Risk of All-Cause Mortality in Alcohol-Dependent Individuals: A Systematic Literature Review and Meta-Analysis☆

Background Alcohol dependence (AD) carries a high mortality burden, which may be mitigated by reduced alcohol consumption. We conducted a systematic literature review and meta-analysis investigating the risk of all-cause mortality in alcohol-dependent subjects. Methods MEDLINE, MEDLINE In-Process, Embase and PsycINFO were searched from database conception through 26th June 2014. Eligible studies reported all-cause mortality in both alcohol-dependent subjects and a comparator population of interest. Two individuals independently reviewed studies. Of 4540 records identified, 39 observational studies were included in meta-analyses. Findings We identified a significant increase in mortality for alcohol-dependent subjects compared with the general population (27 studies; relative risk [RR] = 3.45; 95% CI [2.96, 4.02]; p < 0.0001). The mortality increase was also significant compared to subjects qualifying for a diagnosis of alcohol abuse or subjects without alcohol use disorders (AUDs). Alcohol-dependent subjects continuing to drink heavily had significantly greater mortality than alcohol-dependent subjects who reduced alcohol intake, even if abstainers were excluded (p < 0.05). Interpretation AD was found to significantly increase an individual's risk of all-cause mortality. While abstinence in alcohol-dependent subjects led to greater mortality reduction than non-abstinence, this study suggests that alcohol-dependent subjects can significantly reduce their mortality risk by reducing alcohol consumption.


Introduction
Alcohol use is one of the greatest risk factors for disease and disability (Rehm, 2011;Nutt et al., 2010;Rehm et al., 2009), and alcohol dependence (AD) seems to account for the majority of this burden (Rehm et al., 2012;Rehm et al., 2013). The risk of mortality has been shown to increase as alcohol consumption increases, both for lifetime risk and absolute annual risk, with absolute annual risk almost doubling as alcohol consumption increases from 10 g/day to 100 g/day . In addition to the clinical burden of AD experienced by individuals (François et al., 2014), AD has wider societal consequences, including substantial direct and indirect economic costs (Rehm et al., 2012;Laramée et al., 2013).
Until the 1970s, alcohol use disorders (AUDs) were widely called 'alcoholism'; by this time, however, it was apparent that AD could be considered as a separate diagnosis (Edwards and Gross, 1976). The current version of the International Classification of Diseases (ICD-10) continues to categorise harmful use and AD as separate diagnoses (World Health Organization, 1992), while the latest edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) has integrated alcohol abuse and AD into a single AUD diagnosis (American Psychiatric Association, 2013). In clinical practice, there is often no formal assessment of diagnoses (ie. alcohol abuse vs AD), but for treatment in specialised healthcare services it is safe to assume that most of the cases would qualify as the more severe form of AUD, corresponding to AD (Rehm et al., 2015a).
Previous systematic literature reviews (SLRs) and meta-analyses have examined the relative risk (RR) of all-cause or cause-specific mortality in people with AUDs compared with the general population or with controlled drinkers Roerecke and Rehm, 2014;. One meta-analysis found an RR of 3.38 (95% CI [2.98, 3.84]) for men and 4.57 (95% CI [3.86,5.42]) for women in clinical settings compared to the general population ; another found that individuals treated for AUDs reduced their mortality risk by more than half if they were able to reduce their alcohol consumption, compared to those individuals who continued to drink heavily . However, to our knowledge there are currently no systematic reviews focusing on the risk of all-cause mortality in alcohol-dependent individuals only.
Treatment for AD, and AUDs more widely, has traditionally focused on promoting abstinence as the only acceptable treatment goal. However, some patients may prefer a goal of non-problem drinking (Wallhed Finn et al., 2014). In recent years, there has been an increased emphasis on an alternative harm-reduction approach that attempts to help alcohol-dependent patients achieve a reduction in alcohol consumption without the need to completely abstain, consequently reducing the risk of harmful consequences associated with alcohol use (European Medicines Agency (EMA), 2010; National Institute for Health and Care Excellence (NICE), 2011). Reduced consumption of alcohol in individuals with AUDs has been shown to be beneficial, resulting in a significant reduction in mortality compared to continued heavy drinking , and is also predicted to improve the associated economic and societal burdens .
In this study, we aimed to conduct an SLR and meta-analysis on the increased risk of all-cause mortality among individuals with AD compared to the general population, individuals without AUDs, and individuals qualifying for a diagnosis of alcohol abuse; and to examine the key factors affecting this risk. We also aimed to review the effect of reduced alcohol consumption among alcohol-dependent individuals.

