School-based interventions TO Prevent Dating and Relationship Violence and Gender-Based Violence: STOP-DRV-GBV systematic review

Dating and relationship violence and gender-based violence: joint constructs

Dating and relationship violence (DRV) refers to physical, sexual and emotional violence (including coercive control) in relationships between young people. Critically, it includes those aged under 16, who are not included in UK government definitions of domestic violence. Gender-based violence (GBV) refers to violence rooted in gender inequality and sexuality, such as harassment or bullying on the basis of gender or sexuality, sexual violence, coercion and assault including rape, within or outside dating relationships. 1 Although longitudinal evidence demonstrates that experience of DRV predicts young people’s later GBV victimisation and that they share common risk factors,2,3 they are rarely considered as joint constructs. 4

Previous systematic reviews of interventions for young people have focused on DRV and have not meaningfully considered intervention impacts with GBV. 5–8 This is important because interventions nominally focusing on DRV may impact GBV and vice versa, underpinned by common mechanisms and structural features that lead to high rates of both in schools. Potentially shared aetiological mechanisms include gender norms at the societal level that are inequity-generating (i.e. creating unavoidable and unfair differences between groups); inconsistent development and enforcement of violence prevention policies at school and classroom levels, and the creation of spaces for violence to occur; and, at the individual level, exposure to and reinforcement of antisocial norms relating to gender, sexuality and violence and, conversely, insufficient exposure and reinforcement to prosocial norms relating to the same. 4,9–11 These multilevel influences are experienced across sexual and reproductive health, of which DRV and GBV are important determinants. 12 That is, DRV and GBV have a common basis in exploiting gender and sexual inequalities and in antisocial norms. 13

Dating and relationship violence and GBV are amenable to change in school contexts via a range of approaches, from didactic (e.g. classroom-based) to structural (e.g. school policy changes). 1 Thus, considering DRV and GBV jointly is essential to develop, implement and adapt efficient interventions for schools.

Dating and relationship violence and gender-based violence: public health issues

Dating and relationship violence and GBV are pressing public health problems with manifold and inequity-generating long-term impacts on health. While boys and girls both experience major burdens of emotional and physical DRV, impacts are disproportionately experienced by girls. A cross-sectional study based on a representative sample of 11- to 16-year-olds in Wales found that 28% of girls report emotional victimisation and 12% report physical victimisation by a partner over the course of adolescence, while 20% of boys report emotional victimisation and 17% physical victimisation. 11 However, age-related trajectories in victimisation are steeper in girls than in boys, suggesting that adolescence is a critical period to arrest inequalities arising from DRV and GBV. Girls are more likely than boys to report sexual dating violence victimisation,14 and girls are more vulnerable to experiencing the detrimental impacts of victimisation such as feelings of fear, distress and post-traumatic stress compared to boys. 15 Nationally representative data from the US Youth Risk Behavioral Surveillance Survey from 2001 to 2019 revealed that rates of forced sex were maintained for girls and decreased for boys, and as girls and boys aged, the risk of forced sex increased. 16 In Britain, the median age for most recent occurrence of sex against one’s will, a form of GBV, is 18 among men and 16 among women. 17 In addition, longitudinal evidence suggests that the onset of physical DRV and GBV peaks in mid-adolescence, while onset of sexual DRV and GBV is greatest in late adolescence. 18 This underscores the importance of interventions during school years.

Young people do not report perceived peer sanctions against GBV behaviours,1 and at an individual level, norms accepting of GBV and harassment strongly correlate with DRV perpetration and victimisation, reinforcing the importance of considering these outcomes jointly. 1,13,19,20 In 2021 in the UK, 67% of girls aged 13–18 years reported sexual harassment at school or college; 18% experienced unwanted touching such as being pinned down, having their bra strap or skirt pulled. This increased from 9% for those aged 13–16 years, to 36% for those aged 17–18 years. 21 While it is well understood that sexual minority adolescents experience higher levels of GBV in terms of homophobic and transphobic bullying and sexual harassment,22,23 these adolescents also experience higher rates of physical and sexual DRV. 23–25

Longitudinal impacts of DRV and GBV are numerous. In adolescence, both perpetrators and victims report increased risky sexual behaviour, substance use and depressive symptoms;8,10,23,26 in adulthood, survivors of DRV and GBV are more likely to be re-victimised27 and more likely to report poorer mental and physical health. 28 In particular, a systematic review of longitudinal studies found that both DRV and GBV experiences as adolescents were predictive of adult experiences of domestic violence. 29 Another way in which DRV and GBV are inequity-generating is in their exacerbation of health inequalities between men and women;30 in particular, earlier onset of intimate partner violence leads to greater impacts on mental and physical health in adulthood. 28 In addition, there are strong intersections with other inequalities, such as race/ethnicity and sexuality. 22,23 DRV and GBV generates inequalities between heterosexual and cisgender young people and their sexual minority peers, such as substance misuse and increased burden of suicidal ideation arising from experiences of DRV and GBV. 22,23 Importantly, a key source of these inequalities in mental health is the shared impact of school context, including prevalence and response to DRV and GBV, both of which point to the importance of school-based interventions. 31

School-based interventions for dating and relationship violence and gender-based violence in the UK

Adolescence is a crucial stage for focusing on the prevention of DRV and GBV. Schools, as important sites of gender socialisation (both in the UK and worldwide), have the potential to promote gender-equitable attitudes. In addition, schools are a critical site for the delivery of universal interventions. However, despite their role in establishing prosocial norms and behaviours and their duty of care to prevent violence between pupils, a significant amount of DRV and GBV occurs in schools (e.g. Ofsted 202132). According to a recent report on sexual abuse in schools in England,32 sexual harassment including online sexual abuse is normalised for children in schools and incidents of sexual harassment are so commonplace that they see little point in reporting them. Herbert et al. (2021)33 observe that it is only relatively recently that there has been a sustained UK public health focus on DRV in younger people,10,11,34 with the first trial of a school-based intervention to prevent DRV in young people, Project Respect,35 trialled in the UK.

UK policy context

Importantly, all of this is directly relevant to UK policy and practice. The Department for Education released the final guidance for statutory RSE requiring all schools in England to deliver statutory Relationships Education (RE) in primary schools and RSE in secondary. It came into effect in September 2020, but schools could start using it from September 2019. Those schools not in a position to implement in September 2020 because of the COVID-19 pandemic and closures, had until summer term 2021 to implement. Implementation analysis in schools that adopted the RSE curriculum early shows that challenges were encountered when developing and delivering the RSE curriculum. 50 According to a recent review of sexual abuse in schools in England, children were seldom positive about their RSE lessons and most felt the curriculum did not give them the information and advice they needed to navigate the reality of their lives. 32 This review also highlighted the normalisation of sexual harassment including online sexual abuse in schools for children and recommended that school leaders should take a whole-school approach (WSA) to create a culture where sexual harassment and online sexual abuse are not tolerated, and where they identify issues and intervene early to better protect their students. In light of these findings and in order to fulfil their statutory duties in relation to RSE, schools are likely to need to develop and implement programmes that address DRV and GBV prevention and improve responses to these outcomes.

Thus, there is a good rationale for a new systematic review that provides usable information for practitioners and policy-makers regarding the design and implementation of school-based interventions for DRV and GBV prevention. By synthesising which intervention characteristics are most important for preventing DRV and GBV, what kinds of issues implementers are likely to face, and how interventions are most likely to function in local contexts can support local decision-making and commissioning. Policy and practice stakeholders and young people consulted in preparation for this review identified the need for evidence that could be used to select, develop and implement locally relevant interventions. For example, the statutory guidance on RSE noted that in secondary school, students should understand ‘how stereotypes, in particular stereotypes based on sex, gender, race, religion, sexual orientation or disability, can cause damage (e.g. how they might normalise non-consensual behaviour or encourage prejudice)’ and ‘what constitutes sexual harassment and sexual violence and why these are always unacceptable’. 51 Despite this guidance, policy and practice stakeholders noted a lack of clarity or clear evidence as to the best ways to achieve these goals. These stakeholders also noted the importance of including sexuality-based bullying as part of GBV, especially in light of the Government Equalities Office campaign against homophobic, transphobic and biphobic bullying.

From a research perspective, syntheses that go beyond intervention effectiveness can shape forward development and implementation of best bet interventions against developing evidence of heterogeneity in effectiveness. 43 This review is especially timely given our experience with the recently concluded National Institute for Health and Care Research (NIHR)-funded pilot trial of Project Respect, a school-based DRV prevention intervention. The first UK trial of an intervention to prevent DRV in young people, Project Respect,35 initially intended to adapt two related interventions found to be effective in the USA: Safe Dates36 and Shifting Boundaries. 40 The adaptation and optimisation process suggested that it would be unlikely to yield an effective intervention in UK school settings. Thus, a systematic review that ‘deconstructs’ existing interventions to better understand intervention functioning would be an essential starting point to develop a relevant evidence base on effective strategies for DRV and GBV prevention specifically in UK school contexts. Indeed, a key finding of Project Respect was that reconsidering types and combinations of components would be critical before undertaking further intervention development.

Finally, the lack of consideration of health inequalities in the evidence base is a major gap as yet unaddressed in a systematic review. This information is vitally important to prevent implementation of interventions that may exacerbate inequalities in respect of a problem already marked by exceptionally large social gradients in long-term impact.

In sum, this mixed-method systematic review including both DRV and GBV as eligible outcomes aims to:

• address inconsistencies across prior systematic reviews;

• ameliorate gaps in the understanding of GBV alongside DRV;

• consider the evaluability of the evidence base across multiple types of evidence;

• generate timely, relevant and innovative evidence for policy, practice and research.

Types of population

We included studies with children in compulsory education (e.g. aged 5–18 years) who are attending school.

Types of intervention

Guided by our logic model, we included evidence relating to interventions implemented in school contexts with students separate from, or as part of, RSE. These interventions could include one or more of:

• individual behavioural intervention (e.g. individual learning modules or apps);

• group or classroom-based intervention or practices (e.g. as part of RSE; delivering DRV and GBV prevention content in other academic sessions;44 delivery of content in groups during school hours);

• network-based approaches, such as public opinion leader interventions;

• staff training and other service provision in schools (e.g. to recognise and respond better to sexual violence34);

• local and school policy change56 to address structural factors relating to DRV or GBV, or to change school responses to DRV or GBV.

Interventions could be single-component or multicomponent, or implemented the same type of approach (e.g. group or classroom-based intervention) in a range of ways, for example by differentiating instruction over a range of school years. Included interventions focused in whole or in part on DRV and GBV, and could be universal, selective or indicated; could be primary prevention (reducing incidence of DRV and GBV) or secondary prevention (improving responses to DRV and GBV); and could focus on gender-specific groups (e.g. boys or girls only).

We excluded interventions that:

• did not seek to address DRV and GBV outcomes, for example interventions focusing on another health promotion topic, such as healthy eating, with an ‘opportunistic’ effect on DRV or GBV outcomes, but that did not describe prevention of DRV or GBV in intervention descriptions;

• were not delivered in compulsory education (e.g. university-based sexual violence prevention, or youth services);

• were not delivered at least in part in school contexts.

Types of control

Comparators included business as usual, waitlist control or another active intervention.

Types of outcome

We included outcomes relating to the full scope of DRV and GBV behaviours. These included:

• DRV perpetration or victimisation, including physical violence; emotional violence, including isolation; coercive control, including internet-mediated DRV; sexual assault in the context of relationships;

• GBV perpetration or victimisation, including harassment and bullying on the basis of gender or sexuality, including homophobic and transphobic bullying; internet-mediated GBV, such as unwanted sexting or forwarding of sexts; unwanted sexual contact, such as groping; sexual assault; sexual harassment and rape;

• DRV and GBV-related behaviours, such as harm reduction behaviours, help-seeking behaviours and bystander behaviours;

• knowledge and attitudes related to DRV and GBV, such as rape myth acceptance, bystander attitudes and GBV-condoning norms.

Outcomes included self-reported behaviours or experiences (e.g. were you groped in the last year; did you call someone names because of their sex or because you thought they were gay), teacher-reported behaviours (e.g. how many times did you see students engaging in sexual harassment) or official reports (e.g. how many sexual harassment incidents were reported in the last year). Included outcome measures were quantitative, expressed as categorical, continuous or count measures. Measures could be composite items (a range of DRV behaviours collected as a count of behaviours) or may be behaviour-specific. Behavioural outcomes could focus on behaviours over a specific period; frequency (monthly, weekly or daily); the number of episodes of a behaviour; or an index constructed from multiple measures. Economic analyses could also include health-related quality of life. We excluded knowledge and attitude outcomes relating to gender, or violence norms generally.

We did not include evaluations where outcomes related only to honour-based violence, forced marriage or female genital mutilation as these outcomes are potentially less amenable to curriculum-based school-based interventions.

Types of study

Types of study included are categorised by research question.

• For RQ1, we used intervention descriptions and descriptions of theories of change across all included evidence.

• For RQ2, we drew on process and implementation evidence from eligible interventions that examined intervention delivery or receipt and how delivery was influenced by provider, user or context characteristics. This evidence could be reported in randomised or single-arm trials, retrospective or prospective evaluation studies, and evidence could be qualitative (e.g. description of acceptability of interventions) or quantitative (e.g. measurement of intervention fidelity).

• For RQ3 and RQ5, we drew on randomised trials, including cluster trials. We also sought any economic evaluations or modelling studies linked to these trials; that is, evidence that seeks to relate intervention costs and savings to health and well-being outcomes or benefits. Finally, we included moderation or subgroup analyses linked to these trials that explore equity-relevant characteristics.

• RQ4 and RQ6 drew on all included evidence.

Database search strategy

In July 2020, the following bibliographic databases were searched from inception and without limitation on date, language or publication type.

• MEDLINE, EMBASE, PsycINFO, Social Policy and Practice (Ovid).

• Cumulative Index to Nursing and Allied Health Literature, Education Resources Information Center, British Education Index, Education Research Complete, EconLit, Criminal Justice Abstracts (EBSCO).

• Cochrane Database of Systematic Reviews and the Cochrane Central Register of Controlled Trials.

• NHS Economic Evaluation Database (via the Centre for Reviews and Dissemination).

• Social Science Citation Index and Conference Proceedings Citation Index (Web of Science, Clarivate Analytics).

• Australian Education Index, ProQuest Dissertations and Theses Global, Sociological Abstracts including Social Services Abstracts, Applied Social Sciences Index and Abstracts (ProQuest).

• Trials Register of Promoting Health Interventions and Bibliomap [Evidence for Policy and Practice Information (EPPI)-Centre].

• Campbell Systematic Reviews (Campbell Collaboration).

The search strategies were developed by an experienced information specialist following extensive scoping searches and consideration of existing systematic review search strategies. 6,8 The Ovid MEDLINE search strategy was peer-reviewed by another experienced information specialist. The search strategy includes both free-text terms and subject headings for the school setting and DRV/GBV outcomes. In order to identify studies to answer review questions regarding outcome and economic evaluations, intervention theory, process and implementation evidence, and mediation and moderation evidence, no filters for specific study designs were applied. Searches were not restricted by date or language of publication. Full search strategies for each bibliographic database are available in Report Supplementary Material 1.

We searched trial registers to identify ongoing or unpublished research [www.clinicaltrials.gov, WHO International Clinical Trials Registry Platform (ICTRP)] and we conducted searches for grey literature including conference abstracts, reports and theses from web searches, as well as searches of websites identified in initial scoping searches (including VAWnet; www.vawnet.org).

The bibliographic database searches were updated in June 2021, with a revised strategy developed to improve precision. The revised strategy was based on analysis of titles, abstracts and index terms of included studies. The updated search strategy also incorporated programme names not identified in the initial scoping searches. We created a search summary table for all included studies to identify database sources that retrieved unique records, and this informed bibliographic database selection for update searches. Results from the update searches were not date limited, but instead de-duplicated against previous result sets to ensure records were not missed. Full search strategies for the update searches are available in Report Supplementary Material 1.

Search strategy for other literature sources

Supplementary search methods were used to identify studies not captured by our sensitive database strategies. The reference lists of existing systematic reviews or relevant reports were reviewed for relevant literature. This was the core of our cluster-based approach57 to capture theory underpinning evaluated interventions and identify any missed ‘sibling’ studies. Forward and backward citation chasing was conducted on included studies identified from the June 2020 bibliographic database searches. Scopus (Elsevier), Web of Science (Clarivate) and Google Scholar were used for citation chasing, and bibliographies of included studies were manually checked where this information was incomplete on Web of Science and Scopus.

Targeted searches were conducted in Web of Science and Scopus using first and last author names for studies identified in bibliographic database searches in June 2020. Specific project names (e.g. Project Respect, Shifting Boundaries or Safe Dates) were included in the update search strategies in bibliographic databases, and Google Scholar. Results were screened in Google Scholar, with the first 200 records scanned for each search string. Websites identified in initial scoping searches were browsed or searched for additional reports (including VAWnet: www.vawnet.org; USAID: www.usaid.gov; AVA – Against Violence and Abuse; UNGEI, National Criminal Justice Reference Service: www.ncjrs.gov). Full details of Google Scholar and website searches are available in Report Supplementary Material 1.

Finally, we planned to hand-search journals that published included studies which we found only via reference checking and which are not indexed on databases we have searched. However, due to the exhaustive cluster-based search methods, this was unnecessary.

Assessment of process evaluations

Process evaluations were appraised using the EPPI-Centre tool. 59 This addressed the rigour of sampling; data collection; data analysis; the extent to which the study findings are grounded in the data; whether the study privileges the perspectives of participants; the breadth of findings; and depth of findings. These assessments were then used to assign studies to two categories of ‘weight of evidence’ (low, medium or high) to rate the reliability or trustworthiness of the findings (the extent to which the methods employed were rigorous/could minimise bias and error in the findings), and to rate the usefulness of the findings for shedding light on factors relating to the research questions. Guidance was provided to reviewers to help them reach an assessment on each criterion and the final weight of evidence. Findings from critical appraisal were used to inform synthesis, including by describing the qualitative strength of findings in evidence syntheses. The tool was used in this review to assess process evaluations regardless of whether these drew on qualitative or quantitative data, because the criteria were judged applicable to both.

