Effectiveness and treatment moderators of internet interventions for adult problem drinking: An individual patient data meta-analysis of 19 randomised controlled trials

Background Face-to-face brief interventions for problem drinking are effective, but they have found limited implementation in routine care and the community. Internet-based interventions could overcome this treatment gap. We investigated effectiveness and moderators of treatment outcomes in internet-based interventions for adult problem drinking (iAIs). Methods and findings Systematic searches were performed in medical and psychological databases to 31 December 2016. A one-stage individual patient data meta-analysis (IPDMA) was conducted with a linear mixed model complete-case approach, using baseline and first follow-up data. The primary outcome measure was mean weekly alcohol consumption in standard units (SUs, 10 grams of ethanol). Secondary outcome was treatment response (TR), defined as less than 14/21 SUs for women/men weekly. Putative participant, intervention, and study moderators were included. Robustness was verified in three sensitivity analyses: a two-stage IPDMA, a one-stage IPDMA using multiple imputation, and a missing-not-at-random (MNAR) analysis. We obtained baseline data for 14,198 adult participants (19 randomised controlled trials [RCTs], mean age 40.7 [SD = 13.2], 47.6% women). Their baseline mean weekly alcohol consumption was 38.1 SUs (SD = 26.9). Most were regular problem drinkers (80.1%, SUs 44.7, SD = 26.4) and 19.9% (SUs 11.9, SD = 4.1) were binge-only drinkers. About one third were heavy drinkers, meaning that women/men consumed, respectively, more than 35/50 SUs of alcohol at baseline (34.2%, SUs 65.9, SD = 27.1). Post-intervention data were available for 8,095 participants. Compared with controls, iAI participants showed a greater mean weekly decrease at follow-up of 5.02 SUs (95% CI −7.57 to −2.48, p < 0.001) and a higher rate of TR (odds ratio [OR] 2.20, 95% CI 1.63–2.95, p < 0.001, number needed to treat [NNT] = 4.15, 95% CI 3.06–6.62). Persons above age 55 showed higher TR than their younger counterparts (OR = 1.66, 95% CI 1.21–2.27, p = 0.002). Drinking profiles were not significantly associated with treatment outcomes. Human-supported interventions were superior to fully automated ones on both outcome measures (comparative reduction: −6.78 SUs, 95% CI −12.11 to −1.45, p = 0.013; TR: OR = 2.23, 95% CI 1.22–4.08, p = 0.009). Participants treated in iAIs based on personalised normative feedback (PNF) alone were significantly less likely to sustain low-risk drinking at follow-up than those in iAIs based on integrated therapeutic principles (OR = 0.52, 95% CI 0.29–0.93, p = 0.029). The use of waitlist control in RCTs was associated with significantly better treatment outcomes than the use of other types of control (comparative reduction: −9.27 SUs, 95% CI −13.97 to −4.57, p < 0.001; TR: OR = 3.74, 95% CI 2.13–6.53, p < 0.001). The overall quality of the RCTs was high; a major limitation included high study dropout (43%). Sensitivity analyses confirmed the robustness of our primary analyses. Conclusion To our knowledge, this is the first IPDMA on internet-based interventions that has shown them to be effective in curbing various patterns of adult problem drinking in both community and healthcare settings. Waitlist control may be conducive to inflation of treatment outcomes.


