Disaggregating Exposure to Violence

Over the past 15 years, civil war scholars have increasingly turned from understanding the causes of war onset to understanding the micro-level impacts of violence on people’s social, political, and economic preferences and behaviors. Studies have shown that direct or indirect exposure to violence can increase prosocial behavior toward ingroup members but aggravates intergroup relations (Bauer et al. Reference Bauer, Blattman, Chytilová, Henrich, Miguel and Mitts2016; De Juan et al. Reference De Juan, Haass, Koos, Riaz and Tichelbaecker2024). Over time, the perceptions of rival outgroups have been shown to improve (Scacco and Warren Reference Scacco and Warren2018) but distrust and negative sentiments toward the government remain (De Juan and Pierskalla Reference De Juan and Pierskalla2016). Several mechanisms have been proposed to explain the observed relationships, such as post-traumatic growth and migration of asocial types (Gilligan, Pasquale, and Samii Reference Gilligan, Pasquale and Samii2013). However, we still know very little about the mechanisms that underpin the relationship between general exposure to violence and sociopolitical behaviors.

Studies often conceptualize and measure “violence exposure” as a simple binary measure of whether respondents reside inside conflict-affected areas (Yaylacı and Price Reference Yaylacı and Price2023). This overlooks the intricate variations in types of violence experienced during civil wars, which differ in terms of lethality, frequency, degrees of exposure, and personal significance (Gutiérrez-Sanín and Wood Reference Gutiérrez-Sanín and Wood2017).

While sexual violence is one form of violence against civilians among many that occur in violent conflicts, it is an extremely brutal and consequential human rights violation. The qualitative knowledge base suggests that it is important to disaggregate sexual violence from other types of violence to understand its impact on social and political behavior (Dumaine et al. Reference Dumaine, Nordås, Gargiulo and Wood2021; Gutiérrez-Sanín and Wood Reference Gutiérrez-Sanín and Wood2017).

The Harmful Legacies of Wartime Sexual Violence

The psychological and social consequences of wartime sexual violence are understood to be distinct from the consequences of other forms of violence. It is often nonlethal, and accompanied by a sense of shame from having been dominated by another individual (Eriksson Baaz and Stern Reference Eriksson Baaz and Stern2009) and the expectation that one’s community will blame the victim for their experience (Koos and Lindsey Reference Koos and Lindsey2022; Woldetsadik Reference Woldetsadik2018). Survivors often feel stigmatized and a sense of self-blame for violating social norms of fidelity, sexuality, honor, and reproduction. Thus, there is a direct harm associated with sexual violence as well as an acknowledged secondary trauma experienced by survivors, including social stigma and shame. These consequences have been described by victims as comparable to the experience of rape itself (Woldetsadik Reference Woldetsadik2018).

Dozens of qualitative studies have documented the pervasive social consequences of wartime sexual violence. Drawing on interviews with women in Mozambique, Sideris (Reference Sideris2003) traces how women who were raped and impregnated by Renamo rebels were rejected by their husbands and children born of rape were not accepted in their communities. In another study from post-genocide Rwanda, Mukamana and Brysiewicz (Reference Mukamana and Brysiewicz2008) document that rape victims were often labeled as prostitutes by their community members which exacerbated social marginalization and exclusion. Kelly et al. (Reference Kelly, Albutt, Kabanga, Anderson and VanRooyen2017) and Finnbakk and Nordås (Reference Finnbakk and Nordås2019) draw on interviews with rape survivors in DR Congo to describe victim experiences of social exclusion and collective suffering.Footnote ²

The majority of these studies were conducted with the aim of exposing and tracing continued harms, social ostracization and exclusion felt by survivors of sexual violence. However, they do not engage with questions about sociopolitical participation explicitly. They were also not designed to make statements about effects of sexual violence on the wider population of survivors since these studies primarily draw from small convenience samples rather than representative populations. Nonetheless, conventional wisdom tends to extrapolate from this literature that sexual violence will decrease sociopolitical engagement.

