Systematic review of structural interventions for intimate partner violence in low- and middle-income countries: organizing evidence for prevention

We conducted a per-protocol review following PRISMA guidelines for systematic reviews [15] (review protocol unregistered, see Additional file 1: Table S1 for PRISMA checklist). We systematically searched PubMed and Web of Science for articles that evaluated structural interventions for IPV. Studies eligible for inclusion were (1) original research, (2) written in English, and (3) published between January 1, 2000 and May 23, 2015 in the peer-reviewed literature. Such studies (4) evaluated the impact of a structural intervention for primary and secondary prevention of male-to-female IPV in (5) World Bank-defined LMIC through (6) quantitative evaluation of the impact on (a) IPV incidence or prevalence or (b) secondary outcomes theoretically linked to IPV incidence or prevalence by the authors. Secondary outcomes included intermediate outcomes hypothesized to explain the impact of the intervention on IPV, such as increased financial autonomy and security among women or decreased acceptability of spousal violence among men.

We defined structural interventions as interventions that address economic, physical, politico-legal, or social environments that produce or reproduce IPV risk, in contrast to individual interventions that target individual knowledge, attitudes, and behaviors [11]. We included structural approaches that target structural risk at local levels to permit inclusion of rigorous study designs [7]: for example, randomized control trials (RCT) of microfinance delivered to individual women to understand the potential impact of ameliorating poverty-related stressors, diminishing intra-household gender inequalities, and affecting community norms at scale. Consistent with the definition of structural interventions [11], however, we excluded interventions that targeted risk behaviors without modifying aspects of the economic, physical, politico-legal, or social environments that produce and reproduce risk for IPV. These included processes within healthcare systems, such as one-stop centers for survivors of gender-based violence [3], and advocacy interventions aimed to empower and link IPV survivors to community services [16].

We defined primary and secondary prevention interventions as those aimed to prevent or reduce IPV, without consideration of prior exposure to violence [3]. Whereas primary prevention aims to prevent initial IPV and secondary prevention aims to prevent ongoing IPV, current epidemiologic measures of IPV have limited sensitivity to measure IPV patterns, and structural interventions address aspects of the environment that likely promote initial and sustained violence. We did not include tertiary prevention strategies, which aim to prevent negative health or social sequelae among victims following IPV, in alignment with the goal to investigate interventions that reduce the incidence or prevalence of IPV.

We focused on peer-reviewed literature to capture the highest quality research, and limited publications to English due to constraints on translation and evaluation of non-English language publications. We did not specify population restrictions, given the disproportionate prevalence of IPV and significant structural challenges that subpopulations, such as adolescents, injection drug users, and sex workers, experience. Similarly, we did not limit study designs, outcomes, or statistical summary measures. There exists active debate about appropriate outcomes for IPV intervention studies (e.g., IPV, proximal measures of social change, or outcomes under the control of women) [17, 18]. Additionally, structural interventions may require diverse evaluation strategies; RCTs may be most appropriate when interventions alter single, proximate risk factors and other influences are known and measured [7]. Although understanding process outcomes is important [7], these issues were beyond the scope of this review. We privileged quantitative measures in order to compare impact across studies.

We performed four searches that combined three search themes (IPV, intervention, and one of four structural categories) with the Boolean operator “and” (Figs. 1 and 2) on March 14, 2013 (January 2000 to March 2013) and May 23, 2015 (March 2013 to May 2015). Each search theme included a comprehensive list of terms intended to account for historical and disciplinary heterogeneity in terminology. We selected search terms through a multistep process. We first identified potential terms through expert identification, published reviews, and key terms in relevant articles. We then evaluated and selected terms based on their sensitivity (ability to identify relevant literature) and uniqueness (ability to identify articles not captured by other search terms). Index terms identified in PubMed were applied in Web of Science as phrases bracketed by quotations.

Secondary search strategies included manually searching reference lists of articles identified for data extraction and contacting 23 experts to recommend additional studies. These experts included prominent researchers and first and senior authors for 12 articles identified for data extraction during the original submission, including two articles [19, 20] that were found ineligible during data extraction: one because the outcome definition did not ensure IPV measurement [19] and the other because statistical analyses of relevant outcomes were not published [20]. Among 23 experts contacted, seven responded. They recommended 26 articles cumulatively, one of which met inclusion criteria. Given time limitations, we did not repeat this step during manuscript revision. We also planned, but did not conduct, hand-searches of key journals identified from electronic searches and reference list reviews due to limited consensus on key journals from these steps.

