In February 2010 we carried out systematic searches of the literature in the following bibliographical databases: MEDLINE, PsycINFO (1997 to present), Science Citation Index Expanded, Social Sciences Citation Index, Arts and Humanities Citation Index (1997 to present), CINAHL, EMBASE, the Cochrane Drug and Alcohol Group Specialized Register, the Cochrane Effective Practice and Organization of Care (EPOC) Group register, the Alcohol and Alcohol Problems Science Database, and ETOH (etoh.niaaa.nih.gov, 1972 to 2003). Searches were conducted with keywords and text words, in which words indicative of eHealth interventions (Internet, Web, online, manual, and computer) were combined with terms indicative of alcohol disorders (alcohol abuse, alcoholism, problem drinking, hazardous drinking, harmful drinking, abstinence, moderation, treatment, brief intervention, self-help, and e-self-help) and our target group (adults). Those search strategies were combined with the optimal search strategy for randomized controlled trials (RCTs) designed by the United Kingdom Cochrane Centre (Cochrane Collaboration 2008). We also scanned Dissertation Abstracts and Digital Dissertations to cross-check references relating to earlier meta-analyses and systematic reviews on eHealth interventions, brief interventions, and e-self-help interventions for problem drinking as well as unpublished literature. Reference lists of retrieved papers were screened, and papers that possibly met inclusion criteria were retrieved and studied (Figure 1). No language restrictions were applied.

In this meta-analysis, we included only those studies that examined e-self-help interventions for adult problem drinkers (aged 18 or older). Studies on e-self-help interventions targeting student populations in college and university settings were excluded, as their effectiveness has been reported in other reviews and meta-analyses [22,26]. From studies that examined both problem and nonproblem drinkers (the latter not exceeding the guidelines for low-risk drinking), we entered the results for the problem drinkers only. We included only those randomized controlled trials that (1) compared e-self-help intervention groups with control groups, (2) reported data that were usable for meta-analytic procedures, and (3) assessed alcohol-drinking behavior (eg, frequency or quantity) as their primary outcome measure (Table 1).

Our initial selection from the first search was based on information derived from titles, abstracts, and keywords; if these yielded insufficient information to assess the inclusion criteria, then the full paper was retrieved. All papers excluded at this stage were independently assessed on their inclusion criteria by two independent raters (authors BB and HR) to ensure an error-free selection procedure. In addition, two independent coders, authors VS and HR, assessed the effect sizes and moderator variables of the included studies; any disagreement was resolved by discussion and consensus (Figure 1).

Some 25 scales are available to assess the validity and quality of RCTs [42]. As there is no evidence that more elaborate scales give more reliable assessments of validity than simpler ones, we chose an approach similar to that suggested by Higgins and Green [42] as well as approaches applied in two reviews of brief interventions for problem drinking in primary care [6,43]. This resulted in four basic criteria that we used to assess the validity and methodological quality of the studies we analyzed: (1) random allocation to condition performed by an independent third party, (2) random allocation concealed to study participants, (3) randomization status concealed to assessors of outcomes, and (4) completeness of follow-up data.

We first examined the effects of e-self-help interventions for problematic alcohol use in comparison with control conditions. The usual approach is to calculate the standardized mean difference, also known as Cohen’s d (the mean difference divided by the pooled standard deviation) that is, d = (m₁ – m₀) / SD, where m₁ and m₀ are the mean scores in the experimental and control conditions.

The pooled standard deviation (SD) is defined as SD = √ ([n₁–1]S₁ ² + [n₀–1]S₀ ²) / (n₁ + n₀ –2) where n₁ and n₀, and S₁ and S₀ are the sample sizes and variances of the experimental and control groups as obtained from the primary study. As this effect size d is subject to small-sample bias, it can be adjusted by using a scaling factor, which is multiplied by d to arrive at Hedges’ bias-corrected effect size g, where g = d (1– [3/4(n₁ + n₀] –9).

The interpretation of g is simple. An effect size of 0.5 indicates that the mean of the experimental group is half a standard deviation larger than the mean of the control group. In second-order meta-analysis, effect sizes of 0.56 to 1.20 may be interpreted as large, those from 0.33 to 0.55 as moderate, and those from 0 to 0.32 as small [44].

