Educational and Skills-Based Interventions for Preventing Relationship and Dating Violence in Adolescents and Young Adults: A Systematic Review

1.1 DESCRIPTION OF THE CONDITION

The term ‘intimate partner violence’ describes actual or threatened physical, sexual or psychological violence that occurs within a relationship or is perpetrated by a current or former partner or spouse (Saltzman 2002). Saltzman 2002 further defines the components of intimate partner violence as follows: •

physical violence is the intentional use of physical force with the potential for causing death, disability, injury or harm;

Physical, sexual and verbal violence can be common responses to conflict within relationships and can have significant effects upon the mental, physical and social well-being of those involved. Although intimate partner violence is often unreported, prevalence within the adult population is estimated to be high, with prevalence rates varying between countries (WHO 2005). For example, one multi-country study conducted by the World Health Organization (WHO) found that between 3% and 54% of women report having experienced physical or sexual violence by an intimate partner in the previous year (WHO 2005). In addition, between 10% and 50% of women report having experienced violence from partners or ex-partners at some point in their lives (Watts 2002).

Rates of relationship abuse vary according to age, sex and previous experience of violence (Foshee 1996; Foshee 1998; Archer 2000). The prevalence of relationship violence is higher in adolescents than in adults, with females aged 12 to 18 years having the highest victimisation rate (Home Office 1999; Wolfe 2003). This form of violence is called dating violence and perpetrators are most likely to be peers (Schewe 2006). Approximately 20% of young women have experienced violence from a dating partner (O'Keeffe 1986; Bergman 1992). Additionally, studies on relationship violence have found that first episodes of violence frequently occur in adolescence (Henton 1983). In younger dating samples, relatively higher proportions of aggression by women against men have been described, although results vary according to the measurement methods used and must, therefore, be interpreted with caution (Archer 2000).

Early experiences of dating violence are linked to poor health outcomes such as sexually transmitted infections (Campbell 2002; WHO 2005; Exner-Cortens 2013), teenage pregnancy (Campbell 2002), substance misuse (Roberts 2003; Tyler 2012), cancer, coronary heart disease, attempted suicide (WHO 2005; Exner-Cortens 2013), depression (Campbell 2002; Roberts 2003; Wolitzky-Taylor 2008), and symptoms of post-traumatic stress disorder (Campbell 2002; Wolitzky-Taylor 2008). Relationship violence during pregnancy is also associated with adverse maternal and neonatal health outcomes including preterm delivery (Campbell 2002), low birthweight (Murphy 2001; Campbell 2002), and foetal death (Campbell 2002). Moreover, adolescents who have experienced dating violence in the past are more likely to be perpetrators or victims of intimate partner violence as adults (Krug 2002; Loh 2006; Chiodo 2012).

1.2 DESCRIPTION OF THE INTERVENTION

This review focuses on educational and skills-based interventions targeted at young people aged 12 to 25 years. It includes primary preventive interventions, where participants may have never experienced or perpetrated relationship violence, and secondary prevention, where participants have experienced or perpetrated relationship violence in the past. This review focuses only upon interventions that actively provide the participants with knowledge and skills aimed at preventing initial or further relationship violence. It does not include ‘screening programmes’ that only offer referral to support agencies. We selected the age group 12 to 25 years to include both adolescents and young adults.

A number of environments can be used to deliver educational and skills-based interventions, including the community, and in particular, within schools and higher education environments (Wolfe 1999). Because schools play an important role in the development of social behaviour, they provide an appropriate environment to target children and adolescents in the prevention of dating violence and subsequently other forms of relationship violence. Previous systematic reviews have focused on the effectiveness of general violence prevention programmes, such as those against aggression and bullying (Mytton 2006; Adi 2007; Park-Higgerson 2008). However, there is further potential to utilise schools and other settings in preventing relationship violence. Studies from the USA suggest that interventions delivered to college-based populations may have an effect on reducing incidences of sexual assault and possibly intimate partner violence (Luthra 2006). Programmes can also be delivered within home (Foshee 2012) and community (Wolfe 2003; Salazar 2006) settings to raise awareness about abuse, promote positive relationships, enable help-seeking and peer support, challenge discriminative viewpoints and encourage the development of protective skills (Wolfe 1999).

1.3 HOW THE INTERVENTION MIGHT WORK

Educational and skills-based programmes aiming to prevent or reduce dating and relationship violence may provide participants with the skills to communicate effectively; deal constructively with stress, disappointment and rejection; resolve conflicts and promote healthier relationships (Wolfe 1999). They may also provide young people with skills to protect themselves from the risk of relationship violence and to improve low self-esteem, which is linked to the likelihood of being a victim of relationship violence (Gidycz 2006).

