Exposure to Armed Conflict and HIV Risk Among Rwandan Women

Abstract

This article focuses on the link between past exposure to violence and a critical public health issue in sub-Saharan Africa: HIV-positive status in women of reproductive age. Specifically, we use biosocial data from the Rwandan Demographic and Health Survey (2005‒2014) to assess how the timing and intensity of women's exposure to the war and genocide in Rwanda (1990‒1994) may be associated with their HIV status. We find significant differences in risk across age cohorts, with the late adolescence cohort (women born in 1970‒1974, who were aged 16‒20 at the start of the conflict) having the highest risk of being HIV positive 10‒20 years after the violence, even after controlling for current socioeconomic and demographic characteristics. Women who reported two or more sibling deaths, excluding those related to maternal mortality, during the conflict years also had higher odds of being HIV positive, net of cohort and control variables. Age at first sexual intercourse and number of lifetime sexual partners partially—but not fully—explain the associations between cohort and sibling deaths and HIV. These findings advance research related to armed conflict and population health and indicate that experiencing conflict during key stages of the life course and at higher intensity may affect women's long-term sexual health.

Introduction

HIV remains a major threat to population health, especially in low- and middle-income countries (De Lay et al. 2021). Sub-Saharan Africa (SSA) has the highest HIV rates in the world, accounting for 71% of HIV infections (Kharsany and Karim 2016). Young women are particularly at high risk of infection, compared with both older women and men, and while HIV can be acquired over the life course through multiple modes of transmission, heterosexual sex is the leading cause of infection among women in SSA (Kharsany and Karim 2016; Melesse et al. 2020; Ramjee and Daniels 2013). Women in this region also care for HIV-positive children at higher rates than men, making HIV risk among women of reproductive age a pressing population health and development concern (Feyissa et al. 2019; Kharsany and Karim 2016). Notably, SSA also has some of the world's highest rates of armed conflict (UCDP 2022). While researchers have hypothesized a potential link between conflict and the spread of HIV, findings to date are mixed (Anema et al. 2008; Bennett et al. 2015; Spiegel et al. 2007). Moreover, there is a significant gap in research on how conflict may affect HIV rates among women, despite documentation that women are typically at high risk of sexual violence during times of conflict (Bennett et al. 2015; Cohen and Nordås 2014).

To begin to address this gap, we focus on the violent conflict in Rwanda (1990–1994). Rwanda experienced one of the deadliest episodes of violence of the twentieth century—a conflict that continues to affect the lives of Rwandans today. While there are myriad effects of such catastrophic violence, we focus on those affecting Rwandan women who lived through the 1990 civil war and the 1994 genocide and, specifically, examine the link between exposure to the violence and HIV status. Although the genocide targeted Rwandan Tutsi, women in the Hutu majority were also exposed to violence during the civil war and genocide (Straus 2019). This violence was widespread throughout the country, although it was more intense in some regions (Nyseth Brehm 2017), suggesting that there was variation in women's direct exposure to violence that can be used to assess the link between conflict and HIV.

In assessing the association between conflict and HIV in Rwandan women, we characterize women's exposure to the conflict in two main ways: timing and intensity. By timing, we refer to the ages at which women experienced the conflict. To do so, we rely on HIV biomarker data gathered by the Demographic and Health Survey (DHS) from Rwandan women in 2005‒2014. We group women by cohort, or five-year age groups that are related to the length of the conflict and key life course stages. The timing of the Rwanda DHS HIV module and age range of the sample women (15‒49 years) provide us with variation in the ages of exposure and include key life course stages of early/middle childhood, early/late adolescence, and entry into/early adulthood. The DHS also includes women's lifetime reports of the years and number of sibling deaths, which we use to construct a measure of intensity of exposure to conflict in line with prior research (Kraehnert et al. 2019). We further assess whether timing and intensity of exposure are associated with HIV status through two key pathways: age at first sexual intercourse and lifetime number of sexual partners.

We find that both timing and intensity of exposure to conflict are associated with individuals' later HIV status among Rwandan women of reproductive age, and these associations are partially mediated by age at first intercourse and number of sexual partners. While our study focuses on the Rwandan context, our findings also inform research on the links between conflict and HIV, as well as broader research on violence exposure and population health. Our approach likewise underscores the potential for HIV prevention and treatment policies to target women whose life course experiences with conflict put them at high HIV risk.

Background

HIV and Armed Conflict in SSA

Much existing research on conflict and HIV has focused on cross-country comparisons of conflicts and HIV in the broad region of SSA, typically resulting in mixed findings. Spiegel and colleagues (2007) reviewed a diverse set of 65 studies from seven countries and concluded that little evidence supports the hypothesis that conflict increases HIV infections in SSA. In line with this, Anema and coauthors' (2008) simulation models predicted marginal increases in national HIV prevalence rates, even when rape and HIV rates were assumed to be high. Other studies—using national-level survey data on conflict and HIV rates in multiple SSA countries—have indicated a potential positive association between conflict and HIV rates (Bennett et al. 2015; Iqbal and Zorn 2010), and simulation models from another study suggest that mass rape likely increased HIV incidence rates (new cases) in Rwanda, as well as in other conflict-affected SSA countries (Supervie et al. 2010).

