Search strategy

We started by identifying previously published meta-analyses and systematic reviews of interventions following self-harm and retrieved the primary studies included in these reviews for further analysis. We then searched the following electronic databases: EMBASE (1969 to February 2005); PsycINFO (1967 to February 2005) and Medline (1966 to February 2005). We used a broad range of search terms based on those developed by Hawton et al (Reference Hawton, Townsend and Arensman2000), and searched the reference lists of all the relevant papers.

Inclusion criteria

Studies were eligible for inclusion in the review if they were randomised controlled trials; involved patients who had harmed themselves in the period prior to entry into the trial; and compared additional or enhanced intervention with a form of control or standard care. Where papers met these criteria but did not provide mortality data, we attempted to obtain these data from the authors.

Selection of studies and data extraction

Two raters independently assessed all papers for possible inclusion in the review. If there was disagreement about whether a study should be included this was discussed in detail with a third reviewer and a decision made on whether or not to include it. Information on the study population and interventions were recorded, and data on study size, follow-up rate and numbers of suicides and other deaths in each arm of the trial were extracted by two raters and double-entered onto a database.

Data analysis

As we were interested in exploring whether additional psychosocial treatment interventions had an impact on the likelihood of suicide, we took an a priori decision to combine data from all psychosocial interventions. Meta-analysis was based on suicide rates; this allowed us to take account of variation in the length of follow-up of studies (from 3 months to 36 months). Also, if we had used an odds ratio to calculate the overall effect size we would not have been able to incorporate studies in which no suicide was reported. Instead, we calculated the rate of suicide in each arm of the study (μ₁ and μ₂) to model the numbers of deaths X and Y in treatment and control arms. With numbers of participants in the treatment and control arms m and n respectively, the proportion of participants in the control arm of the study who died by suicide q was calculated using the formula q=n/(m+n), and the proportion who died by suicide in the treatment arm by 1–q=m/(m+n).

To test for a difference in death rates within each study we can compare a value of a test statistic T=2[(X+3/8)^1/2 q ^1/2–(Y+3/8)^1/2(1–q)^1/2] to critical values of a standard normal distribution (Reference HuffmanHuffman, 1984). This statistic is especially useful for our data with low death rates because it uses a variance stabilising transformation. An effect size associated with this statistic is θ=μ₁ ^1/2–μ₂ ^1/2. Its estimate from a study with a follow-up time t is \batchmode \documentclass[fleqn,10pt,legalpaper]{article} \usepackage{amssymb} \usepackage{amsfonts} \usepackage{amsmath} \pagestyle{empty} \begin{document} \(\underline{{\hat{{\theta}}}}[(X+3{/}8){/}mt]^{1{/}2}-[(Y+3{/}8){/}nt]^{1{/}2}\) \end{document} . The standard deviation of this estimate is s=2^–1 [(m+n)/mnt]^1/2.

The values of θ for each study were entered into RevMan 4.2.7 for Windows (Cochrane Collaboration, Oxford, UK; see http://www.cc-ims.net/RevMan) as ‘generic inverse variance data’ and can be routinely analysed. A fixed-effects model was used. Heterogeneity between studies was examined using the χ² test. We then repeated the analysis using death from any cause as the outcome. Finally, funnel plot asymmetry was examined in order to assess the possibility of publication bias (Reference Egger, Davey Smith and SchneiderEgger et al, 1997).

Types of interventions

Most of the interventions involved a limited number (between 2 and 50) of sessions of individual psychotherapy such as cognitive–behavioural therapy, interpersonal psychotherapy and dialectical behaviour therapy. In other studies changes were made to the organisation of services to enhance uptake of psychosocial treatment or facilitate contact with services at times of crisis.

Impact on suicide

Data from the 18 studies appeared to be homogeneous. Since the value of the test statistic for heterogeneity is less than its degrees of freedom, there was no need to consider the random-effects model. Root differences in rates of suicide in the 18 trials of psychosocial intervention are presented in Fig. 2. Overall there was no evidence of any difference in death rates.

Fig. 2 Meta-analysis of effect of psychosocial interventions on completed suicide.

There were 37 suicides among the 3918 participants in the trials of psychosocial interventions, 18 per 2165 person-years among those who received enhanced treatment and 19 per 2237.5 person-years among those who received control treatment. Randomised trials of manual-assisted cognitive therapy (Reference Tyrer, Thompson and SchmidtTyrer et al, 2003), out-patient cognitive–behavioural therapy (Reference Brown, Ten Have and HenriquesBrown et al, 2005), and intensive out-patient care (Reference Van der Sande, van Rooijen and BuskensVan der Sande et al, 1997b ) showed a trend towards lower levels of suicide. The remaining studies all showed no difference, or slightly higher levels among those in the active arm of the trial. Funnel plots of effect size estimates against inverse standard error (an indication of study size) were symmetrical, suggesting that publication bias was unlikely to have affected study findings (Fig. 3).

Fig. 3 Funnel plot of studies reporting suicides among those offered experimental and control treatment following episodes of self-harm.