One author (CB) screened all records from the database search, reference lists, and expert recommendations. Each author (CB, WW, EB, and RS) independently assessed a selection of full-text articles for eligibility using a form developed and piloted on five randomly selected articles. WW and EB extracted data independently using a form developed and piloted on three randomly selected articles. These authors extracted data in duplicate for the original submission; research assistants checked extracted data during the review update. We resolved discrepancies at each stage by consensus. In instances where multiple articles reported on the same study, we extracted data from each article separately due to varied theoretical frameworks, analytic samples, and outcome measures. Planned data extraction included intervention characteristics (intervention and comparison exposures, allocation or targeting, and exposure period) and study characteristics (data source, study design, sampling strategy, sample size, inclusion criteria, follow-up period, exposure assessment, coverage, and unintended harms). We recorded data on primary outcomes (IPV incidence or prevalence) and secondary outcomes (intermediate outcomes linked to IPV incidence or prevalence by study authors) to facilitate comparison across studies. Statistics included those from the most saturated models presented by authors, privileging intention to treat analyses. Risk of bias evaluation differed for RCTs and other study designs. Questions captured selection, performance, detection, analysis, and reporting biases and IPV-specific issues, reflecting PRISMA [21], Cochrane Collaboration [22], and WHO recommendations for IPV research [23]. This evaluation focused on study level risk of bias, with selective reporting within studies recorded and outcome level risk included in the discussion of intervention effects. We collected information on risk of bias as descriptive data; these did not influence data synthesis, as meta-analysis was not planned for statistical, methodological, and programmatic reasons [21]. The diversity of study designs, outcome measures and definitions, and contexts inhibited quantitative synthesis.

Planned data differed from collected data for risk of bias analyses. Heterogeneous study designs and data presentation styles limited the ability to calculate intervention coverage and attrition consistently. Calculations of intervention coverage alternately reflected the percentage of interviewed or enrolled participants among potential participants, all volunteers, or eligible volunteers or the percentage meeting varied participation thresholds among randomized or enrolled participants. Attrition similarly described loss to follow-up among randomized, interviewed, or participating individuals.

Although we did not contact authors for further information, we invited first authors of the 10 studies included in the original submission to comment on any aspect of the manuscript, including data presentation and interpretation. As only one author responded, without expressed concerns, we did not repeat this step during the review update.

The articles described interventions conducted between 1992 and 2012 in nine countries: Bangladesh [24], China [39], Côte d’Ivoire [28, 30], Ecuador [29], Ethiopia [40], India [26, 36], Mexico [25], South Africa [31‐34, 37, 38, 41], and Uganda [6, 27, 35, 42] (Additional file 2: Table S2). Eleven of 20 articles used rural samples [24, 25, 27, 28, 30, 31, 33, 34, 37, 38, 42], by intervention design or analytic restriction. Eight articles, representing seven studies, engaged men and women [6, 27, 28, 31, 32, 35, 41, 42]; seven articles (four studies) targeted women exclusively to address outcomes of interest [24, 25, 29, 33, 34, 37, 38]; and five articles (five studies) involved men exclusively [26, 30, 36, 39, 40]. Common restrictions included socio-demographic vulnerabilities, such as women or households with low socioeconomic status [24, 25, 27, 29, 33, 34, 37, 38] and communities with high marginality [25], informal land tenure [32], or increased exposure to armed conflict [27]. Twelve articles (11 studies) restricted interventions or analytic samples to reproductive-aged or young persons [6, 24, 25, 27, 29, 31, 32, 35, 36, 39, 40, 42]. The mean sample size across studies was 1061.5 (IQR 489.5–1274.3) for outcome models or study samples, where authors did not report sample sizes for outcome models.

All articles described social, economic, or social and economic interventions. Economic interventions examined the utility of microfinance and cash transfers, including BRAC-led microfinance and skills programs in rural Bangladesh [24], cash transfers and microenterprise training in post-conflict northern Uganda (Women’s Income Generating Support, WINGS) [27], the national conditional cash transfer program in Mexico (Oportunidades) [25], and the national unconditional cash transfer program in Ecuador (Bono de Desarrallo Humano, BDH) [29].