Our effect size calculations were performed on alcohol consumption measures only. If means and standard deviations were not reported, we used other statistics (F values and P values) to calculate effect sizes. If more than one measure was used in a single primary study, then the mean of the effect sizes was calculated so that each study was represented with only one effect size in our meta-analysis. In one study [45], two experimental conditions were compared with a control condition. In this case, the number of respondents in the control condition was divided equally over the two experimental conditions so that each respondent figured only once in the meta-analysis. In calculating pooled mean effect sizes, we used Comprehensive Meta-analysis (CMA) software, version 2.2.021 [46].

As we did not know before the analysis whether to expect heterogeneity among the studies, we used both the fixed effects model (FEM) and the random effects model (REM) to calculate pooled effect sizes. Heterogeneity was evaluated with the Q statistic and the I² statistic. A significant Q rejects the null hypothesis of homogeneity, indicating that the variability among the effect sizes is greater than what would likely result from sampling error alone [47]. The I² statistic describes the percentage of total variation across studies, that is, attributable to systematic heterogeneity rather than chance alone. An I² value of 25% is associated with low heterogeneity, 50% with moderate, and 75% with high heterogeneity [48].

In subgroup analyses, we tested for significant differences between the effect sizes in different categories of studies using mixed effects analyses in CMA. We analyzed the following attributes: (1) type of treatment (single session personalized normative feedback versus extended self-help); (2) venue where intervention and assessments took place (research, health care, or workplace setting versus participants’ homes); (3) type of analysis (intention-to-treat versus completers only); (4) type of control condition (information, assessment-only, waiting list); and (5) small sample sizes (n < 100) versus large sample sizes (n > 100). Publication bias was tested by funnel plot and by Duval and Tweedie’s trim-and-fill procedure, which yields an estimate of effect size after publication bias has been taken into account (both procedures implemented in CMA).

A total of 492 studies were retrieved. Of these, 483 did not meet the inclusion criteria and were excluded (see Figure 1). A total of 9 trials with 9 comparisons and 1553 participants were ultimately included. Selected characteristics of those studies are summarized in Table 1. All 9 studies had been conducted in high-income countries: the United States (3), Canada (1), Netherlands (4), and Germany (1). Of these studies, 5 involved either single-session personalized normative feedback interventions while 4 involved more extended self-help interventions (see Table 1).

All studies used well-validated alcohol consumption measures and well-described, theoretically based interventions [56]. In 4 studies [49,53,50,57], treatment was allocated by an independent third party; in 2 studies [31,45] it was not, and in 3 studies [51,52,54] this was unclear. Concealment of random allocation to participants was not applicable as study participants were recruited by self-reference and informed about the study conditions. In all but 3 studies [51,52,54] all outcome measures were self-administered by participants, making it irrelevant whether the researchers assessing the outcomes were blinded to group assignment. Dropout rates differed widely from 0% to 42% (Table 1).

We then performed subgroup analyses to examine the contrast between extended e-self-help interventions and e-single-session personalized normative feedback interventions (omitting the outliers). Based upon the mixed effects model, these yielded a mean effect size of g = 0.27 (95% CI 0.11-0.43) for the e-single-session interventions and g = 0.61 (95% CI 0.33-0.90) for the extended e-self help interventions, a significant difference of P = .04 (Table 2). We found no significant differences between (1) where interventions and assessments were performed (research, health centre, or workplace versus places from where participants’ accessed the Internet, such as at home); (2) type of control condition (information, assessment-only, waiting-list); (3) small samples (n < 100) versus large samples (n > 100); nor (4) type of analysis (completers-only versus intention-to-treat).

The overall mean effect size was maintained even after exclusion of the largest study (N = 450) [50] in a random effects model (g = 0.44, 95% CI 0.24-0.64, Q = 6.87, I² = 27.23, P = .23). Duval and Tweedie’s trim-and-fill analysis did not detect publication bias (observed at g = 0.38, 95% CI 0.25-0.51, adjusted g = 0.35, 95% CI 0.23-0.47), nor did the funnel plot analysis detect bias. In view of these findings, we assume our post-treatment results to be robust up to 9 months [59].