We chose the EPPI-Centre tool given our prior experience with this tool in other NIHR-funded systematic reviews of qualitative research. 39,44,59 A key strength of this tool is that it generates appraisal in terms of both study relevance and study trustworthiness, both of which are important in qualitative evidence synthesis. In addition, the tool reflects the degree to which qualitative findings privilege participant voices, which is important given our focus on generating evidence that speaks to local implementers’ needs.

A final step in quality assessment was to assign studies two types of ‘weight of evidence’. First, reviewers assigned a weight (i.e. low, medium or high) to rate the reliability or trustworthiness of the findings (i.e. the extent to which the methods employed were rigorous/could minimise bias and error in the findings). Second, reviewers assigned an additional weight (i.e. low, medium or high) to rate the usefulness of the findings for shedding light on factors relating to the review questions. The criteria guiding these decisions are shown in Table 1. Quality was used to determine the qualitative weight given to findings in our synthesis, with none of the themes represented solely by studies judged as low on both dimensions. This is relevant because our experience with systematic reviews of process evaluations is that studies relating to exceptionally relevant interventions may not provide rich data; conversely, process evaluations of interventions that only address knowledge and attitudes, rather than behaviours, may provide especially meaningful data that illuminate both how interventions function and how interventions were implemented. Thus, distinguishing between these two weights of evidence sensitised our analysis and helped in prioritising most relevant and most trustworthy findings in synthesis.

Assessment of outcome evaluations

Critical appraisal of the cluster randomised controlled trials (cRCTs) and RCTs was conducted using the Cochrane Risk of Bias Tool version II. 60 For each study, reviewers judged the likelihood of bias for sequence generation and allocation concealment; blinding (of participants, personnel, or outcome assessors); incomplete outcome data; selective outcome reporting; other sources of bias (e.g. recruitment bias in cluster-randomised studies); and intensity/type of comparator. Each study was subsequently identified as ‘high risk’, ‘low risk’ or ‘unclear risk’ within each domain. Decisions were guided by the guidance associated with the tool,53 with the following review-specific adaptations:

• Where trials did not report baseline characteristics of the included clusters or arms, it was assumed that there were no confounding factors unless the reviewer was concerned that the sample size would undermine effective randomisation.

• It was assumed that students would likely guess the aims of the intervention, unless evidence to the contrary.

• There appeared to be a low rate of switching in the trials, but there was limited information about the risk for other protocol deviations that may lead to performance bias. The following were not downgraded for performance bias if there was no switching and an intention to treat (ITT) analysis was reported, as the risk of contamination was considered low:

  • cluster trials where schools were the unit of randomisation

  • interventions lasting one class only.

• An absolute participant attrition rate > 30% was considered problematic.

• A scale outcomes measure was judged to be acceptable where items were considered by the reviewer to have face validity, and if internal reliability (generally the only psychometric outcome reported) was ≥ 0.7.

• Trials were not downgraded for unblinded outcome assessors within the outcome measurement domain. Due to the nature of the interventions evaluated, it was not possible for the vast majority of trials to fully blind or obscure students from the intervention they were receiving. This was a limitation noted by many trial authors, who considered there to be a risk that outcomes completed by students could be subject to desirability bias, where students complete outcomes in a way they think is desired by researchers. Trials were not downgraded for this to avoid a floor effect in critical appraisal ratings across the evidence base. However, all trials included in the review are nevertheless considered to be at risk of this bias, in addition to the overall risk of bias across the other tool domains.

Finally, we assessed reporting bias in trials according to Cochrane Handbook guidance. 53 We reduced the effect of reporting bias by focusing synthesis on studies rather than publications and avoiding duplicated data. We attempted to detect duplicate studies and, if multiple articles report on the same study, we extracted data only once. We prevented location bias by searching across multiple databases. We minimised language bias by not excluding any article based on language.

Assessment of economic studies

No formal health economic evaluations were identified by the review. Were we to have identified any eligible economic evaluations and modelling studies linked to randomised trials, we would have appraised them using the Drummond61 or Philips62 checklists, respectively. These checklists require the analyst to answer 24 questions regarding each study, ranging from the type of economic evaluation (e.g. cost–utility analysis) to the time horizon and rationale for the choice of modelling approach.

Research question 1: theories and components

We used intervention components analysis63 to analyse intervention descriptions across all included evidence. Intervention components analysis is an inductive approach to comprehensively describing and categorising intervention components in a target body of evidence. This is an appropriate method to describe intervention components when these components do not fit into pre-existing taxonomies of behaviour change, which is especially the case in this review given the diverse, contextually situated and frequently multilevel nature of included interventions. Two reviewers used open coding to generate a comprehensive list of possible intervention descriptors from five different intervention descriptions. The two lists were compared and combined. Using principles of axial coding, the two reviewers proceeded through the remaining intervention descriptions, collapsing codes and adding new ones as required and meeting periodically to compare codes, determine if new axial codes were required and organise axial codes into categories. The final result was a comprehensive list of descriptors to characterise included interventions, organised by relevant categories. Our intervention components analysis ultimately included a set of descriptors for included interventions (type, level, focus), which was analysed across all included evidence, and a set of descriptors for intervention components specifically, which was applied only to outcome evaluations given the size of the evidence based and differences in depth of intervention description.

We also synthesised theories of change based on intervention descriptions and accounts embedded in individual studies’ theories of change. Drawing on methods used in previous theory syntheses,64 this inductive analysis included a lines-of-argument synthesis. 65 Lines-of-argument synthesis is an appropriate method based on its understanding that each included study examines a ‘part of the whole’; that is, each study’s account of theory of change represents one possible part of how school-based interventions can work to reduce DRV and GBV broadly. Analysis was undertaken using thematic synthesis and line-by-line coding. 66 We intended to analyse interventions within type first; however, this was not done as it would have limited the breadth of interventions analysed in initial coding, and ultimately did not generate any analytical advantages. Two reviewers began the theory synthesis with a subset of studies first to understand and agree a common approach, prioritising a selection of interventions that (a) were multiply evaluated and (b) represented the breadth of interventions. Subsequently, both reviewers proceeded through each set of studies, meeting to agree findings. The findings from each intervention (second-order constructs) were then compared against each other using a ‘higher-level’ iteration of these thematic grids to produce an overarching theory of change (third-order constructs). Coding unfolded over several cycles, proceeding from initial coding to saturation, followed by confirmatory coding on remaining papers in the analysis. The outcome of this synthesis was a hierarchically organised account of the mechanisms by which interventions are theorised to function in reducing and preventing DRV and GBV and the contexts in which these mechanisms are most likely to operate, broadly across interventions and specifically within intervention types. Specific methods for how the analysis ‘unfolded’ are presented in Chapter 6.

For both the intervention components analysis and the synthesis of theories of change, two reviewers undertook analysis in parallel, with periodic meetings among the analysis team to validate the developing framework. We additionally checked findings with our advisory and stakeholder groups for face validity. These checks focused on identifying which intervention components are most promising in the UK context on the basis of relevance and fit with existing school curricula, intervention strategies and cultures, and which contexts and mechanisms are most likely to be relevant to the UK educational system.

Research question 2: implementation and process evaluations

We synthesised data from implementation and process evidence to identify salient factors affecting intervention implementation and how these relate to context. We used thematic synthesis67 to synthesise these data. Thematic synthesis is apposite as it is a flexible method that seeks to understanding findings across studies, without necessarily imposing a theoretical framework or ‘theorising’ the data, while still preserving the value of reciprocal translation65 in understanding patterns of findings across studies rather than merely summarising these. Thematic synthesis includes both descriptive, or in vivo, themes that describe the specific content in included studies, and analytical themes that cut across findings from multiple studies.

Synthesis of implementation and process evidence began with a tabulation and descriptive coding phase. First, reviewers created a table of included studies reporting information on study type, methods, context and sample; interventions evaluated; and summarised findings. Two reviewers working in parallel then undertook pilot analysis of five reports that were appraised as being of high quality and relevance. The reviewers read and reread the results from these reports, applying line-by-line codes to capture the content of the data, and drafted memos explaining these codes. Coding then proceeded with descriptive codes which closely reflect the words used in findings sections. The reviewers then grouped and organised codes, applying analytical codes reflecting higher-order themes. Subsequently, the reviewers coded remaining studies drawing on the agreed set of codes but developing new descriptive and analytical codes as these arose from the analytical process, and again writing memos to explain these codes. At the end of this process, the two reviewers met to compare their sets of codes and memos. They identified commonalities, differences of emphasis and contradictions with the aim of developing an overall analysis. The outcome of this synthesis was an account of barriers and facilitators that intervention implementers could match to their own context. We sensitivity-analysed all findings by considering whether findings relate to high-income country (HIC) contexts or low-income and middle-income country (LMIC) contexts, in order to better understand the applicability of findings to the UK context.

In addition to routine auditing of findings by the investigator team, we presented findings to our advisory and youth stakeholder groups for feedback. The goal of this was to ensure that findings are relevant to intervention implementers in the UK, and to identify which findings are especially salient in the UK education context.

Research question 3: intervention effectiveness, moderation, mediation and economics

Included trials were organised by intervention type as informed by findings from RQ1, and relevant outcomes were hierarchically categorised by type of DRV or GBV behaviour or experience. We distinguished between different intervention follow-up times (up to 1 year from baseline; more than 1 year from baseline). All time points reported by a study were extracted.

We then meta-analysed study findings where possible. We first analysed within intervention and time-point categories to estimate specific intervention effects on broad outcome categories and then meta-analysed across intervention categories and across outcome categories. Meta-analyses then considered different behaviours, experiences and forms of perpetration and victimisation separately across all intervention types. We originally sought to meta-analyse by intervention type, specific outcome (e.g. physical DRV victimisation) and time point but found that this would be inappropriate given the small number of studies included in any one meta-analysis.

We sensitivity-analysed meta-analyses by whether evidence arose from high-income settings or low-income or middle-income settings, using a standard subgroup difference-based test to identify any statistical differences in effectiveness that may shape the relevance of overall results to the UK context.

The key metric for all meta-analyses was defined in the protocol to be odds ratios (ORs); we ultimately used a standardised mean difference to analyse knowledge and attitude outcomes (i.e. not victimisation or perpetration outcomes). Where victimisation and perpetration outcome measures were continuous, these were converted to ORs using the logistic transformation. We used a random-effects robust variance estimation meta-analysis model to synthesise effect sizes. This is because many outcome evaluations included multiple measures of conceptually related outcomes. Robust variance estimation meta-analysis improves on previous strategies for dealing with multiple relevant effect sizes per study, such as multilevel meta-analysis, meta-analysing within studies or choosing one effect size, by including all relevant effect sizes but adjusting for inter-dependencies within studies. 68 Unlike multivariate meta-analysis, it does not require the variance–covariance matrix of included effect sizes to be known. Where meta-analyses were performed, we included pooled effect sizes in forest plots, with the individual study point estimates weighted by a function of their precision.

We checked that cluster randomised trials have accounted for unit of analysis issues. Prior to synthesis, we checked for correct analysis (where appropriate) by cluster and reported values of ICCs, cluster size, data for all participants or effect estimates and standard errors (SEs). Where proper account was not taken of data clustering, we corrected for this by inflating the SE by the square root of the design effect. 58 Where ICCs were not reported and were relevant to estimating an intervention effect, we imputed an ICC of 0.05, based on values reported in other studies. We planned to undertake sensitivity analysis with respect to ICC values, but this was ultimately unnecessary given that the ICC value was ‘known’ from included studies.

We then produced a narrative account of intervention effectiveness. This narrative account included a comprehensive table of included studies reporting information on study design, methods, context and sample, interventions evaluated and summarised findings, alongside forest plots to describe visually the range and precision of estimates of intervention effectiveness. We then examined the extent of heterogeneity among the studies (as determined both by a Cochran’s Q test and by inspection of the I²). If an indication of substantial heterogeneity was determined (e.g. study-level I² value > 50%) that cannot be explained through metaregressions, we investigated this further using subgroup and sensitivity analyses.

We used the Grading of Recommendations, Assessment, Development and Evaluation (GRADE) approach as described in the Cochrane Handbook for Systematic Reviews of Interventions53 to present the quality of evidence, focusing on primary victimisation and perpetration outcomes. The downgrading of the quality of a body of evidence for a specific outcome was based on five factors: limitations of study; indirectness of evidence; inconsistency of results; precision of results; and publication bias. The GRADE approach specifies four levels of quality (high, moderate, low and very low). If sufficient studies are found, we drew funnel plots to assess the presence of possible publication bias (trial effect vs. SE). While funnel plot asymmetry may indicate publication bias, this can be misleading with a small number of studies. 69 We considered possible explanations for any asymmetry in the review in light of our number of included studies. We assessed the impact of risk of bias in the included studies via restricting analyses to studies deemed to be at low risk of selection bias, performance bias and attrition bias; however, risk of bias in included studies was generally poor on these domains and thus a restricted analysis would have been uninformative.

Next, we planned to produce a narrative account of findings from economic evaluations by intervention type. However, as we did not identify any relevant economic evaluations, we instead provided a narrative summary of costing studies linked to included randomised trials. Measures of costs and indirect resource use were summarised using tables. These data were used to inform a narrative synthesis of economic analyses and applicability to the UK context. Because no full economic evaluations were identified, we did not appraise economic studies.

Subsequently, we narratively synthesised evidence of mediation of intervention impacts to demonstrate the degree to which DRV and GBV victimisation outcomes in interventions are mediated by intervening variables.

As a final synthesis step, we assessed equity effects by drawing on moderation analyses to illustrate how interventions impact health inequalities, focusing on ethnicity, socioeconomic position, gender, sexuality and age. Our analysis focused on priority outcomes of DRV and GBV victimisation and perpetration. Moderation analyses considering differential intervention effectiveness on these equity-relevant characteristics were organised by intervention type and outcome category and narratively synthesised. Harvest plots were used to graphically depict how interventions ameliorate or worsen social gradients on specific outcomes by equity-relevant characteristics. 70 Where possible and where sufficient data existed, we extended our random-effects robust variance estimation meta-analyses using metaregression to estimate how equity-relevant characteristics of study populations relate to intervention effectiveness.

We presented the findings from meta-analyses to stakeholders in our advisory and youth groups to understand which interventions are ‘best bets’ in the UK context based on both effectiveness and relevance to the UK educational system.

Research question 4: metaregression and qualitative comparative analysis

The next step of the synthesis integrated findings across RQ1–3 to identify how components, mechanisms and implementation factors relate to effectiveness across DRV and GBV outcomes.

First, we used metaregression and pattern matching to understand which components are most effective across both DRV and GBV. We revisited meta-analyses undertaken in RQ3 and used components identified in RQ1 to estimate how inclusion or exclusion of a specific component is associated with intervention effectiveness. These analyses were undertaken by outcome type and time point, but not intervention type, and were estimated using random-effects robust variance estimation meta-analyses. We considered the magnitude, precision and direction of the regression coefficient associated with presence of an intervention component and the reduction in between-study variance resulting from using an intervention component as a metaregressor. We examined the same meta-regressor across a range of outcomes and used pattern-matching71 to identify how specific components are consistently associated with improved or reduced effectiveness across a range of DRV and GBV experiences and behaviours.

Second and where appropriate, we used qualitative comparative analysis (QCA) to identify how different implementation and intervention characteristics combine to form pathways to effectiveness, including on health inequalities. QCA focuses on understanding configurations of conditions that form pathways to effectiveness in interventions. 72 Given that interventions are likely to report diverse types of outcomes using different measurement approaches, QCA is especially apposite as it transforms numerical estimates into a calibrated measure of whether an intervention is effective. QCA focuses on how different implementation and intervention characteristics act in concert to ‘unlock’ pathways to effectiveness. Informed by findings from RQ1 to 3, we developed several candidate groups of implementation and intervention characteristics and considered how these characteristics form pathways to effectiveness. We coded included outcome evaluations as to the presence or absence of these implementation and intervention characteristics where this has not been done already, and classified outcome evaluations as to their effectiveness. We planned to describe effectiveness in several ways, each corresponding to a separate model:

1. interventions that are effective in reducing DRV versus interventions that are not effective in reducing DRV;

2. interventions that are effective in reducing GBV versus interventions that are not effective in reducing GBV;

3. interventions that are jointly effective on DRV and GBV versus interventions that are only effective in one or neither domain.

However, the overlap in interventions testing both DRV and GBV outcomes was insufficient to proceed with this analysis.

We generated truth tables to understand combinations of characteristics by each outcome, and sought to resolve any contradictory configurations (i.e. where combinations of characteristics span both effective and ineffective interventions). We used Boolean minimisation to identify common pathways to effectiveness across each of the three models, and compared findings across models to note which pathways are specific to DRV or GBV outcomes and which pathways are common to effectiveness of DRV and GBV outcomes, and thus should be considered when seeking to address both types of outcomes jointly.

We presented the findings of this analysis to the advisory and youth stakeholder groups to consider how best bet components are relevant in the UK context, and how identified pathways from QCA may be especially important in shaping forward evaluation and implementation in UK schools.

Research question 5: network meta-analysis

Drawing on randomised trials analysed in RQ3, we used the findings from our analysis of intervention typology and intervention components to estimate network meta-analyses (NMAs). For all NMAs, we considered as separate outcomes DRV victimisation, DRV perpetration, GBV victimisation, and GBV perpetration. We considered our intervention typology analyses as primary and our intervention component analyses as exploratory.

All NMAs were implemented in a frequentist framework drawing on methods proposed by Rücker et al. (2012)73 and Melendez-Torres et al. (2015),74 using random effects for all analyses. Implementation used – network – in Stata for typology NMAs and – netmeta – in R for component NMAs. Where trials reported multiple effect sizes for the same outcome (i.e. different estimates of DRV victimisation drawing on, e.g. emotional, physical, sexual DRV), we assumed that outcomes were correlated with ρ = 0.8, consistent with standard assumptions used in robust variance estimation meta-analyses. Analyses were checked for inconsistency statistically using standard design-by-treatment interaction models, and conceptually by considering transitivity with regard to key effect modifiers (principally age, sex, high-income vs. low- and middle-income context). Where necessary, we used network meta-regression to account for plausible effect modifiers unevenly distributed between network comparisons. NMAs were then presented using the standard pairwise comparison league table, with intervention types ranked using a resampling method and surface under the cumulative ranking curve (SUCRA).

