: Harris (2.03%), Pence (1.37%)); (minority race slurs: Harris (.05%), Pence (.04%)); (masculine slurs: Harris (.32%), Pence (.11%)). This confirms hypothesis one, that tweets regarding Harris will contain a greater number of race and gender slurs than those regarding Pence. For feminine slurs, as well as white race words and white race slurs, there is no significant difference in the proportion of tweets containing these terms. Interestingly, while the Pence dataset contains slightly fewer masculine slurs compared to feminine slurs (.11% versus .16%), the Harris dataset contains just over twice as many masculine slurs as feminine slurs (.32% versus .16%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Descriptive statistics.

https://doi.org/10.1371/journal.pone.0280828.t003

Multivariate analyses

From our multivariate analysis results, shown in Table 4, we see that logged retweets are significantly, and negatively related to tweet sentiment in Model 1. Next, we find that the sentiment of tweets containing minority racial slurs, as compared to those without such slurs, is significantly more negative (Model 2), which confirms our second hypothesis. Tweets with white race words are significantly more negative than those that do not contain such words, while those with white race slurs are more positive. The control variable, logged friends, is significant and negative, logged followers is significant and positive in the model, whereas logged likes is nonsignificant. In Model 3, we see that gender slurs contribute significantly and negatively to the regression.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Full dataset regression of tweet sentiment.

https://doi.org/10.1371/journal.pone.0280828.t004

Next, according to findings in Model 4, the tweets including a mention of either Pence or Harris are significantly more negative in sentiment than those without these names, although the model fit is not as good as the previous ones that include race or gender variables. Moreover, while the sentiment score of tweets with a Harris-keyword are more negative than those without such a keyword, tweets with a Pence-keyword are even more negative, while controlling for other variables, which is contrary to our third hypothesis. Finally, including all key variables in Model 5 represents an improvement over earlier models, with the lowest AIC and BIC values, and the highest adjusted R². Our variables of interest remain significant in Model 5, with much of the effect on sentiment deriving from the presence of minority racial slurs, feminine slurs, and masculine slurs, which negatively relate to sentiment, and white racial slurs, which have a positive relationship to sentiment. This is consistent with patterns in prior models.

To elucidate the possible differences between the Pence and Harris corpora of tweets, we split this dataset, filtering tweets related to each candidate. This was accomplished by using regular expressions to search for instances of the candidates’ first or last names. In this way, tweets could be assigned to the Pence dataset if they contained “mike” or “pence,” to the Harris dataset if they contained “kamala” or “harris,” or both if they contained either the first or last names of both candidates. As seen in the best fitting model, Model 5 (Table 5), Pence-related tweets that contain minority racial words are significantly more positive in sentiment than those without such words, whereas those with white racial words are more negative. Tweets with masculine and feminine slurs remain significantly more negative. Likewise, for tweets regarding Harris (Model 5, Table 6), those including feminine and masculine slurs are significantly lower in sentiment. In addition, the inclusion of minority racial slurs and white racial words in Harris-related tweets results in significantly more negative sentiment, while those tweets with white racial slurs are significantly more positive. Therefore, we find that tweets in the Harris corpus drove the effects of white race slurs in the full dataset. In both split datasets, as in the full dataset, moreover, minority race slurs and white race words are linked significantly to lower sentiment scores. To test if the two hashtags “#pencelies” and “kamalalies” drove these results, in analyses not shown here we dropped these tweets from the dataset, finding that our results did not substantively change. Therefore, we include them in the analyses and models shown and discussed here.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Regressions of sentiment for tweets regarding Pence.

https://doi.org/10.1371/journal.pone.0280828.t005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Regressions of sentiment for tweets regarding Harris.

https://doi.org/10.1371/journal.pone.0280828.t006

Finally, note that the negative coefficient for number of retweets (logged) is larger for Harris (-.057) than that for Pence (-.005) in Model 5 for each of the separate analyses (Tables 5 and 6), and this difference is significant (p < .001). Not only does the average tweet in the Harris dataset have close to four times as many retweets (2,124) as the average tweet in the Pence dataset (546; p < .001, Table 3), but the retweeted messages in her dataset tend to be more negative in content. These results support our final hypothesis, Hypothesis 4. In sum, the relatively high retweet rate for tweets mentioning Harris, and the elevated negativity of these retweets, suggests that derogatory messages about Harris diffused especially widely in the virtual, Twitter sphere.

Illustrations of popular tweets

We turn now to examples of tweets from the dataset, as well as demonstrations of particularly “viral” tweets: those with some of the highest retweets counts that spread far online. These can be positive in content, but frequently, are quite negative. These examples illustrate the distance these popular tweets can travel within Twitter, dispersing well beyond the original poster, often to tens of thousands of people or more. Illustrations of tweets are paraphrased, and identifying material removed, to preserve anonymity.

Pence.

The top two tweets within the Pence dataset concern two very different topics: the first one references the fly that landed on Pence’s hair during the debate and is comedic in tone. As paraphrased, it reads:

There is a fly on the top of Pence’s head, i.e., “flygate” [URL].

In contrast, the 2^nd is concerned with Pence’s relationships, and seeks to claim that his spousal relationship is not what it should be. This tweet (paraphrased) is:

This video describes VP Mike Pence and his bizarre and disturbing marriage with his wife, Second Lady "Karen" Pence [URL]

The network of these two tweets is shown in Fig 3. Interestingly, we can see that there is one individual who connects these two otherwise disparate hubs of activity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The two most retweeted tweets relating to Pence.

https://doi.org/10.1371/journal.pone.0280828.g003

Harris.

Several popular tweets involving Harris were opposed to her, and their networks of interactions are depicted in Fig 4. Two of these cases are as follows (paraphrased):

Wait until Election Day. This was an incredible con really. The riots were used to make Biden choose Harris for all incorrect reasons. Next the socialists will assume the presidency.

Of course, some U.S. citizens and many illegal ones will cast votes for these two China tools. Harris and “WhatsHisName”

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Several popular tweets related to Harris.

https://doi.org/10.1371/journal.pone.0280828.g004

These figures depict the extensive reach of certain instances of critical, online content.

Illustrations of tweets and gender and race stereotypes

There were many instances of racial and gendered themes in tweets pertaining to Pence and Harris, some of which contain the slurs shown in the descriptive details of Table 2. Next, we describe illustrations of specific tweets for both the Pence and Harris datasets and note gender and race themes when they occur. Tweets are again paraphrased and identifying material removed.