Wartime Sexual Violence and Survivor Mobilization

At the same time—and taking inspiration from earlier quantitative work that finds increased political participation after exposure to armed violence—new work highlights survivor agency and community mobilization in response to sexual violence.Footnote ³

Kreft (Reference Kreft2019) finds that countries affected by sexual violence have a higher likelihood of women’s protests. While informative, aggregate-level analyses like these cannot unpack sociopolitical consequences at the subnational or individual level. For example, Kreft’s (Reference Kreft2019) work does not speak to survivor mobilization in rural areas, but rather women in urban areas whose risk of wartime sexual violence is comparatively low.

Turning attention to research on survivors, Koos (Reference Koos2018) finds that sexual violence victimization increases prosocial behavior in post-war Sierra Leone. Yet, the survey measure of sexual violence exposure leaves questions about measurement and underreporting. In separate studies, González and Traunmüller (Reference González and Traunmüller2023) and Koos and Traunmüller (Reference Koos and Traunmüller2024) also find evidence of increased social activity among survivors of sexual violence in several contexts. The authors use a list experiment to analyze effects, but leave central questions open about whether and how known psycho-social processes associated with sexual violence—such as stigma and self-blame—relate to survivor mobilization.

While there is emerging quantitative evidence that wartime sexual violence may lead to increased social and political participation, this finding is not well established for survivors, their families and communities at large. Existing studies also do not engage with the psychological and ethnographic literature that demonstrates a range of harms that survivors and their families experience and which are—at first glance—at odds with the notion of sociopolitical mobilization. A full understanding of the effects of wartime sexual violence on social and political attitudes, preferences and behaviors among survivors and their families requires a new theory that rectifies these seemingly conflicting sets of evidence.

Theoretical Motivation

How does sexual violence by armed groups affect social and political action among survivors and their households? Given stigmatization of and shame felt by survivors, the most established view suggests that survivors and their families will be ostracized and excluded from sociopolitical participation in their communities, or at best ignored (Kelly et al. Reference Kelly, Albutt, Kabanga, Anderson and VanRooyen2017; Koos and Lindsey Reference Koos and Lindsey2022; Woldetsadik Reference Woldetsadik2018). This notion is also reinforced by the framework advanced by the United Nations and other organizations that sexual violence is a weapon of war that destroys and weakens the social fabric of communities (Amnesty International 2010; Médecins Sans Frontières 2021; United Nations Secretary-General 2014).

Indeed, sexual violence is often perpetrated with immense brutality leading to long-lasting physical and psychological effects not only for the survivors themselves but also for their families and community members. Survivors and their families often feel shame, even guilt, and experience stigmatization, which makes it challenging to reintegrate into communal life and foment social reconnections.

We offer a theory that builds on the new literature on survivor agency while also emphasizing the relevance of the described social harms for how survivors and their families respond and engage in their communities. The theory directly addresses a tension in the literature between research on survivor experiences of stigma and exclusion (Dumke et al. Reference Dumke, van der Haer, Koos and Hecker2021; Kelly et al. Reference Kelly, Albutt, Kabanga, Anderson and VanRooyen2017; Koos and Lindsey Reference Koos and Lindsey2022; Woldetsadik Reference Woldetsadik2018) and some early quantitative evidence on mobilization and participation (González and Traunmüller Reference González and Traunmüller2023; Koos Reference Koos2018; Kreft Reference Kreft2019). Rather than viewing these as competing understandings, we describe how psychosocial harms can beget social and political mobilization.

Our theory of survivor sociopolitical engagement centers on the different ways survivors and their families navigate the social harms that they experience. We first describe how people in weak, conflict-affected states primarily rely on their local communities, a background condition under which we expect our theory to operate. We then describe how psychosocial factors associated with sexual violence are essential mechanisms within our theory of survivor-driven mobilization.

Community Reliance and Access to Local Public Goods

The value of belonging to a community is an important condition for understanding people’s sociopolitical engagement in areas of insecurity and ongoing fighting. During violent conflict, the state rarely provides security, justice, or basic services. Rather, communities collaborate to generate local public goods to address pressing needs (Migdal Reference Migdal1988). Community acceptance is important for accessing these locally generated public goods, which in turn can incentivize stigmatized survivors and their households to seek connections with their communities.