Impact on mortality

Complete mortality data were available in only 12 studies. Difference in root rates for suicide in the 18 trials of psychosocial intervention are presented in Fig. 4. There was a total of 29 deaths among people in these trials: 18 per 1422 person-years among those receiving control treatment and 11 per 1527.5 person-years among those receiving additional psychosocial interventions.

Fig. 4 Meta-analysis of psychosocial interventions on all-cause mortality.

Level of suicide

Few papers provided information about how suicide deaths were ascertained. Given that a proportion of participants in each study were lost to follow-up it is possible that some suicide deaths were not recorded.

It has been suggested that participation in a trial might of itself affect patient outcomes (Reference RothwellRothwell, 2005). Participants in these studies all received a higher level of assessment as a result of participation in a trial, and most trials excluded people who were judged to be at high risk of subsequent suicide. Five trials we included followed participants for less than a year, but among those that followed people for 12 months the overall rate of suicide during this period was 0.016. This is similar to the rate of 0.018 reported in cohorts of people followed up after an episode of self-harm (Reference Owens, Horrocks and HouseOwens et al, 2002).

Interpretation of results

The relative lack of statistical power in this meta-analysis means that caution needs to be used when interpreting the study findings. Many of the trials of psychosocial interventions we examined were too small to detect differences in the level of repetition of self-harm, and suicides occurred in only 11 of the 18 studies included in the review. During the course of this review we identified three trials of pharmacological interventions (Reference Hirsch, Walsh and DraperHirsch et al, 1982; Reference Montgomery, Roy and MontgomeryMontgomery et al, 1983; Reference Verkes, Van der Mast and HengeveldVerkes et al, 1998). With a combined study population of 243 and one suicide by a member of the control group in the trial by Verkes and colleagues, a meta-analysis was not appropriate for these interventions.

A total of 4004 person-years (2002 in each arm) was sufficient to achieve power of 80% to detect a difference between a rate of suicide of 1.8% in the year following an episode of self-harm, and 50% reduction in the level of suicide among those receiving additional psychosocial interventions. Had trials demonstrated reductions in levels of suicide of the magnitude reported by Tyrer et al (Reference Tyrer, Thompson and Schmidt2003) we would have had ample power to detect statistically significant differences in our primary outcome, with 4397.5 person-years in this meta-analysis. However, we had insufficient power to detect smaller – but still clinically significant – reductions in the level of suicide. Study power would have been increased had we been able to obtain suicide data from the nine remaining randomised trials that met our inclusion criteria. These nine trials included a combined population of 2914 participants.

None of the studies we identified set out to examine whether intervention would lead to a reduction in the rate of suicide, and the relatively low rate of suicide among people in the period following self-harm would make this an odd choice of outcome in a single trial. However, it has been argued that interventions to reduce repetition of suicidal behaviour following self-harm could have an impact on suicide rates (Reference Gunnell and FrankelGunnell & Frankel, 1994; Reference Mann, Apter and BertoloteMann et al, 2005), and publication of detailed information on suicides and other deaths occurring among participants in future trials of interventions for people who self-harm would enable a more precise estimate to be made of the impact of these interventions on suicide.

Because we were interested in exploring whether enhanced psychosocial treatment following self-harm affects the likelihood of subsequent suicide, we combined data on a variety of different types of psychosocial intervention. In doing so it is possible that the impacts of specific forms of intervention have been minimised. Differences in the rate of suicide in the trial of manual-assisted cognitive–behavioural therapy by Tyrer et al (Reference Tyrer, Thompson and Schmidt2003) are of particular note. The single death (in the control arm) of the more recent trial of cognitive–behavioural therapy by Brown et al (Reference Brown, Ten Have and Henriques2005) adds little to the statistical power required to examine whether this therapy reduces the subsequent rate of suicide. However, the relatively modest reduction in repetition of self-harm in such trials cautions against the view that this therapy is likely to be associated with marked reductions in levels of subsequent suicide.

Although we found no difference in rate of suicide among those randomised to standard and enhanced treatment, a trend towards a reduction in the overall rate of mortality was observed. This could be a chance finding, but it could reflect a real difference in mortality between those offered standard care and those offered additional psychosocial interventions. It is known that some suicide deaths are sometimes misclassified as due to other causes (Reference Phillips and RuthPhillips & Ruth, 1993; Reference Neeleman, Mak and WesselyNeeleman et al, 1997). This finding emphasises the need to interpret the absence of difference in rate of suicide in studies with caution.

Implications for suicide prevention

This study did not examine the impact of standard care on levels of suicide following self-harm and it is possible that interventions currently provided by mental health services do much to limit the likelihood of suicide (US Department of Health and Human Services, 2001). However, we found little evidence to support the view that additional psychosocial treatment for people who self-harm leads to reductions in levels of suicide.

Individual randomised trials of psychosocial treatments have demonstrated statistically significant reductions in the likelihood of repetition of non-fatal self-harm, but such findings do not necessarily mean that these treatments would reduce the likelihood of subsequent suicide. We believe that these data support the view that a range of public health measures for preventing suicide should be pursued, including population-based strategies such as restricting access to means of suicide and encouraging responsible discussion of suicide in the media.