Research question 6: realist integration of findings

We used implementation and process evidence, mediator and moderator analyses, and ‘informal’ evidence from included studies (e.g. discussion of how interventions were implemented in outcome evaluation discussion sections) to identify mechanisms by which interventions impact DRV and GBV outcomes and associated contextual contingencies. Mechanisms focused on understanding causal chains by which interventions are likely effective, including proximal and antecedent steps (e.g. increasing engagement, addressing social norms, knowledge and attitudes) leading to distal effects on outcomes. Informed by realist synthesis75 and best-fit framework synthesis,76 reviewers used findings from RQ1 as a framework to infer and induce mechanisms from studies. Best-fit framework synthesis is a flexible form of framework analysis for the synthesis of diverse study types. Importantly, it accounts for the possibility that previously ‘untheorised’ findings may emerge as relevant from included study findings. Our approach is informed by realist synthesis in that we aim to understand the links between contexts and mechanisms in generating outcome patterns.

Using the context-mechanism findings from the theory synthesis in RQ1 as a template, reviewers analysed findings from the subsequent syntheses in RQ1–5, working independently and together to understand which causal propositions from the theory synthesis are supported, refuted or unevidenced. This synthesis unfolded collaboratively through meetings of the reviewer team, with an audit trial kept of synthesis decisions made. Analysis could also generate additional context-mechanism-outcome configurations that were not present in the theory synthesis but that reflected learning from other syntheses.

Findings were periodically discussed in meetings of the investigator group and were presented to the advisory and youth stakeholder groups. These groups provided additional insights in revisiting contexts and mechanisms previously discussed in RQ1, and in identifying which previously untheorised contexts and mechanisms are especially salient in the UK context.

Study designs

Process and implementation evaluations (RQ2)

One hundred and thirty-seven studies reporting process or implementation outcomes were included. Twenty-nine (21.2%)49,77–101 of the studies were associated with a RCT or cRCT, while others were non-randomised, naturalistic or quasi-experimental studies. Eighty-two studies reported quantitative data and 122 studies reported qualitative data.

Many of the included studies reported incomplete information about the sample included in process evaluations. Where sample size was reported or could be estimated, student sample size ranged from 6 to 5937 (median = 85). Studies evaluating interventions to target both DRV and GBV were generally larger (median = 307) than those targeting DRV only (median = 87) or GBV only (median = 44). Those studies including staff in their evaluations reported a sample size ranging from 1 to 1900 teaching staff or other professional stakeholders (median = 26).

Two studies (1.5%)102,103 were conducted solely with children of primary school age (i.e. < 11 years), and an additional five studies (3.6%)104–108 included students of primary school age or their teachers amongst a broader sample. The other studies were conducted in secondary/high school settings.

The majority of studies (n = 75, 54.7%) were conducted in North America, with the remaining split across Europe (n = 32, 23.4%),49,79,88,91,94,102,105–107,109–131 Asia (n = 3, 2.2%),78,85,87 Africa (n = 11, 8.0%),90,132–141 South America (n = 7, 5.1%)84,104,142–146 and Australasia (n = 9, 6.6%). 103,147–154 A breakdown of the specific continents and countries in which studies were conducted is provided in Figure 3 and Table 2.

FIGURE 3.

Location of process/implementation evaluations.

TABLE 2 - Continent/countries of included process evaluations

Location	Number of studies	% of total
Africa	11	8.03
Botswana	1	0.73
Botswana, Malawi and Mozambique	1	0.73
Cameroon, Senegal and Togo	1	0.73
Malawi	1	0.73
Namibia, South Africa, South Sudan, Eswatini, Tanzania, Zambia and Zimbabwe	1	0.73
South Africa	4	2.92
Tanzania	1	0.73
Zimbabwe	1	0.73
Asia	3	2.19
Bangladesh, India, Vietnam	1	0.73
India	1	0.73
Taiwan	1	0.73
Australasia	9	6.57
Australia	8	5.84
New Zealand	1	0.73
Europe	32	23.36
Bosnia and Herzegovina, Croatia, Serbia	1	0.73
England	10	7.30
England and Wales	1	0.73
Germany	1	0.73
Ireland	1	0.73
Netherlands	1	0.73
Portugal	1	0.73
Scotland	1	0.73
Serbia	1	0.73
Slovenia	1	0.73
Spain	7	5.11
Sweden	2	1.46
UK	3	2.19
UK, France, Spain and Malta	1	0.73
North America	75	54.74
Canada	14	10.22
Hawaii	2	1.46
Mexico	1	0.73
USA	58	42.34
South America	7	5.11
Brazil	5	3.65
Colombia	1	0.73
Peru	1	0.73

Outcome evaluations (RQ3)

Of the 68 included studies, 14 were RCTs (20.6%) and 54 were cRCTs (79.4%). The majority of studies (n = 61, 89.7%) compared two intervention arms, though five (7.4%)78,94,155–157 and two (2.9%)100,158 studies evaluated three and four intervention arms, respectively. cRCTs ranged in size from 4 to 151 clusters (median 23).

Sample size (ITT population) of the studies ranged from 34 to 89,707 participants, including all groups (median 816). Trials targeting both DRV and GBV were generally larger (median = 2403) compared to trials targeting GBV only (median = 1306) or DRV only (median = 660).

School settings for the included studies are shown in Table 3. Only one study (1.5%)159 was conducted solely in a primary school setting; rather, the vast majority of included studies were conducted in secondary school settings.

TABLE 3 - School settings of included outcome evaluations

School setting	Number of studies (%)
Primary school only (< 11 years)	1 (1.5)159
Primary and secondary school (≤ 16 years)	3 (4.4)160–162
Secondary school only (11–16 years)	57 (83.8)
Secondary school and sixth form (11–18 years)	4 (5.9)49,84,163,164
Sixth form only	2 (2.9)88,165
Unclear	1 (1.5)166

Where reported, recruitment or data collection for the included studies spanned 1994–2017; there was no clear difference in the age of studies evaluating DRV or GBV.

The majority of studies (n = 42, 61.8%) were conducted in North America, with the remaining split across Europe (n = 9, 13.2%),35,49,79,88,94,164,167–169 Asia (n = 8, 11.8%),78,85,87,170–173 Africa (n = 6, 8.8%)90,155,160,174–176 and South America (n = 3, 4.4%). 84,163 A breakdown of the specific continents and countries in which studies were conducted is provided in Figure 4 and Table 4.

FIGURE 4.

Location of outcome evaluations.

TABLE 4 - Continent/country of included outcome evaluations

Location	Number of studies	% of total
Africa	6	8.82
Kenya	1	1.47
Malawi	1	1.47
South Africa	3	4.41
Uganda	1	1.47
Asia	8	11.76
India	2	2.94
Iran	2	2.94
South Korea	2	2.94
Taiwan	1	1.47
Vietnam	1	1.47
Europe	9	13.24
England	3	4.41
France	1	1.47
Germany	1	1.47
Spain	3	4.41
The Netherlands	1	1.47
North America	42	61.76
Barbados	1	1.47
El Salvador	1	1.47
Haiti	1	1.47
USA	39	57.35
South America	3	4.41
Brazil	3	4.41

Intervention characteristics

Outcome evaluations (RQ3)

Seventy-nine active interventions were evaluated across 68 included studies. A description of the interventions identified is provided in Report Supplementary Material 1.

Seven interventions were identified as having been evaluated across more than one study: Coaching Boys into Men (CBIM) (n = 2),77,180 Expect Respect (n = 2),95,159 IMPower (n = 2),160,174 Media Aware (n = 2),97,98 Safe Dates (n = 2),36,166 Shifting Boundaries (n = 2)100,156 and TakeCARE (n = 2). 38,181 Other interventions were found to share the same title (e.g. ‘dating violence prevention program’), but were not clearly identified as being the same intervention.

The vast majority of studies (n = 65) compared interventions against a control arm, namely an active control (n = 13),36,38,87,92,160,162,165,169,170,176,181–183 usual practice (n = 12),35,83,98,99,174,184–190 waitlist (n = 17),49,79,84,88,89,94,95,97,167,168,172,175,191–195 no intervention (n = 23)77,78,81,82,84,85,90,96,100,155,157–159,161,163,164,171,173,180,196–198 or ‘control’ (n = 1). 199 For the purpose of this review, one intervention was considered an active control, even though the content of the intervention was targeted towards DRV mechanisms. 36 This was because the intervention was almost entirely conducted outside the school setting, and therefore was not relevant for inclusion in the review. Limited information was typically reported about the nature of non-active control interventions, but often students received their usual classes (excluding topics related to DRV or GBV) over a comparable time period.

Only eight studies included a head-to-head comparison (Table 5). This included comparisons with different methods of delivering the same intervention (e.g. Sabella 1995),158 and comparison of different intensity levels of the same intervention (e.g. Taylor et al. 2017). 156

TABLE 5 - Head-to-head comparisons evaluated for outcome (efficacy)

GEA, group education activities.

Author, date	Comparison	Intervention 1	Intervention 2	Intervention 3
Achyut 201178	GEA + CAMPAIGN vs. CAMPAIGN vs. no intervention	GEA + CAMPAIGN	CAMPAIGN	–
Jewkes 2019155	Skhokho vs. Skhokho + caregivers vs. no intervention	Skhokho	Skhokho + caregivers	–
Muck 201894	Scientist-Practitioner Program vs. Practitioner Program vs. control	Scientist-Practitioner Program	Practitioner Program	–
Niolon 2019166	Dating Matters vs. Safe Dates	Dating Matters	Safe Dates	–
Sabella 1995158	Sexual harassment classes: Peer-led vs. Adult-led vs. Self-led vs. control	Peer-led	Adult-led	Self-led
Taylor 2010200	Interaction curriculum vs. law and justice curriculum vs. control	Interaction curriculum	Law and justice curriculum	–
Taylor 2011100	Shifting boundaries: Building + classroom vs. building only vs. classroom only vs. control group	Shifting boundaries: Building + classroom	Shifting boundaries: Building only	Shifting boundaries: Classroom only
Taylor 2017156	Shifting boundaries: 6th grade only vs. 6th and 7th grade vs. 6th, 7th and 8th grade	Shifting boundaries: 6th grade only	Shifting boundaries: 6th and 7th grade	Shifting boundaries: 6th, 7th and 8th grade

Intervention target

Interventions were mostly intended to target DRV (n = 43, 54.4%). Fifteen92,96,157,158,160,168,171,174,175,180,184,192 interventions targeted GBV (19.0%), and 14 interventions78,81,82,85,98,100,156,197 were identified as targeting both (17.7%). The intended target was unclear for eight interventions79,94,170,176,191,195,198 (10.1%), though these were included in the review as the intervention content included topics relevant to either DRV or GBV.

Delivery of interventions

A breakdown of intervention facilitators is provided in Table 6. The majority of interventions were facilitated by external staff only (35.4%) or teaching staff only (31.6%). A total of 38 interventions (48.1%) included at least some role for teaching staff in the delivery. A small number of interventions incorporated a peer-delivery component (n = 11, 13.9%),36,49,82,95,158,165,166,168,175,187,199 or did not require a facilitator (n = 4, 5.0%). 36,88,100,169 There was no clear difference in choice of facilitator between DRV and GBV interventions.

TABLE 6 - Intervention facilitators (outcome evaluations)

	Number of interventions (% of total)
External staff only	28 (35.4)
External staff and teaching staff	7 (8.9)
External staff and peer delivery	3 (3.8)
External staff, teaching staff and peer delivery	2 (2.5)
Teaching staff only	25 (31.6)
Teaching staff and peer delivery	4 (5.1)
Peer delivery only	2 (2.5)
No facilitator	3 (5.0)
Unclear	5 (6.3)

The method through which interventions were delivered is outlined in Table 7. The majority of interventions were delivered wholly (73.4%) or partially (16.5%) in person. A minority of interventions included a self-study (12.7%)88,158,163,169,171,184,188,189,195,198 or digital (15.2%)49,79,96,97,161,166,170,171,173,182,188,189 component (e.g. the use of virtual reality game format). One intervention100 was delivered entirely as a community intervention. There was no clear difference in the method of delivery between DRV and GBV interventions.

TABLE 7 - Method for delivering the intervention (outcome evaluations)

	Number of interventions (% of total)
In-person delivery only	58 (73.4)
In-person with some self-study	4 (5.1)
In-person with a digital component	6 (7.6)
In-person with both a self-study and digital component	3 (3.8)
Self-study only	3 (3.8)
Digital delivery only	3 (3.8)
Community delivery	1 (1.3)
Unclear	1 (1.3)

The duration of the intervention was not clearly reported in many of the included studies. Overall, reviewers categorised nine interventions as being single sessions (11.4%), 20 interventions as lasting 1 month or less (25.3%), and 47 interventions as lasting longer than 1 month (59.5%). The duration of interventions according to intervention target is shown in Figure 5; these were similar, though short interventions lasting < 1 month were more common for interventions targeting DRV only. Full details for the duration of interventions as reported by studies are provided in Report Supplementary Material 1.

FIGURE 5.

Duration of interventions according to intervention target (outcome evaluations).

Participant characteristics

Process and implementation evaluations (RQ2)

Sample characteristics for the included process evaluations are reported in Report Supplementary Material 1. Across the 137 studies reporting process evaluation outcomes, data were reported for 109 independent samples of students. This includes one study that analysed three separate student samples (ICRW 201785 in India, Vietnam and Bangladesh), while 28 studies did not include students (i.e. the sample was school staff and teaching professionals only).

Demographic data for students were frequently not reported in included studies. In many instances, demographic data were only reported for the full sample of students included in the evaluation, and not reported separately for the intervention of interest, or specifically in those students relevant to process evaluation outcomes. Demographic information was commonly not reported for students involved in qualitative research.

Age

Where reported, students ranged in age from 7 to 19.3 years, though mean ages were all above secondary school age (mean age range 11.45–16.7 years, median mean age 14.5). No difference in age was noted between studies evaluating interventions for DRV and GBV.

Gender and sexuality

The vast majority of studies asked students to provide their sex or gender as a binary outcome. Five studies91,97,144,201,202 permitted students to respond using alternative gender categories, although one of these (Scull et al. 2021)97 did not report the number of students who identified with another gender category. In the remaining four studies, the number of students identifying with another gender ranged between 0.2% and 10%. One of these studies (Watson 2012)202 evaluated an intervention intended to target GBV. [Note: when calculating sex/gender across studies, when a study reported data for a single sex only (e.g. 55% females), we assumed that the study offered a binary option for sex and lumped the remaining participants into the other sex category.]

Students reported their own sexuality in six of the included samples only (5.5%);91,202–206 in four of these cases, the study was evaluating an intervention for GBV. Students not identifying as heterosexual ranged from 0% to 28% (median 15.8%). An additional study207 reported that 100% of students were heterosexual, although this estimate was considered unreliable as it was based on staff judgement rather than self-reporting from students. There was no notable difference in the proportion of students not identifying as heterosexual between studies evaluating an intervention for GBV and DRV.

Ethnicity

Race or ethnicity information was reported for students or the school setting in 33 samples (24.1%) and is displayed in Figure 6. In 15 studies,97,98,123,147,201,204,206,208–212 the majority student population was white or Caucasian. In 10 studies, ethnicity was mixed and no one race or ethnic group accounted for more than 50% of students. 91,101,114,126,202,207,213–216 The remaining studies reported a majority black or African American80,137,178,217 or Hispanic or Latino96,218–220 population (four studies each). Other ethnicity groups represented a minority of students in the included samples.

FIGURE 6.

Majority ethnicity/race group of process evaluation samples (where reported).

Socioeconomic status

Information about schools’ or students’ socioeconomic status (SES) was reported for 18 of the samples (16.5%),78,80,84,91,94,111,120,124–126,136,154,207,211,217,219,221,222 although in some cases this was limited to a general statement about the district or school. Of these samples, nine samples80,124–126,136,154,207,217,221 were identified as including a majority of students from lower SES backgrounds (e.g. free or subsidised school lunches, or in areas with high economic deprivation).

Baseline risk

Interventions in six of the included studies (4.4%)203,213,218,223 were targeted towards students who had experienced or witnessed domestic violence or were considered to be at a high risk of DRV. One of these studies (Mgalla et al. 1998)138 evaluated an intervention for GBV. An additional study (Kervin and Obinna 2010)224 began with a universal sample, and then the intervention was moved to one school site where students were at high risk of not graduating and were considering to be at a higher than risk of domestic violence than the general population. All other samples were identified as universal, or abuse history was not reported (and therefore was considered likely to have included a universal sample).

Outcome evaluations (RQ3)

Cluster and sample characteristics for the included studies are reported in Report Supplementary Material 1.

Age

Where reported, students in included studies ranged in age from 9 to 19 years, though all reported mean ages were above secondary school age (mean age range 12–16.5, median mean age 14.25). There was no clear difference in the age of students between studies evaluating interventions for DRV and GBV.

Gender and sexuality

The vast majority of studies asked students to provide their sex or gender as a binary outcome. Ten studies were conducted solely with a single-sex group: four77,170,180,183 with males and six88,96,160,171,172,174 with females. Where studies included mixed sex/gender samples, slightly more female students were included overall (mean 53.4%, range 41–76.3%). One study (Meiksin et al. 2020)91 permitted students to record their gender as non-binary, or disclose that they were unsure about their gender, while another (Scull et al. 2021)97 permitted students to record their gender as non-binary but numbers were not reported (numbers were lumped with students who did not disclose their gender). None of the studies that evaluated an intervention targeting GBV reported gender identities beyond the binary. [Note: when calculating sex/gender across studies, when a study reported data for a single sex only (e.g. 55% females), we assumed that the study offered a binary option for sex and lumped the remaining participants into the other sex category].

Only six studies35,49,81,167,182,197 included in the review reported data for the sexuality of participants (8.8%). In five35,49,81,182,197 of these studies where students self-reported their sexuality, the number of students not identifying as heterosexual ranged from 4.8% to 20% (mean 12.7%). In the sixth study,167 it was reported that 3.6% of students had previously dated someone of the same sex. No studies that evaluated an intervention for GBV reported student sexuality.