For example, in DR Congo, traditional authorities within communities mete out justice for many within-community disputes. In addition, communities are often tasked with providing their own means of protection from armed groups and people band together to gain access to security information (Lindsey Reference Lindsey2022). Being integrated and networked within a community is essential for having access to such essential public goods; thus there are clear risks that accompany exclusion. The increased risk of exclusion provides survivors and their families an instrumental incentive to engage with their community despite and perhaps even because of the psychosocial harms that they anticipate.

Evidence suggests that people under severe duress have become more politically and socially engaged. In the case of Rwanda, Berry (Reference Berry2018) describes how the genocide-imposed need for women’s access to land to support their own lives and the lives of children drove women to lobby for those rights and achieve greater representation. Survival needs can thus be understood as both threatening to people’s well-being and politically mobilizing. In the same vein, we suggest that survivors and their families operate under the potential threat of exclusion from their communities and thus seek to offset this threat by engaging with and embedding themselves in their communities rather than retreating or remaining passive.

Psychosocial Factors Motivating Sociopolitical Action

In this theory, we suggest that sexual violence survivors and their families very often face a threat of social exclusion due to prevailing gender norms. However, some survivors and their families will perceive a greater risk of exclusion than others. We argue that this depends on several psychosocial factors.

Empirical Expectations

Summarizing the overlapping needs-based framework and psychosocial consequences that we believe push survivors and their families toward achieving reconnection with their communities, we propose the following directional hypothesis.

Empirically, we assess our hypothesis by examining the effects of sexual violence exposure within households on several measures of sociopolitical engagement. We then examine the plausibility of the psychosocial mechanisms described within our theory. We use an original survey conducted in eastern DR Congo. Before turning to the description of the survey design and results, we introduce the study context and discuss the existing evidence base on sexual violence and its social and political implications in this critical case.

List Experiments and Direct Questions

Measuring exposure to sexual violence at the individual and household level comes with significant measurement and ethical challenges. Many respondents who have been victims and fear stigmatization will choose not to report sexual violence when asked directly in a conventional survey question. Thus, conventional direct questions are most likely to underestimate prevalence rates of wartime sexual violence.

Moreover, nondisclosure to direct questions can be correlated with both psychological well-being such as shame and stigmatization but also social and political behavior. For instance, if survivors are socially engaged and this makes them more confident to report wartime sexual violence (to a direct question), there will be measurement bias toward the hypothesis that wartime sexual violence victims are more sociopolitically active. Alternatively, if survivors are socially engaged and this makes them less likely to report sexual violence exposure (to a direct question), the results are biased against the hypothesis that sexual violence survivors are more socially engaged and demonstrate resilience. Thus, it becomes vital to consider measures that account for nondisclosure bias baked into direct questions commonly used.

The list experiment is an alternative to uncover truths about behaviors, attitudes and experiences that respondents may prefer not to disclose in direct survey questions. List experiments assign each survey respondent randomly to a treatment or a control group. The control group receives several (often three) nonsensitive item questions carefully formulated such that few people will respond “yes” (ceiling) or “no” (floor effects) to all three items. The treatment group receives one additional, sensitive question item considered more intrusive or implicating than the other items in the list. Each respondent only answers the total amount of items to which the answer is “yes.” This gives anonymity to each respondent so that each feels less inhibited from revealing true answers to an unfamiliar enumerator. Estimates of the sensitive behavior are then calculated as the mean difference in the count of affirmed question items between the treatment and control groups.

This “statistical truth serum” (Glynn Reference Glynn2013) has been used to estimate a wide range of outcomes, including people’s true preferences for supporting insurgents in civil wars (Blair, Imai, and Lyall Reference Blair, Imai and Lyall2014), vote-buying by political parties (Corstange Reference Corstange2018), and intimate partner violence (Cullen Reference Cullen2023).Footnote ⁷

Our Two Survey Questions

Given the high probability of disclosure bias and our focus on sociopolitical mobilization, we used a list experiment to estimate levels of wartime sexual violence in our survey sample. We administered the list experiment in the survey as follows.

