CAF - Women in Office The Impact of Female Politicians on Gender Based Violence Reporting

C A F - W O R KI N G P AP E R # 2 0 2 5 / 1 2

T h i s v e r s i o n : O c t o b e r 3 1 , 2 0 2 5

Women in Office: The Impact of Female Politicians on Gender-Based Violence Reporting Veronica Frisancho1 | Evi Pappa 2 | Camila Ramírez 3 |

Chiara Santantonio4 1CAF - Development Bank of Latin America and the Caribbean. vfrisancho@caf.com 2Universidad Carlos III de Madrid and CEPR.ppappa@eco.uc3m.es 3World Bank. cramirez3@worldbank.org 4University of Bath. cs2983@bath.ac.uk Gender-based violence in the U.S. is a silent epidemic. Twenty percent of women experience rape, yet only one in three reports it. Using FBI data and a regression discontinuity design, we examine the impact of female U.S. House Representatives on reported rapes and femicides. Our findings suggest an increase in reporting, rather than higher levels of violence. Our setting and additional analysis allow us to rule out policy channels. We argue that female politicians serve as role models, influencing reporting through symbolic and social pathways. Congressional speech data support this argument: female legislators advocate more against gender-based violence, and their speeches correlate with higher reporting in their districts. K E Y W O R D S mixed-gender race, gender-based violence, crime data Small sections of text that are less than two paragraphs may be quoted without explicit permission as long as this document is acknowledged. Findings, interpretations and conclusions expressed in this publication are the sole responsibility of its author(s) and cannot be, in any way, attributed to CAF, its Executive Directors or the countries they represent. CAF does not guarantee the accuracy of the data included in this publication and

is not, in any way, responsible for any consequences resulting from its use. ©2024 Corporación Andina de FomentoC A F - D O C U M E NT O D E T R AB A J O # 2 0 2 4 / 0 1

E s t a v e r s i ó n : O c t u b r e 3 1 , 2 0 2 5

Mujeres en el Cargo: El Impacto de las Mujeres en Política en el Reporte de Violencia de Género Veronica Frisancho1 | Evi Pappa 2 | Camila Ramírez 3 |

Chiara Santantonio4 1CAF - Development Bank of Latin America and the Caribbean. vfrisancho@caf.com 2Universidad Carlos III de Madrid and CEPR.ppappa@eco.uc3m.es 3World Bank. cramirez3@worldbank.org 4University of Bath. cs2983@bath.ac.uk La violencia de género en EE. UU. es una epidemia silenciosa. El veinte por ciento de las mujeres sufre violación, pero solo una de cada tres lo denuncia. Utilizando datos del FBI y un diseño de regresión por discontinuidad, examinamos el impacto de las Representantes femeninas de la Cámara de EE. UU. sobre las violaciones y los femicidios denunciados. Nuestros hallazgos sugieren un aumento en la denuncia, más que niveles más altos de violencia. Nuestro entorno y análisis adicionales nos permiten descartar canales políticos/normativos (policy channels). Argumentamos que las mujeres políticas sirven como modelos a seguir, influenciando la denuncia a través de vías simbólicas

y sociales. Los datos de discursos del Congreso respaldan este argumento: las legisladoras abogan más en contra de la violencia de género, y sus discursos se correlacionan con un mayor nivel de denuncias en sus distritos.

K E Y W O R D S

carrera de género mixto, violencia de género, datos de criminalidad Pequeñas secciones del texto, menores a dos párrafos, pueden ser citadas sin autorización explícita siempre que se cite el presente documento. Los resultados, interpretaciones y conclusiones expresados en esta publicación son de exclusiva responsabilidad de su(s) autor(es), y de ninguna manera pueden ser atribuidos a CAF, a los miembros de su Directorio Ejecutivo o a los países que ellos representan. CAF no garantiza la exactitud de los datos incluidos en esta publicación y no se hace responsable en ningún aspecto de las consecuencias que resulten de su utilización. ©2024 Corporación Andina de FomentoFRISANCHO ET AL. 2 1|INTRODUCTION Gender-based violence (GBV) is a pervasive and often unspoken crisis in the United States (U.S.). Nearly 20 percent of women in the country have experienced rape during their lifetime, and one in four has endured severe physical violence from a partner (NCADV, 2014). Despite its prevalence, rape remains one of the least reported crimes, with roughly 63 percent of sexual assaults going unreported to law enforcement (NSVRC, 2015). Underreporting is not unique to the U.S.: ample evidence documents similar patterns in both developed and developing countries (Palermo et al., 2014; Agüero and Frisancho, 2022; Cerqua et al., 2024). Reporting is nonetheless vital for prevention and deterrence, as formal complaints enable

