In recent years, many women in office have reported high levels of online abuse and called on social media companies to protect them (Brown Reference Brown2021; CCDH 2024; Frankel Reference Frankel2020; Inter-Parliamentary Union 2018; Phillips Reference Phillips2017). Despite these widespread self-reports, observational studies and survey-based research consistently yield contrasting results on whether women politicians experience a harsher online environment than men (Holm et al. Reference Holm, Bjarnegård and Zetterberg2024). Computational text analyses of social media data typically find that men politicians receive a higher volume of online abuse (Fuchs and Schäfer Reference Fuchs and Schäfer2021; Gorrell et al. Reference Gorrell, Bakir, Roberts, Greenwood and Bontcheva2020; Greenwood et al. Reference Greenwood, Bakir, Gorrell, Song, Roberts and Bontcheva2019; Theocharis et al. Reference Theocharis, Barberá, Fazekas and Popa2020; Ward and McLoughlin Reference Ward and McLoughlin2020). In contrast, politicians’ self-reported accounts frequently suggest an opposite pattern of women receiving more online hostility (Collignon and Rüdig Reference Collignon and Rüdig2020; Håkansson Reference Håkansson2021; Herrick and Thomas Reference Herrick and Thomas2022; Herrick et al. Reference Herrick, Thomas, Franklin, Godwin, Gnabasik and Schroedel2019; Pathé et al. Reference Pathé, Phillips, Perdacher and Heffernan2013; Zeiter et al. Reference Zeiter, Pepera and Middlehurst2019). Notably, computational text analyses detect larger volumes of abuse against women only when the analysis focuses on the highest-profile and most visible politicians (Guerin and Maharasingam-Shah Reference Guerin and Maharasingam-Shah2020; Oates et al. Reference Oates, Gurevich, Walker and Di Meco2019; Rheault et al. Reference Rheault, Rayment and Musulan2019).
I argue that this discrepancy stems from a mismatch in what each method captures. While self-reports often reflect the content and impact of abuse, large-scale text analyses generally measure only the volume of hostile messages. To understand why the content of hostility matters, we must first ask why trolls would want to disproportionately target women politicians. Researchers conceptualize online and offline violence against women in politics (VAWiP) as a form of backlash against women who are perceived to violate traditional gender roles (Eagly and Karau Reference Eagly and Karau2002; Felmlee et al. Reference Felmlee, Rodis and Zhang2020; Krook and Restrepo Sanín Reference Krook and Restrepo Sanín2016; Wilhelm and Joeckel Reference Wilhelm and Joeckel2019). Drawing on the gendered violence literature, I hypothesize that backlash to perceived gender-role violations – such as public visibility – will manifest in gendered content rather than simply higher volumes of abuse (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020; Erikson et al. Reference Erikson, Håkansson and Josefsson2023; Holm et al. Reference Holm, Bjarnegård and Zetterberg2024). I focus on visibility as an existing proxy for perceived role incongruity (Rheault et al. Reference Rheault, Rayment and Musulan2019; Håkansson Reference Håkansson2021) and introduce two novel indicators – legislator tone and the presence of women in the chamber – to further validate the relationship between perceived gender-role congruity and content.
I test the relationship between perceived gender-role violations and gendered content by examining tweets directed at US state legislators. I collected over three million public messages referencing all lower-house state representatives on the social media site Twitter (now known as X) between 2015 and 2018. To identify hostile and gendered content, I developed a replicable and cost-effective approach that combines hand coding, GPT-4 zero-shot classification, and fine-tuning a pre-trained BERT model. I calculated differences in the distribution of hostile and gendered mentions directed at men and women politicians using multivariate regression with state-fixed effects.
My results demonstrate that large-scale text analyses must account for gendered language in hostile content to capture the full scope of women politicians’ online environments. Had I only compared the frequency of hostility directed at men and women politicians, I would have wrongly concluded that social media vitriol does not uniquely target women. Accounting for gendered language reveals that women perceived as less gender-role congruent receive nearly twice as much hostile gendered abuse as men. Although men and women receive similar volumes of general abuse, the downstream consequences may be greater for women’s political engagement, as they remain globally underrepresented not only in elected office but also in political power, leadership, and agenda-setting roles (Bochel and Bochel Reference Bochel and Bochel2008; UN Women 2025; National Press Foundation 2024). Moreover, gender-based hostility is likely more harmful than general criticism, making it essential to examine the content of abuse in addition to frequency (Krook Reference Krook2020; Håkansson Reference Håkansson2024; Sobieraj Reference Sobieraj2020).
