Women in Committees Although it is certainly important to consider how women’s speech during floor debate differs from that of men, we believe that committee action is simultaneously more consequential for the goals of many Members and is also more likely to result in gendered differences in members’ speech patterns. Ensuring equal ability of Members to participate in committees is important for a number of reasons, not the least of which is that most congressional business is done there (Deering and Smith Reference Deering and Smith1997). In addition, as committees can hold considerable power over not only which bills will advance to the floor but also the form in which they do, committee members can wield disproportionate influence over policy matters in which they are expert (Fenno Reference Fenno1973; Hall Reference Hall1987). Hearings are a central aspect of this process. Committees frequently use hearings to exercise oversight of executive departments’ activity, but they are a crucial part of the legislative process as well; committees hold hearings to get feedback on proposed legislation from experts, interest groups, and the public. During hearings, committee members can question witnesses or use their time to air opinions. They may also choose to engage their colleagues in debate. Thus, hearings serve an important function in both an investigatory and legislative sense and bring members into close contact. Compared with floor debate, committee hearings are also less restrictive in ways likely to be consequential with respect to how members might be treated, given the gendered communication styles described above. Especially in the House, floor speeches are highly regulated affairs in which members control the floor during their allotted time and during which strong norms might dissuade frequent interruption. Indeed, during many floor speeches in the modern Congress, there are often few other members present at all. Members’ time is generally regulated during committee hearings as well, but all members are usually guaranteed the opportunity to speak and the interplay between witness and Member offers more opportunity for one party to interrupt, talk over, or dodge the other. Thus, committees are the ideal theater in which to examine whether women do indeed experience more difficulty communicating their ideas in Congress.
Expectations If interruptions are a sign of conflict in Congress, they could signal conditions in which policy is more contested. As noted above, gender dynamics are a powerful determinant of a group’s communication dynamics. Moreover, committee work is an important theater in which women Members can support policies that will benefit women, which may be especially likely to create policy conflict. Although the dynamics surrounding gender, power, and conversation are complex (e.g., Itakura and Tsui Reference Itakura and Amy2004), a long-standing body of work has shown that men employ a number of tactics in conversation—including interruption—to exert social dominance (e.g., Zimmermann and West Reference Zimmermann, West and Konrad Koerner1996). Meanwhile, women have been substantially underrepresented in American legislatures, and Congress as an institution tends to diminish in various ways Members who are not white, male, or both (Hawkesworth Reference Hawkesworth2003; Kanthak and Krause Reference Kanthak and Krause2012). These dynamics are likely to collide in legislative proceedings, with potentially important implications for women Members. Some previous work on state legislative committees has uncovered gendered legislator styles in committee, with women allowing more of the hearing to elapse before speaking, taking fewer turns speaking for a lesser duration, and exhibiting a lower likelihood of interrupting their colleagues than men (Kathlene Reference Kathlene1994). In short, there are myriad reasons to believe that women in Congress face hurdles to effective communication that men do not. Thus, our primary hypothesis is that women Members are more likely to experience interruption in Congressional committees than men. We further expect that the pattern described above will hold when we focus on not only single instances of interruptions but also rapid-fire clusters in which Members fight for control of the conversation, which we believe is an especially aggressive form of interruption. Guided by previous work, (Bilous and Krauss Reference Bilous and Krauss2010; Hirschman Reference Hirschman1994; Palomares Reference Palomares2008), we also expect higher instances of interruption in mixed-gender Member pairings. Yet, a number of factors beyond gender are likely to affect interruptions—either on their own or by moderating the probability that women are interrupted. As noted above, women Members are powerful advocates for so-called women’s issues (e.g., Gerrity, Osborn, and Mendez Reference Gerrity, Osborn and Mendez2007), and legislative speech-making of men and women in Congress can depend on the topic of a debate (Osborn and Mendez Reference Osborn and Mendez2010). The fact that men and women in Congress are often advocating different policies is itself a potential mechanism of interruption; although some men will also support these policies, the position of others is likely to range