Social interventions tested a variety of participatory learning, community mobilization, and multimedia approaches. Participatory learning programs aimed to improve sexual health among young men and women in South Africa (Stepping Stones) [31], enhance gender equitable attitudes among young men in China [39], and promote bystander behaviors among adolescent male cricket athletes in India (Parivartan) [36]. Community mobilization interventions focused on critical analysis and discussion of power inequities among men and women in urban Uganda (SASA!) [6, 35] and IPV-related norms and behaviors in rural Uganda (Safe Homes and Respect for Everyone, SHARE) [42] and included a compilation of workshops and campaigns on gender relations and violence against women in North India, led by and targeted to men (Men’s Action to Stop Violence Against Women, MASVAW) [26]. Interventions in rural Côte d’Ivoire [30] and urban Ethiopia (Male Norms Initiative) [40] combined community-based programming with discussion groups for men. A media campaign promoted implementation of domestic violence legislation and social norm change related to domestic violence in South Africa (Soul City Fourth Series, SC4) [41].

Four studies, reported in seven articles, combined economic and social approaches: (1) microfinance and gender training in South Africa (Intervention with Microfinance for AIDS and Gender Equity, IMAGE) [33, 34, 37, 38]; (2) livelihoods training (Creating Futures) and sexual health training (Stepping Stones) in South Africa [32]; (3) cash transfers, microenterprise training, and gender and couples training in post-conflict northern Uganda (Women Plus) [27]; and (4) group savings for women with gender dialogue groups for couples (or women and their male family members) in rural Côte d’Ivoire [28].

Thirteen articles discussed prospective RCTs, with a mean follow up period of 24.3 months from study initiation. Three articles used longitudinal designs: a non-randomized trial, in which researchers allocated urban middle schools to intervention and wait list control conditions purposively to reduce potential contamination due to geographical proximity [36]; an interrupted time series evaluation, which measured outcomes successively among intervention participants [32]; and a before-after analysis, in which researchers collected data from a unique stratified random sample at baseline and follow-up and defined intervention and comparison groups retrospectively [41]. Additionally, three articles reported on cross-sectional studies, two of which used study-specific surveys with non-random or unspecified sampling [24, 26] and one of which used a nationally representative sample [25]. Exposure periods in non-RCTs ranged from a measured seven months [41] to a potential maximum of seven years [24].

All articles, except one [32], reported comparisons with unexposed or standard-of-care comparators. Although this article reported a proof of concept study, rather than an impact evaluation [32], we included it because it represents a unique contribution to the literature. The study evaluated the combination of Stepping Stones, a widely implemented social intervention, and Creating Futures, an economic intervention providing livelihoods training without supplementing participant income. Two articles, by contrast, included multiple comparators. Kim et al. [34] reported a three-way comparison to disentangle IMAGE intervention components, comparing combined microfinance and social components, microfinance only, and no exposure. Green et al. [27] reported two trials: first comparing cash transfers and microenterprise training to waitlist control and then comparing cash transfers and microenterprise training to its combination with gender and couples training.

Outcome measures varied across studies (Tables 2 and 3, Additional file 3: Table S3). Sixteen articles reported IPV as an outcome, including single types of IPV (i.e., physical IPV [6, 25, 28‐30, 32, 39, 42], psychological IPV [25, 29, 39, 42], sexual IPV [6, 25, 28, 30, 32, 35, 42], economic IPV [28], and controlling behaviors [27, 29, 33, 34, 37]) and aggregate measures (i.e., physical or psychological IPV [24, 27, 39], physical or sexual IPV [25, 28, 30‐34, 37, 40], and physical, sexual, or psychological IPV [40]). Authors described using behavioral measures of IPV in 15 articles [6, 24, 25, 27‐30, 32‐35, 37, 39, 40, 42], although several reported behaviors incompletely [24, 25, 34] and few discussed validation of selected questions in the study context. Nearly all studies measured recent IPV: differentially defined as three [32, 39], four [24], six [40], eight [27], or 12 months [6, 25, 28, 30, 31, 33‐35, 37, 42]. Only one study measured lifetime physical IPV or did not specify a recall period [29]. In many articles, self-reported experience of IPV was measured among married or partnered women [6, 24, 25, 28‐30, 34, 35, 37], by sampling or analytic restriction. By contrast, several articles summarized male-reported perpetration of violence [31, 32, 39, 40, 42].