This paper makes two primary contributions. First, I help explain inconsistent findings between self-reported and observational approaches to measuring online hostility. Building on extensive survey evidence that documents gendered forms of abuse (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020), I argue that accurately capturing the online environments of women politicians through computational text analysis requires attention to gendered patterns in content as well as volume. My research indicates that when computational text analysis incorporates gendered language, it produces findings that closely align with women’s self-reported experiences (Bjarnegård et al. Reference Bjarnegård, Håkansson and Zetterberg2022; Erikson et al. Reference Erikson, Håkansson and Josefsson2023; Kosiara-Pedersen Reference Kosiara-Pedersen2023). Second, I deepen our understanding of violence against women in politics by illustrating the disproportionate abuse directed at women who are perceived as less congruent with traditional gender roles. Importantly, I show that backlash against women perceived as less gender-role congruent is not only more frequent but also more gendered in content. By incorporating gendered content and introducing new measures of role incongruity, this study provides observational evidence that supports women politicians’ reported experiences.
Why Does Visibility Matter?
Self-reported and observational studies of gendered online hostility show consistent agreement that high-profile women receive more abuse than similarly prominent men (Håkansson Reference Håkansson2021; Holm et al. Reference Holm, Bjarnegård and Zetterberg2024; Rheault et al. Reference Rheault, Rayment and Musulan2019). As visibility increases – whether in media appearances, office status, or social media followers – a gender gap in online abuse emerges. This recurring pattern suggests a rare point of alignment between methods, with visibility helping to explain when and why women politicians face disproportionate online hostility.
Role congruity theory offers a valuable framework for understanding why visible women politicians receive more online hostility than their less visible peers (Rheault et al. Reference Rheault, Rayment and Musulan2019). In societies that excluded women from public life, the traits linked to competent leadership – assertive, confident, dominant, ambitious – have been historically coded as masculine (Eagly et al. Reference Eagly, Makhijani and Klonsky1992; Eagly and Karau Reference Eagly and Karau2002; Heldman et al. Reference Heldman, Conroy and Ackerman2018; Katz Reference Katz2016). Women who exhibit the desirable ‘masculine’ leadership traits – including political ambition – may be perceived as threats to the standing social order and penalized (Brescoll et al. Reference Brescoll, Okimoto and Vial2018; Eagly and Karau Reference Eagly and Karau2002; Fulton Reference Fulton2012). This role incongruity helps explain why women politicians experience unique harassment. Women’s presence in public life, particularly in positions of power, can be perceived as violating gender roles (Eagly and Karau Reference Eagly and Karau2002).
There are at least two mechanisms by which visibility may intensify the perception that women in positions of power are violating gender roles. First, high-profile women may attract more abuse because increased visibility exposes them to a larger audience, heightening the number of observers who choose to act on gender biases (Rheault et al. Reference Rheault, Rayment and Musulan2019). In this view, all women politicians are gender role incongruent, but visibility increases the salience of this deviation and the likelihood that others will respond to it (Krook and Sanín Reference Krook and Restrepo Sanín2016; Mansbridge and Shames Reference Mansbridge and Shames2008).
Secondly, some observers may infer that women who attain public notability are substantially more role incongruent than their less visible peers. Achieving prominence in politics often requires demonstrating high levels of ambition, assertiveness, and confidence: traits historically associated with masculinity (Eagly and Carli Reference Eagly and Carli2003; Folke and Rickne Reference Folke and Rickne2016; Okimoto and Brescoll Reference Okimoto and Brescoll2010). Whether or not prominent women exhibit these traits, their visibility may lead observers to perceive them as more agentic and, therefore, as more role-violating than less visible women politicians (Håkansson Reference Håkansson2021). In this view, visibility does not merely amplify awareness of a role violation; it intensifies perceptions of that violation by signaling a deeper departure from stereotypical femininity.
These two mechanisms are observationally equivalent and likely mutually reinforcing. Recognizing that visibility amplifies the perceptions of gender role incongruity which trigger backlash helps explain why text-based approaches tend to detect greater hostility towards women only among the most visible politicians (Guerin and Maharasingam-Shah Reference Guerin and Maharasingam-Shah2020; Oates et al. Reference Oates, Gurevich, Walker and Di Meco2019; Rheault et al. Reference Rheault, Rayment and Musulan2019). Understanding the link between visibility and perceived gender role incongruity underscores the need for text-based research to engage more fully with online abuse as a form of gendered violence.
