Intimate partner violence is associated with HIV infection in women in Kenya: A cross-sectional analysis

Funded by the US Agency for International Development (USAID), DHS has conducted national household surveys on health-related issues in more than 90 countries since 1984. The Kenya 2008-2009 DHS surveyed women aged 15 to 49 in selected households. In every second selected household, women were offered voluntary HIV testing [34], p. 8. Of the 4,418 women eligible for HIV testing, 8.2% refused to provide blood and 3.5% were not interviewed or were absent from the blood collection [34], p.212. One woman per household was selected to participate in a domestic violence module. Questions on IPV were administered to women completing the domestic violence module who had ever had an intimate partner, with questions asked about either the current or most recent partner [34], p. 254. Our sample for this study is women who are: currently married or living with someone, completed the IPV questions, and obtained HIV testing.

Women participating in the DHS survey gave their informed consent before data collection and before giving blood for HIV testing [34]. For the domestic violence survey, privacy was ensured before the interview was conducted [34]. The blood collection and analysis was based on a protocol developed by the DHS and revised and approved by the Scientific and Ethical Review Committee of the Kenya Medical Research Institute, and the National AIDS Control Council [34]. Permission for using the data for this study was authorized by ICF International, which coordinates the DHS.

There were 12 IPV questions in the DHS IPV survey questionnaire, grouped into four IPV types: psychological, severe physical, less severe physical and sexual. The participants were asked if they had ever experienced any of these, and if so, about frequency in the last 12 months. Dried blood spots were collected during the survey and tested for HIV using ELISA (see [34], p. 10).

Basic sociodemographic factors analyzed in this study were those commonly considered in the aformentioned studies, and included age, education, working status (whether respondents are currently working), religion, geographic residence (urban/rural), wealth and age gap between the respondents and their male partners [13, 20, 24, 25, 27‐29, 33, 36‐40]. On the basis of theoretical considerations and previous research [23‐28, 41‐46] (though limited by the data collected), we also selected certain HIV risk factors a priori to assess: partner’s alcohol consumption, condom use at last sex, partner’s number of other wives, respondent’s of sex partners with the last 12 months and in lifetime. In addition, location (cluster ID and province) and ethnicity were treated as random effects to properly control for correlations between people from the same geographic area and ethnicity. Women who reported no sexual activity within the preceding 12 months were not asked about condom use; we coded these women as “not asked” rather than excluding them from the study.

We used generalized linear mixed models (GLMMs) to examine the association between IPV and HIV infection. The GLMM framework allows us to model a binary response variable (HIV test result), and to take random effects (location and ethnicity) into account. To simplify interpretation and to account for relationships between predictors, we assessed all the selected sociodemographic variables with the IPV predictor together in a single GLMM. Given theoretical and empirical uncertainty about whether HIV risk factors may mediate and/or confound the association between IPV and HIV, we included HIV risk factors in a model with sociodemographic variables to see whether and how they affected the magnitude of association between IPV and HIV, without presupposing which of these roles they might have.

To construct a simple model, we combined the 12 IPV questions (see Table 1) by converting responses into scores (see Table 2), and used the first principal component from a scaled, uncentered principal components analysis (PCA) as an index to describe the overall IPV experience of each woman, using a single variable.

In a follow-up model, we constructed a separate PCA index for each of the four DHS categories of IPV, using the same methodology as we used for the overall index.

Variable-level p values were calculated by sequentially dropping each variable and comparing the restricted models to the original model.

We are not able to make our data available, but researchers can request them from DHS. All of the R scripts use to analyze the data and produce the figures are available for download at http://​lalashan.​mcmaster.​ca/​theobio/​Kenya_​IPV_​risk/​.

The analyses included 1904 women, after dropping those who were not currently in a relationship, and those with missing data. Their overall HIV prevalence was 7.5%, compared to 8.3% for the whole national DHS survey, and 6.3% for the 350 otherwise-eligible women who were not selected for the domestic-violence module.

Table 2 presents the four types of IPV categorized in the DHS surveys, their sub-categories, and the proportion of women in our sample reporting each type of IPV. Less severe physical IPV is the most common type of IPV experienced, followed by psychological IPV, sexual IPV and severe physical IPV.

Sociodemographic variables and their relationships with HIV prevalence and IPV (measured by the first component of a PCA decomposition, see Methods) are shown in Table 1.