While we cannot reconcile these divergent findings, important differences may exist depending on the type of conflict (e.g., international or internal) and other country-specific characteristics. For example, internal conflicts (like the conflict in Rwanda) have been linked with higher HIV rates only when accounting for country-specific characteristics, such as economic openness and the number of refugees (Iqbal and Zorn 2010). Bennett and colleagues (2015) further suggested that the five years prior to violent conflicts in SSA may be an important period to study, as their study documented that HIV incidence rates increased more in the five years preceding a conflict than during the conflict years. Iqbal and Zorn (2010) argued that the lack of evidence of an increase in HIV rates during conflict years may be because of several other factors as well, including but not limited to the potential error in estimating HIV infections during times of conflict and the use of broad conflict measures across countries that do not capture the specific geographies of conflict and HIV incidence in a given country (Bennett et al. 2015). They suggested that more research is needed to make concrete conclusions about the potential for conflict to affect HIV rates, and that subsequent research of specific populations (e.g., women) is particularly important to advance research in this area.

HIV and Conflict in Rwanda

The HIV/AIDS epidemic reached East Africa by the early 1980s. In 1986, Rwandan national HIV prevalence rates were estimated at 18% in urban areas and 1.3% in rural areas, with particularly high rates among women and adults aged 26‒40 (Rwandan HIV Seroprevalence Study Group 1989). The National AIDS Program was swiftly established in 1987, although little progress was made in the subsequent years, in large part because of the onset of widespread conflict in the country and the resulting collapse of the health care system.

On October 1, 1990, the Rwandan Patriotic Front's (RPF) Tutsi-led army invaded Rwanda from the Ugandan border, challenging the Hutu-led government.¹ After several years of civil war skirmishes, peace negotiations in August 1993 resulted in a fragile ceasefire amid high tensions and fears. However, on April 6, 1994, the plane carrying the Rwandan president was shot down by unknown assailants, beginning a genocide targeting Tutsi with hundreds of thousands of civilians engaged in violence that unfolded within neighborhoods (Nyseth Brehm 2017; Straus 2006). The RPF's army reinitiated the civil war the next day, ultimately toppling the government in July 1994, bringing an end to the genocidal violence.

Throughout this conflict, numerous forms of violence occurred simultaneously. Although the genocide of Tutsi resulted in many more deaths than the civil war, Hutu were also killed between 1990 and 1994,² and estimates suggest that over a million people perished (Lemogne et al. 2013; Verpoorten 2020). Moreover, rape was pervasive (Bijleveld et al. 2009; Des Forges 1999; Mullins 2009), and purposeful HIV transmission was used as a weapon (Nduwimana 2004). Furthermore, national military officials and international peacekeeping forces offered some women food, clothing, shelter, and protection from attack through partnership, but because these men often had numerous partners, such partnership may have further increased the risk of HIV among women during these years (Burnet 2012a, 2012b; Nduwimana 2004). Refugee populations fleeing to camps in nearby countries also faced harsh conditions in which women were at high risk of rape and other gender-based violence (Adelman 2002).³

Under these violent conditions, along with the collapse of HIV prevention and other health services, it has been assumed that the Rwandan conflict (like others in SSA) resulted in increased HIV prevalence and incidence rates—during the violence, as well as after the violence ended (Anema et al. 2008). However, no studies have examined the link between the violent conflict in Rwanda and the risk of HIV among Rwandan women. Given the continued importance of violent conflict and HIV as critical public health issues in the SSA region, particularly for women, we aim to further research in this area by focusing on women's exposure to conflict in terms of timing and intensity via a case study of Rwanda.

Conceptual Framework and Research Questions

We draw on life course theory to postulate the links between violent conflict and HIV. In doing so, we address three key research questions: (1) How does the timing of conflict exposure relate to HIV risk among women of reproductive age? (2) How does the intensity of conflict exposure relate to HIV risk among women of reproductive age? (3) Do sexual risk behaviors (age at first intercourse, number of sexual partners) explain associations between timing and intensity of conflict exposure and HIV status?

In developing our first research question, we hypothesize that associations between conflict and HIV risk may depend on women's ages at exposure to conflict. We conceptualize age at exposure by cohorts, or groups of women who were in key age groups at the time of the violence. Women born in closely spaced years, often referred to as “birth cohorts,” live through similar social and historical events at the same ages (Reither et al. 2009). In this case, women in the same birth cohort lived through the conflict—as well as preconflict and postconflict contexts—at similar ages.

Focusing on HIV as a sexually transmitted infection, we posit that women exposed to the armed conflict in adolescence and when entering adulthood will be at highest risk for HIV infection; we consider “entering adulthood” as the period that includes women's median age at first intercourse and first union (near 20 for both events) (Barrère et al. 1994). Armed conflict has been linked with earlier and premarital first sex among Rwandan women (Lindskog 2016), which would increase their exposure to sexually transmitted HIV. Furthermore, during the conflict, young women were targets of sexualized violence and purposeful HIV transmission (Burnet 2012b; de Brouwer et al. 2009). Recent research related to conflict and HIV in SSA found that the years leading up to conflict are characterized by increasing HIV incidence rates (Bennett et al. 2015). This suggests potentially higher HIV rates at the start of the conflict in Rwanda than in the prior years. As a result, those exposed to armed conflict in late adolescence/entering adulthood may have initiated their sexual debut earlier and under riskier conditions, increasing their lifetime risk of HIV.