Gendered Backlash Takes Gendered Forms
Violence against women in politics differs from general political violence through gendered motives, forms, or impacts (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020). Bardall et al. (Reference Bardall, Bjarnegård and Piscopo2020) argue that a higher volume of online hostility directed at women may signal gendered motives. As discussed above, research comparing the frequency of abuse towards men and women politicians produces divergent findings depending on the methodological approach. Self-reported data often indicate greater volumes of abuse towards women, whereas observational studies frequently find the opposite, with the important exception that such studies consistently detect disproportionate hostility directed at the most visible women in politics (Holm et al. Reference Holm, Bjarnegård and Zetterberg2024).
In contrast, self-reported data provides strong evidence for gendered forms of abuse (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020). Women politicians self-report higher rates of sexualized harassment, gender-based hate speech, and sexist abuse than men (Bjarnegård Reference Bjarnegård2018; Erikson et al. Reference Erikson, Håkansson and Josefsson2023; Esposito and Breeze Reference Esposito and Breeze2022; Herrick et al. Reference Herrick, Thomas and Bartholomy2022; Holm et al. Reference Holm, Bjarnegård and Zetterberg2024; Kosiara-Pedersen Reference Kosiara-Pedersen2023). Computational text analyses that examine gendered content find the same pattern (Gorrell et al. Reference Gorrell, Bakir, Roberts, Greenwood and Bontcheva2020; Southern and Harmer Reference Southern and Harmer2019; Ward and McLoughlin Reference Ward and McLoughlin2020). Understanding visibility as a proxy for perceived gender-role congruity helps explain why observational researchers find support for gendered motives in the volume of hostility against high-profile women (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020; Rheault et al. Reference Rheault, Rayment and Musulan2019). The same logic suggests that gendered forms of abuse are also more likely to appear when women deviate farther from traditional gender expectations (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020; Erikson et al. Reference Erikson, Håkansson and Josefsson2023).
By incorporating gendered language, these forms of abuse go beyond opposition to an individual politician and instead challenge women’s right to participate in politics at all (Krook and Restrepo Reference Krook and Restrepo2019; Krook Reference Krook2020). Online violence against women in politics often operates as semiotic violence, using language and imagery to mark women’s political presence as transgressive (Krook and Restrepo Reference Krook and Restrepo2019; Krook Reference Krook2020).Footnote 1 Trolls force gender into the conversation through slurs, stereotypes, and sexualized imagery to reassert gender roles that frame politics as the domain of men (Jankowicz Reference Jankowicz2017, Reference Jankowicz2022; Sobieraj Reference Sobieraj2020).
The gendered content of this form of hostility is central to its intended function to punish women’s perceived gender-role violations and reduce their political participation (Chadha et al. Reference Chadha, Steiner, Vitak and Ashktorab2020; Frankel Reference Frankel2020; Holm et al. Reference Holm, Bjarnegård and Zetterberg2024; Kjøller and Pedersen Reference Kjøller and Pedersen2025; Krook Reference Krook2020; Krook and Restrepo Reference Krook and Restrepo2019; Pedersen et al. Reference Pedersen, Petersen and Thau2024; Ramachandran et al. Reference Ramachandran, Lee, Kornberg, Peeler-Allen, Edlin, Fishman, Park and Yuthok Short2024; Women’s Media Center 2017; Yan and Bernhard Reference Yan and Bernhard2024). The use of gender-based hostility against women politicians is well-documented in self-reported data (Bjarnegård Reference Bjarnegård2018; Erikson et al. Reference Erikson, Håkansson and Josefsson2023; Esposito and Breeze Reference Esposito and Breeze2022; Herrick et al. Reference Herrick, Thomas and Bartholomy2022; Holm et al. Reference Holm, Bjarnegård and Zetterberg2024; Kosiara-Pedersen Reference Kosiara-Pedersen2023). Large-scale text-based analyses should account for gendered language in hostile content to capture the full scope of women politicians’ online environments (Gorrell et al. Reference Gorrell, Bakir, Roberts, Greenwood and Bontcheva2020; Southern and Harmer Reference Southern and Harmer2019; Ward and McLoughlin Reference Ward and McLoughlin2020).