Ethnicity

Race or ethnicity information was reported for students in 36 samples (52.9%). In a third of these studies, ethnicity was mixed and no one race or ethnic group accounted for more than 50% of students. In the remaining studies, more than 50% of students belonged to a white/Caucasian (37.8%),35,36,82,88,95,97,98,159,180,186,191,196,197,200 Hispanic or Latino (18.9%),38,81,96,185,188,193,195 or black or African American (10.8%)89,92,155,166 demographic (Figure 7). Other ethnicity groups accounted for a minority of samples only.

FIGURE 7.

Majority ethnicity/race group of outcome evaluation samples (where reported).

Socioeconomic status

Information about schools’ or students’ SES was reported for 35 of the samples (51.5%), although in some cases this was limited to a general statement about the district or school. Of these samples, 11 studies38,81,85,166,175,180,181,187,189,192,195 were identified as including more than 50% of students from lower SES backgrounds (e.g. free or subsidised school lunches, or in areas with high economic deprivation).

Baseline risk

None of the identified studies exclusively included students who had experienced DRV or GBV. One study (Miller 1998)162 included students with ‘experience’ of domestic violence, which may include DRV or experience of violence in the home (as victims, perpetrators or witnesses). A further study (Silverman 2000)99 included students who were considered to be at high risk of DRV, which included students who had witnessed couple violence, experienced child abuse, reported current substance use, or expressed beliefs in the acceptability of couple violence or rape myths. All other studies included a universal sample of students regardless of their abuse history, although four studies166,169,184,186 limited inclusion only to those students with ‘dating experience’. The authors of these studies proposed this criterion either to identify those students who they considered more appropriate to receive a DRV intervention, or to best identify a change in DRV outcomes pre and post intervention.

Intervention components analysis

The intervention typology was informed by all included studies, while analysis of components focused on interventions reported in outcome evaluations (n = 68). Drawing on intervention descriptions from the body of the text of each study, its supplementary appendices and, when necessary, additional information from the authors, we undertook initial inductive coding on a subset of 27 intervention descriptions, selected to capture a range of surface-level diversity. This generated an initial list of 143 different codes, organised around 11 preliminary groups that related to: content topics; skills discussed or practised; intervention activities; intended result of those activities; deliverer; location; the presence of others; how students are guided through the intervention; medium(s) of delivery; how school staff were involved with the student portions of the intervention and/or received their own material; and how the extended community (parents, families, local groups, etc.) were involved with the student portions of the intervention and/or received their own materials.

Through discussions between the co-investigators, and based on the principles of axial coding, these 11 groups were re-organised into an intervention typology:

• intervention type, the essential intention or ‘spirit’ of the programme

• intervention focus, the broad content included

• intervention domains of implementation, using the socioecological model to account for the levels over which the intervention was implemented.

We also described intervention components, defined as the activities in which students, school staff and the extended local community experience as part of the intervention.

Synthesis of theories of change

The synthesis of theories of change drew on different sets of included studies in cycles. After an initial coding round of 5 interventions described in 10 reports, we undertook exploratory coding between 2 reviewers informed by template analysis. 226 Using a coding template with codes exchanged by two reviewers in a subset of interventions (n = 14) reported in two or more reports (n = 38), focusing as well on breadth of intervention type. The template reflected the focus on programme theories and included intervention inputs, components, mechanisms, outcomes and key theoretical concepts (Table 8). An example of the coding template as applied to interventions is provided in Report Supplementary Material 1.

We then synthesised programme theories using thematic synthesis. 67 A subset of interventions that were reported in two or more papers were prioritised: we identified similarities, differences of emphasis and contradictions between multiple papers and developed summaries of the theories of change for the interventions, describing inputs, mechanisms of change, underlying theoretical assumptions, and proximal and distal outcomes. For the remaining interventions, summaries of the programme theories were produced. The analysis proceeded with another 37 reports analysed, at which point thematic synthesis yielded no new insights. The remaining included reports were analysed using a confirmatory coding approach.

The findings of the thematic synthesis were transformed using meta-ethnography to develop a line of argument for an overarching theory of change. The key concepts identified during coding were treated as ‘first-order constructs’227 which we summarised as logic models and compared between each intervention (see Report Supplementary Material 1 for examples). This generated second-order constructs, which were our interpretations of the first-order concepts. Third-order constructs were then developed (based on key concepts and second-order interpretations) to produce a line of argument that aimed to describe the overall mechanisms of change.

Intervention type

The intervention ‘type’ captures the essential intention of the intervention. Categorisation depended on the intervention’s domain, focus and components, as well as most importantly a consideration of the specific change mechanisms being targeted by the intervention. Programme type is a classification of the level of sophistication and complexity in how the intervention intends to accomplish the goal of reducing DRV/GBV. This classification allows us to look across the highly varied landscape of DRV/GBV prevention programmes and to group them in ways that make for more useful comparison.

Intervention focus

Intervention content addressed two broad areas: a focus on perpetrators and victims and/or a focus on bystanders. Most interventions (86%, n = 168) focused on potential perpetration and victimisation but this content varied considerably. For example, one programme in Brazil focused on perpetration reduction by teaching content on coping skills, including ‘practical aspects of how to behave in situations with aggressive potential in a non-violent way’ (Filho 2017, p. 4384). Adolescents could use those coping skills to replace their own violent tendencies or normalised violent reactions in those types of situations in future encounters. Other programmes, such as the Fourth R programme, addressed perpetration by targeting social norms about violence that students might see and ‘counteracting pro-abuse messages from peer culture’ (Cissner and Ayoub 2014, p. 3281). Such messages can normalise the idea that violence is an acceptable response in some situations, so some programmes address it directly by correcting myths and misperceptions about violence, the use of violence and the impact of violence on both perpetrators and victims (Knowles 1997;232 Madsen 1994;233 Maphosa 2017;136 Rogers et al. 2018127). Content focused on victimisation also varied considerably. In some interventions this consisted of directly addressing adolescents as potential victims through teaching them active resistance or self-defence skills (e.g. Baiocchi et al. 2017;160 Decker et al. 2018;174 Ekhtiari et al. 2014;234 Jordan and Mossman 2018147). Others focused on teaching adolescents to avoid such situations by talking about ‘red flags’ – an indication which might signal that a relationship or a situation is unhealthy or could potentially become dangerous (e.g. Busch-Armendariz et al. 2008;235 Cascardi and Avery-Leaf 2014;80 Genereux 2020;236 Taylor et al. 2017156).

Fewer interventions (27%, n = 52) focused on bystanders as a change pathway. Even within this smaller group the nature of how bystander helping is addressed was highly variable. Some programmes provided examples of bystander intervention in a variety of scenarios, stories or video clips (e.g. Cissner and Ayoub 2014;81 Jouriles et al. 2019;181 Merrell 200492) so that students develop self-efficacy around using those concrete strategies in future situations they might encounter. Other interventions provided the scenarios but asked students to brainstorm or role-play different ways the situation might be risky and how to best respond those risks by overcoming obstacles to helping (Coker et al. 2017;82 Lee et al. 2018;87 Pacifici et al. 2001191). More general promotion of bystander behaviours was also included in many programmes. Relationships Without Violence214 encouraged bystander behaviour with content that promoted ‘pro-social behaviours, positive peer culture, and healthy masculinities/femininities’ (p. 4). Green Dot82 and BITB197 both included direct skills practice for bystander intervention as well as reflective content addressing rape myths and attitudes supportive of or tolerant towards violence. Addressing myths and attitudes is an indirect way to increase bystander behaviour through reducing barriers to action. Bystander intervention can also take place after violence has occurred. The Women’s Self Defense Network-Wahine Toa courses in New Zealand147 not only taught skills for responding to attempt assaults but also focused on how to support others if they have experienced violence by helping them to seek support and access local resources.

Intervention domains of implementation

Intervention implementation can be divided into student intrapersonal, student interpersonal, staff, structural-social and/or structural-environmental domains. The student individual level was addressed in two ways: the most common form was improving students’ intrapersonal traits (86%, n = 168), such as their own attitudes, opinions and behaviours. In addition, two-thirds of programmes focused on improving students’ interpersonal traits (64%, n = 126), such as incorporating skill-building to help students create healthier and more supportive relationships with their peers (e.g. Bando et al. 2019184).

About half of the interventions focused on staff (45%, n = 95) as a domain of interest. This most often included preparing the school staff to deliver programme content to students as the key recipients. Additionally, some interventions targeted school staff specifically, to support them in how to respond to violence and improve their capacity for addressing violence that takes place in school settings. For example, the BITB-HSC197 includes a 60-minute workshop designed to help school personnel learn skills to be positive bystanders when they see situations of adolescent interpersonal violence.

Interventions addressed the structural level in two ways. Just under a third of programmes (30%, n = 58) focused on implementing components to change the social structure of a school by improving policies and influencing social norms in the school setting to be less tolerant of violence and more supportive of prosocial behaviours such as believing survivors and intervening in risky situations. Fewer programmes (13%, n = 25) focused on changing the environmental structure of the school, including addressing things in the physical landscape of the school and/or who is present in school spaces. This included alterations in staff practices (e.g. increasing hallway monitoring) or changing the visual landscape of the school (e.g. posters that list local survivor support resources).

Starting theoretical propositions

Shaped by Markham and Aveyard’s (2003)228 theory, we developed two starting hypotheses, broadly summarised as that school-based interventions designed to reduce DRV and GBV have the greatest potential to be effective when based on the principles of weakening classification and framing. Specifically, and first, we hypothesised that interventions aimed to weaken the institutional boundaries between the school and surrounding communities and strengthen relationships within the school. For example, several interventions with curricular and classroom-based components had additional components directed at the community and family/parents to support and reinforce change initiated in the classroom. Many of the interventions were delivered by teachers, some in collaboration with students, which opened opportunities for increased interaction and positive relations between teachers and students. In some cases, teachers and other school staff were trained to model positive behaviours to reinforce behaviour change among the students. Importantly, as our analysis progressed, our focus shifted from ‘weakening boundaries’ to ‘improving commitment’, based on an understanding of the diverse contexts in which interventions were implemented. As discussed below, weakening boundaries, for example with communities can be undesirable in the context of violent communities and harmful gender norms. Second, we hypothesised that interventions aimed to involve students in their learning and in school-level decisions and increase belonging, thereby weakening the framing of the schools. For example, many of the interventions utilised student-centred techniques such as group work and role-playing to encourage interactive and experiential learning with student input to their own learning, which in turn, promoted positive interactions and relations between students and teachers and among the students themselves. This facilitates students’ capacity for practical reasoning and critical reflection.

Intervention inputs

Intervention mechanisms of change

In Figure 8, we have depicted mechanisms as lower and higher level: the lower-level mechanisms are where individuals react or respond to the resources provided by the interventions that trigger change (e.g. the knowledge gained from lessons about sexual violence), and higher-level mechanisms are where mechanisms interact with one another and evolve within the school system of relationships (e.g. students interacting in positive and prosocial ways with peers).

Two key themes were generated from the theory synthesis mechanisms leading to student commitment to prosocial values and behaviours:

1. strengthening relationships between the school and surrounding communities, and within the school (between students and teachers, and between students)

2. increasing belonging in the school by student-centred approaches.

The higher-level mechanisms of change through which this happens are depicted in the logic model. These occur through interactive and experiential learning, positive and prosocial student interaction with peers, positive relationships between students and teachers, normalisation of pro-social behaviour and increased school connectedness (see Figure 8 dark blue boxes). Lower-level mechanisms facilitating change occur through the changes in student, teacher and family knowledge, attitudes, behaviours and skills (see Figure 8 pink/orange boxes). These are discussed below through the broader themes of ‘strengthening relationships’ and ‘increasing belonging’.

Hypothesised intervention outcomes

Most interventions aimed to reduce violence: some focused on DRV victimisation and perpetration, some on GBV and some on both interventions. For many of the interventions, proximal outcomes (see Figure 1, light grey box) led to reduced dating-based violence and GBV through the development of prosocial skills, increased empathy, increased intent to intervene, report and seek help, and increased social and communication skills. These changes were realised through the mechanisms discussed in the previous section.

School resources and infrastructure → ease of implementation and determined necessity of modification

In 32 studies, school and programme staff cited the absence of necessary school infrastructure as both a barrier to implementation and a cause for programme modification. School infrastructure in these instances was defined in regard to physical infrastructure (building space and materials and resources) (n = 20) along with school organisational infrastructure and management (n = 13, 14). One such physical limiting factor was room availability (n = 9, 9), with rooms frequently limited by either number or size. These limitations in space necessitated modifications in intervention delivery, with the extent of the modification ranging in accordance with the severity of the limitation. In one example, limited rooms (and teaching staff) required that a same-gender programme be adapted to a mixed-gender programme due to the school’s inability to run two simultaneous programmes. 217 In another study, small room sizes led to the omission of a physical activity as the activity was unsuitable for the 30 students to complete in the available classroom. 257 Similar themes were observed with regard to missing or non-functional materials and resources, which led facilitators to modify or entirely drop the delivery of intervention components (n = 12, 12). In one such case, an audio drama intended to be played on an MP3 instead had its script read by facilitators when the MP3 players were found to be broken or missing. 134 In regard to organisational infrastructure and management, unsurprisingly, strong pre-existing health programmes or prior experience with similar interventions aided implementation and fidelity, while intervention components being outside of existing procedures hindered their delivery (n = 5). Furthermore, poor organisational infrastructure reduced the already limited time available for the interventions, with poor communication leading staff to be uninformed and unprepared for training91 or delayed pupil authorisation slips, leading students to be late for classroom-based interventions. 130 Ultimately, although on-the-ground modifications were often done with relative success (as claimed by the studies), programme staff expressed hesitancy in implementing prevention interventions in resource-poor schools due to structural limitations:

the schools that need us the most – like the C schools [rating on the NYC DOE Report Card] … they’re in such bad shape … Because if they’re a C or a D, they’ve got a lot of structural problems. Not just the students being poor, they just have poor management. 81

Time constraints → dosage and fidelity

Fifty-four studies discussed barriers relating to school-level time constraints, which were shown to lead to curriculum modifications, impacting intervention dosage and fidelity. Limited space in the curriculum restricted the class time that was available for DRV/GBV interventions (n = 22, 23). The time instead was prioritised for academic performance in many cases (n = 16), such as relating to standardised testing, with school administrators believing that time out of class would impact test scores as described by the following Green Dot facilitator:

when I talked to our control school, you know, he [the principal] was like these [results] are wonderful but, you know, test scores … If your test scores aren’t here or you don’t have this much class time then you’re losing funding and you’re not going to do this. 258

Additionally, limitations in the time of school staff (n = 27, 28) impacted both the time they had available to facilitate interventions (n = 12, 13), along with the time they had available to participate in staff training (n = 7), ultimately reducing both the dosage of intervention delivered and the attendance at staff training. Additionally, class session lengths often had to be modified to fit the block schedule of the school, along with accommodating half-days or other schedule modifications (n = 17, 18). In some cases, these time limitations were anticipated by the intervention designers and led to the planned modification of the intervention prior to implementation (n = 8). In other cases, or even with the aforementioned anticipated curricular modifications, dosage was reduced while the intervention was implemented, reducing intervention fidelity to varying degrees as lessons were abridged or removed (n = 17). These limitations ultimately reduced the willingness of school personnel to implement longer interventions, instead preferring one-time lessons in order to save time:

I don’t think we can commit that amount of curriculum time to it, particularly in year 9 … I would say whoever’s organizing the package, they need to remember that everybody, so drugs awareness, smoking, tobacco-awareness you know, all the resources you can get are about five, six weeks. 91

Dating and relationship violence/gender-based violence comprehension and perception → influences school and community support

In addition to time and resource limitations, the school staff’s comprehension and perception of DRV/GBV, both as issues more broadly and in relation to their specific student body, were cited in 42 studies as factors impacting a school’s willingness to implement an intervention. A belief among school personnel that DRV and/or GBV was not a problem within their school community (n = 14) was a major barrier to school personnel engagement, along with programme adoption in the first place. For school personnel and particularly administrators, this lack of recognition was often tied to worries about school reputation (n = 8), as discussed below by a facilitator of one such intervention:

any school will say we don’t have a problem with bullying, we don’t have a problem with violence, and I think that’s just to say, you don’t want it out there that your school is violent. Who wants to go to a school that’s full of violence?258

While for teen dating violence focused interventions, this manifested as a refusal to admit that DRV or GBV was a problem for their students (n = 5), for lesbian, gay, bisexual and transgender (LGBT)-bullying focused interventions, this was both an assertion that LGBT-bullying was not a problem in their school or even that there were no LGBT students in the school (n = 3). 259 Furthermore, although interventions addressing DRV and GBV could both come with political connotations, interventions addressing GBV (in regard to both LGBT-bullying and more broadly) generated more pronounced controversy than DRV prevention education alone. Specifically, the term GBV and any teaching about the gendered component of DRV alienated some male staff (n = 3) and, as discussed in a later section, male students (n = 15), reducing their engagement and investment in the intervention. In one school, the researcher was pulled aside by the leader to say that they were struggling with delivering the intervention to make teaching staff:

there was almost a divide between the male and female staff about their views on it … there was a clear distinction between who got that it’s an issue and who didn’t, as in like the males, sort of didn’t as much. 91

Furthermore, some school personnel (both male and female) were averse to teaching LGBT-bullying due to personal homophobic views (n = 5) or due to a fear of a negative community response (n = 8):

You’ve got to remember, especially, I’m talking about the [name of the area] community. It’s a very white, working class, not so intellectual group … The staff fear that they will be accused [by parents/community] of promoting [‘transgender’] … or that … they are exposing their child to notions of that. Of transgenderism. Big fear there. 260

On the other hand, while negative perceptions and misunderstandings of DRV/GBV were linked to reduced staff engagement and likelihood of programme adoption, an informed comprehension of the issue and its relevance to the student was shown to have the opposite effect. In fact, comprehension of the importance of prevention education on these issues was shown in some cases to overcome other factors, such as time limitations (n = 10, 10), that may otherwise have dissuaded school personnel from engaging with the intervention:

Often my attitude with MVP is ‘pain in the arse, pain in the arse’ when it comes to organizing things together because I’ve got so many other things to do. But there’s not a hint of resentment in anything I’m saying … I do not mind the time because it’s worthwhile. 131

One potential method of increasing DRV/GBV awareness (and thus increasing staff engagement and buy-in) was staff training, both for facilitators and for school personnel more broadly (n = 9):

[The training for our staff] has significantly raised awareness in groups of staff … there is a belief that we need to do something about this. That it’s such an important issue for young women and we need to make sure everyone in the school gets to know more. 109

At the same time, however, staff who had a lower level of initial knowledge, and thus could potentially benefit more greatly from the training, were less engaged in them (n = 4). Modifications of training based on pre-existing staff knowledge and/or additional outreach for hesitant staff may be one potential way of combatting variable staff investment. 204

Ultimately, a parallel fourth theme illustrated a point that is already well-established in school-based intervention literature, namely, that whole-school investment was critical for both programme adoption and successful implementation. As such, a major factor impacting implementation was how all levels of school personnel, ranging from head teachers to janitors, viewed DRV/GBV prevention (n = 21, 22). Investment in and understanding of these issues could thus be a deciding factor in an intervention’s success and was even able to overcome other barriers to investment, such as time or resource limitations.