Systematic Literature Review
An SLR was conducted in accordance with PRISMA guidelines (Moher et al., 2015) to identify studies reporting on mortality in alcohol-dependent subjects. MEDLINE, MEDLINE In-Process, Embase and PsycINFO were searched using the Ovid SP platform, and the Cochrane Library was searched using the Wiley Online platform. Search strings included terms relating to AD and mortality (Supplementary  Tables 1 and 2). All searches were conducted on 26th June 2014; databases were searched for studies published from database conception up to that date.
Titles and abstracts of all studies identified in the database searches were screened using pre-defined eligibility criteria. Full texts for all potentially eligible studies were acquired and screened again. Screening at both stages was performed independently by two reviewers, with disagreements resolved by consensus or third-reviewer arbitration.
Studies were included if they were published in English and met the following criteria: they reported on subjects with AD; the study design was a randomised controlled trial (RCT), non-RCT, prospective observational study, retrospective cohort study, nested case-control study, systematic review or meta-analysis; mortality outcomes were reported for alcohol-dependent subjects; mortality in alcohol-dependent subjects was compared to mortality in an appropriate comparator population (including the general population, subjects without AUDs, subjects qualifying for a diagnosis of alcohol abuse, or alcohol-dependent subjects with differing levels of alcohol consumption); and a measure of association (hazard ratio [HR], odds ratio [OR], RR, standardised mortality ratio [SMR]) with 95% confidence intervals (CIs), or sufficient data to calculate these, was reported.
The "general population" comparator subgroup represented an unselected population of individuals in terms of drinking behaviour. This control group could therefore include a mixture of alcohol-dependent subjects, subjects qualifying for a diagnosis of alcohol abuse, abstinent subjects or individuals with any other level of pathological or nonpathological drinking. On the other hand, "subjects without AUDs" could be defined in a study as "non-alcoholics", "subjects without AD or alcohol abuse diagnosis" or any similar definition.
Studies involving alcohol-dependent subjects were included irrespective of whether a formal definition of AD (e.g. ICD or DSM) had been used to identify them. For studies involving "alcoholics", the definition of alcoholism was reviewed to determine whether it was operationally similar to a diagnosis of AD (included) or alcohol abuse (excluded).
The reference lists of all included full texts were scanned for further potentially relevant studies. These studies then underwent full text review using the same criteria as studies identified in the database searches.
The study design, methodology, patient population parameters and outcomes for all studies included in the SLR were extracted into a prespecified grid. Data extraction was performed by a single individual with independent verification by a second reviewer, with disagreements resolved by consensus or third-reviewer arbitration. It was planned that the quality and potential for bias of included RCTs would be assessed using the criteria provided by the York Centre for Reviews and Dissemination (Centre for Reviews and Dissemination, 2009) and the quality of non-RCTs would be assessed using the TREND checklist (Des Jarlais et al., 2004). The quality of observational studies was assessed using a checklist designed by the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) Good Research Practices Taskforce, which includes domains for relevance and credibility. Credibility questions related to study design and data analysis, among others (Berger et al., 2014). Quality assessments were conducted from the perspective of the populations and outcomes of interest to this review. All studies found to be relevant and credible were eligible for meta-analysis.