I extend existing research on online VAWiP by examining gendered language outside hostile contexts. Some messages use forms of gendered language in explicitly supportive messaging, such as the hashtag #Vote4Women. Nevertheless, by heightening the salience of a politician’s gender, even well-intentioned gendered language may inadvertently activate stereotypic expectations associated with gender roles (Bigler and Leaper Reference Bigler and Leaper2015). For example, calls to elect more women, as expressed in #Vote4Women, may subtly reinforce the perception of women candidates as interlopers in a man’s domain (Puwar Reference Puwar2004). In light of this, I analyze how non-hostile gendered language may still reinforce gender roles.
To address the discrepancy between self-reported and text-based findings, I apply the gender-role congruity framework to the analysis of online abuse towards women politicians. Drawing on the literature on gendered violence, I expect that backlash to perceived gender-role violations will manifest in gendered forms (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020; Erikson et al. Reference Erikson, Håkansson and Josefsson2023). I focus on visibility as an existing proxy for role incongruity (Håkansson Reference Håkansson2021; Rheault et al. Reference Rheault, Rayment and Musulan2019) and introduce two novel indicators – legislator tone and the presence of women in the chamber – to further validate the relationship between perceived gender-role congruity and content. I expect the following.
Hypothesis 1: Visible women receive more hostile and non-hostile gendered content than less visible women and similarly visible men.
Hypothesis 2: Women perceived as less gender-role congruent receive more hostile and non-hostile gendered content than women perceived as more congruent.
Hypothesis 3: The online environment between men and women will differ more in gendered hostility than in general hostility.
Measuring Hostile and Gendered Content
I examine hostile and gendered language towards men and women state representatives with differing levels of visibility. Although most research on online hostility highlights national politicians, I focus on state legislators for two reasons. First, given that most national politicians begin their careers at the local level, understanding the experiences of these state representatives is crucial for research on the pipeline to national politics (Holman Reference Holman2017; Manning Reference Manning2024; McCrain and O’Connell Reference McCrain and O’Connell2023). Survey research of state and local politicians indicates that gendered exposure to online or offline political violence may deter progress along this pipeline (Herrick et al. Reference Herrick, Thomas and Bartholomy2022; Herrick and Franklin Reference Herrick and Franklin2019; Herrick and Thomas Reference Herrick and Thomas2023, Reference Herrick and Thomas2022; Ramachandran et al. Reference Ramachandran, Lee, Kornberg, Peeler-Allen, Edlin, Fishman, Park and Yuthok Short2024).
Secondly, the abuse of state lower-level politicians warrants special attention as there are reasons to worry that it has more severe ramifications than it does for national counterparts. State legislatures have less staff, security, and funds available to screen abuse, which increases the likelihood of violence making direct contact with the targeted politicians (Wood Reference Wood2021). Online threats are also closer to home in the literal sense for local politicians. Their online violence is more likely to come from their constituents – whose physical proximity raises the risk of offline assault (Bjørgo Reference Bjørgo2022). This study is the first to use a computational text analysis approach on social media data to examine online hostility towards non-national politicians.
This study is the first to apply computational text analysis to social media data to examine online hostility directed at subnational politicians in the United States. The comparatively low visibility of state legislators relative to members of Congress makes them a valuable test case for the relationship between political visibility and gender-based abuse. The identification of systematic gendered patterns in this setting strengthens the claim that such abuse is widespread and structurally embedded. As women ascend to higher office, they typically become more visible and less gender-role congruent, suggesting that abuse may intensify with political power.
Using Butler et al.’s (Reference Butler, Kousser and Oklobdzija2023) list of state legislator handles on Twitter (now known as X), I collected all public mentions of state representatives available in August 2022 and originating between October 2015 and July 2018. This time frame covers the campaign period and the full legislative term. Butler et al. analyzed the tweets these state representatives posted during the time frame, allowing me to control for the legislator’s behavior on the social media platform.
I randomly sampled five to ten mentions directed towards each legislator to build a training corpus of over 41,000 tweets. Using the partition and validation method detailed by Park and Montgomery (Reference Park and Montgomery2025), I give each mention binary labels for containing hostile or gendered content. I classify the mentions with the ‘GPT-4’ model on the OpenAI API.