Methods and findings
Systematic searches were performed in medical and psychological databases to 31 December 2016.A one-stage individual patient data meta-analysis (IPDMA) was conducted with a linear mixed model complete-case approach, using baseline and first follow-up data.The primary outcome measure was mean weekly alcohol consumption in standard units (SUs, 10 grams of ethanol).Secondary outcome was treatment response (TR), defined as less than 14/21 SUs for women/men weekly.Putative participant, intervention, and study moderators were included.Robustness was verified in three sensitivity analyses: a two-stage IPDMA, a one-stage IPDMA using multiple imputation, and a missing-not-at-random (MNAR) analysis.We obtained baseline data for 14,198 adult participants (19 randomised controlled trials [RCTs], mean age 40.7 [SD = 13.2],47.6% women).Their baseline mean weekly alcohol consumption was 38.1 SUs (SD = 26.9).Most were regular problem drinkers (80.1%,SUs 44.7, SD = 26.4) and 19.9% (SUs 11.9, SD = 4.1) were binge-only drinkers.About one third were heavy drinkers, meaning that women/men consumed, respectively, more than 35/50 SUs of alcohol at baseline (34.2%, SUs 65.9, SD = 27.1).Post-intervention data were available for 8,095 participants.Compared with controls, iAI participants showed a greater mean weekly decrease at follow-up of 5.02 SUs (95% CI −7.57 to −2.48, p < 0.001) and a higher rate of TR (odds ratio [OR] 2.20, 95% CI 1.63-2.95,p < 0.001, number needed to treat [NNT] = 4.15, 95% CI 3.06-6.62).Persons above age 55 showed higher TR than their younger counterparts (OR = 1.66, 95% CI 1.21-2.27,p = 0.002).Drinking profiles were not significantly associated with treatment outcomes.Human-supported interventions were superior to fully automated ones on both outcome measures (comparative reduction: −6.78 SUs, 95% CI −12.11 to −1.45, p = 0.013; TR: OR = 2.23, 95% CI 1.22-4.08,p = 0.009).Participants treated in iAIs based on personalised normative feedback (PNF) alone were significantly less likely to sustain low-risk drinking at follow-up than those in iAIs based on integrated therapeutic principles (OR = 0.52, 95% CI 0.29-0.93,p = 0.029).The use of waitlist control in RCTs was associated with significantly better treatment outcomes than the use of other types of control (comparative reduction: −9.27 SUs, 95% CI −13.97 to −4.57, p < 0.001; TR: OR = 3.74, 95% CI 2.13-6.53,p < 0.001).The overall quality of the RCTs was high; a major limitation included high study dropout (43%).Sensitivity analyses confirmed the robustness of our primary analyses.

Conclusion
To our knowledge, this is the first IPDMA on internet-based interventions that has shown them to be effective in curbing various patterns of adult problem drinking in both community and healthcare settings.Waitlist control may be conducive to inflation of treatment outcomes.consultant to/on the scientific advisory boards of Sanofi, Novartis, Minddistrict, Lantern, Schoen Kliniken, and German health insurance companies (BARMER, Techniker Krankenkasse).He is also a stakeholder of the Institute for health training online (GET.ON), which aims to implement scientific findings related to digital health interventions into routine care.RH's company owns the IP of the web applications that their data came from, as a basis for their contribution to the overall data set.HDV is the scientific director of Vision2Health, a company with the goal to implement evidence-based eHealth interventions.HR received a fee for participating in a brainstorm session on ehealth and bipolar disorders/PTSD organized by Sandoz.The remaining authors declare no competing interests.