These women also experienced the postconflict period during early adulthood (ages 21‒25). Following the violence, there were fewer men than women, as indicated by national data demonstrating a drop in the sex ratio from 95.7 in 1978 to 95.1 in 1991 and 91.3 in 2002 (National Institute of Statistics of Rwanda 2014), with the latter drop being attributed largely to the high level of male deaths during the conflict. These lower sex ratios—in combination with the need for women survivors to form sexual or marital partnerships as an economic survival strategy—have been associated with increased risk of women entering lower quality and higher risk partnerships in the postconflict years (Dude 2011; La Mattina 2017). Research conducted before the conflict indicated high levels of sexual coercion (33%) among women in stable partnerships and that such coercion was associated with higher risk of women testing positive for HIV (van der Straten et al. 1998). This suggests that relationship violence—exacerbated by the conflict and the quality of postconflict partnerships—may be a main predictor of women's risk of HIV. This may have been particularly true for women in their early 20s—the normative ages at marriage and first sex—during and after the conflict. While we characterize the conflict as occurring between 1990 and 1994, a cohort approach allows us to capture risks that women in different cohorts faced in other key years, such as the immediate pre- and postconflict environments.

Our second research question asks whether intensity of exposure to conflict affects HIV risk. Women with more intense exposure may have been more likely to be direct victims of sexual violence or exposed more generally to trauma, which can lead to long-term psychosocial and physiological changes (e.g., immunosuppression) that affect both behaviors and biological susceptibility to sexually transmitted infections, including HIV (Hillis et al. 2001; Lindskog 2016; Tsuyuki et al. 2019).

We build on past literature by using sibling deaths during the conflict as an indicator of the intensity of violent exposure (Jayaraman et al. 2009; Kraehnert et al. 2019). While Kraehnert et al. (2019) found that experiencing at least one sibling death during the Rwandan conflict was tied to reduced postconflict fertility, they did not consider health outcomes. We consequently assess associations between sibling deaths during the Rwandan conflict and women's HIV status. We categorize the intensity of exposure as sibling death categories (zero, one, and two or more), hypothesizing the strongest association between HIV and high intensity of conflict exposure (i.e., multiple sibling deaths vs. one or none). We exclude maternal mortality causes of death for sibling deaths to more precisely approximate exposure to violence.

As an extension of our first and second research questions, we also assess whether there is evidence of an interaction between timing and intensity of exposure to conflict. Specifically, we consider if the association between cohort and HIV differs by number of sibling deaths and, correspondingly, whether the association between sibling deaths and HIV differs by age cohort. We do so to clarify if the associations of timing and intensity of exposure with HIV work independently (i.e., additively) or if they work together in some way to increase the risk of HIV for specific groups of women by both cohort and sibling death exposures.

Our third research question asks whether key sexual risk behaviors explain the associations between timing and level of conflict exposure and HIV status. Research in the United States has indicated that adverse events experienced earlier in the life course reduce women's age at first intercourse and increase their number of sexual partners—two key pathways to increased risk of HIV (Hillis et al. 2001). Additionally, a recent systematic review conducted on the continent of Africa—including the SSA region—indicates that age at first intercourse and number of sexual partners are key predictors of HIV among adolescents and young adults (Bossonario et al. 2022). Thus, women's experiences of sexual violence, their exposure to trauma more broadly, and their consequential partnerships may all contribute to higher risk of HIV through early ages at first sex and higher numbers of lifetime sexual partners.

Taken together, addressing these research questions will advance our understanding of how armed conflict and its aftermath are associated with women's risk of HIV. Unlike past studies of conflict effects on national HIV rates, we focus on a critical population: women of reproductive age. We further specify cohort (timing) and intensity of conflict exposure in predicting women's risk of acquiring HIV, as well as earlier age at first intercourse and higher number of sexual partners as key mechanisms. Importantly, we use high-quality biomarker data that are nationally representative of women of reproductive age—a group at high-risk of HIV in SSA, even in nonconflict settings.

Methods

Data and Sample

To test our hypotheses, we use Rwandan DHS data from 2005, 2010, and 2014. The DHS is a cross-sectional household survey that is nationally representative of women of reproductive age (15‒49 years); the survey is implemented every five years in low- and middle-income countries. Households are selected using a stratified, two-stage cluster design, and a new cross section of women is selected using the same sampling process for each wave. In each household, all women of reproductive age are interviewed. The interviews include questions regarding fertility, reproductive health, nutrition, socioeconomic status, and household demographics using the same format and questions across waves; however, only some modules are collected in select waves, including the HIV biomarker data, which were collected only in 2005, 2010, and 2014.