protective interventions and reduce future incidents. For example, Amaral et al. (2023) show that arrests triggered by reporting reduce future domestic violence calls by as much as 51% in the ensuing year, demonstrating a significant deterrence effect. Understanding what encourages survivors to come forward is thus a critical policy priority. This paper investigates whether women’s political representation can help close the GBV reporting gap through a role model effect. We ask whether the visibility and leadership of female politicians can shift social norms, legitimize survivors’ voices, and empower women to report abuse. Using a regression discontinuity (RD) design in close mixedgender U.S. House races from 1970 to 2017 and official FBI crime data, we find that districts narrowly electing a female Representative experience a roughly 40 percent increase in reported rapes relative to those electing men. This rise does not reflect higher violence, as fatal crimes involving women remain unchanged, but instead greater empowerment and willingness to report. Because legislative authority and funding decisions occur at the federal level, while crime data vary at the district level, we interpret this increase as arising from women’s visibility and influence as public role models rather than from direct policy action. 1 Although Members of Congress lack direct authority over local law enforcement, their public engagement can shape social attitudes and institutional responsiveness indirectly. We therefore examine changes in police activity following the election of a Congresswoman (Iyer et al., 2012; Miller and Segal, 2018; Amaral et al., 2021) and find no evidence of increased law enforcement effort in gender-based crime cases, if anything, the share of solved female homicides declines slightly. Supporting the role model effect, we analyze 1.8 million Congressional speeches from 1971 to 2012 (Gentzkow et al., 2018) and find that Congresswomen are substantially more likely than men to address GBV-related issues. The frequency of such speeches correlates

positively with rape reports in their districts, consistent with greater awareness and survivor empowerment rather than local policy changes. Finally, using survey data from the American National Election Studies (ANES), we show that women in districts represented by a Congresswoman express stronger support for reporting harassment and more favorable views towards the police. For men, on the other hand, attitudes toward the police remain unchanged, while some evidence points to more favorable views on reporting harassment. We find no indication of backlash or rising violence, reinforcing the interpretation that the observed effects operate through social and cultural channels rather than political polarization. To ensure robustness, we conduct extensive sensitivity tests. Our results remain consistent across alternative outcome definitions, samples, and bandwidths. We also rule out confounding by partisanship: although women are more likely to be Democrats, electing a 1The role model channel is vividly illustrated by public testimonies. For instance, during the legislative debate over Ohio’s fetal heartbeat bill on April 10, 2019, State Representative Lisa Sobecki shared her own experience of sexual assault. Following her speech, a man later approached her to say that his wife, deeply moved by Sobecki’s words, had disclosed her own rape for the first time in forty-one years (Times, 2019).FRISANCHO ET AL. 3 Democrat over a Republican does not significantly affect reported rapes. Evidence from developing countries shows that electing female leaders often improves outcomes for women. Female politicians can challenge gender-biased norms, inspire girls’ aspirations, and promote redistributive and child-oriented policies (Beaman et al., 2009, 2012; Priyanka, 2020; Brollo and Troiano, 2016; Bruce et al., 2022; Clots-Figueras, 2012; Bhalotra et al., 2022; Bhalotra and Clots-Figueras, 2014). They also prioritize public goods that

