A systematic review of prevention and intervention strategies for populations at high risk of engaging in violent behaviour update 2002-8

Other reviews

This review was conducted in the context of a number of other reviews of research on evaluations of violence reduction interventions by various teams around the world (for a survey of those pertaining to the field of criminal justice, see McGuire35,36). Given the breadth of the inclusion criteria adopted for the review here, 17 of these reviews are most relevant to the focus adopted here. However, these previous reviews cover specific populations and treat them distinctly, whereas the review reported below attempts to integrate the literature across these distinct groups. Four of the reviews focus exclusively on younger offenders in the age range of 12–21 years;37–40 seven address the problem of sexual offending;41–47 one deals with DV;48 one includes both adolescent and adult samples;49 and four are focused on studies of individuals diagnosed with personality disorders. 50–53

The broadest review was a meta-analysis (MA) by Dowden and Andrews. 49 This subsumed a range of studies that included mixed samples of adult and adolescent offenders and target offence behaviours, including general violence, sexual and domestic assaults. These authors found 34 evaluations of interventions to reduce violence, yielding 52 effect-size tests. Approximately 70% of the included studies focused primarily on work with adults. Unfortunately, they do not report separate outcome data for the two age groups. The overall mean effect size (r) was relatively low at +0.07, although there was enormous heterogeneity in the findings: effect sizes ranged from a low of –0.22 to a high of +0.63. The effect size for interventions based on the risk–need–responsivity model54 was better than the overall mean, at +0.12. This corresponds to recidivism rates of 44% for experimental and 56% for control groups. Possibly the most notable finding to emerge from this review was evidence of a close correspondence between the number of criminogenic needs targeted in interventions and the associated effect size: a correlation coefficient of 0.69 (p < 0.001).

Previous review

The review reported here is an update of one part of an existing review. In 2002 the Department of Health commissioned the Liverpool Violence Research Group (LiVio) to complete a broad-ranging systematic review of interventions and risk assessment strategies for the management of violence in a widely defined population (offenders, people with mental health problems and offenders with mental health problems). The aim was to provide a picture of the literature that was broad enough to inform future improvements in research, policy and clinical practice, while strictly adhering to the main criteria for high-quality reviews. The original review55 covered publications released between 1955 and 2001 (with partial update to 2004) on a population that consisted of (1) adult offenders (> 17 years) with or without a mental disorder; (2) adults with a diagnosable mental disorder but no offences; and (3) adults in the general population exhibiting indictable acts of aggression without actual indictment (e.g. DV). Substance abuse alone was not deemed sufficient to constitute a diagnosis of mental disorder. Any pharmacological, psychological or other intervention targeted at the individual patient/offender and delivered individually or in small groups was included. Organisational interventions (e.g. ward-level changes) that did not report individualised outcomes were excluded. Changes on any outcome measure that was an actual or proxy assessment of aggression (e.g. observed aggression, self-reported hostility) were included. There were no exclusions based on design, language or publication format.

Update of the review

This update uses the same search strategy and the same databases were searched where possible. The four senior reviewers involved in the original review were also involved in this update and are referred to in this document as ‘expert reviewers’.

Owing to the size of the original review it was decided that the update would be split into two distinct elements: this intervention review and, secondly, a review of risk assessment approaches with the same population (to be published at a later date). It is important to emphasise that the two processes are closely linked. Estimates of predictive validity from a risk assessment tool are of little use on their own if they are not used to design and target effective interventions. The structured clinical (or professional) judgement approach59 is important in this context, as this approach is recognised as encouraging practitioners to focus on risk management and flexibility in choosing appropriate interventions.

Research question

Which interventions are the most effective in reducing violence and which key variables are associated with a significant outcome?

Inclusion stage one

At stage one inclusion, six reviewers independently applied the inclusion criteria to 200 citations and a Cohen’s kappa (Fleiss–Cuzick extension) was calculated [κ = 0.63 (SE = 0.019), z (for k = 0) = 34.24, p < 0.0001]. Each new reviewer who joined the team was required to look at 100 citations that had previously been looked at by a reviewer and a > 80% agreement had to be reached. At this stage an ‘inclusive’ attitude was taken, i.e. where there was doubt a citation was included. Given the high level of agreement and the inclusive approach, further citations were screened by only 1 of the 11 reviewers.

If a citation was excluded then it was possible to mark the citation as either a review that needed the reference list checked (‘check’) or a paper of particular interest that should be obtained (‘obtain’).

Acquiring papers

Electronic copies of papers were then downloaded where possible by the University of Liverpool’s interlibrary loans team. Where electronic copies were not available, paper copies were either obtained from the University of Liverpool’s library or through interlibrary loans at the British Library.

Inclusion stage two

At stage two, the inclusion/exclusion criteria were applied to the full papers identified from stage one. To aid with this a Microsoft Access database (Microsoft Corporation, Redmond, WA, USA) was developed using a front page form with drop-down menus and tick boxes. It was at this stage that included papers were categorised into each of the two reviews: intervention or both risk and intervention. Furthermore, studies not reporting any statistical analysis, mainly due to qualitative designs, were excluded from the review though retained for future analysis.

As a quality control measure, all seven of the reviewers applied the inclusion criteria to 50 papers and a Kappa score was calculated [Cohen’s kappa (Fleiss–Cuzick extension): κ = 0.62 (standard error, SE = 0.032) z (for k = 0) = 19.46, p < 0.0001]. Investigation of individual pairs of inter-rater agreement (Table 3) revealed that one reviewer (G) had poorer reliability scores owing to being more inclusive than other reviewers. Therefore, it was decided that there was high enough agreement to continue with single reviewer application of inclusion criteria.

Bivariate and multivariate analyses

Meta-analysis

Design of studies

Of 198 studies, 51 (25.8%) were RCTs, one-third (33.3%) were concurrent/cross-sectional group comparisons and 68 (34.3%) were before/after study comparisons. The remaining 13 studies were crossover comparisons, correlational studies and experimental case studies (Table 5).

Length of follow-up

The maximum length of follow-up was reported by 179 studies and ranged from half an hour to 14 years, with the average length of follow-up being: mean = 524.26 days, median = 183.40 days and mode = 365 days (Figure 2).

FIGURE 2.

Total length of follow-up in days.

Attrition

Attrition was calculable for 189 of the studies: 67 (35.4%) reported no attrition and four (2.1%) more than 80% attrition (Figure 3). The mean attrition was 20.0% and the median was 9.9%.

FIGURE 3.

Attrition rates.

Intention to treat

The 198 studies included in the review reported on 728 comparisons. Of these, 31.9% were analysed on an ITT basis, 59.1% were not analysed on an ITT basis and 9.1% did not state whether they were ITT analysis or not (Table 6).

Baseline equivalence

Of the 120 studies comparing different study groups, equivalent baseline measures of aggression were reported for 51 (42.5%) studies. A further 11 reported equivalence on some measures of aggression and 16 (13.3%) reported non-equivalence. Twenty studies reported the baseline levels of aggression for each group but did not compare them statistically and 22 (18.3%) did not report any baseline measure of aggression (Table 7).