The other half of respondents were randomly assigned to the treatment condition in which the sensitive item “rape by armed groups” was added to the list above:

We emphasize that the validity of list experiments depends on a series of assumptions. First, respondents in control and treatment groups must not differ on average (randomization). Second, the inclusion of the sensitive item must not affect responses to the nonsensitive items (no design effect). Third, respondents must respond truthfully to the list experiment and must not strategically choose not to disclose.Footnote ¹⁰ Strategic nondisclosure can occur in the presence of ceiling (choosing all items) or floor effects (choosing no items) that threaten anonymity, the hallmark of list experiments. In Section 9 of the Supplementary Material, we describe these assumptions in detail and employ a variety of descriptive and statistical tests to assess the extent of design effects, ceiling and floor effects, strategic nondisclosure, and monotonicity (Aronow et al. Reference Aronow, Coppock, Crawford and Green2015; Blair and Imai Reference Blair and Imai2012; Buckley et al. Reference Buckley, Marquardt, Reuter and Tertytchnaya2024), and a simulation-based sensitivity analysis in which we test the robustness of our results to the inclusion of potential floor effects. Overall, we find no evidence that would threaten the validity of the list experiment estimate and the inferences we make in the analysis.

To validate our assumption about nondisclosure bias, we also asked respondents a direct question on wartime sexual violence. First, we compare the direct question to the estimated prevalence of the list experiment. Then, we use both measures in our empirical analysis to gain insights into mechanisms of sociopolitical mobilization.

In the direct question, all respondents were asked:

Household-Level Exposure

Both the list experiment and the direct question asked people whether they or someone in the household experienced wartime sexual violence, and not only the respondent. In addition to providing greater privacy to respondents, this approach captures the social and political implications at the household level, a highly relevant social unit in rural DR Congo.

Despite these advantages, there is likely some variation in whether respondents know about the victimization of other household members. Based on our impressions from field research in eastern DR Congo, there were two likely scenarios. First, sexual violence has been public, publicly known or so severe that household members had accurate information to respond to the question. Second, sexual violence has not been public so that survivors have chosen whether to share it with their spouses and family. Nondisclosure might allow the survivor to avoid negative repercussions such as stigmatization or spousal rejection. In our context, we would expect that surveyed survivors who did not disclose it to their partners would be unlikely to report their exposure in a direct question. However, they would be more likely to report under the conditions of privacy afforded by the list experiment. Overall, when someone has been affected personally or as a family, we expect it to have an impact on social and political attitudes and behavior of the household members collectively.

Prevalence Estimations in Eastern DR Congo

Figure 2 shows the estimated prevalence rates of sexual violence for both measures. The list experiment suggests that 12.4% of households in our population-based sample have experienced sexual violence by armed groups. By comparison, the direct question suggests a point estimate of only 6.4%. Consequently, a conventional psychological or public health survey with direct questions would have missed almost half of the share captured by the list experiment.Footnote ¹¹

Figure 2. Estimates of Direct and List Measures of Rape

Note: Bars represent 90% and 95% confidence intervals. Estimates reported from linear regression models. Difference in means and confidence intervals available in Table A1 in the Supplementary Material.

Because the list experiment adds random noise to provide anonymity and relies on a comparison between responses in the treatment group with responses in the control group, standard errors tend to be much larger than in direct questions. Thus, we estimate a large substantive but not statistically significant difference in estimates.Footnote ¹² Scholars have highlighted important trade-offs between reduced disclosure bias and increased variance in list experiments (Blair, Coppock, and Moor Reference Blair, Coppock and Moor2020) and often recommend prioritizing precision unless there are strong theoretical reasons to expect bias in disclosure. Our approach is to conduct all our analyses using both the list experiment and direct measures of sexual violence, which ultimately shows that disclosure bias is essential to account for when considering the effects of sexual violence on sociopolitical outcomes despite increased variance.Footnote ¹³

Although we cannot say how closely the estimate of 12.4% reflects the true value of household-level prevalence of wartime sexual violence in South Kivu, tip-of-the-iceberg arguments on reporting sexual violence suggest that higher estimates tend to better reflect true levels.Footnote ¹⁴ The results also align with previous literature that finds higher estimates of intimate partner violence against women in list experiment measures (Cullen Reference Cullen2023).