We pool the 2005, 2010, and 2014 data, resulting in a total sample of 19,364 Rwandan women with HIV biomarker data. We limit the sample to women with a sibling (because of our interest in sibling deaths), resulting in the removal of 286 women without a sibling. We use case-wise deletion for the small amount of missing data (for seven women) on independent variables. Our final analytic sample is consequently 19,071 women. Notably, women in the final sample do not differ significantly from women removed from the sample across such key variables as HIV status, mean age, and mean age in 1990 (age at exposure).

For both descriptive statistics and regression models, we use survey weights (in this case, the HIV weight). These weights are calculated by DHS for each wave to represent the population of Rwandan women aged 15‒49 for a particular year. Sample weights are produced using the sample selection probabilities of each household and the response rates for households and for individuals for each year (Rutstein and Johnson 2004). In all analyses, we assign women the survey weight that is relevant for the year they participated in the survey.

Measures

Our dependent variable is HIV status as indicated by the HIV biomarker provided in the DHS. Survey enumerators asked participants to provide blood spots, which were dried and later analyzed for HIV. Across all surveys, response rates for the HIV biomarker were almost 100%—rates that are notably higher than those in other DHS surveys (Parkhurst 2010). We measure the outcome as a dummy variable indicating a positive (1) or negative (0) test result for HIV.

Our first independent variable of interest is cohort. To assess the significance of cohort effects, we construct measures of five-year birth cohorts using a woman's date of birth.⁴ Two key confounders of our cohort effects are women's current age and the survey year (or period) at which they participated in the study. Age trends indicate that HIV prevalence rates increase with age among women of reproductive age in Rwanda, although they drop slightly for the oldest age group (National Institute of Statistics of Rwanda 2014). Thus, we control for age and age squared in the models to account for this nonlinearity. We also consider period effects, which are indicated by the survey year. Past research has indicated that the best modeling technique for assessing age, period, and cohort effects in the same model is to use a modeling strategy in which period and cohort are modeled as random effects and age is included as an independent variable, because of the collinearity across these three measures of time.

In creating our cohort measure, we use five-year groupings of women based on their birth years. We do this for several reasons. Five-year age groups allow us to distinguish among important life course stages—such as early adolescence, late adolescence, early adulthood, and middle adulthood (in addition to early and middle childhood)—to better assess HIV risk when exposed to conflict during one or more of these stages. Wider cohort intervals (e.g., 10-year groups) would obscure potentially important differences across these stages when women have different probabilities of experiencing key sexual events, such as first intercourse and first union. Smaller cohort intervals (e.g., every two years) would have little meaning given that they are too narrow to reflect key stages of women's development and would also lead to much smaller groups of women per cohort, potentially amplifying the effects of measurement error for birth year. In addition, the five-year time period captures the years of the conflict, 1990‒1994. The five-year cohorts can consequently be used to reflect ages at exposure to the conflict better than wider ranges, which could include both exposure and nonexposure years. Thus, the five-year intervals allow for better identification of the timing of exposure risk patterns. Finally, this cohort measure allows us to follow methodological research on age‒period‒cohort (APC) models, which recommends using shorter rather than longer intervals (Tarone and Chu 1996; Yang 2008; Yang and Land 2008).

Our second independent variable of interest is the number of sibling deaths experienced during the conflict, excluding pregnancy and birth complication causes (i.e., maternal mortality), following past research in this area (Kraehnert et al. 2019). While Kraehnert and colleagues dichotomized any sibling death (i.e., one vs. none), we categorize our variable into none, one, or multiple sibling deaths. This allows us to more precisely measure higher levels of intensity of exposure to conflict with potential behavioral, psychological, and physiological consequences related to the risk of contracting HIV.

After assessing the fit of the base models, we add variables measuring the current sociodemographic conditions of the women in the sample: household size (number of people living in the home), women's educational attainment (number of years), marital status (married/cohabiting, never married/never cohabiting, or separated/divorced/widowed), and a dummy variable for whether the household is in Kigali (the capital and urban center of the country). In addition, to operationalize household wealth, we use the DHS household wealth categorical variable. In developing the wealth score, the DHS employs a weighted index of household assets using principal components analysis and categorizes the resulting scores as quintiles from poorest (quintile 1) to richest (quintile 5) (Rutstein and Johnson 2004). Controlling for these factors allows us to assess cohort and sibling death effects net of sociodemographic characteristics to better isolate the impact of past exposure to conflict for women of comparable characteristics.⁵

In addition to the cohort and sibling deaths effects, we are interested in whether the relationship between exposure to violence and HIV is mediated by two key variables: age at first sexual intercourse and number of lifetime sexual partners. Age at first intercourse is measured as a continuous variable (a range of 6–35 years). To measure lifetime number of sexual partners, the DHS asks each woman who reported ever having had sex how many different people they have had sexual relations with throughout their lives (top-coded at 95). Given the small number of women reporting more than two partners (less than 9%), and following past research, we categorize this measure as one, two, or three or more partners (Yaya and Bishwajit 2018).