1.4 WHY IT IS IMPORTANT TO DO THIS REVIEW

The high prevalence of relationship violence and the severity and duration of its health consequences render this area an important public health issue. To date, many review studies have focused either on intimate partner violence or domestic violence in adult populations (Ramsay 2009; Wood 2010), or on the prevention of sexual abuse (Zwi 2009), or general violence (Mytton 2006), in children. We have found only one systematic review of interventions to prevent dating and relationship violence in young people (Whitaker 2006). Whitaker 2006, which reviewed only primary prevention programmes, included non-randomised (e.g. pre- and postinterventions) as well as randomised studies, and summarised results narratively. Our review builds upon these findings by limiting included studies to randomised controlled trials (RCTs) and quasi-RCTs, including primary and secondary prevention programmes, and summarising results in a meta-analysis. Given that interventions to prevent relationship violence based in schools and universities are becoming increasingly widespread, this review is important for strengthening the evidence base, providing a clearer idea of what works, and helping to inform future policy, practice and research in this area.

3.1 CRITERIA FOR CONSIDERING STUDIES FOR THIS REVIEW

3.2 SEARCH METHODS FOR IDENTIFICATION OF STUDIES

We considered both published and unpublished work to be eligible for inclusion in the review. The Cochrane Developmental, Psychosocial and Learning Difficulties Group (CDPLPG) Trials Search Co-ordinator advised on and carried out the search. There were no restrictions on language or date of publication. We planned to assess articles published in languages other than those spoken by the review authors using the assistance of translators.

3.3 DATA COLLECTION AND ANALYSIS

4.1 DESCRIPTION OF STUDIES

4.1.1 Results of the search

The literature search identified 22,184 articles. Of these, 95 appeared to meet our inclusion criteria based on titles and abstracts. For these 95 articles, we sought to obtain the full text to establish eligibility for our review. Correspondence with authors yielded an additional three eligible studies that had not been identified by our search strategy. Of the 98 articles identified, full texts were unavailable for 17 articles despite the use of all the methods listed in the Selection of studies section. The Characteristics of studies awaiting classification table lists the missing studies. For nine of these (Bernardo 1994, Brown 2002, Chrappa 1991, Holcomb 1993, Lawson 2006, Murphy 1997, Northam 1997, Sanchez 2011, Walther 1986) no email or postal details for authors could be found. Of the remaining eight studies for whom authors' contact details were found, five responded but did not have a copy of their study available (Bond 1995, Deiter 1994, Heimerdinger 2006, Hill 1995, Layman-Guadalupe 1996) and three (Abrams 1992, Avina 2005, Halvorson 2007) did not reply. A more detailed evaluation of the full text of the remaining 80 articles revealed 41 articles (representing 38 studies, details listed in the Characteristics of included studies table) that were eligible for inclusion in our review and 40 that were excluded (with reasons provided in the Characteristics of excluded studies table). Of the 38 studies included in our review, 33 were included in the meta-analysis. Figure 1 shows a study flow diagram.

OPEN IN VIEWER

1

Study flow diagram

4.2 RISK OF BIAS IN INCLUDED STUDIES

We summarise the risk of bias for each study in the Characteristics of excluded studies table. We have presented authors' assessments of the six domains of bias as a percentage across all included studies in Figure 2 and by each individual study in Figure 3. These figures show that for the majority of studies, the risk of selection bias (due to inadequate random sequence generation or allocation concealment) is unclear, the risk of attrition and reporting bias is low, and the risk of performance bias (due to inadequate blinding of participants and personnel) is high.

OPEN IN VIEWER

2:

Risk of bias graph

OPEN IN VIEWER

3:

Risk of bias summary

4.2.5 Other potential sources of bias

We assessed the risk of publication bias by drawing funnel plots for the two meta-analyses involving 10 or more studies: attitudes towards relationship violence (Figure 4) and knowledge of relationship violence (Figure 5). We found no significant asymmetry and we therefore assume that there is no significant publication bias in the studies included in our meta-analysis. Based on the results of the funnel plots, we compared results of fixed-effect and random-effects models only for attitudes towards relationship violence, which displayed moderate heterogeneity.