I utilized hand coding, zero-shot classification, and BERT feature representation to label the social media mentions corpus for hostile and gendered content. I first classify the mentions with the ‘GPT-4’ model on the OpenAI API.Footnote 2 A team of expert coders, given identical instructions to GPT-4, validate the zero-shot classification labels.Footnote 3 Next, I used the training corpus to fine-tune a pre-trained BERT model.Footnote 4 Validation of these labels indicated high accuracy and convergence.Footnote 5 A full description of this process is available in the Supplementary Materials.
Briefly, I consider a mention ‘hostile’ if it contains impolite/rude/derogatory comments, vulgarities, threats, hate speech, dismissive tones, sexual harassment, racism, ad hominem attacks, or calls for a legislator’s resignation. A mention is ‘gendered’ if it references a legislator’s physical appearance, relationship status, parental role, or competence based on gender; utilizes gendered pejoratives; comments on stereotyped gender traits; or directly mentions the legislator’s gender. Table 1 shows a randomly chosen tweet from each combination of labels. These confirm that GPT-4 accurately picks up on hostile and gendered content.
Zero-shot labels show high face validity
I test my hypotheses using three dependent variables: the percentage of received mentions that are hostile (% Hostile), the percentage of hostile mentions that are gendered (% Hostile Gendered), and the percentage of mentions that are gendered but not hostile (% Gendered, Not-Hostile). My independent variable of interest is the interaction between Woman and Visibility.
I follow the literature and measure a legislator’s Visibility by the number of mentions she receives (Cha et al. Reference Cha, Haddadi, Benevenuto and Gummadi2010; Gorrell et al. Reference Gorrell, Bakir, Roberts, Greenwood and Bontcheva2020; Rheault et al. Reference Rheault, Rayment and Musulan2019; Theocharis et al. Reference Theocharis, Barberá, Fazekas and Popa2020; Ward and McLoughlin Reference Ward and McLoughlin2020). The number of mentions a legislator receives in this time frame correlates strongly with her legislative leadership position, how often the news references her, and how many times she is searched for on Google (see Supplementary Materials, part C). I control for the legislator’s partisanship, ideology, and behavior on Twitter (now known as X). I model the relationship with linear regression and include state-fixed effects with robust errors.Footnote 6
Women Receive Disproportionate Gender-Based Content
Figure 1 illustrates the importance of considering gendered content when evaluating the online environments of men and women politicians. The leftmost plot depicts the positive relationship between Visibility and % Hostile. Had I followed much of the computational text analysis literature in examining solely the frequency of online abuse, I would conclude that trolls do not disproportionately target women in politics. The central plot in Figure 1 belies the error in this conclusion.
Visibility impacts gender-based content for women.
Note: the x-axis shows Visibility, measured as the logged number of mentions a legislator received (for example, a score of x equals e x mentions). From left to right, the y-axes represent the percentage of (1) mentions that were hostile, (2) hostile mentions that were gendered, and (3) mentions that were gendered but not hostile. I obtained the predicted values from Table A.1 in the Supplementary Materials.
Going from low to high Visibility increases a woman’s percentage of hostile mentions with gendered content by ∼22 percentage points and a man’s by ∼10. At the highest levels of Visibility, a full quarter of the hostility women receive relates to their gender, compared to 15 percent for men. Visible women also receive more non-hostile gendered content than less visible women and similarly visible men. The much smaller y-axis in the rightmost panel suggests that gendered content is more likely to appear in hostile than non-hostile messaging. These results support my hypothesis that visible women receive more hostile and non-hostile gendered content than less visible women and similarly visible men (Hypothesis 1).
These findings illuminate the unique challenges women in politics endure because of gender-role expectations. Gender-based hostility is a form of semiotic violence that challenges the right of women to participate in politics (Krook and Restrepo Reference Krook and Restrepo2019). As such, the damaging effects of gendered hostility extend far beyond the individual targets to undermine democracy itself (Frankel Reference Frankel2020; Holm et al. Reference Holm, Bjarnegård and Zetterberg2024; Sobieraj Reference Sobieraj2020).
Role Incongruity Correlates with Gendered Content
I theorized that if a backlash against visible women took the form of gendered content, this would further support the conceptualization of public visibility as violating gender-role expectations for women politicians (Rheault et al. Reference Rheault, Rayment and Musulan2019). The results support this theory by indicating that online hostility towards visible women state representatives takes the gendered form of disparate content rather than volume. To lend further support to the centrality of perceived gender-role congruity in online violence against women in politics, I demonstrate that backlash towards two additional indicators of perceived gender-role incongruity also manifests as gendered hostility rather than an increased frequency of generic hostility.