Intervention interactivity → improved student engagement and acceptability

Forty-eight studies reported that interactive components within the intervention helped improve student engagement and acceptability. Students tended to be more engaged and rated activities more positively that were in some way interactive, whether through discussion (n = 10, 11), role-play (n = 12, 13) or games (n = 11, 12), as discussed by a teacher below:

The participatory and interactive nature of the work – not just talking or preaching at them through theatre but really involving them. The moments where the pupils were allowed to stop the action and take over were particularly effective. 130

Some studies found that, within their samples, the favourite type of activity varied by gender, specifically demonstrating a preference for role-plays among boys and a preference for discussion among girls (n = 2). However, the small number of papers supporting this claim suggests this merits further investigation. In contrast to more interactive intervention components, more ‘boring’ tasks, such as homework, were found to have a lower level of student engagement (n = 4, 5). Critically, the level of engagement and acceptability seems to rely not only on the ‘fun’ nature of these tasks, but on the interaction with their peers and facilitators, as shown by the fact that other activities outside the realm of normal school programming, such as media clips257 or video games,251 were reported to have a higher student engagement and often acceptability when followed by a discussion. Similarly, student ownership of the intervention (n = 9), which likewise involved student interaction, was shown to increase student engagement, as discussed by a student:

We were all happy that we would be the ones to distribute the condoms … It was as if we belong to this team and, like, it somehow left the best impression on me … We are now, you know, like some collective that fights together for something important. 124

Poor student attendance among certain populations → reduced intervention reach

While intervention attendance was reported to be relatively high across the reviewed studies, a small percentage of students had inconsistent attendance or did not attend at all, as directly discussed in 22 studies. Numerous barriers to intervention attendance were cited, including dropping out of school (n = 6), truancy (n = 3) and parental permission forms (n = 7). For voluntary programmes, attendance was frequently far more variable and often lower (n = 4). This was especially true for after-school programmes (n = 5) due to competing evening responsibilities (n = 2) or a long travel time from school to home (n = 3). Critically, the students absent from the intervention may be among those who need it most due to the perceived correlations between dropping out/truancy and DRV/GBV, as stated by a nurse below:

At first I thought they choose the best children to come to the project until I realised it was voluntary. Basically the children that came weren’t the children I wanted to see. The children that we know needed the programme were not there. 261

While ultimately those students who do not attend the intervention may be a minority of the total student body, they may be particularly vulnerable to DRV/GBV and thus may require innovative methods of outreach.

Positive relationship formation → improves student comfort and engagement over time

Another major factor that operated at the level of the student, both in an intrapersonal and interpersonal sense, was how positive relationship formation, both with peers and facilitators, improved student comfort and willingness to engage in the face of embarrassing or challenging material, as cited in 38 studies. Student embarrassment and discomfort when learning about DRV/GBV was consistently seen across student populations and is thus a barrier that teachers across any school may initially encounter. At the same time, a number of studies discussed how students who were initially hesitant to participate in activities due to the private and ‘awkward’ nature of the topic, became increasingly comfortable and engaged over time (n = 23):

At first when I started group, I was kind of scared to talk to people. I didn’t know anyone in group, so I kind of had to work up a trust relationship with them … Now I can talk to anyone in group. 203

One factor that may mediate this increase in student comfort is relationship formation, both with facilitators (n = 13) and with peers (n = 14):

As time passed, well, they gave me confidence … [that] I can ask them for help, it won’t be a problem. At the beginning I felt a bit embarrassed [talking to them], but afterwards, no, I would feel relaxed. 205

As a result, the fostering of positive relationships may be a key mechanism that facilitators can use to increase both student comfort and engagement. This is particularly critical for facilitators from outside organisations who will generally not have a pre-established relationship with the student body. For example, two student participants discussed how an outside facilitator built a positive relationship with them:

He learned how we act towards each other. He’d get out there, and he’d wrestle with us. 124

He was in the weight room, and I was like, ‘Who’s this bulky guy? … This is our weight room’. And then towards the end of the season, I was like, ‘This is a cool guy’. The longer he was there, the more I listened. 124

Facilitator relationship formation was shown to be enabled by a recognition of student experience and identity, as well as a willingness to be vulnerable at times and share perspectives from their own lives:

Several students also said that when the Mexfam health educators shared their own personal experiences during the sessions it helped the course participants to open up about their own lives. 205

Similar benefits stemming from positive relationship formation were seen in peer interaction. Increased comfort and engagement due to peer social relationships were especially apparent in the case of interventions that targeted students with a shared background (n = 18, 19). In one such example, a group of students who all had lived experience of DRV/GBV victimisation found comfort in their shared experience:

It helped me because I kept it in for that long, for like seven years. It was just like a wound that never healed. It was shocking to see that I wasn’t the only person who’s been through it. It made me feel more comfortable. 218

Additional bonding was found in other group compositions as well, including LGBT students (n = 2) or in single gender groups (n = 15), the latter of which is discussed below by an intervention facilitator:

The conversation with the young ladies when you separate [boys from girls] is awesome. They can talk freely and know that they’re not going to be teased because usually a fellow girl won’t do something like that. 262

At the same time, the question of gender composition is not so clear cut, as many students likewise preferred mixed-gender environments (n = 8).

Programme fit to student population → student and staff acceptability, student engagement and programme fidelity

Fifty-nine reports examined the extent to which interventions fit the needs of the student population as a factor impacting implementation. Critically, the fit of the intervention to student needs was shown not only to impact student and staff acceptability, but was a major cause for programme modification, both prior to and during implementation. The literature discussed several different aspects of student background that may require intervention modification, namely, cultural context (n = 21), student age (n = 14), background knowledge (n = 14), sexual identity (n = 7), prior trauma expose (n = 6) and disability (n = 2). In many studies, students struggled to engage with material that was not relevant to them or their experiences. This could be due to factors such as age; in the quote below, some students found materials either too young or too old for them, as discussed by one high school student:

I thought it was more towards adult relationships than towards teenagers. Or like early adult … There were some text messages we couldn’t really relate to … Usually, when you’re an adult relationship it’s like kind of different because you’re with each other more. 215

For students with prior exposure to trauma, the material could be triggering and cause discomfort, leading to a decrease in student engagement, as discussed by another student: ‘I found that it stirred up a lot of memories and emotions and really invaded my comfort zone, I wanted to leave after a while’. 177

As a result of such disconnections between student profile and programme contents, facilitators frequently chose to modify the intervention, for example, in one study where the facilitator translated and added cultural adaptions, specifically through changing the names in a role play, to a pre-existing intervention to better suit the needs of their Spanish-speaking students. 263 Ultimately, the modification of interventions to improve student fit was shown to be a fairly frequent phenomenon across studies and should be expected when an intervention is implemented in a new context or with a population group different to the original design/test.

Student immaturity and disruptive behaviours → inhibited student engagement and intervention delivery

Thirty studies discussed student immaturity and disruptive behaviours as a barrier inhibiting student engagement and intervention delivery, as discussed by a teacher: ‘Students [not] taking the scenarios seriously [was a challenge]. Some were laughing and making jokes because they had a hard time taking it seriously and being vulnerable to the rest of the class’. 81

Immature behaviours were found in participants of a wide age range, ranging from elementary schoolers laughing at words such as ‘private parts’264 to high-schoolers bullying their peers. 265 One common challenge that facilitators faced was deciding how to properly respond to these behaviours, with one option being detention:

The teacher tells me about the consent lesson. He relates that a male student responded that in such a case ‘I’d give her a slap!’. The teacher says there were intakes of breath from other students and that he gave the student an hour’s detention – ‘I had to, it was unacceptable …’257

But in other cases, facilitators instead chose either to allow the student’s peers to respond to their comments or to respond to it themselves. 266 Facilitators often had to balance zero-tolerance policies on bullying with the imperative to educate disruptive students who may need DRV/GBV prevention education more than their peers. While disruptions and joking around were described as occurring regardless of gender, some studies did link these behaviours particularly to male students, some of whom could be particularly disruptive in the case of discussions on GBV or the gendered component of DRV more broadly due to a fear of the intervention being perceived as ‘anti-men’ (n = 15):

Initial reaction to the project was a lot of caution on, ‘Is this a male-bashing kind of thing?’, from some students … I think it was very smart on our parts to promote it not specifically as violence against women because there’s a defensiveness that happens with a lot of males. 250

As a result, several studies discussed the inclusion of a discussion on male victims in the curriculum in order to prevent the alienation of male students (n = 3).

Facilitator content knowledge → intervention fidelity and student acceptability

Thirty studies identified facilitator knowledge of DRV/GBV as a key factor impacting implementation, shown to influence both intervention fidelity and student acceptability. At the same time, facilitator content knowledge and delivery comfort were consistently cited as being key factors impacting implementation success for both school-based (n = 21, 23) and outside facilitators (n = 13). While both facilitators from outside organisations and facilitators from within the school were discussed as having sufficient and insufficient levels of content knowledge, variation in skill and preparation was mentioned far more frequently for school-based facilitators (n = 21, 23), for example as discussed by the deputy head of a school:

I think that there are some staff that are absolutely fabulous at delivering stuff like that. And then some others who should not be allowed anywhere near it. Because it can … be quite damaging if it’s not done the right way. 91

In contrast, outside-facilitators were generally described as having high content knowledge and high confidence when delivering the material (n = 11): ‘The advocate’s like well-versed in all the topics that he’s talking about, and our coach like would just be making it up as he went along’. 178

For both school-based facilitators and outside facilitators (n = 5), selecting the most appropriate individuals to facilitate the programme was just as critical as sufficient training of the facilitators after selection. This was particularly apparent for school-based facilitators who ranged greatly in their initial confidence and competence in delivering DRV/GBV prevention materials. Teacher type was discussed as playing a role in facilitator comfort and confidence. Generally, school-based facilitators with a health background were identified as being more comfortable with leading DRV/GBV prevention due to their increased knowledge both in terms of the material itself and the style of teaching required to teach it, (e.g. participatory and activity-based approaches) (n = 8, 9), along with the increased ease of incorporating this material into the existing health curriculum (n = 2), as opposed to one focused on English or Maths. On the other hand, for LGBT-bullying prevention, LGBT teachers themselves were unsurprisingly identified as being the most knowledgeable and prepared to teach the topic (n = 3). At the same time, however, ‘coming out’ was discussed as being a challenge in certain work environments and may not be what is wanted by every LGBT teacher:

[Even now,] if you’re gay and you’re a teacher you may still be on guard that someone’s going to find out about you. I’m not in this situation now but other teachers in my school are, and in every other school … Most gay teachers are so closeted. They’re scared to death. 232

Support from the ‘other side’ → facilitate implementation and increase facilitator comfort

For both school-based and outside facilitators, having a support structure from the ‘other side’ was viewed as essential in successfully implementing the intervention as discussed in 27 studies. For teachers and other school-personnel, this mainly involved support from violence prevention experts with whom they could discuss delivery and resources (n = 20, 22). For outside facilitators who likely already had substantial knowledge on violence prevention, having support from teachers and/or other school contacts could assist in the logistics of implementation, through tasks such as managing student behaviour or student referrals (n = 10). On the other hand, having disengaged or unsupportive school contacts could make implementation more logistically difficult (n = 2), as discussed in the example below in which teachers remained passive at the sidelines of the intervention rather than helping manage students as planned:

We had to spend half our time and energy dealing with crowd control … the whole thing we do best – talking to them on their level, not like teachers, not like they’re used to – that was being totally undermined. 130

Ultimately, whether a school-based or outside facilitator was used to deliver the intervention, communication and support between the school staff and any other participating parties appeared critical for a well-implemented intervention.

Embedding intervention into school curriculum → increase intervention sustainability

One less frequently mentioned but still critical theme is the importance of embedding DRV/GBV prevention into the school curriculum in order to increase sustainability long-term, a task that is more easily accomplished via the use of teacher facilitators (n = 3) as discussed by the Start Strong Bronx staff:

it wasn’t just this extra, side thing that they did once in a while, but it was woven into the curriculum in a lot of ways. The big thing, though was, that if the teachers were trained, then it stays. 81

Ease of programme delivery → increased programme fidelity

For facilitators, especially those who were school-based, the more straightforward and lesser amount of additional time the intervention took to deliver, the better the programme fidelity and staff acceptability. Because school-based facilitators often had competing demands on their time, they could have limited availability to invest in intervention delivery:

As teachers, you’re busy, you already have your own lessons, and it’s like, ‘geez, I got to have another one.’ So, it gave you a lot of stuff to pick and choose from. 256

As such, making the process as straightforward as possible, such as through providing robust programme materials so the teachers do not need to spend time developing their own curriculum (n = 5, 6), increased the likelihood that the intervention would be delivered as intended. Despite restrictions in teacher time, another factor found to facilitate programme implementation was teacher training, which helped ensure that teachers felt comfortable and ready to implement the intervention (n = 26, 27).

Ease of programme modification → increased programme fidelity and effectiveness

Another theme that emerged at the level of the intervention was the ease with which the programme could be modified, as discussed in 19 studies. Across the interventions, programme modifications were consistently made for all of the reasons listed above. As a result, flexibility in delivery timing (n = 8) was highly acceptable to school personnel. 106 At the same time, while some modifications could greatly improve the fit of the intervention to the needs of the student population, informed by the knowledge of the facilitator, other modifications could occur with less thought, due instead, for example, to running out of time at the end of the lesson. 267 As a result, providing support and guidance on modifications in the programme materials themselves was cited in several studies as a potential method of easing implementation and improving fidelity to function (n = 11). Ultimately, programme designers needed to strike a balance between providing enough materials to ease the implementation of interventions for often overburdened facilitators and, at the same time, giving enough flexibility in the programme so that it could be appropriately modified to fit within a variety of different contexts.

Dating and relationship violence victimisation and perpetration outcomes

Dating and relationship violence victimisation outcomes (i.e. victimisation by a dating partner, such as a boyfriend or girlfriend or intimate partner) were reported in 26 included trials (19 with short-term outcomes and 15 with long-term outcomes). Within this category, we distinguished between several DRV outcomes.

• First, we considered ‘omnibus’ measures of DRV victimisation, which included scales that captured multiple types of DRV victimisation. These were reported in 13 trials (11 with short-term outcomes and 6 with long-term outcomes).

• Second, we considered measures of emotional DRV victimisation, including behaviours such as insulting, demeaning or spreading rumours about a dating partner. These were reported in 14 trials (8 with short-term outcomes and 9 with long-term outcomes).

• Third, we considered measures of physical DRV victimisation, including behaviours such as hitting, kicking or slapping a dating partner, or threatening physical violence against a dating partner, but not sexual violence. These were reported in 11 trials (6 with short-term outcomes and 7 with long-term outcomes).

• Fourth, we considered measures of sexual DRV victimisation, including unwanted sexual touching, forced sex and other forms of sexual violence by a dating partner. These were reported in 12 trials (8 with short-term outcomes and 6 with long-term outcomes).

• Because several trials jointly measured physical and sexual DRV victimisation, we created a separate category for these outcomes. A total of six trials reported mixed physical and sexual DRV victimisation outcomes, of which four reported short-term and five reported long-term outcomes.

• Finally, a small number of trials reported cyber DRV victimisation outcomes, defined as dating violence via electronic means (e.g. threatening or demeaning behaviour towards a dating partner, disseminating intimate images of a dating partner). Of the four trials that reported these outcomes, three reported short-term outcomes and two reported long-term outcomes.

Classification of DRV perpetration outcomes followed a similar pattern. Of the 28 trials that reported DRV perpetration outcomes, 20 included short-term outcomes and 18 included long-term outcomes. Omnibus DRV perpetration outcomes were reported in 14 trials (12 with short-term outcomes and 8 with long-term outcomes), emotional DRV perpetration outcomes were reported in 15 trials (9 with short-term outcomes and 9 with long-term outcomes), physical DRV perpetration outcomes were reported in 14 trials (8 with short-term outcomes and 8 with long-term outcomes), sexual DRV perpetration outcomes were reported in 12 trials (9 with short-term outcomes and 5 with long-term outcomes), physical and sexual DRV perpetration outcomes were reported in 5 trials (3 with short-term outcomes and 5 with long-term outcomes) and 3 trials reported cyber DRV perpetration outcomes (2 with short-term outcomes and 2 with long-term outcomes).

Gender-based violence victimisation and perpetration outcomes

Gender-based violence victimisation outcomes (i.e. violence on the basis of gender, sex or sexuality, whether physical, verbal or sexual) were reported in 20 trials. These included 15 trials reporting short-term outcomes and 13 trials reporting long-term outcomes. Within this category, we distinguished between several GBV outcomes.

• First, we considered ‘omnibus’ measures of GBV victimisation, which included scales that captured multiple forms of GBV victimisation. In total, 12 trials reported omnibus measures of GBV victimisation, of which 9 included short-term outcomes and 9 included long-term outcomes.

• Second, we considered measures of physical GBV victimisation, which included sexual violence of different forms, such as sexual touching and forced sex. Of the 14 trials reporting these outcomes, 11 reported short-term outcomes and 8 reported long-term outcomes.

• Third, we considered measures of verbal GBV victimisation, which included verbal sexual harassment but not homophobic harassment. Of four trials reporting this outcome, two reported short-term outcomes and three reported long-term outcomes.

• Finally, we considered measures of homophobic GBV victimisation, which included GBV on the basis of sexual orientation or gender identity. This was only reported in one trial, which provided both short-term and long-term outcomes.