In subsequent analyses and discussion of our theoretical argument, we prioritize the list experiment measure, because (i) the literature prioritizes it as a measure of this sensitive subject, and (ii) the nature of our outcome variable (sociopolitical mobilization) is inherently social, as is survey disclosure (Lindsey Reference Lindsey2023). For scholars interested in comparing our household-level measures of sexual violence with event-based geospatial data, we provide correlations between our measures of sexual violence and geospatial data from two widely used, publicly available conflict event datasets in the Supplementary Material.Footnote ¹⁵

Model Specification

New statistical estimation procedures allow researchers to move beyond a pure difference-in-means estimate of list experiments to predict each individual’s likelihood of experiencing the sensitive item, in our case sexual violence, based on a set of covariates. Blair and Imai (Reference Blair and Imai2012) have offered a maximum likelihood estimation procedure that uses logistic regression to model the probability that a respondent has experienced a sensitive item or not. The procedure also accounts for the potential relationship between the covariates and the three nonsensitive items in the list experiment. The result is a more precise estimate of sexual violence with smaller confidence intervals, since each individual has a predicted value of experiencing sexual violence available for the analysis. It also appropriately models the sensitive item as a dichotomous outcome.

Imai, Park, and Greene (Reference Imai, Park and Greene2015) propose a one-step maximum likelihood estimation procedure that estimates the sensitive item at the individual level taking into account the covariate distribution and nonsensitive items (as described above), and then estimates the effect of the sensitive item—considering uncertainty bounds—on the outcomes. The modeling strategy uses the same covariates to predict sexual violence as those within our sociopolitical outcome models: gender, age, household size, other forms of violence experienced, and pre-exposure social exchange (described in the Section “Mitigating Threats to Causal Inference”).Footnote ¹⁶

We implement this one-step estimation procedure to analyze the effects of household exposure to sexual violence on sociopolitical outcomes using the ictreg.joint function in the R package “list” (Blair et al. Reference Blair, Imai, Park, Coppock, Chou and Blair2016; Imai, Park, and Greene Reference Imai, Park and Greene2015). This fully efficient one-step method requires making assumptions that there are no design effects and no strategic nondisclosers within the list experiment. We find no evidence for violations of these critical assumptions after conducting a range of state-of-the art statistical tests (see Section 9 of the Supplementary Material).Footnote ¹⁷

Outcome Variables

We test our hypotheses on several outcome variables that reflect various dimensions of local social and political mobilization and engagement. Personal exchange reflects the frequency of how often a respondent meets with people in their community (mean = 1.29, SD = 0.79, min = 0, max = 3). Event engagements measures how frequently a respondent participates in communal work such as road cleaning, maintenance of irrigation channels or providing labor for communal construction projects (mean = 1.47, SD = 0.938, min = 0, max = 4). Donate amount is a behavioral measure which records the amount respondents donated to a school project in South Kivu (mean = 297, SD = 303, min = 0, max = 5,000).Footnote ¹⁸ As local political engagement, we asked respondents whether they were leaders of a local association including local councils, farmers’ associations, savings groups, education and health committees (mean = 0.13, SD = 0.34, min = 0, max = 1). The membership variable measures the number of memberships instead of leadership roles in these organizations (mean = 0.54, SD = 0.83, min = 0, max = 4).Footnote ¹⁹

The models include conventional controls for gender, age, education level, household size, and assets. Importantly, we also added measures for witnessing sexual violence and having lost a family member, relative or friend due to violence. We control for these conflict experiences in order to convincingly disentangle the effect of sexual violence from other experiences of violence, which may co-occur but have different sociopolitical effects. We also control for the level of pre-exposure social exchange to reduce the risk of reverse causation and include territoire fixed effects (ADM2) to account for unobserved spatial heterogeneity.Footnote ²⁰

Mitigating Threats to Causal Inference

One important empirical limitation of our study is that we cannot claim to estimate causal effects. Causal claims are inherently difficult to make, in particular for our theory. Our approach is to transparently assess and mitigate risks to causal inference. While our list experiment measure relies on respondents’ random assignment to a treatment and control condition, the estimation of the effect of wartime sexual violence on social and political outcomes is correlational.