Based on the most saturated models, privileging intention to treat analyses, and standardizing statistical significance at p < 0.05 or 95 % confidence intervals not including unity, nine of 16 articles reported a statistically significant difference in the desired direction for at least one IPV indicator. Among economic interventions, WINGS, which included cash transfers and microenterprise training in post-conflict Uganda, was associated with decreased odds of controlling behaviors compared to waitlist control [27]. BDH, which analyzed the national unconditional cash transfer program in Ecuador, was associated with decreased odds of controlling behaviors in the full sample [29]. Among social interventions, a participatory gender learning program in China was associated with decreased psychological and physical or psychological IPV perpetration among workers and students at follow-up [39]. SHARE, a community mobilization and HIV intervention, also noted decreased female-reported physical and sexual IPV experience at 35 months and male-reported psychological IPV perpetration at 16 months [42]. Combined economic and social interventions included group savings for women and gender dialogue groups for couples in rural Côte d’Ivoire [28]; these were associated with decreased odds of economic IPV compared to group savings only. The combination of Stepping Stones, a participatory learning program, and Creating Futures, a livelihoods intervention, was associated with decreasing sexual IPV and physical or sexual IPV among female participants over time [32], and IMAGE, a microfinance and participatory learning program, was associated with decreased risk of physical or sexual IPV [33, 34, 37].

By contrast, SASA! in Uganda [6, 35], BRAC in Bangladesh [24], Oportunidades in Mexico [25], Stepping Stones in South Africa [31], the Male Norms Initiative in Ethiopia [40], and the addition of male discussion groups to community-based prevention in Côte d’Ivoire [30] did not report significant associations at the p < 0.05 level for any IPV indicators included in this review. Interventions with multiple comparators also showed non-significant effects for specific intervention components. Cash transfers and microenterprise training provided in WINGS were associated with reduced IPV, yet the addition of gender and couples training to this (Women Plus) was not associated with further IPV reduction [27]. Conversely, IMAGE was associated with reduced risk of past-year physical or sexual IPV compared to waitlist control [33, 34, 37], but a disentanglement study showed no significant effect for the intervention when IMAGE was compared to microfinance alone [34].

Other outcome measures included varied indicators of economic wellbeing [27, 32‐34, 37], IPV-related help seeking or receipt [6, 41], attitudes toward IPV [28, 30, 33, 34, 36, 37, 41], gender norms [26‐28, 32, 36, 39, 40], and other social pathways related to women’s equity, autonomy, or agency at varied levels of the social ecology, including improved relationship quality, personal empowerment, greater social capital, and collective action [27, 30, 32‐35, 37, 38, 41]. Economic interventions with positive effects included WINGS, which was associated with improved economic wellbeing and relationship quality [27].

Among social interventions, SASA! was associated with improvement in nine of 10 indicators measuring relationship quality among men [35], although only one of 11 indicators improved among women [35] and the prevalence of appropriate community responses to women experiencing IPV did not improve significantly [6]. In MASVAW, a community action program targeted toward men in North India, activist men (active intervention members) and influenced men (non-activists in intervention communities) demonstrated statistically greater agreement with gender equitable norms compared to controls for all scales, except knowledge of women and child laws among influenced men [26]. In Côte d’Ivoire, men assigned to discussion groups and community-based prevention programming reported improved ability to control hostility and manage conflict and greater participation in gendered household tasks than men assigned to community-based prevention programming only [30]. The between-group comparison on change score similarly differed for gender attitudes among adolescent males in Parivartan, a bystander intervention in India [36]. The Gender Equitable Men scale revealed more equitable attitudes among male workers and students assigned to participatory health programming in China [39] and group education and community engagement in Ethiopia [40]. Further, increased Soul City media exposure was associated with higher prevalence of collective action and help seeking behaviors, and eight of ten indicators suggested reduced acceptability of IPV at follow-up than baseline [41].

Combined economic and social interventions included Women Plus in Uganda, which was associated with improved relationship quality, general partner support, and partner support of household activities among women assigned to receive the economic (cash transfers and microenterprise training) and social intervention (gender and couples training) versus the economic intervention only [27]. The combination of group savings for women with gender dialogue groups for couples in rural Côte d’Ivoire demonstrated greater improvement in attitudes condoning spousal abuse than group savings alone [28]. Similarly, the combination of Stepping Stones, a participatory learning program, and Creating Futures, a livelihoods intervention, was associated with improvement in gender norms reported by men and women, relationship equity reported by men, and club or group involvement reported by women as well as half of indicators measuring economic wellbeing [32]. IMAGE was associated with increased economic wellbeing [33, 34, 37] and social capital or empowerment [33, 34, 37] compared to unexposed controls, in addition to less endorsement of attitudes condoning IPV compared to microfinance only participants [34]. Only two articles did not report statistically significant associations for any economic or social indicators included in this review [6, 38]. These articles, however, reported on SASA! in Uganda and IMAGE in South Africa; both interventions were associated with positive effects in other articles [33‐35, 37].