GBV perpetration outcomes were similarly organised. Of the 16 trials reporting GBV perpetration outcomes, 13 reported short-term outcomes and 11 reported long-term outcomes. Of the 12 trials reporting omnibus measures of GBV perpetration, 11 reported short-term outcomes and 8 reported long-term outcomes. Four trials reported verbal GBV perpetration, of which all reported long-term outcomes, and two reported short-term outcomes as well. Two trials reported homophobic GBV perpetration, including both short-term and long-term outcomes.

Dating and relationship violence and gender-based violence knowledge, attitudes and behaviours

Dating and relationship violence and GBV knowledge, attitudes and behaviours were classed into seven constructs measured for each of DRV and GBV, and one construct, bystander behaviours, that was measured across DRV and GBV.

Violence acceptance includes outcomes based on scales that measure participants’ acceptance of DRV or GBV. The Illinois Rape Myth Acceptance Scale (Payne et al. 1999)268 or the Attitudes Toward Dating Violence Scale (Price et al. 1999)269 are archetypal scales for this construct. Of the 28 trials presenting DRV-related violence acceptance outcomes, 23 presented short-term outcomes and 8 presented long-term outcomes. Of the 22 trials presenting GBV-related violence acceptance outcomes, 17 reported short-term outcomes and 7 reported long-term outcomes.

Knowledge includes outcomes based on knowledge or recognition of DRV or GBV, or resources to support victims or perpetrators of DRV or GBV. Scales used to measure this construct were often bespoke in nature. Of the 14 trials reporting DRV-related knowledge outcomes, all reported short-term outcomes, but only 3 reported long-term outcomes. Of the 12 trials reporting GBV-related knowledge outcomes, all but 1 reported short-term outcomes and reported long-term outcomes.

Attitudes to intervening includes outcomes based on participants’ attitudes and self-efficacy to intervene when witnessing DRV or GBV; in some trials, these are described as ‘bystander intentions’. Archetypal scales for this construct include the Readiness to Help Scale (Banyard et al. 2014)270 and the Pros and Cons of Bystander Helping Scale (Edwards et al. 2017). 271 Eight trials presented DRV-related outcomes in this domain, of which all reported short-term outcomes and two reported long-term outcomes. Of the eight trials reporting GBV-related outcomes in this domain, all but one reported short-term outcomes and three reported long-term outcomes.

Attitudes to personal help-seeking includes outcomes based on participants’ attitudes and intentions to seek help when experiencing DRV or GBV. While scales were primarily bespoke in nature, an indicative outcome measure is ‘belief in need for help’ from Safe Dates (Foshee et al. 1998). 36 Eight trials presented DRV-related outcomes in this domain, of which all reported short-term outcomes and three reported long-term outcomes. Four trials presented GBV-related outcomes in this domain, of which three reported short-term outcomes and one reported long-term outcomes.

Individual self-efficacy includes participants’ intentions for, and confidence in, their ability to avoid DRV or GBV as either victim or perpetrator. Scales used to measure this construct were often bespoke in nature, but measured, for example, participants’ responses in hypothetical DRV or GBV situations (Taylor et al. 2008, 2011). 100,248 Five trials reported DRV-related individual self-efficacy outcomes, and two trials reported GBV-related outcomes. All were short-term.

Social norms include outcomes related to participants’ perceptions of injunctive norms and peers’ behaviours relating to DRV and GBV. Scales were bespoke in nature but measured, for example, whether participants would ‘[believe] others would approve of responding assertively to pressure’ (Arnab et al. 2013). 272 Three trials reported DRV-related social norm outcomes, of which two reported short-term outcomes and two reported long-term outcomes. Only one trial reported GBV-related social norm outcomes and these were short term.

A seventh construct that we did not meta-analyse relates to DRV and GBV-related behaviours. These related to help-seeking, harm reduction and preventive behaviours, and outcomes were heterogeneous in nature. Of the six trials reporting DRV-related behaviours in this category, five presented short-term outcomes and one presented long-term outcomes. Of the three trials reporting GBV-related behaviours in this category, two presented short-term outcomes and two presented long-term outcomes.

The eighth and final construct, bystander behaviours, was meta-analysed combining DRV and GBV estimates. This was because included bystander outcomes frequently did not distinguish between whether bystander behaviours were with respect to DRV or GBV. Bystander behaviours included proactive bystander behaviours, such as initiating conversations on DRV or GBV with peers; reactive bystander behaviours, such as intervening when witnessing abuse; and supportive bystander behaviours, such as offering support to victims of DRV or GBV. Of the nine trials presenting bystander outcomes, six presented short-term outcomes and six presented long-term outcomes.

Dating and relationship violence

Gender-based violence

Role of knowledge and attitudes

With regard to DRV, there was limited evidence that interventions led to improved knowledge and attitudes. Where evidence of an improvement was shown, this was in a subset of trials, and the evidence was greater at short- rather than long-term follow-up. The majority of attitudinal outcomes were related to perceived acceptability of violence, with fewer studies evaluating the impact of interventions on group norms, or attitudes towards help-seeking and intervening in risky situations.

Conversely, the majority of interventions for GBV showed effects in reducing violence acceptance and increasing student knowledge about GBV, though there was uncertainty in the magnitude of these effects. Effects for violence acceptance and knowledge were also greater in the short-term compared to long-term follow-up. There was limited evidence for other attitudinal outcomes, and the evidence available suggested that interventions were not effective at increasing willingness to intervene in GBV violence, or in personal help-seeking. However, there was some limited evidence that GBV interventions may increase individual self-efficacy in responding to GBV. Only one trial evaluated whether an intervention affected group norms to GBV, which did not show effectiveness.

Surprisingly, few interventions that were associated with a reduction in DRV and GBV victimisation or perpetration also reported knowledge and attitudinal outcomes. Two DRV interventions (Teen Choices,182 Katie Brown186), each associated with reductions in DRV perpetration and victimisation, reported an improvement in attitudes towards DRV. Of these, only the evaluation of Katie Brown also measured knowledge about DRV, which was also found to increase following the intervention. Notably however, two other interventions with mixed or inconclusive evidence for DRV violence nevertheless reported an improvement in attitudes. 36,49 Furthermore, GBV interventions that showed improvements in attitudes to violence and in knowledge did not all also show a reduction in violence perpetration or victimisation. This suggests that changing knowledge or attitudes to violence may not be sufficient for affecting violent behaviour. The potential mediating role of attitudes and knowledge for DRV and GBV victimisation and perpetration is discussed in Chapter 9.

Bystander outcomes

Meta-analyses of bystander outcomes across DRV and GBV did not suggest a short-term effect [d = 0.17, 95% CI (−0.01 to 0.34), I² = 76%] based on six studies and 23 effect sizes. This was similarly the case for long-term meta-analyses [d = 0.15, 95% CI (−0.32 to 0.61), I² = 89%] based on six studies and 25 effect sizes.

Comparability of dating and relationship violence and gender-based violence evidence

Few interventions reported both DRV and GBV outcomes. Those interventions associated with a reduction in DRV victimisation or perpetration did not report GBV outcomes, and therefore it was not possible to compare findings for GBV outcomes. Overall, current interventions were more effective at reducing DRV than GBV. This may suggest that the mechanisms involved in reducing GBV are less well understood, or else that GBV victimisation and perpetration may be less amenable to change than DRV.

Stratification by country context

Metaregressions comparing HIC contexts with LMIC contexts were estimated for DRV victimisation, DRV perpetration, GBV victimisation and GBV perpetration both in the short term and in the long term; and for DRV-related violence acceptance, DRV-related knowledge, GBV-related violence acceptance and GBV-related knowledge in the short term. Findings are presented in Table 17, where intercept represents effectiveness in high-income contexts and coefficient represents the difference in effectiveness in low-income and middle-income contexts. Further description of findings is presented in Report Supplementary Material 1. While effectiveness was comparable in the short term, the analyses showed variation in the effectiveness of interventions for long-term GBV outcomes: in the long term, interventions were associated with a reduction in GBV perpetration and victimisation in HIC contexts, but not in LMIC contexts. This may reflect the availability of resource to implement interventions, though gender inequality is also more pronounced in LMIC contexts. 273

Publication bias analyses

Publication bias analyses are undertaken for each of DRV victimisation, DRV perpetration, GBV victimisation and GBV perpetration both in the short term and in the long term, and for short-term DRV-related violence acceptance, short-term DRV-related knowledge, short-term GBV-related violence acceptance, and short-term GBV-related knowledge. Analyses for victimisation and perpetration are plotted on the log odds scale, whereas analyses for violence acceptance and knowledge are plotted on a continuous-scale parameterised by standardised mean differences. The findings are provided in Report Supplementary Material 1. Overall, the analyses showed evidence of publication bias in short-term DRV victimisation, DRV perpetration and GBV victimisation, and in long-term DRV perpetration and GBV victimisation. In most cases, bias was towards publication of positive intervention effects by smaller trials, though for GBV victimisation smaller trials were more likely to report negative effects.

Characteristics of included mediation evidence

Three analyses186,241,274 relating to three outcome evaluations considered mediation for DRV victimisation, five36,186,238,241,274 analyses relating to four outcome evaluations considered mediation for DRV perpetration, one analysis274 considered mediation for GBV victimisation and three238,239,274 analyses relating to two outcome evaluations considered mediation for GBV perpetration. One analysis181 considered mediation for bystander outcomes.

Analytical methods used to test and estimate mediational relationships could be broadly classed as structural equation modelling-based (e.g. path analysis or conditional process analysis, used in five papers)181,186,238,239,274 or causal steps-based (e.g. checking for attenuation of intervention effect on the outcome after inclusion of a mediator; used in three papers36,238,241). Three of the included papers238,241,274 considered outcomes longitudinally (i.e. looking at mediation impacts over time), of which two238,274 included only long-term follow-up periods and one241 included both short-term and long-term follow-up periods. Five36,181,186,238,239 of the included papers considered outcomes cross-sectionally (i.e. looking at end-point differences), of which three36,181,186 tested mediation for short-term outcomes only and two238,239 tested mediation for long-term outcomes only.

Mediation of dating and relationship violence victimisation outcomes

Of the three analyses36,186,274 considering mediation of DRV victimisation, two186,274 considered omnibus DRV victimisation outcomes and one36 considered physical DRV victimisation. One analysis186 considered only those reporting experience of dating at baseline.

Mediation of DRV victimisation outcomes via violence acceptance was assessed in two analyses. 186,241 In one analysis241 considering longitudinal mediation in the Safe Dates trial, prescribed norms mediated the longitudinal impact of the intervention on physical DRV victimisation, based on a criterion of attenuating intervention effects to non-significance (intervention effect p = 0.01 to p = 0.46). In another analysis186 considering short-term mediation in the KBEP, approval of retaliatory aggression [indirect effect −0.32, 95% CI (−0.70 to −0.07)] and, marginally, attitudes towards dating violence [indirect effect −0.30, 95% CI (−0.67 to 0.00)] mediated impacts on total DRV victimisation.

Mediation of DRV victimisation outcomes via knowledge was also assessed in these analyses. In the longitudinal analysis of Safe Dates,36 inclusion of awareness of community services as a mediator in regression models attenuated the intervention effect (p = 0.01 to p = 0.08). However, in the short-term analysis of the KBEP, the indirect effect of the intervention on DRV victimisation through knowledge outcomes was not significant [indirect effect −0.12, 95% CI (−0.94 to 0.28)].

The longitudinal analysis of Safe Dates36 tested a number of additional mediators. Gender stereotyping (intervention effect p = 0.01 to p = 0.29), but not conflict management skills (intervention effect p = 0.01) or belief in need for help (intervention effect p = 0.03), mediated longitudinal impacts on physical DRV victimisation. Finally, the longitudinal analysis of Second Step274 tested longitudinal change in school belonging over 3 years post baseline as a mediator of longitudinal change in DRV victimisation over 4–6 years post baseline, finding a non-significant indirect effect.

Mediation of dating and relationship violence perpetration outcomes

Of the five analyses36,186,238,241,274 considering mediation of DRV perpetration outcomes, two186,274 considered omnibus DRV perpetration outcomes, three36,238,241 considered emotional DRV perpetration outcomes, three36,238,241 considered physical DRV perpetration outcomes and two36,241 considered sexual DRV perpetration outcomes. Two36,186 analyses considered subgroups defined by baseline dating experience, and one analysis36 considered subgroups defined by baseline experience of DRV perpetration.

Four analyses36,186,238,241 relating to three outcome evaluations tested mediation of impacts on DRV perpetration via violence acceptance. In one longitudinal analysis238 testing mediation of DRV perpetration in Green Dot over 4 years, violence acceptance (defined as rape myth acceptance and acceptance of dating violence) did not attenuate intervention effects on physical DRV perpetration (intervention effect p < 0.0001 to p = 0.0001) but did attenuate intervention effects on emotional DRV perpetration (intervention effect p = 0.0005 to p = 0.03). In the short-term analysis of mediation in Safe Dates,36 violence acceptance was assessed by prescribed norms, proscribed norms, perceived positive consequences of using DRV and perceived negative consequences of using DRV. Mediation by violence acceptance was assessed in the full sample for emotional DRV perpetration, physical DRV perpetration in the current relationship and sexual DRV perpetration using all mediators; for emotional DRV perpetration in the sample of baseline daters, using prescribed norms, proscribed norms and negative consequences of DRV as mediators; and for emotional DRV perpetration and sexual DRV perpetration in the sample of those with baseline experience of DRV perpetration, using negative consequences of DRV as a mediator. Prescribed norms attenuated intervention effects by more than 20%, meeting authors’ criteria for mediation, in all DRV perpetration outcomes tested in the full sample. Perceived negative consequences of DRV mediated intervention impacts on sexual DRV perpetration in the sample with baseline experience of DRV perpetration. No other violence acceptance variables were found to mediate any of the DRV perpetration outcomes tested. In the longitudinal analysis of mediation in Safe Dates,241 violence acceptance was measured via prescribed norms. Based on a criterion of attenuating intervention effects to non-significance, prescribed norms attenuated longitudinal impacts over 4 years post baseline for all DRV perpetration outcomes tested: emotional DRV perpetration (intervention effect p = 0.0005 to p = 0.16); physical DRV perpetration, both moderate (p = 0.02 to p = 0.51) and severe (p = 0.04 to p = 0.09) and sexual DRV perpetration (p = 0.04 to p = 0.81). Finally, in the short-term mediation analysis186 of the KBEP, approval of retaliatory aggression [indirect effect −0.37, 95% CI (−0.78 to −0.10)] and attitudes towards dating violence [indirect effect −0.30, 95% CI (−0.70 to −0.02)] mediated intervention impacts on omnibus DRV perpetration.

Mediation of DRV perpetration outcomes by knowledge was assessed in three analyses36,186,241 covering two outcome evaluations. In the short-term mediation analysis of Safe Dates, awareness of community services was tested as a mediator in the full sample for emotional DRV perpetration, physical DRV perpetration in the current relationship and sexual DRV perpetration; and in the sample of those with baseline DRV perpetration, for emotional DRV perpetration and sexual DRV perpetration. Awareness of community services was a mediator for all outcomes tested except for physical DRV perpetration in the current relationship. In the longitudinal mediation analysis of Safe Dates,241 awareness of community services mediated intervention effects over 4 years on moderate physical DRV perpetration (intervention effect p = 0.02 to p = 0.13) and sexual DRV perpetration (p = 0.04 to p = 0.15) but not emotional DRV perpetration (p = 0.0005 to p = 0.01) or severe physical DRV perpetration (p = 0.04 to p = 0.05). In the short-term mediation analysis of the KBEP, DRV knowledge did not mediate intervention impacts on omnibus DRV perpetration [indirect effect −0.09, 95% CI (−0.63 to 0.26)].

The short-term36 and longitudinal241 mediation analyses of Safe Dates considered several other mediators, specifically gender stereotyping, belief in the need for help and conflict management skills. Gender stereotyping was tested as a mediator in the short-term mediation analysis of Safe Dates for emotional DRV perpetration, physical DRV perpetration in the current relationship, and sexual DRV perpetration; and in the sample of those with baseline dating experience, for emotional DRV perpetration, mediating intervention effects on emotional DRV perpetration and sexual DRV perpetration only. In the longitudinal analysis of Safe Dates, gender stereotyping mediated intervention effects over 4 years for emotional DRV perpetration (intervention effect p = 0.0005 to p = 0.06), moderate physical DRV perpetration (p = 0.02 to p = 0.19) and sexual DRV perpetration (p = 0.04 to p = 0.34), but not severe physical DRV perpetration (p = 0.04 to p = 0.05). Belief in the need for help did not mediate any intervention effects in the short-term mediation analysis of Safe Dates, where it was tested in the full sample; in the longitudinal mediation analysis of Safe Dates, belief in the need for help mediated intervention effects for sexual DRV perpetration (intervention effect p = 0.04 to p = 0.07) but not emotional DRV perpetration (p = 0.0005 to p = 0.001), moderate physical DRV perpetration (p = 0.02 to p = 0.001) or severe physical DRV perpetration (p = 0.04 to p = 0.05). Finally, conflict management skills did not mediate any full-sample intervention effects in either the short-term mediation analysis or in the longitudinal mediation analysis, where intervention effects were not attenuated in any analysis (all p-values unchanged).

Bystander actions were tested as mediators in the longitudinal analysis of Green Dot. 238 Bystander actions, both proactive and reactive, did not mediate intervention impacts on physical DRV perpetration (intervention effect p < 0.0001 to p = 0.0001) or emotional DRV perpetration (intervention effect p = 0.0005 to p = 0.001).

Finally, the mediational analysis of Second Step274 tested longitudinal change in school belonging over 3 years post baseline as a mediator of longitudinal change in DRV perpetration over 4–6 years post baseline, finding a non-significant indirect effect.

Mediation of gender-based violence victimisation outcomes

Only one analysis,274 relating to the trial of Second Step,192 tested mediational impacts on GBV victimisation. In this analysis, longitudinal change in school belonging over 3 years post baseline was tested as a mediator of longitudinal change in sexual harassment victimisation and homophobic teasing victimisation over 4–6 years post baseline. In both cases, an indirect effect was significant at p < 0.10 (indirect effect −0.004, p = 0.088 for sexual harassment victimisation; indirect effect −0.004, p = 0.074 for homophobic teasing victimisation). This indicated that the intervention reduced longitudinal increases in GBV victimisation by causing longitudinal improvements in school belonging.