The first alternate measure of perceived gender-role congruity is the positive or negative tone of the legislator’s tweets during this time frame.Footnote 7 Many citizens still expect women in politics to express a more positive viewpoint than men (Bauer Reference Bauer2020). Consistent with this expectation, women legislators in the dataset tweeted more positively than their partisan counterparts (Butler et al. Reference Butler, Kousser and Oklobdzija2023), and women members of Congress posted more images of themselves as happy than neutral or upset (Boussalis et al. Reference Boussalis, Coan and Holman2022).
Women tweeting in a non-positive tone are thus using a communication style that is less congruent with traditional gender roles. As shown in Figure 2, the corresponding backlash is evident in the content rather than the volume of their hostile mentions. Legislators who express positive emotions tend to receive less hostility. However, women who tweet in a negative – or even neutral – tone receive double the percentage of hostile gendered content that negative men do. Again, gendered content is expressed more frequently in hostile than non-hostile terms.
Women's tone corresponds to the gendered content they receive.
Note: the x-axis shows Legislator Tone, ranging from –1 (entirely negative) to 1 (entirely positive), with 0 indicating neutrality. From left to right, the y-axes represent the percentage of (1) mentions that were hostile, (2) hostile mentions that were gendered, and (3) mentions that were gendered but not hostile. I obtained the predicted values from Table A.2 in the Supplementary Materials.
The second alternate measure of perceived gender-role congruity is the growth rate of women’s representation in the state legislature.Footnote 8 The gender composition of a chamber may shape whether observers perceive women politicians as interlopers or insiders (Puwar Reference Puwar2004). In legislatures with low proportions and limited growth of women officeholders, women may stand out as more visible deviations from the male norm. Conversely, a sudden increase in women’s representation may itself prompt backlash, as suggested by research on violence against women in politics (Herrick et al. Reference Herrick, Thomas and Bartholomy2022; Sanín Reference Sanín2022).
Once again, perceived gender-role incongruity is associated with greater backlash in the form of gendered content rather than overall volume. In states where the proportion of women is declining, women legislators receive approximately twice as much gender-based vitriol as men legislators. At the same time, the data reveal a more hopeful trend. As women’s representation increases, gender remains a salient attribute but becomes less frequently tied to hostility and more often linked to non-hostile messaging. This trend persists even in states where the proportion of women representatives remains below the national median. The model in Figure 3 indicates that as the proportion of women officeholders increases, hostility towards them becomes less gendered in content.
Increasing women's presence decreases gendered hostility.
Note: the x-axis for each graph shows the growth rate of women in the state legislature. From left to right, the y-axes represent the percentage of (1) mentions that were hostile, (2) hostile mentions that were gendered, and (3) mentions that were gendered but not hostile. I obtained the predicted values from Table A.3 in the Supplementary Materials.
These results demonstrate that backlash to women perceived as less gender-role congruent often takes the form of distinctly gendered content. Patterns across visibility and other measures of perceived congruity reinforce the framing of public visibility as gender-role incongruent for women (Rheault et al. Reference Rheault, Rayment and Musulan2019). Across the three measures of perceived gender-role congruity, general hostility does not vary systematically by legislator gender.Footnote 9 In contrast, gendered hostility is closely tied to perceived gender-role incongruity and disproportionately directed at women. As women’s perceived role congruity decreases, they face significantly steeper increases in gendered hostility than comparably situated men.
These findings support the expectation that women who are perceived as defying gendered roles are more likely to be targeted with gender-specific abuse (Hypothesis 2) and affirm that online environments diverge more starkly for men and women in the volume of gendered hostility than general hostility (Hypothesis 3). While gendered language does appear in non-hostile contexts, it is nearly three times more likely to occur alongside hostility, and non-hostile gendered language shows little systematic relationship to measures of perceived gender-role congruity. These patterns underscore the need for computational text analysis research to move beyond volume-based measures and account for both gender roles and gendered content to capture the full scope of online political violence against women.