When controlling for the sociodemographic variables only (the base model), there was a strongly significant positive association between IPV and HIV status Table 3. The estimated slope of the response of logistic predictor to our IPV index was 0.047; this means that, compared to a zero-IPV baseline, the odds ratio for HIV risk of a woman experiencing the mean amount of IPV was 1.13 and that of a woman at the 95th percentile for IPV was 1.79. Of the sociodemographic variables, only working status was a significant predictor of HIV in the multivariate model: women reported currently working were more likely to be HIV positive (OR estimate 1.66, P=0.016). Effect sizes for all variables in the model are shown in Additional file 1: Table S1 (fixed effects) and Additional file 2: Table S2 (random effects). When controlling for risk factors as well as sociodemographic variables in the full model, the significant positive association between IPV and HIV risk remained (see Table 3). The estimated slope of the response of logistic predictor to our IPV index was 0.036. The results mean that, compared to a zero-IPV baseline, the odds ratio of HIV risk was 1.10 for a woman experiencing the mean amount of IPV and that of a woman at the 95th percentile for IPV was 1.55. The effect of working status remained the only significant sociodemographic predictor (OR estimate 1.62, P=0.023). Of the HIV risk factors, only lifetime number of partners was significant in the multivariable model. The estimated OR for an additional lifetime partner was 1.21 (P=0.014).

Figure 1 shows the estimated effect on HIV infection of various levels of IPV, compared to a relationship with no IPV. Other factors held equal, women who report IPV were more likely to test positive for HIV than those who did not, when controlling for sociodemographic variables (black lines), or sociodemographic variables and risk factors (blue lines). The estimated effect of IPV on HIV risk is similar, but not as strong, when risk factors are added to the model.

When we replaced our IPV index with separate indices for each of the four DHS categories of IPV, the response to socio-demographic variables and risk factors remained the same, but none of the individual categories indicated a statistically significant response.

An earlier version of our model used a quadratic, rather than linear, response of HIV risk to IPV. We discarded this model because it presented similar significance levels to the current version, but was more difficult to interpret.

Based on a nationally representative sample, we have identified a positive association between a PCA-based index of overall IPV and HIV infection, after controlling for basic sociodemographic variables. This association persists when HIV risk factors are included in the model. Working status and reporting a larger number of lifetime sexual partners were also positively associated with women’s HIV risk. Our study combines four types of IPV: psychological, less severe physical, severe physical, and sexual IPV. This comprehensive approach may allow a more accurate estimation of the association between IPV and HIV. A follow-up model that tested the four categories of IPV separately did not find a significant effect of any category, which may be due to a lack of sufficient power.

A bivariate analysis of data on people attending an STI clinic in Nairobi, Kenya [37] found a similar association, while other Kenyan studies did not find any association [30‐32]. Of particular interest, a prospective study found no significant association between IPV and HIV seroconversion in discordant couples, though they did find a significant association between IPV and prevalent HIV infection [33]. This study examined data from seven nations in East and Southern Africa and did not report data stratified by country, which may be problematic if there is geographic variation in the magnitude or very existence of the association between IPV and HIV. These differing results may in part be due to the fact that different studies analyzed data from different population subgroups (e.g. pregnant women [31] and couples [30, 33] in clinical sites), examined fewer or different types of IPV [30, 32, 33, 37] than our study, and used univariate analysis [37] or different sets of basic and risk-factor variables.

Our study has several limitations. Our reliance on cross-sectional data precludes direct investigation of causal relationships. Some relevant potential confounders were not available, such as partner’s sexual risk behaviours and partner’s HIV status. Another potential source of bias is the self-selection of individuals for both of the violence module and the DHS HIV screen. Participants may also underreport their experience of IPV due to potential stigma. There may also be error in the measurement of sociodemographic and HIV risk behaviours, as well as in IPV.

Although there are programs to test for HIV during counseling for gender-based violence in Kenya [47] and the government passed a Sexual Offenses Act to penalize sexual assault [48], enforcement of the Act is a challenge because of fear among victims and perceived stigma from the community, and domestic violence is not yet formally recognized by law [49].

Although our cross-sectional study does not provide direct evidence that IPV contributes to HIV risk, our findings are consistent with several hypothesized mechanisms for a causal association between IPV and HIV, including that women who experience IPV may be unable to negotiate safer sexual behaviours such as condom use [15‐18], may have riskier sexual behaviours [4, 13, 19‐22], or may have relatively compromised immune systems due to the stress of IPV [7]. Additionally, it is possible that sexual IPV may directly lead to HIV infection. Causality in the other direction – i.e. that women with HIV may be more likely to experience IPV [33, 50] – is also possible. Regardless of causality, the finding of association is important: knowing that IPV is associated with HIV risk could be useful for secondary prevention in nations where both IPV and HIV prevalence is high.