from disinterested to hostile, leading to policy conflict. Thus, although we believe that, all else equal, women will be more likely to experience interruptions, we also expect that this effect will be significantly larger in committee hearings where women’s issues are being discussed. Furthermore, we expect chamber-based differences in the probability of gendered interruption. The Senate is traditionally characterized as one of the world’s great deliberative bodies, where personal relationships are especially important and procedural rules are often looser than in the House. These factors could combine to make Senate proceedings seem less formal to Members, which itself might encourage men to assert dominance in committee. However, the Senate is also likely to be a more difficult space for women to navigate. Historically, there have been fewer women in the Senate than in the House, and there is some evidence that men (but not women) in the Senate project their power via floor speech (Brescoll Reference Brescoll2011). Thus, all else equal, we might expect the gendered interruption gap to be larger in the Senate than the House. Contextual factors like party, institutional power, seniority, and even a Member’s unwelcome long-windedness might also affect the likelihood that women will be interrupted in congressional committees. Jacobi and Schweers (Reference Jacobi and Schweers2017) found that liberal Supreme Court justices are more likely than their conservative colleagues to experience interruption; if a similar dynamic is at work in Congress, we would expect Democrats to be interrupted at higher rates than Republicans. We also expect that being a member of the majority party or holding a committee chair, should tend to reduce instances of interruption. Existing work has argued that interruption is a method of asserting dominance (e.g., Zimmermann and West Reference Zimmermann, West and Konrad Koerner1996). As the people in charge of their committees, we should expect committee chairs to experience a lower likelihood of interruption. Power does not always trump gender when it comes to women being interrupted, however (Jacobi and Rozema Reference Jacobi and Rozema2018; Jacobi and Schweers Reference Jacobi and Schweers2017). Men and women wield the chair’s gavel differently; although men chairing committees tend to use their position in a unilateral fashion to steer witness testimony and policy conversation, women approach the job more as a moderator (Kathlene Reference Kathlene and Nielsen1990). Therefore, we expect that although chairpersons experience a lower probability of interruption than other Members, relative to men chairing committees, women in the chair will still be more likely to be interrupted. Perhaps a common-sense expectation is that as debates go long, we will see more interruption. One way to examine this is with the length of continuous remarks. As Members drone on, they could make themselves ripe targets for other Members looking to cut them off. Similarly, we consider the total time elapsed in the hearing prior to the speech chunk. If Members begin to lose interest in the session, they might become more likely to interrupt their colleagues as they grow agitated. Alternatively, if members pay less attention later in proceedings, they might interrupt less as the hearing drags on. Finally, we consider how seniority affects the likelihood that Members experience an interruption and also how it affects the probability that they will interrupt their colleagues. In their analysis of Supreme Court arguments, Jacobi and Schweers (Reference Jacobi and Schweers2017) found that as they gain experience, women develop communication strategies to thwart interruption and maintain command. As they become more seasoned, women Members might similarly develop more aggressive tactics intended to retain (or gain) control of the floor. If behavior of women in Congress mirrors that of women on the Supreme Court, we would expect the gendered interruption gap to narrow as Members serve longer tenures in office and also that women become increasingly more likely to interrupt their colleagues compared with men with the same seniority.
Data: Congressional Hearing Transcripts We analyze 24,103 congressional hearings published by the Government Printing Office (GPO) from 1994 to 2018 (105th–115th Congresses).Footnote 1 This sample of transcripts represents the entirety of those available electronically through the GPO. We merged the congressional hearings data to a database of committee assignments extended from Stewart (Reference Stewart2017) to attribute text to Members. Next, we merged NOMINATE ideology data from Lewis et al. (Reference Lewis, Poole, Rosenthal, Boche, Rudkin and Sonnet2020). The resulting hearings data include speech from 1,269 Members of Congress. Appendix Table A1 contains characteristics of the Members in our sample for each Congress and chamber. As is evident in Figure 1, the number of hearing transcripts available per Congress varies, with fewer transcripts in the more distant past.Footnote 2 The 105th Congress—the first one in our frame—contains the smallest number of transcripts (626). However, the number of available transcripts increases with each subsequent Congress, peaking with the 110th (3,238 transcripts), which was in session between 2007 and 2009; transcript availability has been fairly level since. In our data, the median number of hearing transcripts per Congress is 2,647 (mean = 2,192).Footnote 3 Each hearing transcript contains four major components: metadata, a header, the hearing’s transcript, and appendices (if any). We parsed the metadata to extract the GPO codes for the Members present. Then, we used a series of regular expressions to process these hearing transcripts into annotated chunks of sequenced dialogue, discarding the header and appendices.Footnote 4 These chunks were naturally clustered and annotated by GPO transcriptionists; we used the natural separation of chunks into new lines to define chunk boundaries, the GPO’s clear speaker attributions to parse title and last name, and the GPO’s machine-readable speaker code annotations to identify Members. We validated and expanded the GPO Member code annotations by matching the parsed titles and last names from the text against an external database of member committee assignments (Stewart Reference Stewart2017).Footnote 5 We discarded extraneous text chunks that were not attributable to a Member through GPO attributions or annotations, such as statements by witnesses. Appendix Table A1 reports additional validations, which suggest no significant difference in the distributions of the coded transcript data and the known Member population on the basis of Congress, chamber, gender, state, party, and seniority. The GPO clearly codes for interruptions. Transcriptionists assigned to the hearings annotate the machine-readable text with a flag for interruption, in the form of en or em dashes at the end of a chunk of speech.Footnote 6 We rely on the GPO’s natural composition and annotation of the chunks to define interruptions, formally coding for an interruption in a chunk if the GPO signal was present within the last 10 characters of the chunk.Footnote 7 For example, consider the exchange below, which includes six chunks.Footnote 8 Five of these chunks ended in the GPO signal for interruption; in Chunk 4, only the final dash is counted as an interruption, and Chunk 6 includes no interruption. 1. Mr. Horsford. Can I ask the point of order as to the reason for— 2. Mr. Jordan. You need unanimous consent to enter— 3. Mr. Cartwright. What would be the rule that— 4. Mr. Jordan. I am gonna recognize—I want to try to move and get to as many of our colleagues as I can. So— 5. Mr. Horsford. Mr. Chairman, under rule nine— 6. Mr. Jordan (continuing). For the next vote. This approach produced a total of 3,081,247 chunks used in this analysis—152,338 of which ended in interruptions (4.9%). As Table 1 illustrates, the rate of interruption overall remains fairly stable across all Congresses.
Chunk-Level Results The first analysis considers ordered chunks of speech as the unit of analysis. Each hearing j is represented as a panel of chunks, and each chunk i is coded for interruption, speaker gender, and an array of control variables. The dependent variable—whether the speaker was interrupted in that chunk—is coded as binary indicator $ {Y}_{ij}\in \left\{0,1\right\} $ , where 1 indicates an interruption. The independent variable of interest (gender) is coded as binary indicator $ {G}_{ij}\in \left\{0,1\right\} $ , where 1 indicates that the speaker is a woman. We also add a number of other variables to the model, not only to control for their independent effects but also because doing so allows us to explore whether the relationship between gender and interruptions (if any) is conditional on other factors. For instance, we include Member-level indicators for whether the speaker is the chair of the hearing (C ij ) and whether the speaker was in the majority party at the time of the hearing (P ij ). We also control for the length of the chunk because long-winded speakers may be more likely to be interrupted. To do this, we use the percentile rank of the chunk length within the hearing (L ij ), which transforms the variable into a metric that is comparable across hearings. We control for the timing of the speech within the hearing because participants may become fatigued or feel time constrained as hearings progress, resulting in higher rates of interruption. We capture timing with the percentile rank of the index of the chunk (T ij ), which reflects when in the hearing the chunk occurred; like the length metric, timing is also comparable across hearings. We also include a control for recent interruptions (R ij ), as interruptions may tend to cluster together conditional on topic or cadence. We use the log of the sum of the number of interruptions in the previous 10 chunks to operationalize this variable. Finally, we include fixed effect dummies for the Congress, committee type, and session (X ij ).Footnote 9 In summary, we model the probability of an interruption with the following regression specification, where the coefficients for gender and its interactions are of primary interest: (1) $$ \mathit{\Pr}\left({Y}_{ij}=1\right)={\displaystyle \begin{array}{l}{\mathrm{logit}}^{-1}\Big({\beta}_1{G}_{ij}+\Big(\sum \limits_{k\in \left\{C,L,T,R,,,P\right\}}{\beta}_k{k}_{ij}\\ {}+\hskip2px {\gamma}_k{k}_{ij}{\operatorname{}}^{\ast }{G}_{ij}\Big)+{X}_{ij}\delta \Big).