Discussion
This article addresses a persistent puzzle in the literature: why are gendered patterns of online abuse against politicians consistently found in self-reported but not observational data? I argue that this discrepancy stems from a mismatch in what each method captures. While self-reports often reflect the content and impact of abuse, most computational research captures only volume (Erikson et al. Reference Erikson, Håkansson and Josefsson2023; Rheault et al. Reference Rheault, Rayment and Musulan2019; Theocharis et al. Reference Theocharis, Barberá, Fazekas and Popa2020). Drawing on gendered violence scholarship (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020), I developed a replicable and resource-efficient method to identify hostile and gendered content in over three million tweets directed at US state legislators. This approach allows for scalable, content-sensitive measurement of gendered hostility. I reconcile prior discrepancies between text-as-data and survey-based approaches and show that backlash against women often takes a distinctly gendered form.
This study makes two central contributions. First, it explains why some observational research has underestimated the prevalence of gendered hostility online. When computational text approaches to measuring online hostility fail to examine the gendered content of abuse, they risk missing the unique form of backlash directed at women who are perceived to violate gender roles (Bardall et al. Reference Bardall, Bjarnegård and Piscopo2020). Incorporating gendered language into observational analyses yields findings that align with women’s self-reported experiences (Bjarnegård et al. Reference Bjarnegård, Håkansson and Zetterberg2022; Erikson et al. Reference Erikson, Håkansson and Josefsson2023; Kosiara-Pedersen Reference Kosiara-Pedersen2023). I show that visible women receive nearly twice as much gender-specific abuse as men despite experiencing similar overall volumes of hostility.
Second, I deepen our understanding of violence against women in politics by illustrating the gendered form of violence directed at women who are perceived as less congruent with traditional gender roles. I extend the operationalization of perceived gender-role congruity by validating visibility as an existing proxy (Håkansson Reference Håkansson2021; Rheault et al. Reference Rheault, Rayment and Musulan2019) and introduce two novel indicators: legislator tone and the presence of women in the chamber. Importantly, I show that backlash against women who are perceived as role incongruent is not only more frequent but also more gendered in content.
Targeting women perceived as less gender-role congruent through gendered abuse undermines core democratic principles. Gender-based hostility goes beyond opposition to an individual politician and instead challenges women’s right to participate in politics (Krook Reference Krook2020; Krook and Restrepo Reference Krook and Restrepo2019). The public denigration of women with slurs, stereotypes, and sexualized images leads both the targeted politician and the women who view this harassment to withdraw from politics (Chadha et al. Reference Chadha, Steiner, Vitak and Ashktorab2020; Frankel Reference Frankel2020; Holm et al. Reference Holm, Bjarnegård and Zetterberg2024; Kjøller and Pedersen Reference Kjøller and Pedersen2025; Pedersen et al. Reference Pedersen, Petersen and Thau2024; Ramachandran et al. Reference Ramachandran, Lee, Kornberg, Peeler-Allen, Edlin, Fishman, Park and Yuthok Short2024; Sobieraj Reference Sobieraj2020; Women’s Media Center 2017; Yan and Bernhard Reference Yan and Bernhard2024). By deterring women’s political participation, gendered hostility threatens descriptive representation, which may, in turn, have negative implications for women’s substantive and symbolic representation (Carroll Reference Carroll2001; Celis Reference Celis2009; Childs and Krook Reference Childs and Krook2009; Jarman Reference Jarman2025; Young Reference Young2002). Gender-based abuse thus has wide-ranging implications for democratic accountability and legitimacy.
Finally, this study opens several avenues for future research. Scholars should delve deeper into the costs of gendered language across hostile and non-hostile contexts for women and men. Additionally, more work is needed on other forms of identity-based hostility, such as racism, homophobia, or religious intolerance. Although gendered content did not differ between ethnic and racial groups (see Supplementary Materials, part D), models trained to identify hostile racial content likely would. Future research should consider how being the first or only member of a minority group may generate visibility, even in the absence of institutional prominence or behavior that is gender role incongruent. While this study focuses on public tweets and US state legislators, the methodological text analysis tools introduced here are readily extendable to other contexts.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0007123426101458.
Data availability statement
Replication data for this article can be found in Harvard Dataverse at: https://doi.org/10.7910/DVN/FQWVOS.
Acknowledgments
I thank Diana O’Brien, Dan Butler, Taylor Carlson, and Lucia Motolina, as well as audiences at Washington University in St. Louis, for helpful comments.
Financial support
This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.
Competing interests
None.
Open access