Meta-Analyses
Meta-analyses were conducted in accordance with MOOSE guidelines (Stroup et al., 2000). Results from the included studies were pooled for meta-analysis by comparator population. Given the methodological heterogeneity of studies identified in this SLR (e.g. differences between studies in mean age, source of the alcohol-dependent population, and reference groups) a random-effects model was judged to be appropriate for this meta-analysis (parallel analyses used fixed-effect models). HRs, ORs, RRs and SMRs were assumed to approximate the same measure of risk (Rothman and Greenland, 1998).
Included studies were pooled for meta-analyses based on measures of association being available for the following comparisons: alcoholdependent subjects vs the general population, subjects without AUDs, or alcohol abusing subjects; or alcohol-dependent subjects who continued to drink heavily vs alcohol-dependent subjects who reduced their alcohol intake (abstainers excluded), alcohol-dependent subjects who reduced their alcohol intake (abstainers included), or abstinent alcohol-dependent subjects.
Where the same patients were included in two or more studies, the study involving the greatest number of alcohol-dependent subjects was included in the meta-analysis.
To test the robustness of the findings, subgroup meta-analyses were performed within studies that compared alcohol-dependent subjects vs the general population, by a number of pre-specified study-and patient-level characteristics. This included a subgroup analysis by the definition of AD used within the study (strictly defined AD, such as DSM or ICD criteria vs another definition vs no reported definition). It was not possible to perform subgroup analyses for the other comparison groups due to the limited number of studies and insufficient power.
Random-effects were estimated using the approach of DerSimonian and Laird (1986), with the estimate of heterogeneity being taken from the Mantel-Haenszel model (Mantel and Haenszel, 1959). Summary RRs and 95% CIs were calculated by pooling the study-specific estimates. Stata® statistical software was used, in particular, the METAN command written for Stata® v11.1 (Anon, 2013). Consistency of findings across individual studies was assessed by standard χ 2 tests and the I 2 statistic (Higgins et al., 2003). Statistical tests were two-sided and used a significance level of p b 0.05. Publication bias was assessed by funnel plots and Egger's tests (Egger et al., 1997).

Role of the Funding Source
The study sponsor was involved in the design of the review and its protocol, the review of the data collection and statistical analysis, and the interpretation of the results. The study sponsor was not involved in the collection or extraction of data, or in the performance of statistical analyses. All authors were involved in the writing of the report and the decision to submit the manuscript for publication.

Results
Of the 4540 records identified through the database searches and hand searches, 177 were selected for full-text review and 47 (including two systematic reviews (Roerecke and Rehm, 2014; and 45 primary research studies) were ultimately included in the qualitative synthesis of results ( Fig. 1). Of the 45 primary research studies, three were unsuitable for inclusion in the quantitative synthesis. Of these, two studies duplicated data on subjects from other included studies (de Lint and Schmidt, 1970;Martin et al., 1985a). The third study was excluded since it compared alcohol-dependent subjects vs abstinent subjects without AUDs; no other studies included in the SLR reported an equivalent comparison, so a meta-analysis was not conducted (Dawson, 2000).
Following a thorough quality appraisal, results from three studies were judged to be insufficiently credible for meta-analysis due to a lack of information presented about key components of their protocols (Bell and Orjasaeter, 1983;Fitzgerald et al., 1971;Kessel and Grossman, 1961). The quality of the included studies varied substantially. Frequently-encountered limitations included the lack of a reported definition for AD and results presented without confidence intervals or without adjustment for confounders. Summaries of the credibility domain of the quality appraisals are presented in Supplementary  Table 3 for all studies included in the SLR. Thirty-nine studies included in the SLR were ultimately eligible for meta-analysis (Table 1).
Subgroup analyses within studies that compared alcohol-dependent subjects with the general population ( Supplementary Fig. 1) indicated evidence of heterogeneity by level of adjustment and length of followup (p-value from meta-regression on each covariate; p = 0.02 and p = 0.03, respectively); a lower level of adjustment or a shorter follow-up corresponded to a higher RR. Furthermore, there was a noticeable difference in mortality between alcohol-dependent subjects selected from general population surveys (RR = 1.76; 95% CI [1.48, 2.09]) and alcohol-dependent subjects selected from treatment centres (RR = 3.65; 95% CI [3.10, 4.30]) or hospitals (RR = 3.49; 95% CI [2.37, 5.12]). The definition of AD used in a study (DSM or ICD AD vs other definition vs no reported definition) did not have a significant effect on the RR (p = 0.86).
There was no evidence of publication bias across studies that reported on alcohol-dependent subjects vs the general population, alcoholdependent subjects vs subjects qualifying for a diagnosis of alcohol abuse, or alcohol-dependent subjects with continued heavy drinking vs alcohol-dependent subjects who reduced their drinking (p N 0.05 in Egger's asymmetry test). However, there was evidence of publication bias for analyses that compared alcohol-dependent subjects to subjects without AUDs (Egger's asymmetry test of associations: p = 0.01) ( Supplementary Fig. 2).