\end{array}} $$ Women Are More Likely to Be Interrupted in Senate (but Not House) Hearings Table 2 reports the estimated coefficients and standard errors for the approach reflected in Equation 1. Models 1 and 2 are additive models of Member interruption, whereas Model 3 includes a number of interactions that allow us to calculate probabilities of gender-driven interruption across a range of moderators. In Model 1, which includes controls for specific traits of the chunk, the “Woman” coefficient is positive and statistically significant, suggesting that women in Congress are more likely to experience interruptions. When we include controls for member attributes in Model 2, however, the “Woman” coefficient is effectively zero. Considering that one of these attributes—an indicator for membership in the U.S. Senate—is substantively large and statistically significant, we next examine the possibility that the overall effect in Model 1 is driven by the experience of women in the Senate. Appendix Table A3 contains results of models that subset the data to House and Senate Committees, allowing us to calculate the predicted probability of interruption for speakers of either gender, in each chamber. We depict those results—as well as some results in Table 2 for members overall—in Figure 2. As Figure 2 demonstrates, the results for Members overall in Table 2 (Model 1) appear to be driven by Senate committees. There is little substantive difference in Figure 2 in the probability of interruption among men and women Members in the House. In the Senate, however, the gap between women and men is approximately one half of a point—a statistically significant effect. Although this may seem like a small effect, given baseline interruption rates for men it represents more than a 10% increase in the probability of interruption when the senator is a woman. Moreover, this effect is exacerbated because women already are at a disadvantage with respect to their total amount of speaking time. Due to their underrepresentation in the Senate, women Senators on average use only 21% of the amount of speaking time in a typical hearing.Footnote 10 Any interruption is an uncomfortable experience with the potential to sidetrack a speaker’s progress, but given that women are already at a disadvantage in terms of overall time, the 10% differential in the interruption rate suggests that women face especially challenging conditions while discussing substantive issues in U.S. Senate committees. Women Are More Likely to Be Interrupted When Discussing Women’s Issues We next evaluate whether the likelihood of interruption for men and women in Congress is conditional on the topic of the committee hearing. We expect that women will experience a higher probability of interruption if the topic of the hearing can be labeled as a women’s issue rather than a non-women’s issue. To conduct this analysis, we employed a computer-assisted procedure to code committee hearings with respect to whether they addressed a women’s issue. We developed a list of keywords following the methodology of Osborn and Mendez (Reference Osborn and Mendez2010), who divide congressional speech into categories that include traditional women’s policy issues, such as health care, family-related issues, education, and social welfare issues, as well as issues of direct relevance to women as a constituency, which include abortion, women’s crime issues, women’s health issues, and women’s family issues. We also consulted the Routledge International Encyclopedia of Global Women’s Issues and Knowledge (Kramarae and Spender Reference Kramarae and Spender2004), which contains a comprehensive bank of keywords on women’s issues.Footnote 11 We coded an indicator $ {W}_{ij}\in \left\{0,1\right\} $ , where 1 indicates that the official hearing title included any of the keywords included in Table A10. Our approach resulted in 3,790 hearings—15.8% of the total hearings—coded as addressing a women’s issue.Footnote 12 We included this indicator and its interaction with gender into the modeling approach articulated in Equation 1, the results of which are contained in Table A8. The models in Table A8 indicate that in hearings addressing women’s issues, the speaker is more likely to be interrupted regardless of their gender. This is consistent with our expectation of heightened conflict in this policy domain. Also consistent with our expectation, the interaction coefficient in Table A8, Model 3 demonstrates that if the speaker is a woman, she is even more likely to be interrupted in hearings devoted to women’s issues. In Figure 3 we depict the predicted probabilities derived from these models, as well as those in Table A9, which include interactions that allow us to calculate these probabilities separately by chamber. The pattern in Figure 3 is striking. Overall, women are on average 2.3 times more likely to be interrupted when discussing women’s issues than when not discussing women’s issues (p = 0.01). In the Senate, the multiplier dips slightly to 1.5 (p < 0.01). In the House, the multiplier skyrockets such that a woman is 6 times more likely to be interrupted when discussing women’s issues than when not discussing women’s issues (p = 0.02). The evidence in Figure 3 is therefore consistent with our expectation