Intimate Partner Violence and Depression Symptom Severity among South African Women during Pregnancy and Postpartum: Population-Based Prospective Cohort Study

Background: Violence against women by intimate partners remains unacceptably common worldwide. The evidence base for the assumed psychological impacts of intimate partner violence (IPV) is derived primarily from studies conducted in high-income countries. A recently published systematic review identified 13 studies linking IPV to incident depression, none of which were conducted in sub-Saharan Africa. To address this gap in the literature, we analyzed longitudinal data collected during the course of a 3-y cluster-randomized trial with the aim of estimating the association between IPV and depression symptom severity. Methods and Findings: We conducted a secondary analysis of population-based, longitudinal data collected from 1,238 pregnant women during a 3-y cluster-randomized trial of a home visiting intervention in Cape Town, South Africa. Surveys were conducted at baseline, 6 mo, 18 mo, and 36 mo (85% retention). The primary explanatory variable of interest was exposure to four types of physical IPV in the past year. Depression symptom severity was measured using the Xhosa version of the ten-item Edinburgh Postnatal Depression Scale. In a pooled cross-sectional multivariable regression model adjusting for potentially confounding time-fixed and time-varying covariates, lagged IPV intensity had a statistically significant association with depression symptom severity (regression coefficient b = 1.04; 95% CI, 0.61–1.47), with estimates from a quantile regression model showing greater adverse impacts at the upper end of the conditional depression distribution. Fitting a fixed effects regression model accounting for all time-invariant confounding (e.g., history of childhood sexual abuse) yielded similar findings (b = 1.54; 95% CI, 1.13–1.96). The magnitudes of the coefficients indicated that a one–standard-deviation increase in IPV intensity was associated with a 12.3% relative increase in depression symptom severity over the same time period. The most important limitations of our study include exposure assessment that lacked measurement of sexual violence, which could have caused us to underestimate the severity of exposure; the extended latency period in the lagged analysis, which could have caused us to underestimate the strength of the association; and outcome assessment that was limited to the use of a screening instrument for depression symptom severity. Conclusions: In this secondary analysis of data from a population-based, 3-y cluster-randomized controlled trial, IPV had a statistically significant association with depression symptom severity. The estimated associations were relatively large in magnitude, consistent with findings from high-income countries, and robust to potential confounding by time-invariant factors. Intensive health sector responses to reduce IPV and improve women’s mental health should be explored. All interviews were conducted in accordance with ethical and safety recommendations promulgated by the World Health Organization [29]. Namely, research assistants were trained on how to administer surveys for gathering sensitive information and provided assurances of confidentiality. The survey was framed generally as being part of a study of family health and well-being, not a study about violence against women. In consultation with on-site supervisors, research assistants provided referrals to local counseling resources and/or child social services as needed, with standardized protocols in place to refer women to emergency services in the case of acutely elevated risk of harm to self or harm from others. All study procedures were approved by the South General Institutional Review Board of the University of California at Los Angeles and the Health Research Ethics Committee of the Stellenbosch University Faculty of Health Sciences. A four-person Data Safety Monitoring Board populated by local and international experts monitored implementation of the study. The secondary analysis described in this manuscript was based on a de-identified dataset and did not require additional approval or consent. Details of the study design, field training, and primary outcome analyses at 6, 18, and 36 mo have already been published [30–35] (ClinicalTrials.gov registration {“type”:”clinical-trial”,”attrs”:{“text”:”NCT00972699″,”term_id”:”NCT00972699″}}NCT00972699). The study was conducted during 2009–2014 in three townships surrounding Cape Town, South Africa. All pregnant women living in 24 neighborhoods (matched on population density, number of bars, distance to health care, and access to public works infrastructure) were identified and recruited into the study, with a 98% participation rate. These matched neighborhoods were randomized in blocks of four to either a home visiting intervention or standard clinic care groups. Standard clinic care was available (within 5 kilometers) to all women living in the study catchment area and generally consisted of tuberculosis and HIV testing, partner HIV testing, antiretroviral therapy, antenatal and postnatal care, well-child clinics, and primary health care [30]. The home visiting intervention was implemented by the Philani Maternal, Child Health, and Nutrition Project, a non-governmental organization that has been operating in the Western Cape of South Africa since 1979 [36] and which has since expanded to the Eastern Cape of South Africa as well as Ethiopia and Swaziland. Philani implements a “mentor mother” program, recruiting women from the community who have successfully raised thriving children despite concentrated adversity and then training these women to serve as paraprofessional community health workers for home visiting among pregnant women and their families [37,38]. For the purposes of the randomized trial, the Philani intervention was standardized and augmented with training in a pragmatic model of problem-solving and cognitive-behavioral techniques to address major community health challenges, including HIV/tuberculosis, malnutrition, and alcohol use [39,40,41]. An independent team of Xhosa-speaking research assistants obtained written informed consent from all study participants and collected survey data through face-to-face interviews conducted at baseline, 6 mo, 18 mo, and 36 mo. Analyses of these data revealed that the “Philani Plus” intervention improved overall maternal and child health across a number of different outcomes, notably those related to HIV-prevention behaviors, breastfeeding, and child growth over 18 mo; and maternal emotional well-being, child language development, and child growth over 36 mo [31–35]. The primary outcome of interest in this secondary analysis was depression symptom severity, which was measured at all time points with the Xhosa version of the ten-item Edinburgh Postnatal Depression Scale (EPDS) [42]. Scale items inquire about depressive symptoms within a 7-d recall period, with responses scored on a four-point Likert-type scale ranging from 0 (“not at all”) to 3 (“all the time”). Among Xhosa-speaking women, the EPDS has been shown to have a coherent internal structure [43], high sensitivity and specificity for detecting major depressive disorder [44–46], and good construct validity [39,47]. In the baseline sample, the EPDS had good internal consistency (Cronbach’s alpha = 0.89), and, using 500 bootstrap replications to compute the standard