Our strategy is to plausibly reduce the main threats to causal inference, confounding bias and reverse causality. Apart from standard confounding variables such as gender, age, assets, and household size, our analyses also control for other types of violence (witnessing sexual violence, homicide in family) that could confound the effect of sexual violence. We also specify local-level fixed effects at the territoire level (ADM2) to control for unobserved heterogeneity. Importantly, as described above we include a measure of respondents’ pre-exposure level of social exchange through recall questions to reduce the potential risk of reverse causality. While acknowledging the limitations of this recall question, it is reassuring to see in the outcome models in Tables A17–A20 in the Supplementary Material that this variable previous exchange is indeed a significant predictor of sociopolitical engagement.

In addition to empirical risk reduction, various pieces of evidence indicate that wartime sexual violence is not systematically targeted. For instance, Bartels et al. (Reference Bartels, Scott, Mukwege, Lipton, VanRooyen and Leaning2010) examined 1,021 patient records from Panzi Hospital in Bukavu and found no distinct patterns, noting that it affected women of diverse backgrounds. Similarly, Imai et al. (Reference Imai, Keele, Tingley and Yamamoto2011) analyzed DHS data and concluded that reports of sexual violence were unrelated to individual characteristics.

As an additional source of information, our survey asked the chiefs and heads of the one hundred sampled villages in South Kivu which households they believed were usually targeted in attacks involving rape by armed groups. A total of 61% answered that it is “almost random,” 34% said “those along the road,” 18% agreed with “those near the center of the village,” whereas 21% answered “those far away from the center.” A total of 44% said that “rich households” are targeted and 12% said “poor households” are targeted. (Multiple responses were allowed.)

Finally, to further explore potential selection effects, we regressed the list experiment on education and wealth. We find that these measures are not significantly associated with sexual violence victimization measured with the list experiment measure or the direct measure of sexual violence (Tables A5 and A6 in the Supplementary Material).

Approach to Analysis

Given strong modeling assumptions (such as sequential ignorability) required for formal mediation analysis (Green, Ha, and Bullock Reference Green, Ha and Bullock2010; Imai et al. Reference Imai, Keele, Tingley and Yamamoto2011) and the complexity of our modeling approach for using the list experiment as an explanatory variable, we assess the plausibility of our mechanisms by examining two sets of correlations. Recall our consistent finding that the list experiment measure, but not the direct measure, revealed that sexual violence affected sociopolitical outcomes. This suggests that the positive relationship between sexual violence and sociopolitical engagement is primarily driven by the share of respondents who disclosed to the list experiment but not to the direct question. Based on this understanding, we argue that the plausibility of a mechanism is strengthened if we find both a positive and significant correlation between the list experiment and the proposed mechanism; and no correlation between the direct question and the proposed mechanism. Our approach thus leads us to pay substantive attention to the determinants of respondent disclosure of sexual violence in surveys.

Measurement of Mechanisms

The first mechanism we have emphasized in our theory is the role of stigma. While many theories suggest that stigma leads to ostracization and exclusion, we suggested—building on social psychological research (Molden and Maner Reference Molden and Maner2013; Shih Reference Shih2004)—that it leads survivors and their families to (re)engage, (re)connect and embed themselves in their communities to reduce the likelihood of exclusion. We measure anticipated stigma with the standardized mean of the sum of three binary questions that ask whether respondents are fearful of being gossiped about, verbally insulted or harassed, and physically harassed or threatened (mean = 0.32, SD = 0.37, min = 0, max =1).Footnote ²³

A second psychosocial mechanism that we have emphasized in our theory is self-blame. When people blame themselves they feel that they deserve to be ostracized, will sense a greater likelihood of being ostracized, and will thus seek reconnection with their community to mitigate this perceived risk of exclusion. In weak, conflict-affected contexts, self-blame may therefore promote mobilization and engagement. This mechanism resonates with psychological studies that have shown self-blame to be associated with self-efficacy and perceived control over situations (Foster, Matheson, and Poole Reference Foster, Matheson and Poole1994). We measure self-blame as the standardized mean of five binary questions including feeling ashamed, feeling guilty, blaming oneself, low self-esteem, and feelings of self-punishment (mean = 0.30, SD = 0.33, min = 0, max = 1).