that women face particularly acute instances of interruption when discussing women’s issues, which suggests greater conflict surrounding those policies. Interruptions across Majority, Chairmanship, and Party We next consider how chairmanship, being in the majority, and being a member of a particular party shape the probability of interruption, conditional on gender. There is some evidence that these factors on their own generally affect interruptions in Congress. For instance, the chunk models in Table 2 indicate that Members in the majority party are significantly less likely to be interrupted than their minority-party colleagues. That model suggests that committee chairs and Republicans are also less likely to be interrupted relative to other Members, holding other variables constant. All of these results are consistent with our expectations. Figure 4 depicts the predicted probabilities from Table 2, Model 3 of Members with these traits experiencing an interruption, by gender, with separate effects depicted for the House and Senate. There is some evidence of moderated effects. For instance, in the Senate Democratic women are more likely than Democratic men to experience an interruption, but that effect is not present for Republicans. The probability of interruption is also slightly larger for women in the Senate minority party than for those in the majority. In the House, the effect reverses for women chairing committees, who are about half a point less likely than male chairmen to experience an interruption; this effect is not present in the Senate, though it is directionally negative. In short, certain legislative characteristics do seem to alter the core dynamics at work with respect to the relationship between gender and interruption.Footnote 13 No Gender Differences across Long-Windedness and Hearing Length Contrary to expectations, examining the coefficient for chunk length in Table 2, Models 1 and 2, we find that overall Members are actually less likely to experience an interruption as they speak for a longer period. Our results further suggest that chunk timing matters; both Models 1 and 2 in Table 2 indicate that Members are more likely to experience an interruption late in the hearing. But how do long-windedness and speech timing shape the probability that women Members will experience an interruption? Figure 5 depicts predicted probabilities of interruption conditional on gender across the range of long-windedness and hearing-length percentiles. These probabilities are derived from Model 3 in Table 2—which includes interactions between gender and each of the relevant control variables. Relative patterns for both impatience (left panel) and long-windedness (right panel) are quite similar in the House (top row) and Senate (bottom row). In both chambers, women and men are similarly likely to be interrupted across the range of both the chunk’s timing within a hearing and how long Members speak. In each chamber, neither chunk timing nor length appear to affect the probability of interruption differently for men and women. Women Do Not Adjust Their Behavior with Experience In the leftmost panels of Figure 6, we plot the predicted probability of interruption for men and women across years of seniority, computed from Table 2, Model 3. Although both men and women Members grow more likely to be interrupted as they gain seniority, this trend is stronger for men in both chambers. For instance, women begin their Senate careers with a higher estimated probability of interruption (bottom-left panel), but due to the steeper slope for men, the gender gap closes around Members’ twentieth year of service. Thereafter, the estimated probability of interruption is lower for women with the same length of service. That said, the confidence regions indicate that the predicted probabilities of interruption for men and women in the Senate are never significantly different at the depicted levels of experience. That is not the case in the House, however (top-left panel). Women and men begin their House careers with effectively the same interruption probability, and although Members of both genders grow increasingly likely to experience interruption as they serve longer, the probability of interruption grows at a significantly slower rate for women. As a result, after about six terms of service House women are significantly less likely to be interrupted than men with the same amount of experience, with the gap between men and women continuing to grow as they serve longer. The slower rate of growth in interruptions across women’s House careers—compared with men serving the same tenure—could be taken as evidence that women adapt over time to fending off interruption, perhaps by developing a more masculine legislative style. Such an adjustment would be broadly consistent with the dynamic that Jacobi and Schweers (Reference Jacobi and Schweers2017) report among women on the Supreme Court. But is there evidence that women in Congress are changing their behavior in this way? To answer this question, rather than predicting the likelihood that a Member is interrupted, the right panels of Figure 6 depict predicted probabilities (from Table A11) that a Member who commits an interruption will be a man or woman, conditional on the number