Discussion
In these meta-analyses, alcohol-dependent individuals were found to have a significantly increased risk of all-cause mortality compared to the general population. However, heterogeneity between studies was high. While in subgroup analyses the definition of AD reported in a study (strict AD vs any other definition vs no definition reported) did not affect the pooled RR estimates, differing levels of adjustment between studies (e.g. adjustment for age, sex and smoking vs unadjusted comparisons) and length of follow-up were identified as significant contributors to heterogeneity.
The general population includes individuals with AUDs, and would therefore be expected to have higher alcohol-attributable mortality than a cohort without AUDs. The relative risk of mortality for alcoholdependent subjects vs the general population should therefore be lower than the relative risk for alcohol-dependent subjects vs a cohort without AUDs. However, in this study the opposite was found; a higher risk estimate for studies comparing alcohol-dependent subjects to the general population than for studies comparing alcohol-dependent subjects to cohorts without AUDs. This may be explained by considering the characteristics of the alcohol-dependent population in each case: of the studies reporting on mortality in alcohol-dependent subjects vs the general population, the majority involved treated alcohol-dependent subjects. However, of the six studies reporting on mortality in alcoholdependent subjects vs subjects without AUDs, only one study involved treated alcohol-dependent subjects. Subgroup analyses of studies comparing alcohol-dependent subjects to the general population showed that the pooled risk estimate for studies involving treated alcohol-dependent subjects was higher than for studies involving alcohol-dependent subjects identified from general population surveys, potentially explaining the result above. This reasoning can be further strengthened by recent studies on characteristics of treated vs untreated people with AUDs: there was a substantial difference with respect to comorbid conditions (both for mental and somatic disorders), alcohol consumption (both average drinking levels and binge drinking episodes), hospital days and disability (Rehm et al., 2015a;Rehm et al., 2014).
Although only one out of four studies reporting on mortality in alcohol-dependent individuals compared to subjects qualifying for a diagnosis of alcohol abuse showed significantly increased mortality for alcohol-dependent subjects (Mattisson et al., 2011), the pooled RR was significant. Reduced alcohol consumption without abstinence was significantly associated with a lower risk of death in alcoholdependent subjects. Although this was a lower reduction in mortality than for abstinent alcohol-dependent subjects, the opportunity to reduce the risk of death combined with the preference among some patients to achieve a non-problem drinking outcome rather than abstinence supports the use of controlled drinking as a valid AD treatment goal (Adamson et al., 2010;Luquiens et al., 2011;Rosenberg et al., 1992). Achieving a reduction in alcohol consumption, including among alcohol-dependent subjects who may otherwise not have entered into treatment, would be predicted to ease the overall public health burden associated with AD (François et al., 2014), as well as reducing both direct and indirect costs (Rehm et al., 2012;Laramée et al., 2014).
The results of these meta-analyses are consistent with recently published meta-analyses on the risk of mortality in individuals with AUDs.    . This is broadly in agreement with the findings of our subgroup analyses, where the pooled RR for treated alcohol-dependent subjects was more than double the pooled RR for alcohol-dependent subjects from population surveys. Although this may seem counterintuitive, only a minority of individuals with substance dependence seek treatment for their addiction, and these individuals tend to have a greater severity of dependence than those who do not seek help (Grella and Stein, 2013). This may explain the increased mortality among treated alcohol-dependent subjects. Our observations that studies with shorter follow-up periods or with results adjusted for fewer confounders reported higher RRs were also consistent with previously published results .