Our third psychosocial mechanism is often posited in the literature as a mechanism for increased political participation but rarely examined empirically. Post-traumatic growth is a psychosocial response that drives survivors and their families to engage with their communities because of a change in their outlook on life as a result of traumas experienced. To measure, post-traumatic growth, we draw on a 10-item battery including, for instance, changed life priorities, understanding of spiritual matters, closeness with others, and ability to handle difficulties (mean = 1.51, SD = 0.70, min = 0, max = 3).Footnote ²⁴

Expectations and Results

Our theory emphasized that anticipated stigmatization, self-blame, and post-traumatic growth lead sexual violence survivors to further embed themselves within their communities by engaging in sociopolitical action. Therefore, as discussed above, we expect a positive and significant correlation between these three factors and the list experiment. Figure 4 shows the results of the correlations between each of the described mechanisms, the list experiment measure of sexual violence exposure, and the direct measure of sexual violence exposure. The correlations are derived from a full model with control variables and territoire fixed effects as implemented in our main analyses.Footnote ²⁵ We describe how we interpret the relationships for each variable in turn.

Figure 4. Psychosocial Correlates of Wartime Sexual Violence

Note: Bars represent 90% and 95% confidence intervals. Correlate estimates from linear models of normalized outcomes. Full model specifications provided in Tables A7, A9, and A11 in the Supplementary Material.

Sexual violence-affected households (as measured in the list experiment) have significantly higher levels of anticipated stigma (+0.64 standard deviations, $ \pm 0.02 $ standard errors) and self-blame (+0.54 SD, $ \pm 0.02 $ SE). Combined with the finding that the list experiment measure of sexual violence is associated with increased sociopolitical mobilization, these correlations lend plausibility to our claim that stigma and self-blame are mechanisms that lead people to re-engage rather than become or remain ostracized.

This interpretation is further strengthened by the absence of a correlation between the direct measure of sexual violence and anticipated stigma and self-blame, respectively. Again, this is based on our understanding that the share of respondents who disclose sexual violence in the list experiment but not in the direct question appears to drive the observed relationship between sexual violence and sociopolitical mobilization in our main analysis.

Survivors and their families who disclose sexual violence in the direct measure of sexual violence are neither participating more in their communities nor anticipating stigma nor blaming themselves at higher levels than average. Rather, survivors and their families that choose not to disclose their experiences in the direct measure (likely due to the presence of stigma and feelings of self-blame) but do disclose in the list experiment measure are both feeling stigmatized and blame themselves—and as a result engage more in their communities politically and socially to mitigate risk of exclusion.

Our argument has also suggested post-traumatic growth as a potential psychosocial mechanism that can lead people to engage socially and politically (Bauer et al. Reference Bauer, Blattman, Chytilová, Henrich, Miguel and Mitts2016). However, post-traumatic growth is not supported as a mechanism. It is neither correlated with the list experiment measure of sexual violence nor the direct measure of sexual violence. The absence of difference in disclosure in the list experiment and the direct question is theoretically logical because changed outlook on life is not plausibly related to disclosure or nondisclosure (+0.07 SD, $ \pm 0.05 $ SE).

We acknowledge that psychosocial factors are not the only mechanisms that can lead survivors and their families to engage more with their communities. One important implication of extreme violence such as sexual violence by armed actors is that people often flee and seek refuge in safer areas. Displacement (mean = 0.608, SD = 0.488, min = 0, max = 1) may make it more important to connect with the social communities in which one lives because of broken ties and to ensure basic needs for one’s family. Accordingly, we find in Figure 5 that having been displaced is strongly associated with sexual violence as measured in the list experiment but not in the direct question—suggesting its importance for understanding sociopolitical mobilization. Again, this pattern lends suggestive support to the notion that displacement and the need to connect to new and old social networks may partially explain the relationship between wartime sexual violence and sociopolitical mobilization.

Figure 5. Other Correlates of Wartime Sexual Violence

Note: Bars represent 90% and 95% confidence intervals. Correlate estimates from linear models of the dichotomous outcome. Full model specifications provided in Table A13 in the Supplementary Material.

Overall, the evidence presented here resonates with our theoretical argument and the proposed mechanisms, stigma and self-blame, that drive sociopolitical engagement and mobilization. We emphasize however, that these mechanisms suggest a mode of mobilization that fails to challenge the very gender norms that perpetuate victim-blaming and stigmatization.