error, the 95% confidence interval (CI) was 0.88–0.90. The primary explanatory variable of interest in this secondary analysis was experience of IPV, measured with a four-item scale. Following Straus’ [48] approach of asking behaviorally specific questions, the IPV scale included items inquiring about the frequency with which a woman’s current or previous intimate partner had, during the past 12 mo, slapped or thrown anything at her; pushed or shoved her; hit her with a fist or another object; or threatened or attacked her with a gun, knife, or other weapon. Responses were scored on a four-point Likert-type scale ranging from 1 (“never”) to 4 (“many”). Together, these four items had acceptable internal consistency, with a Cronbach’s alpha of 0.75 (95% CI, 0.71–0.80) in the baseline sample. To generate an omnibus measure of the intensity of IPV across all four items, following Kling and colleagues [49] we defined a summary IPV index as the equally weighted average of the four z-scores (i.e., each item was standardized to a mean of 0 and standard deviation of 1, and then the summary index was defined as their average value). While the absolute values of the index carry no meaning, higher values denote greater intensity of IPV. We adjusted our estimates of the association between IPV intensity and depression symptom severity for a number of potentially confounding time-invariant and time-varying covariates. Time-invariant covariates were elicited at the baseline interview and included binary indicators denoting whether the participant had been assigned to the intervention or standard clinic care arm, age at baseline, and whether the participant had completed high school. Household asset wealth was elicited by asking participants a series of 13 questions about household assets and housing characteristics (e.g., whether there is a flush toilet in the home, whether a household member owns a radio, etc.). Then, following the method of Filmer and Pritchett [50], we applied principal components analysis to these variables. The first principal component was retained and used to define the asset wealth index, and participants were sorted into quintiles of relative asset wealth. Time-varying covariates were elicited at each interview. Time elapsed since the baseline interview was measured in months. We included binary indicators denoting whether the participant was employed (either full- or part-time), whether the father of the child was staying with the participant, HIV serostatus (classified as HIV-positive, HIV-negative, or unknown/refused testing), and whether the participant had been diagnosed with high blood pressure or diabetes. Household monthly income was measured in South African Rand. Alcohol abuse was measured with the three-item consumption subset of the Alcohol Use Disorders Identification Test (AUDIT-C) [51–53]. We did not publish or pre-register a plan for this secondary analysis. The analysis plan is described below, with any deviations noted in S1 Text. Given the repeated-measures design, we sought to estimate the association between IPV intensity and depression symptom severity, adjusted for the time-invariant and time-varying covariates described above. We fitted a linear regression model to the pooled cross-sectional data, specifying the EPDS score as the continuous dependent variable, using cluster-correlated robust estimates of variance [54–56] to correct standard errors for clustering within participants over time. To ensure the correct temporal sequence of the exposure and outcome, IPV measurements were lagged by one time point (an average of 12 mo). The estimated regression coefficients therefore provided information about the association between IPV intensity at one time point and depression symptom severity at the subsequent time point. We sought to determine whether the adverse impacts of IPV were experienced to a greater extent by women at the upper end of the conditional depression distribution. To do this, we fitted quantile regression models [57] to estimate the association between IPV intensity and the 20th, 40th, 60th, and 80th percentiles of the conditional depression distribution, using a covariance matrix of the asymptotic distribution of the quantile regression estimator that permits within-participant correlation over time [58]. Although the regression models included adjustment for a number of potentially important confounding variables, it is possible that some important variables were not observed. For example, in the specific setting of this randomized trial, no data on participants’ histories of child sexual abuse were obtained. Child sexual abuse could potentially confound our estimates of the association between IPV intensity and depression symptom severity [23]; even in the lagged-covariate models where the exposure precedes the outcome, the confounding influence of child sexual abuse would precede both the exposure and the outcome. Estimates could be similarly biased by omitting other types of childhood adversities, such as paternal incarceration or orphanhood [59,60]. We therefore fitted a fixed effects regression model to the data, using within-participant variation over time to identify the estimated associations [61]. The estimated regression coefficients are interpreted as providing information about the association between changes in IPV intensity and changes in depression symptom severity. A substantial advantage of the fixed effects regression model is that the procedure adjusts for confounding, whether observed or unobserved, that is time-invariant over the period of study (such as history of child abuse). The principal disadvantage of the fixed effects regression model is that, because the fixed effects sweep out all time-invariant confounding, only associations between changes in the outcome and changes in time-varying covariates can be examined. Because time-invariant covariates, by definition, do not change over time, they are eliminated from the model. To determine whether changes in IPV intensity were differentially associated with changes in depression symptom severity at different points in the conditional depression distribution, we used Canay’s [62] fixed effects quantile regression model. We conducted a number of ancillary analyses to assess the robustness of our main findings. First, to confirm that our findings were robust to the specification of the IPV variable, we generated dichotomous exposure variables indicating the presence or absence of any exposure to each of the four types of IPV. These four dichotomous exposure variables were included in the multivariable regression models both independently and jointly as lagged covariates. Second, because caseness for (probable) depression is frequently of clinical interest, we defined probable depression as EPDS ≥13 [45,46,63–65]. We fitted logistic regression models as above, instead specifying probable depression as the binary dependent variable. We also report marginal effects [66] so that the logistic regression coefficients can be interpreted as the percentage-point probability difference in the outcome associated with the covariates. Third, to determine the extent to which the association was potentially bidirectional [23,67], we re-fitted the regression model with IPV intensity as the dependent variable and depression symptom severity as a lagged covariate, thereby estimating the association between depression symptom severity at one time point and IPV intensity at the subsequent time point.