benefit women, advance women-friendly laws, and actively address gender-specific issues (Chattopadhyay and Duflo, 2004; Clots-Figueras, 2011; Gerrity et al., 2007). In contrast, studies from advanced economies provide mixed evidence. Hessami and da Fonseca (2020) and Carozzi and Gago (2023) find limited effects of female representation on spending patterns, while Bagues and Campa (2021) show that gender quotas alone do not overcome institutional barriers to women’s influence. Yet more recent work demonstrates substantive impacts: Lippmann (2022) finds that, in France, legislators are more active on women’s issues, while Brogaard et al. (2024) shows that female politicians increase the share of U.S. government procurement contracts awarded to women-owned firms. Within this broader literature, our paper contributes to understanding how female representation affects GBV . Prior research has largely focused on specific policy interventions, such as cash transfers, alcohol restrictions, enhanced property rights, or the creation of women’s police stations (Hidrobo et al., 2016; Luca et al., 2015; Amaral, 2017; Amaral et al., 2021), while the political dimension has received less attention. A small but growing set of studies explores this link. In Indonesia, Kuipers (2020) finds a significant negative relationship between women’s presence in local councils and the share of women who approve of a husband assaulting his wife, showing that female political representation can lower acceptance of GBV . In India, Anukriti et al. (2024) show that exposure to female legislators increase intimate partner violence. The authors attribute this link to expanded access to modern contraception, which exacerbates spousal conflict. In Brazil, Delaporte and Pino (2022) and Bochenkova et al. (2023) find that electing female mayors reduces domestic violence by nearly half. These previous studies analyze local executives with direct control over policy implementation. In contrast, we study U.S. Representatives who legislate federally but have no authority over local policing or welfare programs—allowing us to isolate role model effects from policy effects. Our work also builds on Iyer et al. (2012), who show that female representation in local governments in India has led to increased recorded crimes against women. They leverage data on fatal crimes and rape records and show that this increase is primarily due to greater reporting rather than greater incidence of GBV . However, the use of the staggered enactment of quotas may conflate the impact of increased representation with the broader effects of policy shifts and changing societal attitudes (Beaman et al., 2009; Clots-Figueras, 2012). For example, quotas may be introduced in a policy environment where there is already a big push for gender equality, which may introduce contextual confounders. Quotas can also reflect the consequences of the policy itself – affecting the characteristics of the women who choose to compete under quotas vs open elections– or the nature of political competition (Chattopadhyay and Duflo, 2004). Our research design has an advantage over the use of quotas, identifying the causal impact of women winning office in a politically competitive environment. Our results thus speak to contexts without formal political interventions. The U.S. context also strengthens the interpretation of our results. Because federal laws apply nationally but outcomes are observed at the district level, any differential reporting across districts following the election of a woman cannot stem from new federal legislation. This institutional “mismatch” provides a rare opportunity to disentangle symbolic and cultural influences from formal policy mechanisms. Finally, we go beyond previous studies by linking Congressional speech records and survey data to GBV outcomes, offering novelFRISANCHO ET AL. 4 evidence on the communicative and attitudinal channels through which female leaders may

empower victims and reshape social norms. The rest of the paper is organized as follows: Section 2 describes the data, Section 3 outlines the empirical methodology, and Section 4 presents the main results and robustness checks. Section 5 explores potential mechanisms, and Section 6 concludes. 2|CONTEXT AND DATA 2.1|The U.S. House of Representatives The U.S. House of Representatives is one of the two chambers of the U.S. Congress alongside the U.S. Senate. It is composed of 435 voting members, known as Representatives (or Congressmen and Congresswomen), each elected to serve a two-year term representing one congressional district.2 All seats are contested every two years, and members are elected directly by citizens. Representatives carry out a broad range of work, including introducing bills, offering amendments, voting on legislation, but also serving on committees, and engaging with constituents and other groups. The House of Representatives shares its legislative powers with the Senate: any member may introduce a bill, but legislation must be approved by both chambers – and signed by the President – to become law. It also holds the so-called “power of the purse”, meaning it originates and approves spending and taxation measures.3 2.2|Elections Data on the U.S. House elections from 1970 to 2016 was sourced from CQ Press, and supplemented with information on the sex of all candidates from the Republican and the Democratic party fromPoliticalParity (2015), the Center for American Women and Politics (CAWP), and Pettigrew et al. (2014).4 As mentioned in Subsection 2.1, elections for the U.S. House of Representatives are held biennially across 435 Congressional districts, with one Representative being appointed per district. This results in a total of 10,440 races during our sample period. Mixed-gender races (female vs. male) account for 18.2% of this sample (1,895 races) and are more often won by