Statistical Analysis

Our first step is to determine which model—APC (age‒period‒cohort), AC (age‒cohort), or AP (age‒period)—is the best fit. These models include cohort and period (year) measures as random effects to allow age, period, and cohort to be estimated independently given that they are often highly correlated. In this case, the highest correlation is between maternal age and cohort (r = .9). We employ hierarchical logistic regression models that include cohorts and period as random effects, and sibling death categories and control variables as fixed effects. We use logistic regression because of the dichotomous nature of our outcome: HIV positive or not. We compare APC, AC, and AP models using the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) scores for model fit (Yang 2008; Yang and Land 2008) and by assessing the significance of the cohort and period random effects. The model fit comparisons—included in Table A2 in the online appendix—indicate that the AC (age‒cohort) model is the preferred model with slightly better fit statistics, and because the cohort, but not period, random effect was significant.

for $i=1, 2, . . . , nj$ individuals within cohort j (⁠$j=1, . . . , 9$ birth cohorts), where, within each birth cohort j, respondent i's probability of being HIV positive is modeled as a function of age, age squared, sibling deaths (excluding deaths related to pregnancy or birth complication), and control variables. $β0j$ is the intercept, or the probability of being HIV positive for individuals in cohort j; $β1−5$ are the Level-1 fixed effects; $eij$ is the random individual effects and is distributed with mean of 0 and a variance of $σ2$⁠; $γ0$ is the model intercept; and $μ0j$ is the residual random effect of birth cohort j, assumed to be normally distributed with a mean of 0 and variance of $τμ$⁠. We center all continuous variables at their means to aid with the interpretation of the intercept and random effects. In doing so, the intercept represents the probability of being HIV positive when all continuous variables are at their mean values, and the random effect represents the average probability of being HIV positive across cohorts. The glmmTMB package used to create the models allowed us to compute confidence intervals for the random effect to assess statistical significance at the p < .05 level. We also calculated predicted probabilities of being HIV positive for each cohort (see Figure 2 and Table A3 in the online appendix). Sibling death variables are interpreted in terms of odds ratios, comparing those experiencing one or two or more deaths with the reference group (no sibling deaths in 1990‒1994). After estimating the base model, we sequentially add our hypothesized mediating variables—age at first intercourse and number of sexual partners—as fixed effects.

We also use supplementary regression models, adding interaction effects between cohort and sibling death variables, to assess the potential moderation of sibling deaths by cohort in predicting HIV-positive status. For these models, we use logistic regression and model cohort as fixed (dummy) variables rather than as a random effect. When modeled as dummy variables, the cohort variables can be estimated with less concern for multicollinearity with age (r = .4, on average). Regarding mediation, we conduct supplemental formal mediation tests using the medeff command in Stata (Hicks and Tingley 2011) and provide the full results of these tests in online appendix Table A4.

Results

Descriptive Results

Table 1 shows the descriptive statistics for the full sample (N = 19,071) and for the sample of women who reported having ever had intercourse (n = 13,184). According to the confidence intervals, the percentage of women who tested HIV positive is significantly higher in the latter sample (4.8% vs. 3.6%). While the latter data are relatively equally distributed across the DHS survey years—with 28% from 2005, 35% from 2010, and 37% from 2014—there is a significantly smaller proportion of women in the youngest cohort in the sexually experienced compared with the full sample (15.6% vs. 19.4%). The number of sibling deaths differs significantly between the samples, with 10.4% of women in the sexually experienced sample and 8.5% of women in the full sample reporting two or more deaths. Given these differences, we conduct regression analyses using the full sample before analyzing the proposed mediators that necessarily limit the sample to those who ever had sex.

At the start of the conflict, 8% of women in the full sample were young adults (aged 21‒25), 11% were in late adolescence (16‒20), and 12% were in early adolescence (11‒15); these percentages were slightly higher in the sexually experienced sample, reflecting the higher mean age of that sample. Given our hypotheses about young women being at higher HIV risk because of exposure to direct violence (e.g., rape, forced sexual partnerships) during the conflict, it is important that we have sufficient sample size in these groups.

Regarding intensity of exposure to conflict, 75% of women in the full sample reported no sibling loss due to nonmaternal death causes during the conflict years; 17% lost one sibling and 9% lost two or more siblings during these years. The mean values of our mediating variables of interest indicate that mean age at first intercourse was almost 20 years and ranged from six to 35 years. Most women reported only one sexual partner in their lifetime to date (71%), while 21% reported two, and 9% reported three or more.

The sociodemographic characteristics for the full sample indicate that the mean education level was 4.5 years, while the mean household size was 5.4 people. Approximately 50% of women were currently in a union (married or cohabiting), and 38% had never been in a union (in the sample of women who had ever had sex, 13%). Approximately 20% of the sample lived in the capital city of Kigali, although we do not know where they lived during the conflict years.

Figure 1 shows the unadjusted, unweighted percentage of women who were HIV positive by our key independent variables of cohort and number of sibling deaths. Based on these unadjusted means, the HIV rate was largest for women aged 16‒20 in 1990 (8.1%) and, secondly, for those who had experienced two or more sibling deaths (7.7%).

Regression Results

Table 2 presents the results of our multilevel logistic regression models. Model 1 illustrates the results of the age‒cohort model with intensity of exposure (sibling deaths) using the full sample. Models 2 and 3 use the sample of sexually experienced women and sequentially add the potential mediating variables of age at first intercourse and number of sexual partners. For ease of interpretation, we show logged coefficients (odds ratios) to interpret the fixed effects (i.e., control variables, intensity of exposure, and mediating variables). The timing of exposure effect is the variance of the cohort random effect (variance across cohorts). We interpret this effect by calculating predicted probabilities for each cohort, indicated in Figure 1 and Table A3 (online appendix).