OPEN IN VIEWER

4

(Analysis 1.3)

OPEN IN VIEWER

5

(Analysis 1.5)

We present results of both models in the Effects of interventions section (under ‘Attitudes towards relationship and dating violence‘).

4.3 EFFECTS OF INTERVENTIONS

The objective of this review was to assess the efficacy of educational and skills-based interventions designed to prevent relationship and dating violence in adolescents and young adults. We included 38 studies in the review, of which we included 33 in the meta-analyses. We excluded five studies from the meta-analyses: Andersen 1992 was excluded because results were analysed using non-parametric (Mann-Whitney U) analyses; Foshee 1998 did not report the number of participants in each arm or any tests of significance; Shultz 2000 did not report the number of participants in each group; and Holcomb 2002 and Woodin 2010 reported F-statistics, which could not be used to extract evidence of effect as the statistic was not associated to a direct comparison of the intervention under study (and therefore not equivalent to a t-test).

Where a study used more than one scale to measure a particular outcome, we included the scale that was most widely used across other studies and the most validated, such as the RMAS (Burt 1980) and the SES (Koss 1982). If a study measured more than one outcome (e.g. episodes of dating violence and change in attitudes towards dating violence), we included all outcomes but ensured each outcome was included in a separate meta-analysis. If studies had more than one follow-up period, we selected the longest period of follow-up (up to one year) for inclusion in our meta-analysis.

We applied an ICC value of 0.15 to results of all cluster-RCTs with the exception of Wolfe 2009, in which the authors had already accounted for the effect of clustering using two-level hierarchical models. In three studies, the number of participants in each study arm was not presented but we considered it was reasonable to make the following assumptions. •

Bradley 2009 described an even split of their 196 participants into two groups. We interpreted this as meaning there were 98 participants in each arm.

We describe the effects of the interventions on the outcomes reported below. Table 1 summarises the outcomes that we planned to assess, the outcomes reported in included studies and the outcomes used in this review.

5.1 SUMMARY OF MAIN RESULTS

The effectiveness of interventions to prevent relationship and dating violence can be quantified by a number of different outcomes. We assessed the effectiveness of these interventions as measured by changes in the number of episodes of relationship violence, changes in behaviours, attitudes and knowledge, and protective skills attained. For all outcomes apart from knowledge change, our meta-analyses show no evidence of a statistically significant effect. In the meta-analysis for knowledge of relationship violence, interventions appeared to have a beneficial effect. However, there was substantial heterogeneity (I² = 57%) between these studies, which is likely to be due to the variation between studies. Furthermore, when we excluded the studies at moderate or high risk of selection bias, only one study remained, which showed no evidence of an effect of the intervention on knowledge. Overall, therefore, this review has found no evidence of an effect of interventions on the outcomes reported.

The outcomes addressed by the studies included in our review can be categorised into two groups: direct measures, which we included as our primary outcomes, and proxy measures, which we included as our secondary outcomes. Examples of direct measures include the number of episodes of relationship violence and the physical and psychosocial health outcomes occurring as a direct consequence of violence. Importantly, health outcomes, which constituted two of our primary outcomes, were not assessed by any of the studies included in our review. The noticeable absence of health outcomes, especially when the associations between violence and health have been extensively documented, is an important finding in itself. We regard the lack of findings on these outcomes as a significant gap in the existing literature rather than a poor choice of indicators on our part. Assessing the frequency of dating and relationship violence is another direct measure. This outcome is challenging to measure for a number of reasons. The stigma associated with relationship violence in some settings may lead to fear of disclosing or reporting episodes of violence. In other situations, violent behaviour may be long-standing such that it has become the ‘norm‘, with victims blaming themselves for the violent behaviour of their partner and not seeking support. Peer pressure or the fear of retaliation may prevent victims from seeking help, and especially among adolescents, there may be poor knowledge about support services available. When cases are reported, information may not be available to researchers due to the non-disclosure of personal and sensitive information by officials. Finally, large studies with long follow-up periods are required in order to capture all events reliably.

Besides direct measures of effect, there are also proxy measures such as changes in attitudes, behaviour, knowledge and skills. These differ from direct measures in that they are important contributors to violence but require translation into action in order for a reduction in violence to occur. For example, even if attitudes towards relationship violence are improved and knowledge is increased, participants may not necessarily be able to apply this new information when faced with real-life incidents. Further research is required to assess whether (and how) changes in these proxy measures translate into reduced rates of relationship violence.