men (1,081 male vs. 814 female wins, see Table A1). These races also appear to exhibit a partisan imbalance, with female winners more likely to come from the Democratic party and male winners more often representing Republicans. A geographical analysis reveals that mixed-gender races are more frequent in coastal states (see Panel A in Figure A1). However, female candidates win mixed-gender elections across the country (see Panel B in Figure A1). We match election data with district-level characteristics (e.g., demographics, income, education, and inequality) for 1972–2016 from Foster-Molina (2017). Additional details on the election data are available in Appendix A.1. 2Congressional districts are smaller geographical areas than states. Each state is assigned a number of districts based on its population, with every state guaranteed at least one. States are responsible for drawing the boundaries of their own districts but cannot determine the number allocated to them. 3“Article I, Section 7, of the Constitution provides that all bills for raising revenue shall originate in the House of Representatives but that the Senate may propose, or concur with, amendments. By tradition, general appropriation bills also originate in the House of Representatives.” (see How Our Laws Are Made) 4These datasets do not include the sex of candidates from other parties. Because of this restriction and because of the strong bipolarity of the American party system, we focus on the sex of Republican (R) and Democratic (D) candidates only. On average, in all elections in which the first and second places correspond to R and/or D, the share of votes for these two candidates is above 98 percent.FRISANCHO ET AL. 5 2.3|Crimes Crime data come from the FBI’s Uniform Crime Report (UCR), specifically the Offenses Known and Clearances by Arrest (OKCA) files (1970–2017) and Supplementary Homicide

Report (SHR) files (1976–2017)(Kaplan, 2019b,a). The OKCA files include monthly crime data by individual law enforcement agency, capturing victim counts and clearances. These records do not include detailed information on the victim or the perpetrator, such as their relationship or their sex. We retrieve the number of female victims raped by men (forcible and attempted) and the number of rape offenses cleared between 1970 and 2012. Throughout this period, such crime was explicitly defined as “carnal knowledge of a female forcibly against her will” by a male (FBI, 2013).5 We also retrieve the number of offenses and offenses cleared for other felonies that are not specifically gender-based, such as assault, robbery, and vehicle theft.6 The SHR provides incident-level data on fatal crimes. From these records, we obtain the number of female victims of murder and non-negligent manslaughter (i.e., women killed). We construct two additional outcome measures by restricting this number to the number of female victims when all offenders are male, and the number of female victims with a relationship to the offender (e.g., wife, girlfriend, or daughter), i.e., number of victims of intimate femicide. In addition, we retrieve data on other fatal crimes, such as the number of men killed and the total number of murder victims. Crime statistics are aggregated at the year–district level using geographical matching, based on agency coordinates and district maps from Lewis et al. (2013) and the Census Bureau. As noted in Appendix A.2, the records available provide data on rape cases using the reporting date while homicides cases are recorded using the occurrence date. While these constraints are imposed by the data, we deem it appropriate given our focus on reporting behavior. If Congresswomen become a role-model that fosters reporting, some rapes from

the past may be reported after the election of a woman. Moreover, we argue that comparing recorded rapes vs occurrences of intimate femicides is also helpful to disentangle actual increases in GBV from changes in reporting behavior. To address data quality challenges (e.g., missing data, frequency of reporting, representativeness), we focus on local agencies (e.g., city police) reporting data consistently for all 12 months of the year.7 In our analysis, crimes are normalized per 100,000 district residents. We also examine the sensitivity of our results when we normalize rape and fatal crimes involving women by the size of the female (voting) population in the district. Crime data are aggregated at the term level for each winning candidate. For instance, total crimes committed between 2011 and 2012 are linked to a 2010 election. We also aggregate crimes at the annual level within each two-year term for additional robustness checks. Additional details on the crime data are available in Appendix A.2 and Appendix B. 5This definition was expanded in 2013, after which the term forcible was dropped, and victims and perpetrators of both genders were considered. Since the OKCA files do not include information about the sex of the victim or the perpetrator, we cannot identify these counts for more recent years. See the FBI website at ucr.fbi.gov for more details. 6For crimes against the person (e.g., homicide, rape), one offense is counted for each victim. For crimes against property (e.g., robbery), one offense is counted for each distinct operation, with the exception of motor vehicle theft for which one offense is counted for each stolen vehicle (FBI, 2013). 7Focusing on local agencies (while excluding, for example, those with state-wide jurisdiction) ensures that crimes are more accurately counted within the boundaries of a given district.FRISANCHO ET AL. 6 2.4|Congressional Speeches