Blinding

Given the nature of many of these studies it is not surprising that blinding was not stated in the majority of papers, as for practical reasons this is impossible to achieve when evaluating psychosocial interventions. Where it was stated, it was most frequently reported for patients and the interventionist, with 10.1% of patients not being blinded and 14.6% being blinded, and interventionists not being blinded in 12.1% of studies and blinded in 12.6% of studies (Table 8).

Number of studies

The number of studies published was relatively steady across the years, with an average of 32 papers being published each full year (Figure 4).

FIGURE 4.

Number of studies by year of publication. a, The year 2008 was a partial year.

Country in which studies were conducted

Studies were conducted in 21 different countries, with only three studies being multinational, i.e. participants from more than one country. The majority of studies were conducted in the USA (55.1%), with the UK being the second most common location (10.6%), followed by Canada (6.6%) (Table 9).

Number of participants

The number of people approached to take part in the studies was reported in 94 (47.4%) of studies and ranged from 1 to 8325. The number of participants enrolled was reported in 191 (96.5%) of studies and ranged from 1 to 10,753. The number of participants at the end point of the study was reported in 196 (99.0%) of studies meaning that two studies failed to report the final number of participants in their study.

The studies reporting the final number of participants described a total of 51,258 individuals, the smallest study having one participant and the largest 10,753 participants (Figure 5). The majority (60%) of studies included ≤ 100 people.

Demographics of participants

The sex of participants was reported in 183 (92.4%) studies, with 95 (52%) studies including only males and 15 (8%) including only females. The percentage of males in the remaining studies ranged from 8% to 95% (Figure 6).

The average age of participants was reported in 166 studies (158 reported the mean age, four reported the median age and four reported both the mean and median age). The mean age ranged from 19 to 80.9 years, with SDs (reported by 118 studies) ranging from 1 to 15.9 years. The range of ages was reported by 70 studies (an additional study reported minimum age only). The minimum age of participants ranged from 13 and 65 years and the maximum age ranged from 32 to 97 years. Therefore, the youngest participant was 13 years and the oldest was 97 years (Table 10).

The percentage of participants who were described as Caucasian was reported in 98 studies, with six (6%) studies not including any Caucasian participants, and one study (1%) including only Caucasian participants. The percentage of Caucasian participants in the remaining studies ranged from 6% to 99% (Figure 7).

Population

Populations included in the review were either participants with a diagnosis of mental disorder, offenders, indictable offenders (i.e. those having committed indictable offences but not having been charged) or forensic participants (i.e. those with a diagnosis of mental disorder and offender/indictable offender status). The numbers of studies looking at each of these population types are shown in Table 11. Participants were mainly people with a mental disorder (38%) or offenders (35%), with those reported to have committed an indictable offence being studied in 15% of studies and offenders with a mental disorder (forensic) being included in 12% of cases.

Studies reporting on individuals with a diagnosis of a mental disorder (including forensic groups) reported a range of diagnostic groups, with patients defined as having an ‘other’ single mental health grouping being the most frequently reported (34%), followed by participants with a ‘mixed diagnosis’ (28%). Participants with personality disorders only were studied in 20% of the studies and participants with a diagnosis of schizophrenia or schizoaffective disorder only were studied in 11% of the studies (Table 12).

There were differences between the diagnosis of participants in the mental disorder group and the forensic group. Almost half of the studies investigating forensic participants reported mixed diagnoses, and a further 37.5% reported participants with an ‘other single mental health grouping’. Participants with a specific mental health diagnosis were reported in only 3 out of the 24 forensic studies (12.5%), whereas 32 out of the 75 studies (42.6%) examining participants with just a mental disorder reported investigating participants with specific mental disorder diagnoses.

The index offences that participants had committed differed greatly between the three groups. Offender participants had been charged with predominantly DV (44.3%), followed by mixed group of offences (28.6%) and sex offending (22.9%). For studies including forensic participants, mixed groups of offences were more frequently reported (41.7%), followed by sex offending (29.2%). A further 20.8% of studies did not report what offences participants had committed.

As expected, in the indictable group, DV was the most reported offence type (65.5%), with other indictable offences being reported in 24.1% of studies (Table 13).

Substance abuse

Substance abuse by participants was poorly reported in most studies, with only 43.4% (86) of papers reporting whether current substance abuse was or was not identified. Of the 86 studies reporting on substance abuse, 21 (24.4%) reported no substance abuse, five (5.8%) identified drug abuse, three identified (3.5%) alcohol abuse and 33 (38.4%) both alcohol and drug abuse. A further 24 (27.9%) studies identified some form of substance abuse, but did not report on the nature, i.e. whether it was drugs or alcohol (Table 14).

Types of interventions

Of the 198 studies, 74 (37.37%) were single-group designs and 124 (62.6%) compared two or more groups. Of the 124 using a comparator group, 29.8% compared two different types of treatment (head-to-head comparisons), 24.2% TAU, 14.5% a placebo, 12.9% compared subgroups of one treatment (e.g. completers vs non-completers) and 8.9% no treatment. The remaining seven studies used a historical control (3.2%) or self as a control (2.4%) (Table 15).

The types of intervention studied are shown in Tables 16 and 17. Half of included studies used a psychological intervention (50.5%) as the primary intervention, one-quarter used a pharmacological intervention (23.7%) and one-quarter another form of intervention (25.8%). The specific categories of intervention by comparison type are shown in Table 16 and the comparators used in head-to-head studies in Table 17.

Psychological studies were more likely to use single-group comparisons and pharmacological studies head-to-head or placebo comparators. Where head-to-head studies were used, the same categories of intervention were compared, i.e. psychological interventions compared with another psychological intervention, and pharmacological interventions compared with another pharmacological intervention.

Setting

The start and end settings are shown in Table 18. The term ‘setting’ here refers to the location where the intervention is conducted and in the case of ‘community’ under what conditions, i.e. a probation order, or under the supervision of a mental health practitioner or neither (e.g. a person concerned about their propensity for violence who is offered a self-help intervention). The most frequently reported setting was community with people on probation (18.7%), followed by penal institutions (16.2%), community (14.6%) and community mental health (12.1%). The majority of studies (87.9%) had the same start and end setting. Of the 24 studies reporting different start and end settings, 12 were studies that started in penal institutions but ended in either the community or in mixed settings.

When start settings for the interventions are examined by intervention type (Table 19), it can be seen that studies in a forensic mental health setting mainly studied behavioural and cognitive therapies (75.0%), as did penal institutions (56.3%), community (44.8%), mixed settings (33.3%) and other settings (40.0%), whereas community probation settings used DV programmes. Pharmacological interventions were the focus of the majority of studies in community mental health settings (58.0%), accident and emergency (A&E) settings (100%), mixed settings (33.3%) and studies where the setting was unclear or not stated (40.0%).

Level of intervention

The levels of interventions for each of the types of intervention are shown in Table 20. Pharmacological interventions were by design at an individual level, whereas the psychological interventions were generally at the small group level.