The aim of our review was to provide a comprehensive and unbiased summary of the existing evidence on interventions to prevent dating and relationship violence in adolescents and young adults. However, our results should be interpreted with caution for two main reasons. First, the studies included in our review varied greatly on several aspects. Participants ranged from general (low-risk) high school and university students to individuals with risk factors for experiencing or perpetrating relationship violence, such as a history of exposure to violence (Wolfe 2003) or adjudication (Salazar 2006). Our subgroup analyses comparing general to high-risk audiences revealed mixed results. Interventions appeared to be more effective at reducing the episodes of relationship violence experienced in high-risk audiences: the RR was not statistically significant in general audiences, but indicated a small but statistically significant reduction in risk in high-risk audiences. However, when assessing changes in attitudes, the SMD was not statistically significant for general audiences, and favoured the control group in high-risk audiences. Notably, for both outcomes, the high-risk ‘subgroup’ consisted of a single study. It is possible therefore that there were factors other than the type of participants that made these particular studies differ from others, and their results may not be generalisable to all studies with high-risk audiences.

Other studies targeted pregnant women, in whom exposure to relationship violence poses additional risks to their own health and that of the foetus. Furthermore, the focus was on couples rather than individuals in two studies (Woodin 2010; Florsheim 2011). The relationship violence experienced by these couples, particularly the pregnant adolescent women and their co-habiting parenting partners participating in the study by Florsheim 2011, may differ from the violence experienced by individuals who are not in an established relationship. For example, the latter group might experience violence in the context of dating and include cases of acquaintance rape. In more established relationships, partner violence may be of a more long-standing nature. The differences between ‘dating violence’ and ‘relationship violence’ require further research, especially as there may be important implications for prevention efforts. Finally, the adjudicated youth included in Salazar 2006 represent a group with complex social histories and backgrounds, which may affect their engagement with interventions. On a related note, there is an established link between alcohol and drug use and likelihood of dating or relationship violence. This was not explicitly controlled for within the studies and, while the focus of the review was on universal delivery of educational and skills-based interventions, drug and alcohol use is likely to have a negative impact on outcomes and this was not adequately described in the studies. While it can be argued that outcomes were still attainable even where drugs and alcohol are indicated, caution is still needed in interpreting the results.

Other variation between studies interventions arose in delivery settings, types of outcomes assessed, duration of the intervention, and duration of follow-up. The subgroup analysis by intervention setting was significant only for one outcome (knowledge of relationship violence) in one setting (university vs. school and community). Given that for the remaining 17 subgroup and outcome combinations the results were non-significant, we feel that the single significant outcome has arisen by chance. Intervention duration ranged between single 50-minute sessions (Bradley 2009; Holcomb 2002; Schewe 1996; Stephens 2009) to 21 75-minute sessions (Wolfe 2009). Subgroup analyses showed no significant differences in effect by total contact hours or total number of sessions. The duration of follow-up varied greatly. The majority of studies assessed outcomes immediately following the intervention, with a further assessment between 1 and 12 months later. In order to determine more reliably whether interventions are effective and if so, whether their effects are long lasting and reduce the incidence of relationship violence in later adulthood, longer follow-up is required.

The second reason for interpreting our results with caution relates to methodological aspects of our review. Individual studies used a wide range of measurement scales to assess outcomes. We presented our results as SMDs, which assumes that each of the assessment measures can be standardised and has comparable SDs. This method also renders the translation of results into a quantifiable improvement or deterioration in outcomes challenging. For example, it is difficult to define what a 0.03-point improvement in the combined skills score means in practice and what level of statistically significant change is clinically significant. Combining results in a meta-analysis requires a number of methodological decisions to be made. We have tried to be as transparent as possible by detailing our methods and providing rationales for excluding any studies from meta-analyses.

5.2 OVERALL COMPLETENESS AND APPLICABILITY OF EVIDENCE

All of the studies included in this review came from high-income countries. With the exception of Yom 2005, all studies were conducted in North America. There are a number of possible reasons for this. First, the term ‘dating’ may be less frequently used outside of North America. However, our search strategy included the term ‘relationship’ and other synonyms, which should have identified a broader range of titles. Second, the concept of dating itself may be less common outside of North America. In some lower-income countries marriage occurs at a younger age, so that if violence in a relationship occurs it is more likely to be classed as intimate partner violence rather than dating or relationship violence. Finally, it is possible that relationship violence is less commonly reported and less researched in other cultures.

