In 2017, the Women’s Marches and #MeToo movement dramatically increased public attention paid to the harassment and silencing of women. In politics, the problem is widespread: relative to men, women politicians face more interruption (Och Reference Och2020), sexual harassment (Folke et al. Reference Folke, Rickne, Tanaka and Tateishi2020; Håkansson Reference Håkansson2019), and psychological abuse (Herrick and Thomas Reference Herrick and Thomas2022; Thomas et al. Reference Thomas, Herrick, Franklin, Godwin, Gnabasik and Schroedel2019). Women activists, journalists, and human rights defenders likewise face more threats and violence (Krook Reference Krook2020; Sobieraj Reference Sobieraj2020). Yet the vast majority of research on violence against women in politics focuses on elite women—politicians, public figures, and federal prosecutors—those that Pitkin (Reference Pitkin1967) might call formal or substantive representatives of women.Footnote ¹

This focus on elite women means that we know little about the treatment of the average woman—the voter, the volunteer, and the protester—when she enters the public sphere in the United States today. Research that considers harassment as a barrier to the average woman’s political participation focuses heavily on countries outside of the United States (e.g., Alam Reference Alam2021; Krook Reference Krook2020; Prillaman Reference PrillamanForthcoming). Research on American women’s participation has examined barriers like income and family responsibilities (Bernhard, Shames, and Teele Reference Bernhard, Shames and Teele2021; Schlozman, Burns, and Verba Reference Schlozman, Burns and Verba1994), but this alone has been unable to explain why women are more likely than men to vote but less likely to participate in other political activities (Beauvais Reference Beauvais2020). One possibility is that activities like protesting and canvassing make individual women into “symbolic” representatives who “stand for” women as a group in the eyes of others (Pitkin Reference Pitkin1967, 92), increasing their risk of harassment.

Likewise, while many studies examine the consequences of incivility between elites for public opinion, here too we know little about gendered incivility between citizens (Karpowitz and Mendelberg Reference Karpowitz and Mendelberg2014). Nor do we know whether incivility is a problem for an immensely popular new form of political activism: text canvassing. Text messages have been a boon for campaigns, allowing volunteers to contact more people, more cheaply, than ever before: more than 80 million political text messages were sent every day in September and October 2020 (Bajak and Burke Reference Bajak and Burke2020). In short, we know little about the experience of political activism in the United States today, let alone how women are treated when they participate. We, therefore, ask a simple question: are women more likely than men to receive hostile messages when they participate in politics?

To examine volunteers’ experiences, we conducted two field experiments in 2018 in which we randomized the apparent gender of volunteers during a texting (“Short Message Service,” or SMS) campaign meant to encourage a liberal organization’s supporters to attend rallies and call their representatives. Volunteers used a software program to rapidly text supporters standardized messages. Each supporter received a text message from a volunteer randomly assigned to use either a common male name, a female name, an ambiguously gendered name, or no name. Combining name manipulation with otherwise identical messages allows us to identify perceived gender as one cause of men and women’s different experiences in politics (Bertrand and Mullainathan Reference Bertrand and Mullainathan2004).

In both studies, female-named volunteers receive more offensive, silencing, and withdrawal responses than male-named or ambiguously named volunteers. This suggests backlash against women participating in political activism. In particular, these messages have a “generic” quality that Sobieraj (Reference Sobieraj2020, Reference Bertrand and Mullainathan5) describes as a hallmark of structural abuse: abuse that treats individual women as representatives of all women. When someone texts “fuck off Jessica you’re a slut,” having no information about “Jessica” and despite “Jessica” using exactly the same text messages as “Michael” and “Taylor,” we have evidence that the hostility is toward women as a group rather than some specific characteristic of this individual woman—who may not even be a woman volunteer, thanks to our experiment.

These findings are concerning since many barriers to participation already exist, especially for historically disadvantaged groups, and they help make sense of women’s reluctance to participate in peer-oriented political activities like canvassing. Yet we also find hints that volunteers may be more effective when assigned to use female names. Our study thus raises new questions about women’s equality in political life in the United States.