In clinical practice, the distinction between alcohol abuse, AD and AUDs is not always clear. However, given the low treatment rates, and the comparisons of treated vs untreated individuals with AUDs (Rehm Fig. 2. Random-effects and fixed-effect meta-analyses of mortality risk in a) alcohol-dependent subjects vs the general population, b) alcohol-dependent subjects vs subjects without AUDs, and c) alcohol-dependent subjects vs subjects qualifying for a diagnosis of alcohol abuse. a HR, OR, RR or SMR (depending on study); b HR, OR or RR (depending on study); c College cohort; d Core city cohort. et al., 2015a; Rehm et al., 2014), it is reasonable to assume that the majority of people in specialised treatment qualify for a diagnosis of AD (Rehm et al., 2015b). The definition of AD was carefully considered in the eligibility criteria for this SLR, and the impact of including studies where formal definitions (such as ICD or DSM) had not been used to identify alcohol-dependent subjects was subsequently evaluated in sub-group analyses and found to not have a significant effect. These factors may explain why the results of this SLR on AD are similar to previously published SLRs on AUDs Harris and Barraclough, 1998). To our knowledge, there are no previous SLRs or meta-analyses reporting on mortality in alcohol-dependent individuals specifically rather than individuals with AUDs more broadly. Our study therefore supplements the existing literature on mortality in individuals with AUDs Roerecke and Rehm, 2014;, supporting the conclusions of previous studies while providing an important focus on the subgroup of subjects with AD.
One limitation of this study was the exclusion of non-English language articles; however, in our subgroup analysis of the studies comparing alcohol-dependent subjects vs the general populationhalf of which were from Europe, a third from North America and the remainder from other locationsthere were no significant differences between regions, implying that the inclusion of further studies from other countries or continents would not necessarily have significantly changed the conclusions of this review. Another limitation is that we did not exclude unadjusted studies from the meta-analyses. This may have affected some results, as unadjusted studies were found to show a significantly larger effect size in the comparison between alcohol-dependent subjects and the general population. However, we were unable to conduct subgroup Fig. 3. Random-effects and fixed-effect meta-analyses of mortality risk in alcohol-dependent subjects with continued heavy drinking vs a) alcohol-dependent subjects who reduced their alcohol intake (excluding abstainers), b) alcohol-dependent subjects who reduced their alcohol intake (including abstainers), and c) abstinent alcohol-dependent subjects. a HR, OR or RR (depending on study). analyses by adjustment level for the other meta-analyses due to the limited number of studies and insufficient power. Therefore, it is not possible to determine the direction of any potential effects in these cases. Finally, it is possible that the risk of mortality in alcoholdependent subjects may vary by how long an individual has been alcohol-dependent for; however, we were unable to investigate this, as this information was not commonly reported in the included studies.
In our meta-analyses, we have established that alcohol-dependent individuals who are able to reduce their alcohol consumption reduce their risk of all-cause mortality, but without further insight into the causes of death among these individuals we can only speculate about the mechanism by which this reduction in mortality occurs. It is not clear whether a reduction in alcohol consumption among these individuals would lead to a proportional reduction in risk across all causes of death, or whether the benefits would be greater among specific causes. Furthermore, though a meta-analysis has been published on causespecific mortality in individuals with AUDs (Roerecke and Rehm, 2014), none has been published concerning cause-specific mortality in individuals with AD specifically. We therefore recommend that further research should focus on the risk of cause-specific mortality in alcohol-dependent individuals, and particularly in individuals who are able to reduce their alcohol consumption.

Conclusions
We present the finding that AD significantly increases an individual's risk of all-cause mortality compared to the general population, subjects without AUDs or subjects qualifying for a diagnosis of alcohol abuse. While abstinence in alcohol-dependent subjects leads to a greater reduction in mortality than non-abstinence, alcohol-dependent subjects who are able to reduce their alcohol consumption can significantly reduce their risk of death. This finding reinforces the validity of harm reduction and the associated societal benefit through reduced alcohol consumption as a valid treatment goal alongside abstinence in the treatment of AD.

Author Contributions
Study design and protocol development: All authors. Collection and extraction of data: Leonard, Buchanan-Hughes. Statistical analysis: Warnakula. Review of the data collection and of the statistical analysis, technical support and supervision: Laramée, Leonard, Rehm. Administrative or material support: Laramée, Leonard, Buchanan-Hughes.
Interpretation of data and drafting of the manuscript: All authors. Critical revision of the manuscript for important intellectual content: All authors.