With the exception of the Safe Dates study (Foshee 1998), which followed participants up to four years post-intervention, Wolfe 2009, which followed participants for 2.5 years post-intervention, and Florsheim 2011, which followed participants up to 18 months post-intervention, most studies followed participants up to a maximum of 12 months. A few studies only conducted immediate post-test assessment. There is, therefore, little evidence on the long-term effectiveness of these interventions. As described above, the majority of studies assessed changes in proxy measures such as attitudes and knowledge, rather than episodes of violence and behavioural change. Further studies are required to explore the interaction between knowledge and attitudes on the one hand, and behaviour, skills and episodes of relationship violence on the other hand.

5.3 QUALITY OF THE EVIDENCE

The quality of most studies in this review was limited by unclear methods of random sequence generation, allocation concealment and assessor blinding, as illustrated in the ‘Risk of bias’ graph (Figure 2) and ‘Risk of bias’ summary (Figure 3). While questionnaire scores (as used by most studies) may be less affected by lack of assessor blinding than other outcomes, we felt that the fact that none of the studies raised this point as a consideration was a significant shortcoming. Blinding of personnel delivering the interventions was not possible. However, some studies implemented ways of minimising the effects of this by providing training sessions for staff, scripts and monitoring a selection of sessions to ensure adherence to the study protocol. Studies not implementing these activities were considered to be at high risk of performance bias. We downgraded the evidence from ‘high’ to ‘moderate’ for all outcomes in our Summary of findings Table 1.

Another significant concern was the cluster or quasi-cluster randomisation methods used in 19 of the included studies. Participants were randomised by schools or classes to decrease the risk of cross-contamination and for practicality. Only one study used an ICC to account for the effect of clustering, though the ICC used was low compared to ICCs of other similar studies. The results of the cluster-RCTs included in this reviews are, therefore, likely to overestimate the effects of the interventions. Finally, many included studies had small numbers of participants (Andersen 1992; Forst 1993; Schewe 1996; Boulter 1997; Davis 1997; Miller 1999; Shultz 2000; Yom 2005; Salazar 2006; and Woodin 2010 had fewer than 100 participants each) and short periods of follow-up (maximum follow-up periods were two weeks or less in Andersen 1992; Forst 1993; Schewe 1996; Avery-Leaf 1997; Macgowan 1997; Foubert 1998; Lanier 1998; Pinzone 1998; Miller 1999; Saberi 1999; Pacifici 2001; Holcomb 2002; Yom 2005; Bradley 2009), which is likely to affect their quality adversely by decreasing their statistical power and increasing the risk of type I and type II errors.

5.4 POTENTIAL BIASES IN THE REVIEW PROCESS

Although a large number of articles were initially identified by our search strategy, it is still possible that some relevant studies were missed. Correspondence with authors yielded three further eligible studies that had not been identified by our search strategy. Many authors that we contacted did not reply and it is possible that further relevant studies could have been missed. Two review authors (GF and CH) independently screened titles and abstracts, which minimised the potential for bias in selecting studies for inclusion.

Eighteen studies that appeared by title or abstract to meet our eligibility criteria could not be assessed because full texts were not available. These studies have therefore been listed as ‘awaiting classification‘. Some of the missing studies are doctoral theses, which we were unable to retrieve despite requesting interlibrary loans, contacting authors and corresponding with universities where the doctoral title was awarded. Other studies were published in journals that we were unable to access. We do not know whether these studies are eligible for inclusion in our review and, if so, how they would have affected our results. Funnel plots for the outcomes that included 10 or more studies showed no evidence of publication bias (Figure 4 and Figure 5). However, it is possible that the studies that we were unable to retrieve are in themselves the result of a publication bias that exists in the general literature; in other words, it may be that the very reason we found these studies difficult to access is that they are studies with inconclusive or negative findings that are less likely to be published or more difficult to retrieve (Scargle 2000). It is possible that the failure to include these studies overestimated the effect sizes of interventions and introduced bias into this review (Higgins 2011).

As outlined in the Methods section, we conducted single pair-wise comparisons to avoid double-counting and unit of analysis issues. In some cases, this may have diluted the effect of interventions, for example in studies where high-risk and low-risk participants, or those with and those without a history of sexual victimisation were combined into single groups. Combining all outcome assessments up to 12 months into one outcome in our analysis may also have had an effect on our results. For example, it is possible that the effectiveness of interventions was greatest immediately following the intervention, with the effects gradually wearing off with time. However, the substantial differences in outcome assessments made it difficult to assess the effects separately. There was an insufficient number of studies to perform a separate analysis of outcomes assessed between 6 and 12 months as we had planned.