In addressing our first research question—whether timing of exposure to conflict matters for HIV risk among Rwandan women—we note the statistically significant cohort random effect when controlling for age and age squared, as well as sibling deaths and control variables (Table 2, Model 1). While this indicates significant differences across cohorts, we calculate predicted probabilities to interpret this effect and identify which cohorts may be at higher risk than others. Figure 2 shows the predicted probabilities of being HIV positive by cohort for each model estimated. These adjusted probabilities are lower than the descriptive average percentage of HIV-positive status; however, consistent with the descriptive data, women aged 16‒20 at the start of the conflict have the highest probability of HIV-positive status, at 2.4%.

Next, we consider how intensity of exposure is associated with HIV-positive status by assessing the size and significance of the sibling death variables in Model 1. Again, consistent with the unadjusted means, we find little difference (in terms of size and significance) for HIV-positive status between those experiencing no or one sibling death during the conflict. However, those women who reported two or more sibling deaths were approximately 36% more likely to be HIV positive than those who experienced no sibling deaths (OR = 1.36). Model 1 also indicates that the timing and intensity of exposure effects are additive, with both remaining significantly associated with HIV-positive status independently.

Table 3 presents the results of our models including all main effects, control variables, and the interaction between cohort and sibling variables. In Model 1, we compare the 16‒20 cohort against all other cohorts by including only the 16‒20 cohort and its interaction with one and two or more sibling deaths. In this interacted model, the cohort variable for ages 16‒20 indicates that women in this cohort who reported no sibling deaths during 1990‒1994 have 42% higher odds of being HIV positive compared with women in all other cohorts who also reported no sibling deaths. The interaction terms indicate greater odds of HIV for women in the 16‒20 cohort (compared with all other cohorts) for both one and two or more sibling deaths, although these effects are not statistically significant.

Table 3 also shows the estimates of the interaction between sibling deaths and cohorts. In Model 1, the reference category for cohorts is all ages except for 16‒20 at conflict. In Model 2, the reference category is 16‒20. The reference category for sibling deaths is zero deaths in both models.

In Model 1, the odds ratio is significant and similar in size to the sibling effect from Model 1 of Table 2 (ORs = 1.32 and 1.36, respectively). Model 2 of Table 3 indicates the odds ratio for multiple sibling deaths for the 16‒20 cohort, which is the reference group in that model. Reporting two or more sibling deaths is associated with a 49% increase in the odds of being HIV positive for this cohort. The results from Model 2 also show more nuanced differences between each of the five-year age groups at exposure cohorts and interactions with the sibling death variables with the reference category set at the 16‒20 cohort. There are no significant differences in the two or more sibling deaths effect by cohort, in part because of the relatively small number of women who fall into each category (Table 3, Model 2). While the results indicate significant cohort differences when the number of sibling deaths is zero, there is little evidence of significant interactions between cohort and sibling deaths.

Our final research question focuses on the role of age at first intercourse and number of sexual partners in determining risk of HIV-positive status and as mediators of timing and intensity of exposure effects. Model 2 in Table 2 shows the regression results when adding age at first intercourse, which is significantly associated with the likelihood of HIV-positive status. The odds ratio indicates that each year increase in age at first intercourse is associated with a 6% decrease in the odds of being HIV positive. Evaluating changes in the cohort and sibling death effects from Model 1 to Model 2 shows only small changes in these effects on HIV, indicating little attenuation of time and intensity of exposure to conflict when accounting for age at first intercourse.

Model 3 of Table 2 adds the number of lifetime sexual partners. Compared with women who reported only one sexual partner, those with two partners have more than twice the odds and those with three or more partners have more than three times the odds of being HIV positive. After adding these variables, age at first intercourse becomes nonsignificant. Both the cohort random effect and the association between experiencing two or more sibling deaths and HIV are reduced slightly with the addition of the number of lifetime sexual partners, but each remains statistically significant. As shown in Figure 2, the pattern of the cohort random effect in terms of predicted probabilities remains the same across Models 1‒3.

Returning to the question of mediation, we conducted formal mediation tests with a focus on whether women aged 16‒20 at the start of the conflict (main cohort effect) and women experiencing multiple sibling deaths (main intensity of exposure effect) have higher risk of HIV through our two proposed mediators: age at first intercourse and number of sexual partners. The main results of interest from these analyses indicate that 5% and 4% of the birth cohort effect (16‒20 years vs. all others) is mediated by age at first intercourse and number of sexual partners, respectively; for sibling deaths, 2% and 1% of the multiple sibling deaths effect is mediated by age at first intercourse and number of sexual partners, respectively. While these mediation effects are significant at the p < .05 level, they are relatively small in size. This is consistent with comparing across models in Table 2, which indicates little change in the birth cohort and sibling deaths effects when adding age at first intercourse and lifetime sexual partners to the base model (see online appendix Table A4).