Quality

On variables used to assess methodological quality of studies, the RCTs reflected what was found in the whole data set (see Quality assessment) for baseline equivalence [χ² = 2.347, degrees of freedom (df) = 3, p = 0.504] (Table 21), and sample attrition (Mann–Whitney U-test, p = 0.568) (Table 22). However, the total length of follow-up reported in studies was significantly longer in the non-RCTs (mean 629 days, SD 932, median 364 days) than in the RCTs (mean 253 days, SD 731, median 84 days) (Mann–Whitney U-test, p < 0.0001) (see Table 22) and the number of studies using some form of blinding to the intervention was far higher in the RCT data set (58.8%) than in the non-RCTs (6.1%) (χ² = 66.486, df = 1, p < 0.0001) (see Table 21). Of the 197 comparisons conducted in the RCT studies, 53.3% used an ITT analysis, whereas only 27.2% of the 464 comparisons reporting whether an ITT analysis was used in the non-RCT data set used an ITT analysis (χ² = 41.578, df = 1, p < 0.0001) (see Table 21).

Trial and participant characteristics

The distribution of the number of papers being published in each year was the same for RCTs and non-RCTs (χ² = 3.629, df = 6, p = 0.774) (Table 23), as was the distribution of the number of participants in the studies (Mann–Whitney U-test, p = 0.422), mean age (Mann–Whitney U-test, p = 0.084), proportion of sample who were Caucasian (Mann–Whitney U-test, p = 0.436) (Table 24), the reporting of substance abuse (χ² = 2.347, df = 3, p = 0.05) and the types of offences offenders had committed (statistical analyses not appropriate) (see Table 23) (see Study characteristics and Participants characteristics for description of whole sample).

The country in which studies were conducted appeared to differ between RCTs and non-RCTs (see Table 23), most notably the proportion of RCTs that were conducted in the UK (3.9%) was lower than non-RCTs (13.0%) and RCTs were more likely to fail to report where the study was conducted than non-RCTs (13.7% vs 3.4%, respectively). RCTs also reported a lower percentage of males (mean 55.02%, SD 38.76) than non-RCTs (mean 83.0%, SD 26.06) (Mann–Whitney U-test, p < 0.0001) (see Table 24).

The populations in the studies also appeared different in the RCTS and non-RCTS (see Table 23). RCTs focused primarily on participants with mental disorder (66.7%) compared with non-RCTs (27.9%), whereas non-RCTs included offenders in 40.8% of studies compared with 17.6% of RCTs. Only one (2.0%) RCT included forensic patients, whereas 16.3% of non-RCTs included forensic patients.

The diagnoses of participants in studies investigating a population of people with a mental disorder seemed to differ between RCTs and non-RCTs (see Table 23), with RCTs focusing on participants with a personality disorder (38.2%) and non-RCT studies participants with other single mental health grouping (38.5%).

Intervention characteristics

As expected, the types of comparisons differed between RCTs and non-RCTs, with nearly half of the non-RCTs reporting single group comparisons (49.3%) (Table 25).

As shown in Table 25, the type of primary intervention being tested in the studies also differed between RCTs and non-RCTs (χ² = 43.611, df = 2, p < 0.0001), with 56.9% (n = 29) of RCTs evaluating pharmacological intervention alone, and only 21.6% of studies (n = 11) evaluating a psychological intervention. This compares with 12.2% of non-RCTs testing a pharmacological intervention and 60.5% a psychological intervention. Within each of the four broad categories of primary interventions, the specific groupings of primary interventions had too few studies for statistical analyses.

The setting that interventions were started in appeared to differ between RCTs and non-RCTs. RCTs were more likely to fail to report the type of setting (11.8%) compared with non-RCTs (2.7%), and were more likely to be conducted in community mental health settings (RCTs 19.6%, non-RCTS 9.5%). In contrast, non-RCTs were more likely than RCTs to be conducted in a penal institution (non-RCTs 19.7%, RCT 5.9%) or on community probation (non-RCTs 22.4%, RCTs 7.8%) (see Table 25).

The level the intervention was conducted on also appeared to differ between RCTs and non-RCTs. RCTs were primarily conducted at the individual level (67%), probably reflecting the focus on pharmacological interventions, whereas non-RCTs were more frequently conducted in small groups (46%) (see Table 25).

Analyses by broad intervention groupings

Significant bivariate associations between broad category of intervention and the statistical significance of an analysis were not observed for either pharmacological or psychological interventions (60.8% of pharmacological and 57.3% of psychological intervention analyses found in favour of the primary intervention arm vs its comparator). Where an impact on outcomes was shown, was in the case of analyses focused on ‘other’ forms of intervention. Analyses focused on such interventions were less likely than analyses focused either on pharmacological or on psychological interventions to result in a statistically significant outcome in favour of the primary intervention arm of the analysis (n = 195, χ² = 6.006, p = 0.011) (Table 26).

Population

Outcome measure

As the majority of studies in the literature use a scale-based outcome measure the impact of this choice on outcomes for the primary intervention was investigated. Although a trend was observed for scale-based outcomes to favour the primary intervention arm (58.1% of scale-based outcomes showed statistical significance in favour of the primary arm in contrast with 48.5% of non-scale-based outcomes), this trend did not reach statistical significance (n = 195, χ² = 1.644, p = 0.129) (Table 30).

Study quality indicators

Perhaps unexpectedly, none of the ‘design-quality’ variables, including whether or not a RCT design was used in a study, showed a statistically significant association with whether or not the primary intervention arm was favoured in terms of outcome (Table 31).

Although direct measures of study quality did not show any association with outcomes, measures that can be seen as indicating the ‘strength’ of the evaluation being made did show such an association. Specifically, both the nature of the comparator (if any) against which the primary intervention was tested and whether the analysis focused on a within- (e.g. single group pre/post) or between-group evaluation showed a statistically significant association with outcomes (see Table 31).

Comparisons involving head-to-head contrasts between two active interventions were significantly less likely to favour the ‘primary’ intervention being evaluated (n = 195, χ² = 6.06, p = 0.01). Similarly, comparisons against TAU were less likely to favour the intervention being evaluated (n = 195, χ² = 5.59, p = 0.01). No significant difference was observed between analyses comparing the primary intervention with ‘no treatment’ [placebo, waiting list control (WL) or no intervention at all] and analyses drawing a comparison with an active comparator or TAU.

Analyses comparing two groups (one allocated to the primary intervention and one to a placebo, TAU or an active comparator) were substantially less likely to show favourable outcomes for the primary intervention than analyses relying on a single within-group comparison (e.g. pre/post evaluations) (n = 195, χ² = 12.84, p = 0.000) (see Table 31).

Summary of bivariate associations

Only six variables showed a significant bivariate association with outcomes based on the key dichotomised variable coding whether or not a study (based on its strongest measure) recorded a significant finding in favour of the primary intervention arm. These were:

• whether the primary intervention was something other than a pharmacological or psychological intervention (outcome is less likely to be positive)

• whether or not the study was carried out in a general mental health population (outcome more likely to be positive)

• whether or not the intervention was carried out in an offenders’ institution (outcome is less likely to be positive)

• whether or not the study design involved a head-to-head comparison between the primary intervention and another active intervention (outcome less likely to be positive if comparator is active intervention)

• whether or not the study design involved comparison against TAU (outcome less likely to be positive if comparator is TAU)

• whether the analysis was within or between groups (outcome less likely to be positive for between-groups analyses).