of years that the interrupter has served in Congress. Because committing an interruption is a purposeful action, it likely provides more information about the communication strategy that Members are attempting to pursue than merely observing whether they are interrupted (Jacobi and Schweers Reference Jacobi and Schweers2017). Thus, if we observe that the effect of seniority on women’s likelihood of interrupting colleagues is significantly higher than that for men, we would have more evidence that women are intentionally adjusting their tactics. Figure 6 yields no support for that scenario. In both chambers, Members grow more likely to interrupt their colleagues as they gain more seniority. In the House, women are significantly less likely to be the interrupter across the range of seniority, and the growth trend is similar for men and women. In the Senate, women have a higher predicted probability of being the interrupter across the range of seniority, though the estimates for men and women Members are not statistically different at any point. Thus, these models do not provide clear evidence that women become more aggressive in their interrupting behavior as they gain experience and, by extension, are not consistent with the notion of women Members changing their tactics relative to men as they serve a longer tenure in office. Women Fight for Their Time We close our chunk-level analysis by providing additional contextual description of the conditions in which women in Congress experience interruption. Jacobi and Rozema (Reference Jacobi and Rozema2018) distinguish “conversational overlap,” where interruptions occur as a conversation is transitioning, from “substantive” interruptions that are more clearly intended to cut off a speaker in the middle of making their point. If the interruptions in our data were primarily of the former variety, perhaps they would not indicate a major hurdle in the legislative process. However, most of the interruptions in our data are likely substantive in nature: in a supplementary analysis reported in Table A2 and Figure A2, we find that about three-quarters of the interruptions happen after the first five words of a chunk and more than half of the interruptions happen after the first 10 words of a chunk. However, we note that “conversational overlap” can also impede substantive points if attempted incursions occur repeatedly at a chunk’s outset. Indeed, an interrupter who is truly determined to silence a speaker is likely to make multiple attempts, and repeated efforts to interrupt a speaker before they fully gain the floor could result in “interruption clusters” that may be taken as a more aggressive attempt to cut the speaker off. Being involved in these clusters—regardless of where in the original chunk they begin—might therefore be a sign that other Members are trying to talk over each other or otherwise mitigate the speaker’s influence. Therefore, understanding whether women are more likely to “fight for their time” can shed light on the dynamics underlying their higher likelihood of interruption demonstrated above. If women in congressional committees are more likely than men to find themselves engaged in rapid-fire interruptions as they battle with someone for control of the conversation, it might constitute evidence that women in Congress are interrupted more often because others are actively attempting to thwart the points they were hoping to make. Therefore, we shift the focus here beyond single instances of an interruption to examine whether women are also more likely to fight for their time. We consider whether the pattern we observe for interruption clusters mirrors that for interruptions reported above. To answer this question we employ the same approach used previously (articulated in Equation 1), but we shift the dependent variable to a binary indicator for whether a speaker has been interrupted more than once during the previous 10 chunks, $ {Y}_{ij}^{\prime}\in \left\{0,1\right\} $ . Table A12 reports the coefficients and standard errors for these models, and Figure 7 depicts the predicted probabilities from these models by chamber and speaker gender. We see a similar pattern here as in Figure 3: women in both chambers are significantly more likely to be involved in an interruption cluster in hearings involving women’s issues than they are during hearings on other topics. Both overall and in the House, men and women are similarly likely to fight for time on non-women’s issues. However, in hearings involving women’s issues, significant gaps between men and women emerge. Specifically, women are 44 times more likely (p < 0.01) to fight for time on women’s issues in Congress overall, and they are 6 times more likely (p < 0.01) to do so in the House.Footnote 14 Women in the Senate are more likely than men to be involved in interruption clusters across the board, but they are 1.3 times more likely (p < 0.01) to fight for time when discussing women’s issues. We take this as further evidence that debate surrounding women’s issues comes with especially high levels of policy conflict, as women are more likely to experience other speakers intentionally and aggressively trying to impede their ability to effectively communicate in these policy areas.Footnote 15