We gather data on Congressional speeches between 1971 and 2012 from the Congressional Records dataset compiled by Gentzkow et al. (2018). This dataset contains processed text from the bound and daily editions of the U.S. Congressional Records. It includes transcripts of individual floor interventions in each chamber of Congress. It also provides speaker information such as name, district, and sex.8 To identify speeches related to GBV , we focus on those that contain predefined key terms: rape, femicide, domestic violence, intimate partner violence, violence against women, gender-based violence, sexual assault, sexual violence, sexual harassment, and battered women.9 However, a speech may mention a key term only circumstantially: for example, a Representative may mention “rape" in the context of general crime trends, rather than as part of a discourse on GBV . To address this, we apply a second filter to these pre-selected speeches. We rely on latent Dirichlet allocation (LDA), a topic modeling technique, to determine which of the speeches are most likely centered on GBV . LDA is commonly used in natural language processing to identify latent topics in textual data. We applied this analysis to speeches containing key terms and not to the full body of speeches because only a small proportion of speeches are associated to GBV (less than 1% across the different definitions used) and, therefore, GBV does not come up as a topic when the analysis is applied to all the speeches in the data without a previous filter. We use LDA to extract topics in the collection of speeches, identify those that are GBV-related, and estimate the probability that a given speech belongs to a GBV-related topic. Finally, we filter out speeches for which this probability, denoted as Qi, falls below the t-th percentile of the overall distribution ofQ. In the results discussed in Section 5, we have set t= 75, so that we keep the 25%

of speeches most likely related to GBV . To select this threshold, we selected a random sample of 50 speeches among those containing key words, read them entirely, and manually classified them as “relevant" or “not relevant". In this random sample, the proportion of relevant speeches was ∼ 25%. Moreover, when evaluating the classification accuracy using the manually coded sample, we found that applying the 75th percentile threshold yielded a more accurate identification of GBV-related speeches than alternative cutoff points. With this procedure, we filter out all speeches that: a) do not contain a key term or b) contain key terms but are among the least likely to be GBV-related. The results discussed in Section 5 are robust to other choices of the threshold. We present a robustness test using the 50th percentile in Table D10 in Appendix D. This approach enables us to better comprehend the content of the speeches and refine our dataset by excluding those texts that incidentally mention key terms but do not focus on GBV . Using this filtered set, we identify districts in which a Representative made at least one GBV-related intervention in a given year. In addition, we calculate the proportion of GBV-related speeches delivered by each Representative annually. 3|EMPIRICAL STRATEGY: REGRESSION DISCONTINUITY DESIGN We estimate the causal effect of female representation using an RD design in a sample of close mixed-gender races. Our identification strategy relies on the assumption that, in close races between a man and a woman, it is mostly random factors that tilt the scale in favor 8We retrieve the speech data from the bound edition (Congresses 92nd-96th) and from the daily edition (Congresses 97th-112th). We use district and year identifiers to match the speeches with the crime data. 9We obtained this list iteratively, by selecting an original list of key terms, categorizing the speeches that contained these terms in topics, and expanding the list with keywords and bigrams that were among the most