In addition to the above, one potential confounder in respect of analysis outcomes was identified in respect of associations between the demographic and study design variables. This was:

• proportion of participants who were Caucasian (increasing proportions associated with increasing age and decreasing likelihood of ethnic minority participants).

The very limited bivariate associations between key variables noted above fails to justify substantive multivariate analysis. To explore the extent of variation explained in respect of the central dependent variable (whether or not a significant outcome in favour of the primary intervention was established) we carried out a single regression analysis based around the above modifiers. The likely impact of the one potential confounding variable was not considered sufficiently strong to justify inclusion in the model.

Further exploration of variance

In an attempt to explore further potential modifying variables accounting for variation in outcome in the data set, we divided the broad intervention (‘pharmacological/psychological/other’) and outcome measure (‘scale/not scale’) categories into smaller component groups. In line with the ‘scattergun’ approach for this literature, the range of both interventions and outcome measures was entirely disproportionate to the number of analyses included. Within the 195 analyses included above, there were a total of 94 distinct types of intervention and 55 different types of outcome measure.

The only individual categories with numbers approaching sufficient for any further consideration were as follows.

Meta-analysis of all randomised controlled trials

Figure 8 summarises the outcomes for this initial exploratory analysis of the RCT data. Both on the assumption of a fixed- and random-effects model, the overall mean effect tends to slightly favour the intervention of interest rather than the placebo or active comparator. A diversity of interventions contribute to this profile, with studies evaluating psychological (e.g. Cavanaugh81), pharmacological (e.g. Krakowski et al. 82) and ‘other’ (e.g. Lasley83) interventions all showing positive outcomes for the primary intervention of interest. For the majority of studies favouring the primary intervention, the CIs are also reasonably small, indicating that the sample outcomes are likely to be representative of the actual population profile. Pooled effect size estimates based on the standardised mean difference (see Table 35) suggest that, overall, interventions for violent behaviour can be expected to show a small to moderate impact.

FIGURE 8.

Meta-analysis of all RCTs. For a full explanation of the comparisons included see Appendix 3, Table 56. BPD, borderline personality disorder; CB, cognitive behavioural; CBT, dialectical behaviour therapy; CMAI, Cohen-Mansfield Agitation Inventory; F, female; M, male; PANSS, Positive and Negative Syndrome Scale; PSE, present state examination; SCL-90, Symptom-Checklist-90.

Heterogeneity

Table 35 summarises the outcome of heterogeneity estimates and the effect sizes for this overall combination of RCTs. Although the likely general profile of outcomes for interventions evaluated in the literature can legitimately be drawn from an overview of this exploratory MA, it is important to recognise that we are combining across a very diverse range of studies. It is readily apparent from the estimates in Table 35 that the RCT analyses in this literature are hugely variable. An I² of 86% is extremely large, indicating that from a statistical viewpoint the ‘averaged’ estimates contained in the MA are not robust. They should therefore be taken, at best, as an indicator of potential outcomes. Subsequent MAs, outlined below, demonstrate that this variance is not a simple function of either intervention or outcome type. Before moving on to explore this variability further, it is worth noting that three studies contributing to the MA have particularly large CIs,84–86 suggesting that these studies’ estimates of likely effect size are unlikely to be wholly representative of the potential outcomes in the population of interest. Studies reporting analyses with larger CIs are also not restricted to the smallest sample sizes, implying that levels of uncertainty are not purely the consequence of sampling error. The impact of study focus and study quality is explored further later on in the report. Removing the three studies reporting notably wide CIs from the analysis did not, however, serve to substantially reduce the observed heterogeneity (Q = 260.88, I² = 89.6%) relative to that reported in Table 35. Neither did the removal of all analyses with CIs wider than the average for this group of analyses (the above three analyses plus Volavka et al. ,87 Nickel et al. 88 and Brown University89) (Q = 232.19, I² = 89.66%).

Taking note of the likely unreliability of this exploratory analysis in respect of results, we calculated effect sizes based on standardised mean differences across all included analyses (see Table 35). The aim of this very broad estimate was to provide an initial indication of the potential efficacy of violence reduction interventions as such. Under a fixed-effects model, the mean effect size was –0.29 and under a random-effects model the effect size increased to –0.57. Although these figures are subject to substantial caveats, as outlined above, they indicate the potential for small to moderate decreases in violent behaviour as a result of some intervention, with the potential for greater gain where modifying factors, such as population group, are taken into account (as the random-effects model identified a larger effect size than the fixed-effects model).

Head-to-head comparisons

Although in the broader range of studies included in the bivariate analyses head-to-head comparisons were significantly less likely to favour the primary intervention, this is not the case for the RCT studies included in the MA (Figure 9). Here, again, both fixed-effects models and, to a lesser extent, random-effects models showed outcomes slightly in favour of the primary intervention. As for the general evaluation outlined above, favourable outcomes ranged across diverse intervention types. For all but one small study, CIs were acceptably small. The one exception to this was a small study86 that also presented as something of an outlier in respect of the general trend, in favouring the comparative intervention (‘psychodynamic psychotherapy’) over the primary intervention being evaluated (CBT). As with the general model above, effect size estimates (Table 36) suggested a likely small-to-moderate impact on violence of the interventions taken as a whole, with some potential for further gains if population groups particularly responsive to the intervention could be identified (fixed-effects model effect size –0.26, random-effects model effect size –0.44).

Restricting the MA to ‘head-to-head’ comparisons had no impact on estimated heterogeneity between studies. With I² remaining at 86%, the outcomes for the analysis must be regarded with caution, as, despite their similar focus on drawing a direct comparison between interventions, the studies included clearly remain very diverse in other respects.

Active treatment versus treatment as usual

A smaller range of studies compared the primary intervention with ‘treatment as usual’ (Figure 10). Predictably, observed heterogeneity for this smaller group of studies, all but one of which (Willner et al. ,84 focusing on CBT anger management) focused on ‘other’ interventions, was less extreme than for the ‘head-to-head’ analyses. Nevertheless, with an I² of 68% (Table 37) overall outcomes for this MA still need to be regarded with caution. In addition, it should be noted that the studies included within this analysis vary quite considerably both in size (with larger studies carrying a greater weighting with respect to model outcomes) and in the length of their CIs. Taken as they stand, the fixed- and random-effects models show comparable outcomes, slightly favouring the primary intervention. Again, this is in contrast with the outcome observed for the broader range of studies included in the bivariate analysis. However, the suggested likely impact, based on the observed standardised mean difference, is very small and is also primarily driven by outcomes from one moderately sized study90 that focused on an individually tailored intervention delivered in a hospital setting.