Discussion

In this article, we aimed to understand how Rwandan women's risk of HIV may be differentiated by their ages at exposure to the Rwandan civil war and genocide, as well as the intensity of their exposure to the violence. A particular strength of our approach is the use of biomarker data on HIV status—which eliminates error due to underreporting—in our nationally representative samples pooled across three waves of DHS data. Specifically, we assessed whether and how age at the start of the conflict and number of sibling deaths during the conflict years were associated with HIV status among Rwandan women of reproductive age. We modeled cohorts as a random effect to separate the effect of chronological age from age at exposure, given the collinearity between these age measures. We added sibling deaths as fixed effects in the models to gain an understanding of how individuals' intensity of exposure may work along with age at exposure in predicting HIV status. We assessed both additive and interactive associations between cohort and sibling deaths in predicting HIV status. We further explored the potential mediating role of age at first intercourse and number of lifetime sexual partners. To our knowledge, our study is the first to assess the cohort and sibling death effects of armed conflict on HIV among women in a contemporary setting in Africa, where conflict and HIV/AIDS remain critical public health issues.

Timing of Conflict Exposure

Our main findings from our age‒cohort models are consistent with our hypothesis that women in late adolescence (ages 16‒20) at the start of the conflict were at particularly high risk for eventual HIV infection, compared with women who experienced the conflict at other ages. There are several potential explanations for this cohort effect.

First, this cohort was exposed to the violence during the life course stage that encompasses the average age at first intercourse in Rwanda. As past research suggests, the Rwandan conflict was associated with increased premarital sex among women, potentially increasing the risk of first sex in this group with implications for later risk of HIV (Lindskog 2016). As we address in more detail below, we did indeed find that 5% of the cohort effect is mediated by age at first intercourse, providing support for this explanation but also indicating that other forces are likely at play.

Second, documentation of the Rwandan conflict suggests that women in this age group may have been targeted for direct gender-based violence during the conflict, including rape driven at least in part by the intent to infect young women with HIV (Jones 2002). Related to this, recent research indicates that HIV incidence rates for adults aged 15‒49 tend to increase in the five years preceding a conflict (based on multiple SSA countries) (Bennett et al. 2015). This suggests that women who may have been exposed to premarital or forced sex during the conflict may have been more likely to have a newly infected partner.

Third, this cohort experienced the immediate postconflict marriage/partner market in their early 20s, when pressure to marry may have been high and sex ratios were particularly low. Recent research indicates that women exposed to violence in their teenage years are less likely to form early unions (Torrisi 2022). Thus, this age group may have been at increased risk for forming unions in the years immediately after the conflict when they were in their early 20s and when women entering unions had higher levels of partner violence and lower sexual rights within heterosexual partnerships (La Mattina 2017), adding to their conflict-induced exposure to HIV. However, accounting for timing of marriage (before and after the conflict) did not substantially change the results.⁶

Intensity of Conflict Exposure

Our findings related to sibling deaths during the conflict support the hypothesis that women who experienced more intense violence during the conflict had higher risk of acquiring HIV. Importantly, we found that reporting two or more sibling deaths (but not one) during 1990–1994 was associated with higher odds of being HIV positive, compared with reporting no sibling deaths. While we know of no other study that explores this association, the findings are consistent with research that suggests a link between adverse exposures/trauma, risky sexual behaviors, and HIV risk (Hillis et al. 2001; Lindskog 2016). Notably, the associations of cohort and sibling deaths with HIV status were significant even when controlling for one another, and again, we exclude maternal mortality causes of death for the sibling deaths measure. We also tested interaction models but found little evidence that associations with the intensity of exposure and HIV status were moderated by cohort. This may be at least in part because of the small number of women in the sample who experienced two or more sibling deaths.

To further assess the role that risky health behaviors play in associations between timing and intensity of women's exposure to conflict and HIV, we assessed whether women's age at first intercourse and lifetime sexual partnerships serve as pathways through which conflict-exposure timing and intensity work to affect their risk of acquiring HIV. We found significant associations between earlier age at first intercourse and multiple lifetime sexual partners with the likelihood of being HIV positive, consistent with past research (Bossonario et al. 2022). However, while mediation was statistically significant, only part of the total cohort (timing) and sibling death (intensity) associations with HIV was mediated by these variables. The remaining direct effects may indicate the lasting physiological changes due to violence-induced stress and trauma that put women at higher risk of HIV. Psychosocial and physical trauma have been associated with physiological changes in women's immune systems, increasing their risk of acquiring sexually transmitted infections, like HIV (Daniels et al. 2022; Tsuyuki et al. 2019). As Tsuyuki et. al. (2019) discussed, chronic stress is associated with lower cell-mediated immunity and higher chronic inflammation, which increase susceptibility to disease and reduce the ability to fight off infections.

Another untested mechanism may be the lack of access to reproductive health care and HIV prevention services during and after the conflict, which may have been particularly limited for young women (Binagwaho et al. 2012; Gervais et al. 2009). Furthermore, women who experienced the highest intensity of exposure to conflict may have been living in areas where the destruction and lack of access to health care were particularly severe.