frequent for speeches related to GBV .FRISANCHO ET AL. 7 of one of the candidates. Therefore, a district in which a man wins against a woman by a narrow margin of victory becomes a good counterfactual for a district in which the opposite occurs. In our setting, close mixed-gender races are defined as those in which the top two places are occupied by a male and a female, irrespective of the total number of candidates, and where the election results are tightly contested. We first compute the female margin of votes, i.e., the difference between the votes received by the female and the male candidates as a share of the total votes, for each race taking place in districtd, states, and yeart: MVd,s,t = (Fd,s,t −M d,s,t Votesd,s,t )×100 We define our assignment to treatment as Wd,s,t = 1[MVd,s,t > 0], an indicator variable that takes a value of one if the winner of the (mixed-gender) race is female. The effect of having a woman in power can then be estimated with the local polynomial of thep-th order: yd,s,t =γ s +λ t + pX k=0 ηkMVk d,s,t +W d,s,t pX k=0 βkMVk d,s,t +X ′ d,s,tα+ε d,s,t (1) in the sample of races in which the female margin of votes lies within the interval [−h, +h]. In equation (1), yd,s,t is our measure of gender-based crimes in district d, in state s, in year t, while γs and λt denote state and year fixed effects. Xd,s,t is a vector of pre-determined district-covariates. The vector η captures the relationship between the running variable (margin of votes) and the outcome, absent treatment. In other words, it determines the

slope of the relationship between y and MV to the left of the cutoff. Finally, our vector of interest isβ k. In particular, the main parameter of interest isβ 0, which identifies the effect of electing a female Representative at the threshold, i.e., alocaleffect. Following the literature (Imbens and Lemieux, 2008; Broockman, 2014; Brollo and Troiano, 2016; Cattaneo et al., 2020), our main and preferred specification is a non-parametric local linear regression. Specifically, we estimate a polynomial of order 1 where h is chosen according to the procedure developed by Calonico et al. (2014b) and Calonico et al. (2014a), using MSE-optimal bandwidths. Since one district can contribute multiple mixed-gender elections over the years, standard errors are clustered at the district level. In this RD design, identification rests on two main assumptions. First, around the threshold MV= 0, randomness determines the election result. Second, candidates are not able to control the probability of falling on either side of the threshold; that is, they are not able to manipulate the outcome of an election. We refer the reader to Appendix C for a discussion of the validity of our approach. In particular, we show that pre-determined characteristics are balanced (see Table C1), and the McCrary test of discontinuity around the threshold indicates that there is no evidence of manipulation (see Figure C2). Recent work by Marshall (2022) shows that, unless additional strong assumptions are invoked, RD estimates may be asymptotically biased. In our setting, isolating the effect of gender would require that the gender of the candidate does not affect vote sharesorthat other predetermined potential confounders (e.g., competence, or social preferences) do not affect the outcome of interest (i.e., rapes). It is hard to argue that gender does not affect

the margins of victory. However, we claim that voters in the U.S. are not likely to select Congress Representatives to reduce GBV , which implies that no compensating differential affects the outcome of interest. Indeed, the Pew Research Center instrument to identify top voting issues in the U.S. Congressional elections does not include sexual abuse or gender issues in general as an option. Data from ANES (2019) also seem to support this claim.FRISANCHO ET AL. 8 When asked to identify the most important problem facing the U.S., overall, only around 13% of the respondents surveyed between 1970 and 2000 mention “public order", a category that includes crime, drugs, women’s rights, and gun control, among other issues. Still, we acknowledge the possibility that this assumption (i.e., compensating differentials induced by variation in gender do not affect the outcome of interest) may fail. Therefore, we interpret our estimates of β0 as a compound treatment effect that incorporates the effects of gender as well as the effects of all compensating differentials induced or altered by conditioning on close races (Marshall, 2022). We thus describe β0 as the causal effect of electing a female Representative as opposed to the causal effect of gender. For instance, one of the main differentials between female and male candidates in close races is their party affiliation. In this case, we conduct a robustness check with a similar RD design to evaluate the effect of party affiliation on rapes (see Section 4.1, and Table 2). Reassuringly, this placebo test yields null impacts, which suggests that the main estimate of the impact of electing a female Representative on rapes is not driven by party affiliation. Another potential challenge is highlighted in Eggers et al. (2015), who document the presence of an incumbent advantage in close election for the U.S. House of Representatives. This advantage raises concerns about the validity of the RD design since winners and

losers in close races may not be comparable. To address this issue, we follow the approach proposed in Eggers et al. (2015) and test for the presence of an incumbency advantage in our sample of mixed-gender races. We define a candidate as an incumbent if they are running for the party that holds the seat, i.e., a party