Active treatment versus true control

The number of studies included in our MA of ‘active versus true control’ evaluations was larger (n = 16) than that in the ‘head-to-head’ evaluation and again showed very high levels of heterogeneity (I² = 87%, Table 38) despite the fact that all of the studies in this category focused on pharmacological intervention. The majority of RCTs in this category found in favour of the primary intervention (primarily antidepressant drugs and novel antipsychotic drugs). The main exception to this trend was one large study91 which found no significant difference between divalproex (Depakote, Sanof-Aventis, UK) and placebo. The outcome of this study notwithstanding, the overall effect size based on standardised mean difference (see Table 38) suggest a likely moderate-to-large impact on violence for the interventions combined in this MA (fixed effects –0.60, random effects –0.98), with potential additional gains to be made by tailoring the intervention to specific population groups.

Pharmacological interventions

Figure 12 summarises the outcomes of a MA combining all RCTs with calculated ORs comparing a pharmacological product with either an active or placebo comparator.

FIGURE 12.

Meta-analysis of RCTs comparing pharmacological products with either an active or placebo comparator. For a full explanation of the comparisons included see Appendix 3, Table 56. BPD, borderline personality disorder; CMAI, Cohen-Mansfield Agitation Inventory; DBT, dialectical behaviour therapy; PANSS, Positive and Negative Syndrome Scale; PSE, present state examination; SCL-90, Symptom-Checklist-90.

Given that evaluations of pharmacological interventions formed the bulk of the RCT data, it is not surprising that outcomes for the data for pharmacological interventions alone broadly track those for the overall data set. For the same reason, they also track the outcomes reported above for studies comparing an intervention with a ‘true control’. Specifically, the combined effect sizes tend to provide support for the likely efficacy of the main intervention rather than the active or placebo comparator. Estimates of overall outcome (pooled standardised mean difference using a fixed-effects model –0.54, using a random-effects model –0.81) suggest moderate-to-large impacts on aggressive behaviour with a likely significant benefit of fitting the intervention to appropriate populations, settings or other clinically relevant potential modifiers. Again, however, this implied profile needs to be considered cautiously, as estimates of heterogeneity (Table 39) for these studies, all focused on a single mode of intervention, remain very high (I² = 87%).

Psychological interventions

Outcomes for the combined RCT data for psychological interventions taken as a whole (Figure 13) were slightly less optimistic overall than the combined outcomes for pharmacological intervention. Estimated potential reductions in aggression were towards the low end of the scale (pooled standardised mean differences based on a fixed-effects model –0.26, based on a random-effects model –0.35) with less promise of improvement to be gained by targeting relevant population groups or other modifiers.

FIGURE 13.

Meta-analysis of RCTs comparing psychological interventions. For a full explanation of the comparisons included see Appendix 3, Table 56. CB, cognitive behavioural; F, female; M, male.

Although the overall effect sizes observed were less than those observed for the model for pharmacological interventions, the outcomes of the MA for psychological interventions are more robust, as there is less observed heterogeneity among the psychological studies, so the observed outcomes are more likely to be ‘real’ than the result of artefact. Significant heterogeneity was still observed within this group of studies (I² = 65%, Table 40), but to a moderate rather than an extreme degree. It is therefore reasonably, if not wholly, safe to conclude that psychological interventions do have the potential for making a small impact on violent behaviour. Given the range of approaches included within this category and the lack of a large and consistent body of work evaluating any single type of psychological intervention, further exploration is clearly needed to establish which form of psychological intervention may be most promising.

That the number of RCTs available in this context is so low is unfortunate, as the error margin entailed by this degree of variance cannot be adequately accounted for by dividing analyses into comparable smaller groupings. In the absence of any forthcoming additional RCT data in this area, any conclusions about the effect of psychological interventions will need to be informed by the existing non-RCT analyses to a greater degree than is the case for pharmacological intervention.

Other interventions

The available prospective RCT data for evaluations of interventions other than either pharmacological or psychological intervention are extremely limited. Within this already limited field, only five of the available analyses contributed data suited to MAs based around ORs (Figure 14). These analyses addressed quite distinct forms of intervention, albeit all compared against TAU. Gottfredson and Exum92 compared a drug treatment court approach with TAU, Walsh et al. 93 compared intensive case management with TAU, Labriola et al. 94 and Lasley83 compared intensified bail supervision with other forms of bail supervision, and MacKenzie et al. 95/Mitchell and MacKenzie77 compared a multimodal ‘boot camp’ approach with TAU.

FIGURE 14.

Meta-analysis of RCTs comparing other active or TAU interventions. For a full explanation of the comparisons included see Appendix 3, Table 56.

Given the diversity in focus, it is not surprising that the analyses showed substantial heterogeneity (I² = 77%; Table 41). In contrast with both pharmacological intervention and psychological intervention, the averaged effect size of these ‘alternative’ forms of intervention is also very low, clustering around an outcome of no significant difference in aggression between treatment and TAU (pooled standardised mean differences based on fixed-effects model –0.09, based on random-effects model –0.08; see Table 41).

Although the overall MAs for the three primary modes of intervention suggest more promising outcomes for pharmacological intervention than for either psychological intervention or ‘other’ forms of intervention, it is important to note that, overall, the sets of data analysed here display either significant amounts of heterogeneity between analyses or, in the case of ‘other’ interventions, numbers too small to be suited to the further exploration of possible outcome modifiers. Additional appropriately large RCTs of psychological and ‘other’ interventions would be needed to allow any direct comparison between outcomes for pharmacological interventions and outcomes for other forms of intervention. Evaluation of pharmacological intervention is a well-established research field, while the evaluation of other modes of intervention for violent behaviour is, relatively speaking, in its infancy.

Pharmacological analyses

Sufficient studies were available to analyse three pharmacological groupings in separate meta-analyses: anticonvulsant drugs against placebo, selective serotonin reuptake inhibitor (SSRI) antidepressants against placebo and atypical antipsychotic drugs against haloperidol or placebo.

Anticonvulsant drugs against placebo

As shown in Figure 15 the majority of analyses evaluating anticonvulsant drugs against placebo reported statistically significant outcomes in favour of the active medication (Hollander et al. ,91 Mattes and Mattes96 being the exceptions). Mean effect sizes suggested a substantive potential reduction in aggression, again notably if treatment can be appropriately applied within relevant populations or other modifier subgroups (pooled standardised mean differences based on fixed-effects model –0.62, based on random-effects model –1.47, Table 42).

TABLE 42 - Heterogeneity estimates and effect sizes of RCTs comparing anticonvulsant drugs to placebo

Heterogenity estimates and effect sizes of all included RCTs		Model
		Fixed	Random
n analyses		8	8
Effect size		0.32	0.07
95% CI	Lower	0.23	0.01
	Upper	0.45	0.32
Two-tailed test of null hypothesis (homogeneous data)	z	–6.56	–3.45
	p-value	0.0001	0.001
Estimates of heterogeneity	Q	105.82
	df Q	7
	p-value	0.0001
	I 2	93.38%
Effect size estimates based on standardised mean difference		–0.62	–1.47
95% CI	Lower	–0.81	–2.3
	Upper	–0.44	–0.63
Two-tailed test of null hypothesis (homogeneous data)	z	–6.56	–3.45
	p-value	0.0001	0.001

FIGURE 15.