Materials and Methods

We conducted two randomized control trials in 2018 evaluating whether female-named volunteers receive more harassing SMS responses than male-named volunteers.Footnote ² To do so, we partnered with a progressive political organization, NextGen America (NGA), to contact individuals who had previously interacted with the organization and agreed to be contacted again.Footnote ³ NGA regularly texts supporters to contact their elected officials, take part in protests, and volunteer. Because these individuals have consented to contact from NGA, we expect them to be more civil toward NGA volunteers. Our estimates of the prevalence of hostile behavior should, therefore, be lower than if we had contacted respondents without their consent.

For each study, NGA selected the largest sample sizes they believed their volunteers could plausibly contact before a given deadline (e.g., before a rally). NGA contacted 60,356 individual supporters in the first trial, and 75,231 in the second.Footnote ⁴ In Study 1, NGA texted supporters to encourage them to attend a local March for Our Lives rally; in Study 2, to call their representative to urge then-Environmental Protection Agency administrator Scott Pruitt to resign. Volunteers were overwhelmingly women in both studies: 77.14% in Study 1, and 85.71% in Study 2, for an overall rate of 80.95%.

Design and Procedures

Following a design common to audit studies, to vary perceived gender, we randomly assigned supporters to receive an otherwise identical text message from a volunteer using either a stereotypically female name (Jessica), male name (Michael), ambiguously gendered name (Taylor), or no name. Both studies were double-blinded to reduce possible demand effects (among volunteers) and biases common to studying sensitive behaviors, such as social desirability bias (among respondents). Respondents knew only that they were being contacted by NGA, which they had previously opted into.

NGA’s texting software ensured every initial SMS message to a respondent was standardized; Appendix B.3 of the SM provides the base message text. After the volunteer successfully texted a supporter, the software prompted them to send a new message to another, until the volunteer finished texting their allotment of supporters. The average volunteer sent approximately 1,300 messages.

Dependent Variables

We measure three main outcomes for replies to these messages: average offensiveness, silencing, and withdrawal.Footnote ⁵ We measure offensiveness using questions we designed.Footnote ⁶ In Study 1, volunteers reported when they felt a text they received was offensive. These were coded as 100 and all others as 0, allowing us to report averages in terms of percentage points. For Study 2, instead of the volunteer-reported measure, we recruited two independent coders to indicate how offensive a response was on a 5-point scale ranging from “non-offensive” (1) to “very offensive” (5).Footnote ⁷

To compare offensiveness across studies, we recoded Study 2’s existing 5-point offensiveness measure as a binary measure. To do so, we recoded a response as “offensive” (100) if either coder marked the response as anything other than “non-offensive,” and inoffensive (0) otherwise. We do not find a significant difference in the ratings for offensiveness between the two studies; we, therefore, depict the results for both studies below using the binary measure.Footnote ⁸

NGA requires volunteers to immediately opt out anyone that asks not to be contacted or harasses a volunteer. In both studies, we use this indicator to measure withdrawals—polite or inoffensive requests not to be contacted again—and silencing: whether a respondent intimidates or harasses the volunteer into ceasing contact. Any response that requested an opt-out but was not coded as offensive is coded as 100 for withdrawal, and 0 otherwise. For a response to be coded as silencing (100), it must both opt out the respondent and be coded as offensive. This means that silencing messages represent the overlap between offensive and opt-out messages. Table 1 displays the rates (conditional on responding) and coded examples of each type of behavior. Offensive and silencing responses are relatively infrequent; withdrawals are more common.

TABLE 1. Sample Responses and Response Rates by Category

Analysis

In the main paper, we provide figures based on ordinary least squares (OLS) regressions.Footnote ⁹ These regressions estimate the treatment effect of assigning someone at phone number i to be contacted by a volunteer randomly assigned a particular name condition on the rate of offensive, silencing, or withdrawal replies.Footnote ¹⁰ We employ the “ambiguous-gender” condition as the control condition because it strictly varies the gender cue relative to the clearly gendered male and female names.Footnote ¹¹ In analyses that pool both studies, we include a dummy variable for whether the respondent was in Study 1 (1) or Study 2 (0). In all tables and figures, we report 95% confidence intervals with heteroskedasticity-robust standard errors; all p-values reported are two-tailed.