Limitations

We recognize several limitations of our study. First, the differences in the predicted probabilities of being HIV positive were relatively small despite their statistical significance. We are further unable to attribute direct causality of cohort and sibling deaths on the probability of HIV-positive status decades after the conflict. While the conflict and sibling deaths occurred before the HIV testing was done, we cannot be sure when women contracted HIV.

Second, our results are likely affected by the exclusion of women who died prior to 2005. Using the 2000 Rwanda DHS, de Walque and Verwimp (2010) estimated that women in their late 20s and 30s were more likely to die in 1994 than those in younger cohorts. On the basis of the years of the DHS we use (2005‒2014), these women would have aged out of our sample; thus, selective mortality among females during the conflict is less likely to influence our results. Yet, there is almost certainly some selective mortality with respect to deaths after the conflict, particularly among HIV-positive women. It is likely that poverty and lack of access to health care accelerated death rates among women with HIV during and especially after the conflict by preventing access to antiretroviral therapy. This suggests that the overall impact of conflict on HIV risk for women is underestimated in models that focus on survivors, such as ours.

Regarding mortality selection by cohort, life expectancy for HIV-positive women aged 15‒30 in Rwanda is approximately 29 years and decreases rapidly after age 35 (Nsanzimana et al. 2015). Women aged 30‒35 in 2005 (the first DHS with HIV data) would have been 15‒20 in 1990, substantially overlapping with the birth cohort we find most at risk for HIV. This suggests that conflict exposure may have had even stronger associations with HIV status among those exposed at ages 16‒20 (compared with younger birth cohorts) than what we find here because of selective HIV/AIDS mortality. It may also indicate an underestimation in the associations between the 11–15 cohort (with the lowest HIV risk) and HIV-positive status because of the high death rates of HIV-positive women at ages 29 and 30. Regarding older cohorts, it is likely that a larger percentage of women missing from the older birth cohorts (compared with the 16‒20 cohort) would have died prior to the conflict or aged out of the DHS survey by 2005.

Relatedly, the associations found between sibling deaths and HIV status may be underestimated given that women experiencing the highest intensity of conflict were likely at higher risk of death (including of AIDS) before 2005. Furthermore, we may have underestimated conflict-related sibling deaths because of excluding maternal mortality causes from sibling deaths. While this follows past research that used sibling deaths as a measure of conflict exposure in Rwanda (Kraehnert et al. 2019), broader research indicates that conflict can also exacerbate maternal mortality because of, among other things, the lack of health care infrastructure during the conflict (Bendavid et al. 2021). While we consequently may underestimate sibling deaths, we do so to better isolate the intensity of the violence to which women were directly exposed. Other studies have been able to use regional variation to measure intensity of conflict exposure, but we were precluded from doing so because of the lack of DHS data on where women lived during the conflict.

Select out-migration may be another source of bias in our results, particularly if those in certain cohorts or those who were exposed to higher intensity of violence were more likely to leave Rwanda and not return by the early 2000s. However, given that many of these migrants were in temporary, unsafe refugee camps and returned in subsequent months (UNHCR 2000: chapter 10), and because of the years between the conflict and the DHS HIV data (10 or more), we suspect that most surviving refugee women returned.

Regarding our ability to capture the mediating role of age at first intercourse and number of sexual partners, the main concern is the potential error in reporting these behaviors. Specifically, both measures may be biased by underreporting of sexual initiation, particularly among unmarried women in the DHS, as well as if women did not include premarital or forced sex in their reports of these measures. Such bias must be kept in mind, although we note that the bias is an unfortunate reality for any self-reported data pertaining to sexual behavior.

Finally, while we controlled for marital status, we did not have measures of partner quality, which may have been reduced for women aged 16‒20 during the conflict or for those who experienced more sibling deaths during and after the conflict. Even in preconflict years, women in stable partnerships (cohabiting and marital) were at high risk of sexual coercion, which, in turn, increased their risk of HIV (van der Straten et al. 1998). During and after the conflict, men who were older, involved with multiple sexual partners, or more violent may have been the only partners available to women because of low availability of male partners and women's economic and social pressures to form a partnership/union (Dude 2011; La Mattina 2017). Partnering with such men would have likely increased women's likelihood of acquiring HIV, even in the context of stable unions. This may explain why the number of partners did little to explain the timing and intensity of exposure associations with women's HIV risk.

Conclusion

Despite these limitations, our study contributes to the literature on the links between conflict exposure and health over the life course by asking how timing and intensity of exposure to conflict are associated with HIV infection. Although we focus on Rwanda, this case of armed conflict is but one of many cases of internal and international conflicts that expose women to sexual violence and trauma, with long-term effects (Garry and Checchi 2020). Thus, knowledge about the link between exposure to conflict at specific stages of the life course (cohorts) and by intensity of violence/family loss and later HIV risk can inform public health efforts aimed to reduce HIV. Our findings indicate the need to provide resources and support to women who experience violent conflict in late adolescence and postconflict settings in early adulthood, as well as those who experience higher levels of family loss. Given the continued high rates of HIV infection among women of reproductive age (compared with men), attention to women's experiences of conflict may be critical to reaching HIV reduction goals in SSA (De Lay et al. 2021).

Acknowledgment

We thank The Ohio State University Department of Sociology for partially funding this work.

Notes

References