Dyad Analysis We close with an analysis of speaker dyads, designed to examine the question of whether interruptions are more likely to occur in mixed-gender pairings. Because Congress is a social system, we would expect that dyadic interactions are not independent of other interactions. For example, some Members might be naturally disruptive or rude. Likewise, a Member might be more likely to interrupt another because the other Member interrupted her first (an endogenous “retaliation” effect). Or perhaps communication follows different patterns when people are speaking with a partner of the same gender than when groups are gender-mixed (Bilous and Krauss Reference Bilous and Krauss2010; Hirschman Reference Hirschman1994; Palomares Reference Palomares2008). To account for these types of interactions, we fit additional models for which the unit of analysis is pairs of Members within each Congress. We treat each Congress as an individual network, where the nodes are Members and the edges connecting them are weighted by the number of times they interrupted each other. We use this network design to explore whether the heightened probability of interruption reported above is conditional on the gender of the person with whom the speaker interacts. Edges are directed, such that the edge weights between any two Members may vary based on how much one Member interrupts the other. A dyadic approach has been successfully used to study Supreme Court interruptions (Jacobi and Rozema Reference Jacobi and Rozema2018). This approach allows us to focus on the factors that contribute to the relationships between Members, conditional on factors such as their gender, institutional status, party, delegation, and chamber.Footnote 16 Using a logit estimator to account for these types of interactions can result in bias and overstate significance (Aronow, Samii, and Assenova Reference Aronow, Samii and Assenova2017; Cranmer and Desmarais Reference Cranmer and Desmarais2011; King and Zeng Reference King and Zeng2001) due to the endogeneity of interruption behavior. Therefore, we employ an exponential random graph model (ERGM) to allow for these types of complex interdependencies as well as to estimate less biased coefficients and measures of uncertainty. Specifically, we employ an ERGM where edges are specified using binary flags—rather than a valued ERGM, where edges are specified using counts (e.g., Krivitsky Reference Krivitsky2012)—because the models are easier to interpret and may be more reliably estimated. We follow Cranmer and Desmarais (Reference Cranmer and Desmarais2011), who use a “thinning rule” to binarize edge data for a network of Members of Congress. We prepared N = 8,206,743 directed edges for Ω = 1,264 Member nodes from the 105th to 115th Congresses.Footnote 17 Directed edges for Members were created using the chunk-level data.Footnote 18 Each node was associated with an array of Member characteristics including gender G i (where 1 indicates the Member is a woman); chamber membership C i (where 1 indicates the Senate); partisanship P i (where 1 indicates a Republican); experience S i , measured as the number of years the Member has served; state delegation D i , an indicator for the Member’s state, and; ideology I i , the Member’s NOMINATE score in that Congress. After applying a binarization threshold of four interruptions and creating one network for each Congress, we ultimately obtained 174,508 nonzero edges (see Appendix Section A2). We report the ERGM results in terms of predicted probabilities for the sake of interpretation (for full results see Appendix Table A15.) Figure 8 depicts predicted probabilities of interruption among the four possible gender dyads within each Congress over time.Footnote 19 The predicted probability of forming a tie is generally consistent within all four dyad pairings over the time series. The patterns are also comparable across in-party (top panel) and out-party (bottom panel) dyads. Figure 8 offers some suggestive evidence that interruptions are more likely to occur in mixed-gender dyads, as same-gender dyads have directionally lower probabilities of forming ties than do mixed-gender dyads. Indeed, the “Men Interrupting Women” dyads typically exhibit the highest predicted probabilities of forming an interruption tie. That said, the point estimates for the “Women Interrupting Men” dyads exhibit substantial confidence interval overlap with those in the “Men Interrupting Women” dyads, so we cannot conclude that men are more likely to interrupt women than vice versa. Figure 8 also suggests that women are about as likely to interrupt women as they are to interrupt men; although as noted the predicted probability of women interrupting men is typically higher than for women interrupting women, there is considerable overlap among the confidence intervals in the women-driven dyads. Among men who interrupt, Figure 8 shows that among the four dyads we examine the estimated probability of forming a tie is typically lowest for “Men Interrupting Men” across Congresses. Moreover, we see little overlap in the confidence intervals around the predicted probability of “Men Interrupting Women” and those around “Men Interrupting Men,” suggesting that men may be more likely to interrupt women than they are to interrupt men.