Five studies noted unintended harms. Study authors discussed that passive and active BRAC members (receiving savings only and savings and credit, respectively) reported increased odds of IPV compared to skilled members (receiving savings, credit, and training). They interpreted this as time-bound risk, which dissipates with longer participation, as women exposed to more intervention components also had participated longer in the intervention [24]. Oportunidades was associated with increased threats of violence [25]; however, this change was not statistically significant. In Uganda, the introduction of gender and couples training to cash transfers and microenterprise training improved couples’ relationship quality without increasing endorsement of gender equitable norms, financial autonomy, or economic success among women or reducing IPV prevalence [27]. The authors interpreted these findings positively, suggesting the couples-focused intervention may have initiated a process of social learning, beginning with improved relationship quality, that might result in increased financial autonomy and decreased use of violence [27]. They also noted, however, that men may have learned new ways to establish control in marital relationships, “[influencing] their female partners … by spending time with them, talking to them, and persuading them to do what they want” (p. 187) [27]. BDH demonstrated disparate intervention effects where absolute and relative inequities intersected; women with less than six years of schooling and education levels equal to or more than their partners had greater odds of experiencing IPV at follow-up at the p < 0.1 level in stratified analyses [29]. Additionally, female participants exposed to the combination of Stepping Stones and Creating Futures reported increased heavy drinking over time, which the authors linked theoretically to their rising incomes [32].

Risk of bias assessment revealed limitations affecting the quality and generalizability of findings (Tables 4 and 5). Reflecting limitations in study design or reporting bias, four of 13 articles describing RCTs [6, 27, 28, 42] reported calculating and enrolling a sufficiently large sample to support statistical inference for IPV outcomes. Only one of these 13 articles described allocation concealment [6], and no authors clearly identified blinding of the outcome assessment. All primary outcome analyses (Additional file 3: Table S3) controlled statistically for potential confounders [6, 27, 29‐31, 33‐35, 37, 38, 40, 42] or confirmed successful randomization by analyzing selected indictors at baseline and follow-up [28]. Nine of 13 articles presented intention to treat analyses; clear descriptions of this approach were missing from articles reporting IMAGE [33, 34, 37] and the Male Norms Initiative in Ethiopia [40]. Although all studies, except two [35, 40], appeared to report each outcome described in the methods among study results, only five articles clearly stated that implementation and analyses proceeded independently of funders [28, 33, 34, 37, 42].

Observational studies, by definition, defined comparison groups retrospectively by exposure [24‐26, 41]. An interrupted time series and before-after evaluation did not include comparison groups, but measured changes among those assigned or exposed to the intervention [32, 39]. No articles reporting non-RCT study designs described a sample size calculation to ensure sufficient statistical power [24‐26, 32, 36, 39, 41], although all studies, except one [24], clearly described equivalent eligibility criteria in intervention and comparison groups. Studies without comparison groups did not control statistically for confounding [32, 39]; in one article, this appears to be consistent with the stated intention to conduct a proof of concept study, rather than an impact evaluation [32]. Two studies similarly did not report each outcome described in the methods among the results [39, 41], and none of the authors clearly stated that implementation and analyses were conducted independently of funders.

Across study designs, many articles did not report adherence to IPV research recommendations fully. Among 16 articles that measured IPV, four did not report restricting IPV questions to persons who are at least 15 years old [27, 29, 35, 37]; eight did not clearly report IPV-specific training for interviewers [24, 25, 29, 31, 32, 35, 39, 40]; and 10 did not report developing referral information or protocols to provide support for persons disclosing IPV [24, 25, 27, 29, 31, 32, 34, 37, 39, 40]. The extent to which reporting limitations suggest intervention or study limitations, however, is unknown across all measures of bias.

As the first review to apply the concept of structural interventions to IPV prevention, this review is subject to limitations arising from the challenge of consistently and rigorously defining structural interventions and differentiating them from non-structural approaches. We selected comprehensive search terms and applied definitions consistently; however, more work is needed to develop the concept of structural interventions in the IPV literature. Similarly, although we used a combination of search strategies to capture relevant literature, our search may have omitted relevant studies. Other studies were omitted because we focused on the peer-reviewed literature, as noted previously. Further, although focusing on quantitative measures increased comparability across studies, sufficient data are not available yet for quantitative synthesis or to compare effectiveness by intervention domain, context, or population, including addressing competing theories of economic empowerment and IPV. Understanding “what works, for whom, and in what situations” additionally likely requires varied evaluation strategies [7], including qualitative and sub-sample analyses that were beyond the scope of this review. Answering questions about research methodology is imperative both to understanding the limitations of this review and the impact of structural interventions for IPV broadly.