Mediation of gender-based violence perpetration outcomes

Of three analyses238,239,274 considering mediation of GBV perpetration outcomes, all considered omnibus GBV perpetration outcomes, two239,274 considered homophobic GBV and one analysis238 considered both verbal GBV victimisation and physical GBV victimisation. Mediators were disparate across included analyses and no two analyses considered the same set of mediators.

A mediational analysis of Green Dot238 included both longitudinal analyses and end-point difference analyses of long-term outcomes. Longitudinal analyses over 4 years considered the mediating impacts of violence acceptance (defined as rape myth acceptance and acceptance of dating violence) and bystander actions on various forms of GBV perpetration. Violence acceptance mediated all longitudinal intervention effects, defined as attenuation of intervention significance to a significance level of p > 0.01, specifically for physical GBV perpetration (defined as sexual violence, intervention effect p = 0.0003 to p = 0.02), two forms of verbal GBV perpetration (sexual harassment: p = 0.002 to p = 0.05; stalking: p = 0.009 to p = 0.06) and an omnibus measure of GBV perpetration (p = 0.0003 to p = 0.02). However, bystander actions did not mediate physical GBV perpetration (to p = 0.003), emotional GBV perpetration defined as sexual harassment (to p = 0.01), or an omnibus measure of GBV perpetration (to p = 0.001). Bystander actions did mediate longitudinal effects on stalking as a form of emotional GBV perpetration (to p = 0.06). A subsequent analysis considered mediation of end-point differences in physical GBV perpetration defined as sexual violence, examining long-term follow-up at the end of the second and fourth intervention years. None of these analyses produced significant indirect effects for violence acceptance or bystander actions.

A mediational analysis of Second Step239 considering outcomes after 3 years considered whether intervention impacts on an omnibus measure of GBV perpetration and on homophobic GBV perpetration were mediated by change in delinquency behaviours. In both cases, indirect effects were substantial, and were significant for omnibus GBV perpetration (indirect effect −0.011, SE = 0.006) and marginally significant for homophobic GBV perpetration (indirect effect −0.041, SE = 0.018, p = 0.065).

Finally, the mediational analysis of Second Step274 tested longitudinal change in school belonging over 3 years post baseline as a mediator of longitudinal change in GBV perpetration over 4–6 years post baseline, finding a non-significant indirect effect for an omnibus measure of GBV perpetration (indirect effect −0.001, p = 0.347) but a significant indirect effect at p < 0.10 for homophobic GBV perpetration (indirect effect −0.005, p = 0.086).

Mediation of bystander outcomes

One analysis181 considered mediation of short-term impacts of an intervention on bystander outcomes, measured as self-reported and observed bystander behaviours. Self-efficacy did not mediate intervention impacts on self-reported bystander behaviours but did generate a significant indirect effect for observed bystander behaviours [indirect effect. 93, 95% CI (0.46 to 19.06)].

Characteristics of included moderation evidence

Moderation analyses were classified by equity-relevant characteristics. These are presented in descending order of frequency, focusing on active versus control comparisons.

The largest category of moderator analyses covered sex, comparing boys versus girls as a binary variable (of note is that no included moderation analyses consider gender variation as opposed to sex). These moderator analyses were reported in 22 publications36,81,82,85,94,157,164,168,175,176,182,185–187,189,190,193,240–242,276,277 covering 20 outcome evaluations. Nine reports36,81,96,182,187,240–242,276 from six outcome evaluations considered prior history of the outcome as a moderator (e.g. whether prior history of DRV perpetration moderates intervention impacts in reducing DRV perpetration).

Seven reports168,182,186,240–242,277 from five outcome evaluations considered ethnicity as a moderator, contrasting majority populations against ethnic minority populations. A further moderator analysis189 from an outcome evaluation presented stratified estimates for different ethnic minority groups, but because this was a ‘majority-minority’ population these analyses were incommensurate with other moderator analyses. Age, defined in various ways (e.g. biological age, grade level), was tested as a moderator in four reports168,174,182,277 covering four outcome evaluations. Three reports36,188,240 relating to two outcome evaluations tested dating history as a moderator of intervention impact. Two reports255,277 relating to two outcome evaluations tested sexuality (i.e. sexual minority status) as a moderator of intervention impact. Acculturation185 and poverty status277 were tested as moderators in one outcome evaluation each.

Two reports166,275 of one outcome evaluation comparing only active versus active interventions considered moderation of intervention impacts by sex.

A common issue in included analyses was the lack of a formal test for moderation, by significance of an interaction term, a Wald test or a likelihood ratio test comparing nested models. This affected 11 reports36,82,85,164,174,185,187–189,240,255 of 10 outcome evaluations comparing active versus control comparisons. Of these, we were able to construct estimates of effect modification for six reports82,164,174,187,189,255 covering five outcome evaluations. For the remaining five reports,36,85,185,188,240 we were able to approximate direction and likely significance of interaction estimates based on additional within-trial information (e.g. information on precision and sample size, and similarity of effect estimates between groups).

Introducing harvest plots

Harvest plots are used throughout this chapter to depict graphically the findings of moderation analyses. These are presented for moderation analyses of sex and of prior history of the outcome because these two moderators were the most populated within outcome categories. Remaining moderators were too sparse for harvest plots to be meaningful (e.g. because reporting of findings was ‘non-significant’ only).

Harvest plots are organised by moderator, with outcomes forming rows. In each row, individual bars represent trials. Bars are thus organised by whether trial evidence suggests greater impact for one group (e.g. boys), greater impact for another group (e.g. girls) or no gradient of impact. This means that analysis of each row is based on the collective weight of evidence based on the distribution of the bars as to whether the intervention is more effective for one of the two groups implied by the moderator, or whether there is on balance no evidence to suggest a definitive difference in effectiveness between groups.

Bars vary by height, colour and shading, and have a number that represents the number of relevant interaction tests for that moderator and outcome. To minimise double-counting, the number of relevant interaction tests is the number of tests for non-overlapping outcome constructs; that is where tests are undertaken for multiple types of violence and for a violence construct that is the sum of all of these types, overall tests are not reflected in the harvest plots. The height of bars represents the significance and direction of moderation presented by a trial. Full height bars represent significant moderation for a given outcome demonstrating greater intervention impact for one group. Three-quarters height bars represent a pattern of moderation estimates including some, but not consistent, significant evidence of greater intervention impact for one group. Half-height bars represent consistently non-significant estimates of moderation trending in one direction, and quarter-height bars are used for studies that only present non-significance as opposed to moderation estimates. The colour of bars represents the time frame of outcomes analysed. Blue bars represent moderation evidence from long-term outcomes, red bars represent moderation evidence from short-term outcomes and purple bars represent moderation evidence from both short-term and long-term outcomes. Finally, the shading of bars represents the nature of the moderation analysis. Dark shaded bars represent formal interaction tests, whereas light shaded bars represent outcome evaluations where formal interaction tests were not presented.

Moderation of dating and relationship violence victimisation

Moderation of intervention impacts on DRV victimisation was tested in 1736,81,82,94,157,182,185–189,193,240–242,255,276 reports of 13 outcome evaluations. Moderators included sex, prior history of the outcome, dating history, age, ethnicity, acculturation and sexuality.

Sex as a moderator of dating and relationship violence victimisation outcomes

In total, 15 reports36,81,82,94,157,182,185–187,189,193,240–242,276 of 12 outcome evaluations reported moderation of DRV victimisation outcomes. Two trials82,157 demonstrated a pattern of non-significant findings where moderation effects tended to favour boys. An analysis of Green Dot82 presented evidence stratified by sex for long-term findings for physical and emotional DRV victimisation; we calculated interaction effects and did not find that the prevalence risk ratio for any outcome measure was significantly different between boys and girls, though the pattern of effects on balance favoured boys (five out of eight interaction tests). An analysis of subgroup effects from the Cleveland trial of law and justice and interaction-based programmes157 drew on short-term evidence and included 24 tests, of which 16 were ‘unique’, covering omnibus measures of DRV victimisation and sexual DRV victimisation. No interaction tests were significant, but 18 of 24 tests (12–16 unique tests) suggested a greater impact on boys.

An additional 8 outcome evaluations in 11 reports36,94,182,185–187,189,193,240–242 did not offer evidence of a gradient for DRV victimisation outcomes. Two of these reports189,193 presented subgroup specific estimates. In the evaluation of JOVEN,193 neither test of moderation over both short-term and long-term time categories yielded a significant result; however, stratified evidence suggested a greater impact of the intervention for girls on emotional DRV victimisation and a greater impact for boys on physical and sexual DRV victimisation. Nine reports36,94,182,185–187,240–242 of six outcome evaluations only reported significance. Moderation analyses of Safe Dates were presented in four reports36,240–242 covering short-term,36 long-term,240,242 and longitudinal241 (both short-term and long-term) analyses of outcomes, representing 24 tests in total. In all reports, moderation tests were presented for physical DRV victimisation, sexual DRV victimisation and emotional DRV victimisation and all were reported as non-significant. A moderation test of Teen Choices182 for long-term physical and sexual DRV victimisation and emotional DRV victimisation also reported both tests as non-significant. The remaining four moderation evaluations all presented evidence in respect of short-term outcomes. Moderation tests for the KBEP186 included five tests, of which four were unique, covering an omnibus measure of DRV victimisation, emotional DRV victimisation and physical DRV victimisation. All tests were reported as non-significant. Three tests of moderation in the evaluation of Ending Violence,185 all of which were omnibus measures of DRV victimisation, were reported as non-significant. Three tests of moderation (two unique) in the evaluation of the School Health Center Healthy Adolescent Relationships Program (SHARP)187 related to cyber DRV victimisation, physical and sexual DRV victimisation, and an omnibus measure of DRV victimisation. All were reported as non-significant. Lastly, an evaluation of two programmes94 against a control condition yielded two tests of sexual DRV victimisation, reporting non-significant results.

A final two evaluations81,276 suggested greater impact for girls. The New York City evaluation of Fourth R81 included two tests of moderation of omnibus measures of DRV victimisation. A short-term analysis yielded a numerical, but not significant, estimate of greater impact for boys, but a long-term analysis yielded a significant test (p < 0.05) suggesting greater impact for girls. Finally, the moderation analysis for Shifting Boundaries included nine tests covering omnibus measures of DRV victimisation and sexual DRV victimisation. While all nine tests were reported as non-significant, five of six tests of sexual DRV victimisation suggested a numerically greater impact for girls. All three tests of omnibus measures were reported as non-significant only.

The resultant harvest plot for this outcome and moderator is shown in Figure 11. Taking all outcomes together, the harvest plot does not provide evidence for a gradient in effectiveness on DRV victimisation outcomes by sex. Because of the size of the body of evidence, we undertook an exploratory analysis stratifying outcomes by type of DRV victimisation. This did not uncover any evidence of sex-specific moderation on different types of DRV victimisation.

FIGURE 11.

Harvest plot, DRV victimisation, boys vs. girls.

Prior history of the outcome as a moderator of dating and relationship violence victimisation outcomes

Eight reports36,81,182,187,240–242,276 relating to five outcome evaluations reported prior history of DRV victimisation as a moderator of intervention effectiveness on DRV victimisation. The evaluation of Teen Choices182 included two tests of long-term outcomes relating to physical and sexual DRV victimisation and emotional DRV victimisation. Both tests were significant at p < 0.01 level, reflecting a substantially greater impact for those with prior history of DRV victimisation in reducing future DRV victimisation. The New York City evaluation of Fourth R81 included two tests of moderation of omnibus measures of DRV victimisation. A short-term analysis yielded a numerical, but not significant, estimate of greater impact for those with no prior history of the outcome, but a long-term analysis yielded a significant test (p < 0.05) suggesting greater impact for those with prior history. Three tests of moderation (two unique) in the evaluation of the SHARP187 related to cyber DRV victimisation, physical and sexual DRV victimisation and an omnibus measure of DRV victimisation. All favoured those with prior history of DRV victimisation, with both the cyber DRV victimisation and the omnibus measure generating significant tests. Lastly, the evaluation of Shifting Boundaries276 included nine tests covering omnibus measures of DRV victimisation and sexual DRV victimisation. While all nine tests were reported as non-significant, four of six tests of sexual DRV victimisation suggested a numerically greater impact for those with prior history of DRV victimisation. All three tests of omnibus measures were reported as non-significant only.

As noted above, moderation analyses of Safe Dates were presented in four reports36,240–242 covering short-term,36 long-term,240,242 and longitudinal241 (both short-term and long-term) analyses of outcomes, representing 24 tests in total. As for sex as a moderator, in all reports, moderation tests were presented for physical DRV victimisation, sexual DRV victimisation and emotional DRV victimisation. In short-term and longitudinal tests,36,241 all results were reported as non-significant. However, long-term effects240 at 1 year post intervention were all imputed as non-significant but favouring those with no prior history of DRV victimisation (four tests), while five out of eight tests at 4 years post intervention resulted in significant estimates demonstrating greater effectiveness for those with no prior history of DRV victimisation.

The resultant harvest plot for this outcome and moderator is shown in Figure 12. Taking all outcomes together, the harvest plot does not provide evidence for a gradient in effectiveness on DRV victimisation outcomes by prior history of the outcome.

FIGURE 12.

Harvest plot, prior history of the outcome, all outcomes.

Moderation of dating and relationship violence perpetration

Moderation of intervention impacts on DRV perpetration was tested in 18 reports36,81,82,94,157,164,182,185,186,188,190,193,240–242,255,276 of 14 outcome evaluations. Moderators included sex, prior history of the outcome, dating history, age, ethnicity, acculturation and sexuality.

Sex as a moderator of dating and relationship violence perpetration outcomes

Sex was considered as a moderator of DRV perpetration outcomes in 16 reports36,81,82,94,157,164,182,185,186,189,190,193,240–242,276 of 13 outcome evaluations. Four evaluations provided some evidence of a gradient in effectiveness favouring boys. A long-term test of the impact of Fourth R190 on physical DRV perpetration found evidence of a greater impact on boys (p = 0.002). Three further evaluations found non-significant evidence of a greater impact on boys. In JOVEN, two longitudinal tests of emotional DRV perpetration and physical and sexual DRV perpetration found numerical but not statistical evidence of a greater impact on boys. Across eight long-term tests of emotional DRV perpetration and physical DRV perpetration from the evaluation of Green Dot,82 five suggested numerically greater impact on boys as compared to girls. Nine short-term tests from the evaluation of Shifting Boundaries276 included three tests of an omnibus measure of DRV perpetration, all of which were reported as non-significant only, and six tests of sexual DRV perpetration, of which four numerically favoured boys, one numerically favoured girls and one did not favour either group.

A further seven evaluations36,81,94,182,185,186,189,240–242 did not provide evidence of a gradient in effectiveness by sex. One evaluation189 included two long-term tests: one of physical DRV perpetration, which numerically favoured girls, and one of emotional DRV perpetration, which numerically favoured boys. The other six evaluations36,81,94,182,185,186,240–242 presented findings as non-significant only. The evaluation of Safe Dates36,240–242 included 24 tests covering short-term,36 long-term240,242 and longitudinal241 outcomes for emotional DRV perpetration, physical DRV perpetration and sexual DRV perpetration. The New York City evaluation of Fourth R81 included one short-term and one long-term test of an omnibus measure of DRV perpetration. Two long-term tests from the evaluation of Teen Choices182 included emotional DRV perpetration and physical DRV perpetration. The remaining three evaluations94,185,186 presented short-term tests only: from the KBEP,186 five tests, four unique, covering an omnibus measure of DRV perpetration, physical DRV perpetration and emotional DRV perpetration; from another evaluation, two short-term tests of sexual DRV perpetration;94 and finally, from one evaluation,185 a test of an omnibus measure of DRV perpetration.

A final two evaluations157,164 provided short-term, non-significant evidence of a greater impact on girls. Across 24 tests (16 unique) including omnibus measure of DRV perpetration and sexual DRV perpetration, 21 tests indicated a numerically greater impact on girls, and three indicated a numerically greater impact on boys. 157 In another evaluation, intervention impacts on both emotional DRV perpetration and physical DRV perpetration showed numerically greater benefits for girls.

The resultant harvest plot for this outcome and moderator is shown in Figure 13. Taking all outcomes together, the harvest plot provides some evidence of a gradient in effectiveness on DRV perpetration favouring greater impacts for boys. This appears to be primarily driven by long-term and longitudinal evidence. Because of the size of the body of evidence, we undertook an exploratory analysis stratifying outcomes by type of DRV perpetration. This suggested that greater impacts on boys were primarily driven by evidence for emotional and physical DRV perpetration.

FIGURE 13.

Harvest plot, DRV perpetration, boys vs. girls.

Moderation of gender-based violence victimisation

Moderation of intervention impacts on GBV victimisation was tested in 13 reports81,82,85,94,96,157,168,174–176,255,276,277 of 13 outcome evaluations. Moderators included sex, prior history of the outcome, age, ethnicity, sexuality and poverty status.

Sex as a moderator of gender-based violence victimisation outcomes

Sex as a moderator of GBV victimisation outcomes was considered in 10 reports81,82,85,94,157,168,175,176,276,277 of 11 outcome evaluations. Four outcome evaluations85,157,175,176 suggested a greater impact for boys. One report of an intervention in South Africa176 considered physical GBV victimisation (specifically, forced sex) over short-term and long-term time frames. All six tests indicated a greater impact on boys than girls, with three generating a significant result. Long-term evidence from an outcome evaluation in India85 included four relevant tests of an omnibus measure of GBV. Three of four tests favoured boys, and two of these were imputed as significant. Long-term evidence from an outcome evaluation of the Good Schools Toolkit175 included three tests of an omnibus measure of GBV victimisation. All were non-significant, but all three numerically favoured boys. Lastly, short-term evidence from the Cleveland evaluation of interaction-based and law and justice-based interventions157 generated eight tests of moderation by sex. All were non-significant, but five numerically favoured boys.