The ICC of 0.15 that we used to adjust for the effect of clustering is likely to lead to conservative estimates of effect. However, we felt this to be the most appropriate ICC based on a review of ICCs used in meta-analyses of similar topics. It would have been preferable if individual studies had ascertained their own ICCs, or if different ICCs were available depending on whether randomisation occurred at class or school level. An insufficient quantity of examples of ICCs was found to enable us to conduct ICC sensitivity analyses.

The decision to include quasi-RCTs in our review was based on an assumption that this could be a common method of allocating participants to groups for logistic reasons, particularly in the case of cluster-RCTs. The inclusion of quasi-RCTs may have introduced bias into our results as they are at high risk of allocation bias. However, sensitivity analyses showed that including these studies made no significant difference to overall pooled results.

We included studies where more than 80% of the sample was within the age range of 12 to 25 years in anticipation that some community-based studies may target a wider range of age groups. The only studies that included participants higher than our stated age range were those set in universities, which captured a number of mature students (Forst 1993; Anderson 1998; Stephens 2009). The percentage of students outside of the age range in these studies was very low. Only one included study had participants younger than our age range (Miller 1999): students were aged 10 to 14 years, with a mean of 12.8 years and the majority were aged 13 years. It is unlikely that these small numbers would have influenced our results. Furthermore, subgroup analyses showed no significant differences in effect by intervention setting. As participant age is directly correlated with setting, we feel that there is little risk of the inclusion of these studies impacting our results.

Non-RCT studies on the topic of relationship and dating violence and those that were excluded from our review were not systematically reviewed.

5.5 AGREEMENTS AND DISAGREEMENTS WITH OTHER STUDIES OR REVIEWS

Whitaker et al conducted a systematic review of primary prevention programmes for the perpetration of partner violence (Whitaker 2006). The authors found that nine of the 11 included studies reported at least one positive intervention effect for knowledge, attitude or behaviour and concluded that such prevention programmes are promising. This finding differs from the lack of evidence of an effect found in our review. However, the review by Whitaker et al included observational (non-randomised) trials, which may have influenced results. Results were summarised narratively, rather than by means of a meta-analysis. To our knowledge, no other systematic reviews on this topic have been conducted.

6.1 IMPLICATIONS FOR PRACTICE

There is currently a wide range of interventions based in educational and community settings aimed at the prevention of dating and relationship violence in adolescents and young adults. The studies included in this review showed no evidence that these interventions reduce episodes of violence or improve attitudes, behaviours and skills related to relationship violence. There was evidence of a small increase in knowledge but this result must be interpreted with caution due to high heterogeneity among studies. Most studies had methodological shortcomings, which may have led to overestimation of their effects, especially when cluster randomisation was used. Importantly, our results show no evidence of effect, rather than evidence of no effect. Therefore, current interventions should not necessarily be stopped, but rather further research and more methodologically sound studies should be conducted.

6.2 IMPLICATIONS FOR RESEARCH

Further evidence is required to assess the effectiveness of interventions to prevent dating and relationship violence. The current evidence is predominantly focused on assessing changes in attitudes and knowledge. Research into the effects of interventions on incidence of relationship or dating violence, and exploration of the relationship between attitudes and knowledge and skills, behaviour and episodes of violence are needed. It is possible that in order to reduce the occurrence of relationship violence effectively, a number of interventions across both educational and community settings as well as within homes and families is required. Exploring these themes will require larger RCTs with longer follow-up periods. Researchers should consider using existing reliable and validated scales such as the RMAS (Burt 1980) or the Revised CTS (Straus 1996). New measurement scales developed by authors have often not been adequately validated, and the use of multiple different scales renders the comparison of results from different studies difficult.

Although RCTs are preferable to cluster-RCTs, in practice it is often more feasible to conduct the latter. In this case, authors should adjust results using an ICC to avoid overestimating the effect of the interventions. We identified ICCs from a number of meta-analyses on similar topics by searching CENTRAL. However, a wider and more systematic search would be helpful in ascertaining the range of ICCs used across cluster-RCTs and establishing the most appropriate ICC figure to use.

With one exception, all studies were conducted in North America. Interventions addressing relationship violence are likely to be highly culturally sensitive and it is important to understand what types of interventions are effective in different settings. Further studies are, therefore, required from high-, middle- and low- income countries in Europe, Asia, Africa and Australasia.

11.1 CHARACTERISTICS OF INCLUDED STUDIES