Meta-analysis of RCTs comparing anticonvulsant drugs to placebo. For a full explanation of the comparisons included Appendix 3, Table 56. SCL-90, Symptom-Checklist-90.

Unfortunately, despite the evident similarities between the analyses in this category, very substantial heterogeneity remained apparent in this combined analysis (see Table 42). As previously, therefore, the promising mean effect sizes cannot be considered robust and should be regarded as indicative at best.

Selective serotonin reuptake inhibitor antidepressant against placebo

A second smaller grouping of analyses97–100 compared three types of SSRI antidepressant (fluoxetine; fluoxetine or nortriptyline and fluvoxamine, respectively) primarily against placebo (against olanzapine in the case of Zanarini et al. 100).

Figure 16 summarises the outcome of this MA. Both fixed- and random-effects models report an equivalently small mean effect size in favour of the primary intervention (pooled standardised mean differences based on fixed-effects model –0.12, based on random-effects model –0.15). The equivalence in outcomes for the two models implies that targeting specific populations or other modifiers is unlikely to result in additional gains.

FIGURE 16.

Meta-analysis of RCTs comparing SSRI antidepressant to placebo/olanzapine. For a full explanation of the comparisons included see Appendix 3, Table 56. BPD, borderline personality disorder; PSE, present state examination.

Unlike the previous analyses, heterogeneity analyses failed to identify any important variances between these analyses (Table 43).

TABLE 43 - Heterogeneity estimates and effect sizes of RCTs comparing SSRI antidepressant to placebo

Heterogeneity estimates and effect sizes		Model
		Fixed	Random
n analyses		4	4
Effect size		0.80	0.76
95% CI	Lower	0.38	0.30
	Upper	1.68	1.93
Two-tailed test of null hypothesis (homogeneous data)	z	–0.58	–0.58
	p-value	0.57	0.56
Estimates of heterogeneity	Q	4.38
	df Q	3
	p-value	0.22
	I 2	31.55%
Effect size estimates based on standardised mean difference		–0.12	–0.15
95% CI	Lower	–0.63	–0.67
	Upper	0.29	0.36
Two-tailed test of null hypothesis (homogeneous data)	z	–0.58	–0.58
	p-value	0.56	0.56

Atypical antipsychotic drugs against haloperidol or placebo

A final grouping of pharmacological analyses with sufficient commonality for evaluation via MA compared a number of atypical antipsychotic drugs (risperidone, aripiprazole, olanzapine, quetiapine and clozapine) with either haloperidol, fluoxetine or placebo. Figure 17 summarises the outcome of this MA. The majority of comparisons drawn found in favour of the active atypical comparator. Mean effect size estimates were very high and roughly comparable between fixed- and random-effects models (pooled standardised mean differences based on fixed-effects model –0.86, based on random-effects model –0.78), suggesting a potentially large impact on aggressive behaviour for these drugs. One characteristic of note is the large effect reported by Suh et al. 101 This study differs from other studies included in the analysis in that it was conducted in a nursing home, with the mean age of patients being 81 years.

FIGURE 17.

Meta-analysis of RCTs comparing atypical antipsychotic drugs to haloperidol or placebo. For a full explanation of the comparisons included see Appendix 3, Table 56. CMAI, Cohen-Mansfield Agitation Inventory; DBT, dialectical behaviour therapy; PANSS, Positive and Negative Syndrome Scale.

As indicated in Table 44, outcomes from this MA were again undermined by the presence of substantial heterogeneity, despite the small number of analyses included and their comparative similarity. No other combination of pharmacological analyses showed sufficient comparability in focus to justify MA.

TABLE 44 - Heterogeneity estimates and effect sizes of RCTs comparing atypical antipsychotic drugs to haloperidol or placebo

Heterogeneity estimates and effect sizes		Model
		Fixed	Random
n analyses		10	10
Effect size		0.21	0.24
95% CI	Lower	0.16	0.14
	Upper	0.27	0.43
Two-tailed test of null hypothesis (homogeneous data)	z	–11.07	–4.87
	p-value	0.0001	0.0001
Estimates of heterogeneity	Q	32.4
	df Q	9
	p-value	0.0001
	I 2	72.23%
Effect size estimates based on standardised mean difference		–0.86	–0.78
95% CI	Lower	–1.02	–1.09
	Upper	–0.71	–0.46
Two-tailed test of null hypothesis (homogeneous data)	z	–11.07	–4.87
	p-value	0.0001	0.0001

Psychological analyses

The only set of psychological analyses with sufficient comparability to combine in a within-groupings MA was a broad set of analyses all comparing CBT (of diverse types) against a range of active and waiting list comparators (prison, anger management, psychotherapy and related interventions). One study102 that focused on meditation is included in this MA, as the approach taken to meditation was relevantly similar to more usual forms of CBT (broadly defined as per the range of approaches taken by this group of studies). Figure 18 sets out the results of this MA and Table 45 explores the degree of heterogeneity within this group of analyses.

FIGURE 18.

Meta-analysis of RCTs comparing CBT to a range of active and waiting list comparators. For a full explanation of the comparisons included see Appendix 3, Table 56. CB, cognitive behavioural; F, female; M, male.

Although the majority of individual analyses within this group were unable to report statistically significant findings in favour of CBT, the combined effect sizes, both fixed and random, indicated a small potential decrease in aggression by applying the CBT model (pooled standardised mean differences based on fixed-effects model –0.27, based on random-effects model –0.27). The size of effect being equivalent between fixed and random models again implied that increases in effectiveness are unlikely to be gained by targeting specific populations or other modifiers.

The outcomes of this MA, indicating a small effect of CBT, are likely to be robust. No significant degree of heterogeneity is indicated in the analysis set out in Table 45 and the number of analyses although small is adequate for the exploration of heterogeneity. This type of intervention would therefore be a promising one to explore further in future RCTs. One caveat here is the need to define the intervention more tightly than has been done to date in the literature. The approaches to ‘CBT’ vary so widely that they could be considered very distinct interventions and it is of substantial importance to be able to identify which components of these diverse interventions actually contribute to the relative ‘success’ of CBT in impacting on aggressive behaviour.

Analyses of ‘other’ interventions

The number of analyses in this final grouping is small, as indicated previously, and the focus of the comparisons is diverse. No further benefit could be gained by attempting to partition these data further in additional MAs.

Clinically relevant factors

Outcome measures

In the context of the RCT analyses, there was considerable consistency in the choice of outcome measure, with the vast majority of included analyses favouring scale-based measures. Contrasting scale versus non-scale (generally behavioural) measures (Tables 46 and 47), scale-based measures tended to be associated with better outcomes under both fixed- and random-effects models, with non-scale-based analyses showing non-significant pooled effect sizes. Both types of grouping showed large heterogeneity between analyses, although the diversity (in all likelihood because of analysis numbers) was smaller between analyses not using scale-based outcomes.