A further five evaluations81,82,94,168,277 did not offer evidence of a gradient in effectiveness by sex. The evaluation of Green Dot provided long-term evidence of moderation by sex over 24 tests relating to physical GBV victimisation and verbal GBV victimisation, 20 of which were unique. While all were non-significant, of the 20 unique tests, 2 demonstrated no numerical difference between groups, 9 suggested a numerical benefit for boys and 9 suggested a numerical benefit for girls. Of the remaining four non-unique tests, three favoured girls, though all were non-significant. A subsequent four evaluations81,94,168,277 only presented evidence of non-significance. The New York City evaluation of Fourth R81 included short-term and long-term tests of an omnibus measure of GBV victimisation. Another evaluation277 presented four long-term tests, of which two were of an omnibus measure of GBV victimisation and two were of verbal GBV victimisation. In the remaining two evaluations, all tests were short-term only. One evaluation168 reported two tests of an omnibus measure of GBV victimisation, and one evaluation94 reported two tests of physical GBV victimisation.

Finally, two evaluations suggested a greater impact on girls. One outcome evaluation85 in Vietnam presented three long-term tests (two unique) of omnibus measures of GBV victimisation and of physical GBV victimisation. While all three tests numerically favoured girls, two were imputed as being statistically significant. Short-term tests from the New York City evaluation of Shifting Boundaries276 included six tests of an omnibus measure of GBV victimisation. All were non-significant, but five of six yielded numerically greater benefits for girls.

The resultant harvest plot for this outcome and moderator can be seen in Figure 14. On balance, the harvest plot suggests the plausibility of a gradient in intervention effects favouring boys, driven primarily by longer-term evidence. An exploratory analysis stratifying outcomes by type of GBV victimisation did not reveal a clear pattern explaining this distribution of study-level moderation findings, though longer-term evidence from omnibus measures of GBV victimisation did appear to favour greater impact on boys.

FIGURE 14.

Harvest plot, GBV victimisation, boys vs. girls.

Moderation of gender-based violence perpetration

Moderation of intervention impacts on GBV perpetration was tested in 10 reports81,82,85,94,157,168,176,255,276,277 of 10 outcome evaluations. Moderators included sex, prior history of the outcome, age, ethnicity, sexuality and poverty status.

Sex as a moderator of gender-based violence perpetration outcomes

Nine reports81,82,85,94,157,168,176,276,277 of 10 outcome evaluations considered sex as a moderator of GBV perpetration outcomes. One report of an intervention in South Africa176 considered physical GBV perpetration (specifically, forced sex) over short-term and long-term time frames. All six tests indicated a greater impact on boys than girls, with four generating a significant result. A further four outcome evaluations81,94,168,277 reported only that results were not significant: one81 including a short-term and a long-term test of an omnibus measure of GBV perpetration, one277 including four long-term tests of an omnibus measure of GBV perpetration and of verbal GBV perpetration, one94 including two short-term tests of physical GBV perpetration, and one168 including one short-term test of an omnibus measure of GBV perpetration. A final set of five outcome evaluations82,85,157,276 showed a pattern of greater effects for girls. In the long-term evaluation of Green Dot,82 findings for verbal GBV perpetration (eight tests) and physical GBV perpetration (specifically, forced sex outcomes; 16 tests, of which 12 were unique) were stratified by sex. All tests were non-significant, but 5 of 8 tests for verbal GBV perpetration suggested greater impacts for girls and 9 of 16 tests for physical GBV perpetration suggested greater impacts for girls (1 test estimated identical effects). One report85 covering included long-term outcomes for two outcome evaluations in Vietnam and India. Three tests (two unique) from the Vietnam outcome evaluation covered physical GBV perpetration, verbal GBV perpetration and an omnibus measure of GBV perpetration. While all three were imputed as non-significant, effects were numerically more beneficial for girls. Two tests from the India outcome evaluation covered an omnibus measure of GBV perpetration. Though both tests were imputed as non-significant, effects suggested the intervention was more beneficial for girls. Eight tests from the Cleveland evaluation157 of law and justice and interaction-based interventions considered short-term physical GBV perpetration outcomes, and though all were non-significant, five of eight tests indicated numerically greater effects for girls. Finally, the evaluation of Shifting Boundaries276 presented six tests of an omnibus measure of GBV perpetration; four of six tests indicated numerically greater effects for girls, though all were non-significant.

The resultant harvest plot is displayed in Figure 15. A plausible interpretation of the collected evidence is that interventions may be more effective for girls than for boys, though this is countervailed by the evidence from one evaluation176 featuring short-term and long-term follow-up and the relatively high proportion of evaluations reporting non-significance only. As an exploratory analysis, we stratified outcomes by type of GBV victimisation. This suggests that evidence for a gradient favouring greater impact on girls is strongest for omnibus measures and verbal GBV perpetration, albeit with relatively few studies supporting each conclusion.

FIGURE 15.

Harvest plot, GBV perpetration, boys vs. girls.

Dating and relationship violence victimisation

For DRV victimisation up to 1 year post baseline, intervention effectiveness does not appear to be significantly moderated by the percentage of girls in the trial sample [OR = 1.02, 95% CI (0.67 to 1.56)]. However, there was some signal of moderation of intervention effectiveness on DRV victimisation at or more than 1 year post baseline. Each additional increase by 10% points of girls was linked at marginal significance (p = 0.073) with increased effectiveness in reducing DRV victimisation [OR = 0.78, 95% CI (0.59 to 1.04)].

Dating and relationship violence perpetration

For DRV perpetration up to 1 year post baseline, intervention effectiveness does not appear to be moderated by the percentage of girls in the trial sample [OR = 1.01, 95% CI (0.89 to 1.15)]. Similarly, intervention effectiveness at or more than 1 year post baseline does not appear to be moderated by the percentage of girls in the trial sample [OR = 1.05, 95% CI (0.70 to 1.57)].

Gender-based violence victimisation

For GBV victimisation up to 1 year post baseline, the percentage of girls in trial samples was linked with increased effectiveness; with each additional 10% points of girls in the trial sample, the odds of GBV victimisation decreased by 9% [OR = 0.91, 95% CI (0.85 to 0.97)]. Of note is that percentage of girls appeared to be an important source of heterogeneity, as the intercept was statistically significant and indicative of a significant impact of interventions in reducing GBV victimisation [OR = 0.86, 95% CI (0.79 to 0.93)], in distinction to the overall pooled effect for this outcome [OR = 0.88, 95% CI (0.76 to 1.02)]. Interestingly, the relationship between percentage of girls and effectiveness was numerically reversed for outcomes at or more than 1 year post baseline, but the coefficient in this analysis was not statistically significant [OR = 1.21, 95% CI (0.71 to 2.06)].

Gender-based violence perpetration

For GBV perpetration up to 1 year post baseline, the percentage of girls in trial samples was not linked with effectiveness [OR = 0.99, 95% CI (0.69 to 1.41)]. A similar finding was in evidence for outcomes at or more than 1 year post baseline [OR = 1.01, 95% CI (0.67 to 1.51)].

Discussion components

The presence of a discussion-based component was in no analysis significantly associated with improved or worsened intervention effectiveness, though analyses were not estimable for long-term GBV victimisation or perpetration. Based on the direction of coefficients, the presence of a discussion-based component appeared linked with numerically worse effectiveness in all DRV analyses, though this was plausibly related to the relatively small number of interventions without these components. In contrast, the presence of a discussion-based component appeared linked with numerically better effectiveness in short-term analyses for GBV victimisation and perpetration.

Visual components

The presence of a visual component was in no analysis significantly associated with improved or worsened intervention effectiveness, though analyses were not estimable for long-term GBV victimisation. The presence of a visual component appeared linked with numerically better effectiveness in DRV victimisation analyses, but numerically worse effectiveness in DRV perpetration analyses and short-term GBV victimisation analyses. In analyses for GBV perpetration, the inclusion of a visual component was linked to no difference in effect (OR = 1.00 for both analyses).

Competition components

The presence of a competition component was in no analysis significantly associated with improved or worsened intervention effectiveness. In all analyses except for short-term DRV victimisation, the inclusion of a competition component was linked with numerically worse effectiveness.

Guided practice components

The presence of a guided practice component was in no analysis significantly associated with improved or worsened intervention effectiveness, though analyses were not estimable for long-term DRV victimisation, long-term GBV victimisation or long-term GBV perpetration. The presence of a guided practice component was numerically linked with improved effectiveness in short-term DRV victimisation, GBV victimisation and GBV perpetration, but numerically linked with decreased effectiveness in short-term DRV perpetration.

Reflection components

The presence of a reflection component was in no analysis significantly associated with improved or worsened intervention effectiveness, though analyses were not estimable for long-term DRV victimisation, long-term GBV victimisation or long-term GBV perpetration. Reflection components were linked with a numerically greater effectiveness for short-term DRV victimisation and short-term DRV perpetration, but with numerically worse effectiveness for long-term DRV perpetration, short-term GBV victimisation [though this was only marginally non-significant, with OR = 1.41, 95% CI (0.99 to 2.02)] and short-term GBV perpetration.

Narrative-based components

The presence of a narrative component was in no analysis significantly associated with improved or worsened intervention effectiveness, though in every analysis it was numerically linked with worse effectiveness.

School personnel components

Analysis of school personnel components was first undertaken overall, and then considering just school personnel as facilitators and just school personnel as participants. An overall analysis of school personnel components was not estimable for long-term GBV victimisation and perpetration, but in every other analysis presence of any school personnel component was numerically, but not statistically, linked to worse effectiveness.

Focusing on just components where school personnel were trained as facilitators, none of the eight analyses suggested a statistical association between this component and effectiveness; however, there was a numerical link with greater effectiveness for short-term DRV victimisation, no numerical difference for short-term DRV perpetration (OR = 1.01), and a highly imprecise CI for short-term GBV perpetration. In all other analyses, presence of a component with teachers as facilitators was linked with numerically worse effectiveness.

Focusing on just components where school personnel were participants, none of the eight analyses suggested a statistical association with effectiveness. All analyses except long-term DRV perpetration suggested a numerical worsening in effectiveness with the inclusion of a component where school personnel were participants.

Parent and family components

The presence of a parent and family component was in no analysis significantly associated with improved or worsened intervention effectiveness, though analyses were not possible for short-term GBV victimisation or GBV perpetration. While the presence of this component was numerically linked with improved effectiveness for all DRV victimisation and perpetration analyses, the converse was demonstrated for long-term GBV victimisation and long-term GBV perpetration.

Structural environmental components

The presence of a structural environmental component was in no analysis significantly associated with intervention effectiveness. In all analyses except for short-term DRV victimisation, the presence of this component was linked with numerically worse intervention effectiveness.

Structural social components

The presence of a social structural component was in no analysis significantly associated with intervention effectiveness. In short-term and long-term DRV victimisation and perpetration and short-term GBV victimisation, presence of this component was linked to numerically better intervention effectiveness; however, for long-term GBV victimisation and short-term and long-term GBV perpetration, this component was linked to numerically worse intervention effectiveness.

Guidance: self, peer, adult

Due to limited variation within analyses, type of guidance could not be analysed at all for GBV victimisation or GBV perpetration. The presence of a peer-guided component could only be tested in short-term DRV victimisation. This suggested a numerical, but not statistical, benefit of including this component. In contrast, an adult-guided component was linked with numerically worse effectiveness for this outcome. Self-guided components were numerically linked with improved effectiveness in short-term DRV victimisation and DRV perpetration, and long-term DRV perpetration, and statistically associated with improved intervention effectiveness in long-term DRV victimisation.

Gender composition

Single-gender components and mixed-gender components were considered separately. An analysis for single-gender components was not estimable for long-term GBV victimisation; however, in remaining analysis, a signal of effectiveness was unclear. While numerically linked with worse effectiveness for both short-term and long-term DRV victimisation, single-gender components were numerically linked with improved effectiveness for both short-term and long-term DRV perpetration. Single-gender components were statistically associated with improved effectiveness for short-term GBV victimisation, but with no difference in effect (OR = 1.00) for short-term GBV perpetration; and numerically linked with improved intervention effectiveness in long-term GBV perpetration.

In contrast, models for the inclusion of mixed-gender components were not estimable for short-term DRV victimisation, short-term DRV perpetration, long-term GBV victimisation or long-term and short-term GBV perpetration. While numerically linked with improved effectiveness in long-term DRV victimisation, mixed-gender components were numerically linked with worse effectiveness in long-term DRV perpetration and short-term GBV victimisation.

Initial calibration and conditions

Our analysis began by calibrating the outcome using pooled intervention-comparison effect sizes generated for the NMAs. As a starting point, we used a threshold of OR = 0.90 to distinguish ‘most effective’ interventions (outcome value of 1) from ‘not most effective’ (outcome value of 0) interventions. This distinction is important because trials with large sample sizes may generate statistically significant effects on an outcome, but with an effect size that may be too small to be a useful, efficient investment of resources by schools. Because we defined our outcome variable as ‘most effective’ rather than ‘effective’, interventions classed as ‘not most effective’ may have statistically meaningful effects on key outcomes but may not be considered best bets for implementers. In addition, an OR of 0.90 appeared to be a consistent breakpoint in the distribution of intervention-control comparisons included in each analysis and is close to the pooled effect size for most overall pairwise meta-analysis models where a significant result was not in evidence. Once we assigned individual intervention-control comparisons to ‘most effective’ or ‘not most effective’, we sense checked the resulting classifications on the basis of the overall pattern of (disaggregated) results from each trial.

We identified a core set of four conditions to enter into each model, based on the previous syntheses. These conditions were:

• country context of interventions, whether low-income and middle-income or high-income, which was found to be a plausible moderator for most outcomes;

• percentage of girl children in the trial sample, which was found to be a plausible moderator for victimisation outcomes, but which was included for perpetrator outcomes for consistency (and because a configurational analysis might nuance the null findings from metaregressions on perpetration outcomes);

• single-component interventions, given the unexpected effectiveness of these interventions in several pairwise meta-analyses;

• structural social domain of implementation, given the theorised salience of these components in activating higher-level change mechanisms, and numerically indicative results from the component metaregression.

Because percentage of girl children was a continuous variable, it required calibration into a binary condition. We calibrated samples with more than 60% girl children as ‘girls plus’ (condition value of 1) on the basis that a sample with more than 60% girl children would be unlikely to arise by chance alone and would also include interventions delivered to girls exclusively. While we included single-component interventions as a condition, we did not include curriculum or multicomponent interventions as conditions as these types were not clearly linked with effectiveness in meta-analyses.

Upon calibration of outcome and conditions, we did not proceed with models for long-term GBV victimisation and perpetration. The number of observations available in each model (n = 12 and n = 10, respectively) compared to the number of conditions, and the number of interventions that could be reasonably classed as most effective, was too small for any QCA to be meaningful.

Dating and relationship violence victimisation

Dating and relationship violence perpetration

Gender-based violence victimisation: short-term

An initial analysis for short-term GBV victimisation with the four core conditions generated contradictory configurations. Based on our experience of the DRV victimisation analyses, we generated a similar condition for most effectiveness in short-term GBV perpetration. This did not, however, resolve all contradictory configurations.

In scrutinising the remaining contradictory configurations, we noticed that a condition that distinguished most effective interventions from not most effective interventions was the presence of narrative components. Based on the consistent numerical patterns seen in component meta-regressions, we believed that this component could plausibly contribute to understanding pathways to effectiveness. When these two conditions were included in the model, we were able to remove conditions for country context and for social structural domain of implementation as these did not improve the model.

The resultant set of conditions generated a fully consistent model with complete coverage of all most effective interventions. This analysis (Table 27), when taken to Boolean minimisation, reflected the importance of reducing GBV perpetration, but also the importance of avoiding narrative components. A potentially surprising finding is that most effective interventions that were also effective at reducing GBV perpetration were not single-component interventions (reduced configuration abCd), and single-component interventions that were most effective did not need to reduce GBV perpetration to be most effective at reducing GBV victimisation (Bcd). Another pathway (reduced configuration Abc) indicated that interventions that were most effective at reducing GBV victimisation were not single-component interventions and were implemented in samples with high percentages of girl children.

Gender-based violence perpetration: short-term

An initial analysis for short-term GBV victimisation with the four core conditions generated contradictory configurations. Based on our learning from models for DRV perpetration, we included conditions for high levels of guided practice and for the presence of student interpersonal components. This generated a completely consistent model; however, conditions for high percentage of girl children in the trial sample, country context and single-component interventions were no longer necessary to explain patterns of most effectiveness.

Thus, the final model included conditions for social structural domain of implementation, student interpersonal components, and a high level of guided practice. As seen in Table 28, pathways to effectiveness required either the presence of social structural components with student interpersonal components (reduced configuration ABc) or the presence of a high level of guided practice (reduced configuration aC).

Feasibility assessment

We undertook type-based NMAs for all outcomes. We ultimately undertook component-based NMAs using the domains of implementation (intrapersonal, interpersonal, staff, structural social and structural environmental) as our components to balance the number of trials and the number of components entered into the analyses. We did not explore interactions between components as the data structure (in which components tended to ‘stack’) did not support reliable estimation of components. Finally, we only undertook component NMAs for DRV outcomes as the data structures for GBV outcomes were not rich enough to support component NMAs.

We proposed to assess potential intransitivity (i.e. the coherence of the network over multiple comparisons and trials) by effect modification using three key factors: age, LMIC context and percentage of girl children in the trial sample. In the course of our feasibility assessment, we did not note any meaningful or systematic variation by comparison in the network on the basis of age. However, we did note on the basis of prior metaregressions presented in Chapters 6 and 7 that while LMIC context was a potential effect modifier for all analyses, percentage of girl children was primarily an effect modifier for victimisation outcomes only. Thus, we undertook network metaregressions for all analyses with the relevant effect modifiers, using a common regression coefficient across the network.

Dating and relationship violence victimisation: short-term

The network diagram for this NMA is displayed in Figure 16. Based on results from this NMA (Table 29), there was no signal of any intervention type being significantly better than either control or any other intervention type. An overall design-by-treatment interaction test did not find significant evidence of inconsistency in the network (χ² = 0.29, df = 3, p = 0.96). Rankings of intervention effectiveness suggested that single-component and multilevel approaches were ranked highest, with SUCRAs of 0.8 and 0.7, respectively. Sensitivity analysis by either LMIC context or percentage of girl children did not change any conclusions.

FIGURE 16.

Network diagrams for DRV outcomes. (a) Represents short-term DRV victimisation; (b) represents long-term DRV victimisation; (c) represents short-term DRV perpetration; and (d) represents long-term DRV perpetration.