Results

We conducted two field experiments to study the treatment of political volunteers. In each study, the only aspect of the message that varied was the gender of the name the volunteer used. In both studies, when volunteers are assigned a female name, they receive more offensive, silencing, and withdrawal responses than any other name condition. However, supporters were also more likely to respond and agree to asks made by female-named volunteers.

Offensiveness

In both studies, respondents were more likely to send offensive replies to volunteers assigned female names than those assigned male or ambiguous names. Figure 1 presents the OLS regression results overall (left-hand panel) and for each study individually (middle- and right-hand panels).Footnote ¹² Across all three panels, we see the same pattern: respondents send more offensive messages to volunteers using female names. Overall, volunteers assigned to use female names are 0.177 percentage points more likely than ambiguously named volunteers to receive offensive messages during a campaign (two-tailed $ p< $ 0.001). Male-named volunteers were 0.097 percentage points less likely to receive offensive messages than the ambiguously named ( $ p= $ 0.002). Unnamed individuals were also more likely (0.148 percentage points) to receive offensive messages ( $ p< $ 0.001). For every 1,000 messages a female-named volunteer sends, she receives on average 1.77 more offensive messages than an ambiguously named texter, and 2.74 more offensive messages than a male-named texter ( $ p< $ 0.001 for both)—a 202% increase over the male-named base rate.

FIGURE 1. Mean Offensiveness

Note: The figure shows the average treatment effect by treatment condition with 95% confidence intervals estimated using ordinary least squares. The comparison category is the ambiguous name condition.

Silencing

Across both studies, supporters are more likely to silence female-named volunteers than all other named conditions, shown in Figure 2.Footnote ¹³ Overall, respondents were 0.117 percentage points more likely to silence volunteers using a female name compared with those using an ambiguously gendered name (two-tailed $ p<0.001 $ ), but 0.053 percentage points less likely to silence volunteers using male names (two-tailed $ p=0.034 $ ). The difference means that respondents force female-named volunteers to end 1.70 more interactions—and all subsequent outreach—out of every 1,000 text messages: a 181% increase over male-named volunteers.

FIGURE 2. Mean Silencing

Note: The figure shows the average treatment effect by treatment condition with 95% confidence intervals estimated using ordinary least squares. The comparison category is the ambiguous name condition.

Withdrawals

In both studies, supporters are more likely to politely end all future outreach when contacted by female-named volunteers compared with male-named and ambiguously named volunteers, per Figure 3.Footnote ¹⁴ Overall, respondents were 0.383 percentage points more likely to withdraw when volunteers used a female name compared with an ambiguously gendered name (two-tailed $ p< $ 0.001). In contrast, respondents were 0.334 percentage points less likely to withdraw from all future outreach when volunteers used a male name relative to those using an ambiguously gendered name ( $ p< $ 0.001).

FIGURE 3. Mean Withdrawal

Note: The figure shows the average treatment effect by treatment condition with 95% confidence intervals estimated using OLS. The comparison category is the ambiguous name condition.

Looking across Outcomes

Female-named volunteers face increased hostility and decreased ability to engage in future activism, especially relative to male-named volunteers. Recall that in our sample, an average volunteer sent approximately 1,300 messages, so a volunteer consistently using a female name—as would be the case for many women—is likely to experience all of these individual behaviors. The unnamed volunteer fares similarly. Yet the volunteers assigned to use a female name are not “anonymous” like those assigned no name.

In the SM, we show that there is no significant difference in how male and female respondents treated the female-named volunteers (see Tables S7–S9 in the SM), though men are more likely to send offensive and silencing texts across all conditions, while women were more likely to withdraw across all conditions. Thus, while most hostile responses come from men, women also reserved more of their hostile responses for other women.

Notably, despite this poor treatment, we find no evidence that female-named volunteers were less effective than volunteers assigned other name conditions. Female-named volunteers obtain higher response rates and supporter commitments to call their representatives than the other name conditions, despite their higher likelihood of receiving offensive, silencing, and withdrawal responses (see Tables S13–S20 in the SM). We discuss the implications below.