Study quality indicators

Clearly, the studies used for these analyses had been selected on the basis of perceived ‘study quality’, following the accepted gold standard of the prospective RCT. This notwithstanding, they of course differed in other aspects of study design and a number of these have implications for the likely reliability and validity of analysis outcomes. The ‘quality modifiers’, for which we have sufficient data for analysis, are explored below.

Model 1: anticonvulsant drugs versus placebo

Four studies were included in this model. 103–106 All four studies focused on an exclusively female, general mental health population. All studies took place in a community setting and used some form of blinding. Sample sizes at the outset ranged from 30 to 64 participants and outcomes for all four studies were scale based. Frankenburg and Zanarini106 evaluated the efficacy of divalproex sodium, Nickel et al. 104,105 evaluated the efficacy of topiramate and Tritt et al. 103 evaluated the efficacy of lamotrigine (Figure 19).

FIGURE 19.

Meta-analysis model 1. For a full explanation of the comparisons included see Appendix 3, Table 56. SCL-90, Symptom-Checklist-90.

Despite the similarities between these studies both in respect of their main focus and in respect of relevant modifiers, substantive heterogeneity was observed (Q = 20.21, p = 0.0001, I2 = 85.16%) (Table 49). Although the model could be viewed as suggesting a potential impact of anticonvulsant medication on violent behaviour, it is consequently not advisable to draw this conclusion because of the substantial heterogeneity observed.

Model 2: atypical antipsychotic drugs versus any active comparator

Three studies were included in this model. 82,101,107 All three focused again on a general mental health population. Participants in all three studies included both males and females, with the proportion of male participants ranging from 25% to 83.8%. Settings were not specified for two of the studies82,107 and the setting for Suh et al. 101 fell into the ‘other/mixed’ category. Some form of blinding was used in all three studies and outcomes for all three were scale based. Citrome et al. 107 compared monotherapy with either risperidone or olanzapine to combination therapy with either of these drugs plus divalproex sodium. Krakowski et al. 82 compared clozapine with haloperidol and Suh et al. 101 compared risperidone with haloperidol (Figure 20).

FIGURE 20.

Meta-analysis model 2. For a full explanation of the comparisons included see Appendix 3, Table 56. PANSS, Positive and Negative Syndrome Scale.

Again, despite selecting studies for apparent similarities in respect of focus and potential modifiers, a significant degree of heterogeneity was observed in this model (Q = 52.96, p = 0.0001, I² = 96.22%) (Table 50). The promising outcomes suggested for atypical antipsychotic drugs therefore fall prey to the unreliability generated within the model by characteristics of the studies that we have yet to identify.

Model 3: atypical antipsychotic drugs versus placebo

Three studies were included in this model. 89,108,109 Again, all three focused on a general mental health population. Participants in all three studies included both males and females, with the proportion of male participants ranging from 19.23% to 59%. Two of the studies took place in community settings,89,108 and the setting for Meehan et al. 109 fell into the ‘other/mixed’ category. Some form of blinding was used in all three studies and outcomes for all three were scale based. Sample sizes ranged from 52 to 272. Brown University89 evaluated the efficacy of aripiprazole, Tyrer et al. 108 evaluated the efficacy of risperidone, and Meehan et al. 109 evaluated the efficacy of olanzapine (Figure 21).

FIGURE 21.

Meta-analysis model 3. For a full explanation of the comparisons included see Appendix 3, Table 56. PANSS, Positive and Negative Syndrome Scale.

This model follows a very similar pattern to the model 2, with promising outcomes undermined by significant heterogeneity in the data (Q = 10.67, p = 0.0001, I² = 81.26%) (Table 51) despite apparent similarities between the studies with respect to focus and key modifiers.

Model 4: cognitive behavioural therapy versus any active comparator

Three studies were included in this model. 86,110,111 All three studies focused exclusively on male participants, none of the studies used any form of blinding, outcomes for all three studies were scale based and sample sizes ranged from 42 to 78. All studies took place in either community111 or ‘other/mixed’ settings. This group of studies were more varied in respect of their target participant group than studies included in the above three models. Lanza et al. 86 focused on a general mental health population, Easton110 on an offender population, and Easton et al. 111 on a group of participants who had committed acts of violence but as yet had not been indicted for these acts (Figure 22).

FIGURE 22.

Meta-analysis model 4. For a full explanation of the comparisons included see Appendix 3, Table 56.

Despite the clear differences in population focus, this model generated the least heterogeneity (Q = 4.70, p = 0.09, I² = 57.4%) (Table 52).

Characteristics of the overall literature

The descriptive analysis in the review provides a summary of the overall literature. Approximately 25–30 new studies are published each year and 70% of these are conducted in three countries (USA, UK and Canada). Inevitably, this restricts the generalisability of findings to less developed and/or non-English-speaking countries. The median length of follow-up is 6 months, with an average attrition rate of 10%. Given the intractability of the propensity to violence compared with more acute problems such as active psychosis and its lifelong nature for many people, this follow-up is not much more than a snapshot of the potential for change or otherwise. A third of the literature focuses on three diagnostic groups exclusively (schizophrenia, dementia and personality disorder). Half of the studies are conducted on males only, with < 10% conducted on female-only populations, and there is a disproportionate emphasis on non-Caucasian populations (i.e. half of the studies have samples in which the majority of participants are non-Caucasian). As, for instance, rates of violence by women with some types of mental health problem are not very different from those for similar men,120 this could be taken to suggest that there is an emphasis on certain groups in society that does not reflect the actual scale of the problem.

Methodologically speaking, it is clear that there are comparatively few RCTs in this research area (51/198, 26%: see Table 53). Compared with the number of trials in many other areas of health care, the proportion is very small. 121 The proportion is smaller still within the psychosocial intervention literature. Possible reasons for this are familiar from other research reviews and include the impracticality of achieving formal randomisation in service delivery settings, within interventions that entail individual contact or of using appropriate placebos, and the absence of means of checking against multiple treatment interference. On the other hand, it is worth noting that, across both literatures, non-RCTs had much longer follow-up periods than RCTs so that gains in causal inference from improved design quality in RCTs have to be weighed against losses in terms of tracking the persistence of an effect. Taken together, these issues reflect the complexity of the violence phenomenon and make it difficult to address the issue of effectiveness with any simplicity.

Within the located studies there are relatively few that focus on measured or recorded aggressive or violent behaviour as outcomes. The majority define outcome in non-behavioural terms, such as hostility or anger (e.g. the STAXI instrument), which can be detected only through self-report. Although these types of outcome measure obviously have value in some respects, they are only ever a proxy or risk factor for the main problem of actual aggressiveness as recorded, for example, on the basis of inpatient ward incident reports or of criminal reconvictions. Much of the literature can therefore provide evidence of effectiveness in terms of only a relatively subjective aspect of the problem.

