This meta-analysis demonstrated an internationally high, three- to eightfold increased risk of drug-related death in the first 2 weeks after release from prison compared with the subsequent 10 weeks. There was heterogeneity in fatal overdose risk across the three continents represented, with one US and the two Australian studies showing a much lower overdose risk in the first 2 weeks post-release than the other US and two UK studies. The risk of drug-related death is also shown to remain elevated beyond the first 2 weeks post-release into the second 2 weeks (relative to weeks 5–12). This reinforces earlier observations [4–6] that, while the risk decreases by weeks 3 + 4, it remains importantly raised.
The increased risk of drug-related death may be explained by a decrease in tolerance to drugs as a result of being in prison where drug use is less frequent and the drugs may be of lower purity ; and there could be a tendency for ‘celebration’ on release. Across the six studies the RRs for weeks 1 + 2 are elevated to different extents, which is perhaps unsurprising given the possible variations in prisons' drug policies and in the nature of illicit drug use more broadly. We highlight here potential explanations for the differently elevated RRs in the first 2 weeks.
The purity of heroin (as well as methamphetamines, cocaine and crack cocaine) could vary according to manufacturing and availability; and associations between heroin purity and the occurrence of drug-related deaths have been found [29,30]. The prevalence of injecting versus non-injecting routes of opioid administration may vary regionally, and injection poses the greater overdose risk . There may also be regional variations in the patterns of co-use of alcohol, benzodiazepines and other depressants of the central nervous system. In combination with heroin, such depressants potentially present an increased overdose risk [21,30,34]. Upon release from prison, regional variations in the above drug use behaviours or cultures could contribute to differential effects on the immediate risk of drug-related death, before tolerance is restored.
In addition, drug treatment programmes may differ in availability, both inside and outside prisons; in methadone maintenance dose (>50 mg); and the protection they afford against injection-related risk behaviours and premature mortality [32,33].
Studies also varied by the age, sex and length of incarceration of their respective ex-prisoners. The Scottish study focused upon 15–35-year-old males who had been incarcerated for at least 14 days, to target the younger, short sentence profile of drug-users . In Binswanger et al. , two-fifths of released prisoners were outside this age range; females were included (3967 of 30 237); and releases were from ‘prisons’ which, by definition in the United States, are for individuals sentenced for more than 1 year (plus those re-incarcerated for violating parole). To try to identify aspects in correctional policies or prisoners' demography that determine their drug-related death risk soon after release, more studies of the type included in this meta-analysis from other correctional systems would be needed.
We note that the large, ineligible, last-release historical study in North Carolina, 1980–2005  yielded a low relative risk comparing weeks 1 + 2 with weeks 3–12, 3.1 (95% CI: 1.7–5.5). This was in agreement with the New Mexico study , 3.1 (95% CI: 1.3–7.1), but considerably lower than the study in Washington State , 8.4 (95% CI: 5.0–14.2). Given that Rosen, Schoenbach & Wohl  introduced upward bias by studying ex-prisoners' most recent incarceration, one might have expected higher absolute risks to be reported.
The Australian studies [7,8] exhibited the lowest RR for weeks 1 + 2. One explanatory factor could be the earlier establishment of methadone maintenance in NSW prisons (readily available for prisoners from the late 1980s  versus from only 2003 in Scotland ). This was tested in an exploratory re-analysis of the NSW data for consecutive 5-year periods. Unexpectedly, the RR for weeks 1 + 2 actually increased from 2.7 (95% CI: 1.7–4.3) in 1988–92 to 5.1 (95% CI: 3.8–6.9) in 1998–2002. Of course, temporal analyses are liable to confounding. Australia's average heroin purity roughly trebled during 1996–2000 , which may have caused an increase in overdose fatalities . Moreover, Australia's heroin drought at the end of 2000  is well known for having complicated the interpretation of its drug-related mortality trends.
Another possible explanation for the observed variation in risks could be studies' different definitions of drug-related death (see Appendix S1). For example, suicides were excluded from the definition of drug-related death adopted by Bird & Hutchinson , but included by Farrell & Marsden [4–6] and Kariminia et al.  (and by Rosen, Schoenbach & Wohl ). The impact of these differing definitions was explored by re-analysing the NSW data according to the definition used by Bird & Hutchinson . This yielded virtually unaltered RRs (data available from authors), suggesting that different definitions did not account for the heterogeneity. Nevertheless, studies' different definitions remain an acknowledged limitation of our meta-analysis.
Inclusion of the study by Verger et al.  would have caused analytical problems because they did not observe any drug-related deaths in weeks 1 + 2 and only one in weeks 3–12 after release. Under the null hypothesis of homogeneity of risk over the 12-week period, the expected numbers of such deaths in weeks 1 + 2 would be so small that the large sample approximation, upon which the χ2 distribution is based, breaks down. In our view, including such studies does not make statistical sense and we suggest that meta-analyses more generally may benefit from the exclusion of such poorly powered studies from the outset.
This meta-analysis includes only six studies, but each is moderately powerful and they conform to robust eligibility criteria. As forewarned by Seaman et al. , our 12-week follow-up period avoids major confounding by periodic changes in ex-prisoners' drug use. The findings do not rely on either ex-prisoners' willingness to self-identify as drug-using or the diagnosis of drug dependence. Analyses are unbiased and estimates have been pooled only where appropriate. The meta-analysis was restricted to studies with associated publications in the English language, so it is possible that studies in other languages were missed. However, as the search retrieved studies from Finland, Denmark and France we hope that studies omitted by restricting to English-language studies are few, if any. Further studies from resource-poor countries and other continents would, however, be highly instructive.
Future studies of drug-related death soon after release from prison should be well designed, along the lines of the eligibility criteria for this meta-analysis. For sufficient statistical power, we suggest that studies should be of at least 10 000 index releases (10 000 unique ex-prisoners), with at least eight drug-related deaths in the first 2 weeks. Based on the calculations of Bird & Hutchinson , studies of this size would have at least 50% statistical power to detect a RR of 4 for drug-related death during weeks 1 + 2 versus weeks 3–12 after release from prison.
Remedial action is required by both prisons and communities to address this observed elevation in risk. Prisoners should be alerted to the high risk of overdose death soon after release, advised not to be alone if they use drugs and to be wary of mixing heroin with other drugs, including alcohol . Transitional care programmes, which provide pre- and post-release treatment and support, are promising interventions but, as yet, the research has not been conclusive. Novel programmes include prison-based education about naloxone (heroin antidote) and its prescription to prisoners with a history of heroin injection .