Because our analysis is based on a rather limited number of studies, the results can be generalized only to self-referred adult problem drinkers in high-income countries recruited via the media; this implies samples of individuals with high readiness to change [64]. Second, some studies had small samples; we dealt with this by analyzing our data with Hedges’ g to adjust our estimates for small-sample bias [65]. Third, the loss to follow-up in some of the studies we reviewed was substantial. High dropout rates are a common feature of both online and offline self-help interventions for problem drinking and for Internet interventions in general [66,67]. Although some studies applied imputation techniques to handle the loss to follow-up, the high attrition may have biased our overall results.

The medium effect size we found for no-contact eHealth interventions could imply a major health impact at the population level, in view of the high percentages of problem drinkers that eHealth interventions might potentially reach. Naturally, eHealth interventions are subject to some constraints, as participants need computer and Internet access and a reasonable degree of literacy. While the costs of disseminating and scaling up no-contact eHealth interventions are low, the costs of developing them can be substantial [68].

Many questions still remain unanswered. We do not know yet whether e-self-help interventions that include professional contacts are more effective against problem drinking than no-contact interventions and, if so, to what types of problem drinkers that might apply (first-time help-seekers in the general population, primary- or secondary-care populations, hazardous or harmful drinkers, or dependent drinkers. (See Textbox 1.) This contrasts with e-self-help intervention studies focusing on common mental health disorders like depression and anxiety [69], which have established a firm evidence base for better clinical outcomes when e-self-help interventions are delivered with some form of professional guidance [18]. By contrast, Doumas and colleagues [45] have observed that a no-contact e-self-help intervention combined with 15 minutes of face-to-face motivational interviewing produced reductions in drinking that were comparable with the results of the no-contact e-self-help intervention alone. Nor could Rooke et al [15] establish any effect moderation attributable to personal guidance in alcohol e-interventions. As most participants in e-self-help interventions for problem drinking are first-time help-seekers [20,21], it could well be the case for this group that timely, anonymous access to e-interventions has a greater impact in altering drinking patterns than professional contact per se [70]. No-contact e-self-help interventions might therefore be an effective first-line choice in a stepped-care approach to problem drinking. A recently published study [71] on the posttreatment (3 months) effectiveness of intensive online treatment with active involvement of a therapist (duration of 3 months with 2 online therapist contacts a week) revealed an effect size of d = 1.21. This may indicate that therapist involvement will lead to higher effect sizes when compared with no-contact self-help interventions. More studies are needed, however, to assess the robustness of this observation and to assess how these more intensive treatments fit into a stepped-care approach.

From an economic point of view, no-contact e-self-help interventions could carry considerable promise as compared to other approaches like screening and brief face-to-face interventions in primary care, especially since the latter have relatively high implementation costs [72]. Studies that rigorously assess this proposition are not yet available [20], but economic evaluation studies on e-self-help for depression show favorable results as shown by a recent study of Gerhards and colleagues [73] in primary care and by Warmerdam and colleagues among depressed persons in the community [74]. The paper of Smit and colleagues in this issue shows that such economic advantages may also be expected from e-self-help intervention when implemented on a broad scale [75].

In light of all the potential benefits of e-self-help interventions in curbing problem drinking, we recommend their further evaluation and implementation as a means to bridging the gap in treating alcohol use disorders in low- and middle-income countries (LMICs). Although some may argue that this is not a viable option given the low Internet penetration in such countries, we would point out that such arguments abounded in high-income countries at the start of the millennium when e-self-help interventions for hazardous drinking were first being introduced. Meanwhile, the digital divide in Internet access between the more and less affluent countries is narrowing. In fact, one in four people worldwide now have Internet access [76]. Some e-self-help interventions have also proven well suited to be shared globally, as exemplified by the work of Munoz [77] on no-contact e-self-help for smoking cessation that are now used by people all over the world.

Our study has sought to synthesize the available evidence about the effectiveness of no-contact e-self-help interventions in curbing adult problem drinking. The data suggest that these are effective. Our findings also highlight the need for more evaluations of the clinical outcomes and the cost-effectiveness of online screening instruments and interventions. The paper of Smit and colleagues in this issue indeed illustrates the economic benefits of e-self-help interventions for curbing adult problem drinking [75]. Future studies should shed light on whether e-self-help interventions produce similar or better results when extended with face-to-face components and on whether they could serve as alternatives or adjuncts to face-to-face treatments in primary care settings.