Finally, a funnel plot of the data set, as a whole, is indicative of the possibility of there being some publication bias within the studies analysed. This is despite significant efforts having been undertaken to secure unpublished studies via extensive search for dissertations. Although this requires the exercise of caution when interpreting the findings, the number of studies conducted and the extent of the review team’s contact with other researchers active in the field, thus ensuring inclusion of a substantial number of unpublished studies, suggests that this bias does not distort the findings as much as in other reviews. It is important to note that the distribution observed here could result from systematic association between sample size and other study characteristics that impact on outcome and that the absence of funnel plot asymmetry does not rule out the possibility of publication bias anyway so the impact of funnel plots should not be exaggerated. Nevertheless, it must be acknowledged that the presence of such funnel asymmetry indicates that the MA results could well be overestimates of the effectiveness of interventions and that the literature is more suited to informing future research strategies than answering clinical questions at this stage.

Regression analysis of the overall literature: correlates of effectiveness

The regression analysis of all 198 studies (RCTs and non-RCTs) was aimed at establishing the design and population factors associated with a statistically significant effect (p < 0.05) in favour of the active intervention, and highlighted a number of issues of interest.

In terms of intervention types, ‘other’ interventions were less likely than psychological or pharmacological interventions to be associated with an effective outcome.

Two population variables showed a significant association with effectiveness. Offender-focused studies were significantly less likely to be effective, whereas mental health-focused studies were more likely to be effective. However, it is worth noting that analysis based on settings rather than populations showed that application in either mental health or offender settings was not relevant to outcome. This gives grounds for optimism with mentally ill samples whether they are located in general mental health or offender settings, but conversely highlights difficulties in implementing therapeutic interventions with offender populations wherever they are located. Furthermore, three design variables were associated with a reduced likelihood of demonstrated effectiveness: head-to-head comparisons, comparison against TAU, and between-groups comparisons.

Further (multivariate) analysis was then conducted to explore the extent of variation explained by outcomes in favour of the primary intervention arm. None of the selected factors had a significant relationship to effectiveness but, in population terms, the relative advantage noted above of studies focused on a mental illness sample over those focused on an offender sample was maintained. The relative disadvantage of ‘other’-type interventions was also maintained. Methodologically, only the use of a within-groups design approached significance in this analysis. Further analysis may help to identify if the lack of effect for evaluations of ‘other’ interventions and those conducted in penal institutions are due to design issues (e.g. insufficient power), implementation issues (e.g. lack of protocols) or the actual inapplicability of these approaches.

Meta-analysis

Heterogeneity became a particular problem when conducting the MA. This underpinned our decision to conduct the above preliminary regression analysis of variables associated with a significant outcome and, when considering the more stringent formal MA, indicates that great caution should be used when interpreting the results. Q-statistics for a series of comparisons in the MA, for instance, were as follows: for all pharmacological interventions comparing two active treatments Q = 111.21 (p < 0.0001), and for all psychological interventions, Q = 21.1 (p < 0.007). I² estimates are similarly large and statistically significant. The exception to this is the comparison between SSRI antidepressants and placebo, for which across four studies Q was exceptionally low, at 4.38 (p = 0.22), suggesting relatively low heterogeneity in this particular set of comparisons. The relatively high heterogeneity overall is a function of the comprehensive scope of the review but does again indicate the absence of a programmatic approach to violence research in which there is consistency in outcome measurement. Also, with regard to non-pharmacological interventions, it indicates a failure to specify the treatment approaches adopted. Even ‘CBT interventions’, supposedly drawing on a consistent theoretical model, adopt a range of different techniques. Although this high degree of heterogeneity limits the conclusiveness of the overall review, the estimation of the size of heterogeneity is an important step in itself in bringing some order to the violence research field.

Notwithstanding the heterogeneity issue, across the data set as a whole there is evidence from the MA of a positive outcome of intervention, and this emerges from both fixed-effect and random-effects analyses of the results. Discussion of mean effects across this entire data set is not, however, likely to be especially informative, given (1) the variability of intervention methods (both with respect to different principal categories, e.g. pharmacological vs psychosocial) and the combinations within those categories and (2) the mixed range of outcome measures used. Comparisons between the same method applied to different populations, in different settings, or using different outcome measures is simply not possible. The exception to this is the set of seven studies of CBT84,86,102,110,111,198 (which are based on a common theoretical model and possess similar operational characteristics). Using Lipsey’s60 ‘rule of thumb’, which, in turn, drew on conventions proposed by Cohen,122 small-to-moderate effects (combined with relatively low heterogeneity) are found for CBT, for all psychological interventions combined and larger effects for atypical antipsychotic drugs. Caution must be exercised when interpreting all results, however, as 95% CIs are relatively wide and sometimes include zero, particularly in the random-effects models.

Analysis of the modifier variables provides some evidence that interventions targeted at mental health populations, and particularly male groups in community settings, are more likely to achieve stronger effects. In terms of future research design, high-quality features such as blinding and ITT analysis are likely to be associated with larger effects, which indicates the detection of a ‘true’ effect. However, the low-quality marker of baseline non-equivalence of comparison groups was also associated with such effects. It is nonetheless encouraging to note that in a review of cognitive behavioural programmes for offenders, examining criminal recidivism as the principal outcome variable, the mean OR effect size for studies with ‘best practice’ design features was considerably higher than that for the collection of studies overall. 123

Scale-based measurement was also associated with stronger effects but we have noted above that these provide only a proxy for actual aggression and thus interventions aimed at such behaviour should include observational measures as well, which may reduce the effect size estimate. It is of course legitimate to study such proxies as anger in their own right using validated scales, as they are likely to be predictive of subsequent violence.

There was evidence, although again not emerging in a clear or entirely consistent pattern, for larger effects to be found in community than in institutional settings, a finding obtained in a number of previous MAs in criminal justice. 123,124 Similarly, larger effects sizes were not from smaller samples, reinforcing the impression that emerges from large-scale dissemination exercises that effects are attenuated possibly because of compromises in integrity of treatment or service delivery. 125

The findings obtained from the present review tailor reasonably well with the trends noted in previous reviews summarised earlier (see Chapter 1, Other reviews), given that, being mainly sited in criminal justice settings, they also focused on aggression and violent behaviour (including sexual assault) rather than adopting a diagnostic approach. Thus, there is further support for the use of cognitive and behavioural interventions. The principal residual difficulty arises from the heterogeneity of specific methods used; thus, although theoretically grounded in the cognitive social learning model there is still insufficient evidence regarding any single treatment programme to identify a ‘treatment of choice’ or make firm selective recommendations. The evidence of effectiveness found here for some forms of pharmacotherapy does not fit well with the recommendations made by NICE against routine use of such treatments for aggression associated with antisocial or BPD, and could be an opportunity to reopen some aspects of this question. The effect observed in this review, for instance, could reflect the general tranquillising or suppressing effect on acting-out behaviour. It does support the conclusions to the Cochrane review52 with regard to the role of antipsychotic drugs with BPD symptoms associated with aggression (e.g. affective dysregulation) and some of the individual studies of neuroleptics cited by Herpertz et al. 53 Nevertheless, the latter review overall was, like NICE, inconclusive with regard to pharmacotherapy and aggression in antisocial personality disorder. On the other hand, the effectiveness of psychological interventions, including CBT, observed here for reducing aggression reinforces the recommendations made by NICE for the role of such interventions in the overall treatment of borderline and antisocial personality disorder.