Introduction
Global estimations continue to show increasing physical and psychological morbidity, allcause and specific-cause mortality, and social harm deriving from all types of alcohol misuse.Usually, a positive and linear association is seen between increased consumption and related health risks [1].A number of factors underlie this mounting health burden.These include increases in the prevalence of alcohol consumers due to population growth and societal ageing, an absolute increase in adult alcohol consumption due to greater wealth and wider acceptance of alcohol use, and escalating alcohol use amongst women and the elderly.At the same time, there are growing insights into health risks connected with even minimal levels of alcohol consumption [2,3].
Brief alcohol interventions (BAIs) in primary care and community settings have been found clinically and cost-effective, with effect sizes in the small to moderate range, for reducing both hazardous drinking (which increases the risk of physical or psychological harm) and harmful drinking (which has already caused some damage) [4].Together, their target groups are referred to as 'problem drinkers' to distinguish them from drinkers with alcohol use disorders, for whom more intensive treatments are recommended [5].Problem drinkers account for the highest prevalence of alcohol misuse.Based on accumulated evidence, many national and professional guidelines now recommend brief interventions for problem drinkers in primary care settings and among community populations [6].These interventions are comprised mostly of brief single or multiple sessions (up to six) and are based on personalised normative feedback (PNF) [7] or combinations of PNF, motivational interviewing (MI) [8], cognitivebehavioural therapy (CBT) [9], or behavioural self-control (BSC) principles [10].Despite the ample evidence available, the actual impact of BAIs on curbing the prevalence of problem drinking in the wider population has been disappointingly low.The main factors in the weak impact include problems with implementation, as relatively few healthcare professionals actually administer BAIs; in addition, only a small proportion of patients who might benefit are actually offered BAIs, and even fewer accept the offer [11].
Internet-based alcohol interventions (iAIs) may overcome some of these problems by virtue of their low-threshold accessibility, their high scalability, and their acceptability to problem drinkers, as was recently echoed by McCambridge and Saitz [11].Major advantages of iAIs, as perceived by many problem drinkers, are reduced stigma and greater comfort about disclosing drinking problems.The majority of iAIs are based on manualised therapeutic principles similar to those in BAIs.They are offered in unguided and guided formats.Unguided iAIs are fully automated interventions that participants can perform without human guidance.Guided interventions provide human support to guide participants through the intervention, mainly via asynchronous secure email contact [12].The support may come from health professionals or trained volunteers.Meta-analytic studies have shown that unguided iAIs, in particular, are now used on a wider scale than conventional BAIs [13].They have been found clinically effective (small effects) in reducing mean weekly adult alcohol consumption as compared with controls [14].As a result, iAIs have been incorporated into some clinical guidelines for treating problem drinking in primary care [15].
All this notwithstanding, various uncertainties still surround the evidence base for iAIs.First of all, still little is known about whether women and older people derive benefits comparable to those seen for male and younger problem drinkers.Such knowledge is important in view of the rising prevalence rates of problem drinking among women and the elderly and their underrepresentation in many intervention studies [16].Secondly, problem drinking actually embraces several different drinking profiles, and only a few iAI studies have investigated whether these might moderate treatment outcomes [17].Such profiles include exceeding the advised weekly alcohol limits to a moderate ('regular drinking') or a serious degree ('heavy drinking') and 'binge-only drinking', whereby alcohol users episodically exceed the maximum advised intakes per drinking occasion.Such divergent drinking profiles may or may not necessitate different interventions.Thirdly, there is the question of whether guided iAIs are more effective than unguided ones-a finding reported for CBT-based internet interventions for common mental disorders such as depression [18].A related question is whether iAI treatment outcomes might vary according to the therapeutic orientation of the intervention.
The few moderator analyses conducted to date had a common limitation: they were statistically underpowered to properly address such questions [14].To overcome this major problem, we conducted an individual patient data meta-analysis (IPDMA) that boosted the number of participants studied and thereby the statistical power.That enabled us to better evaluate the overall effectiveness of iAIs in reducing alcohol consumption, as well as to explore statistically significant differences within the data by performing moderator analyses on treatment outcomes, with a focus on participant, intervention, and study design characteristics.

Identification and selection of randomised controlled trials
PsycINFO, Science Citation Index Expanded, Social Sciences Citation Index, Arts and Humanities Citation Index, CINAHL, PubMed, and EMBASE were searched up to 31 December 2016.All papers retrieved were evaluated by independent assessors (HR, EK, or NBo) (for search string, see S1 Data).

Eligibility criteria
Randomised controlled trials (RCTs) were eligible if they (1) studied people aged �18 with quantifiable levels of alcohol consumption that exceeded recommendations for low-risk drinking; (2) compared an iAI with a control condition (e.g., assessment only, waitlist, or minimal intervention); (3) studied an iAI based on therapeutic principles such as PNF, BSC, CBT, MI, or combinations thereof; and (4) studied either an unguided or a guided intervention or both.RCTs in populations of students or pregnant women were excluded.Primary authors of identified trials were asked to provide their raw RCT data for a set of pre-identified variables (HR/ EK, S2 Data and see S4 Data for data access contact list of original studies) and were queried as to whether they were aware of ongoing RCTs that met our inclusion criteria; two more RCTs were thus identified [19,20].No study protocol for this study has been developed.

Risk-of-bias assessment and data extraction
Five criteria from the Cochrane Collaboration risk-of-bias assessment tool were applied (by EK, HR, and NBo): (1) adequate random sequence allocation, (2) concealment of allocation to the different conditions, (3) blinding of participants and therapists to the study condition, (4) blinding of assessors to outcomes, and (5) handling of missing data [21].

IPDMA
Primary outcome measure.The primary outcome was mean weekly alcohol consumption, expressed in standard units.As RCTs differ in the quantification of alcohol in beverages, based on national custom (ranging from 8 to 14 grams of ethanol per unit [22]), we recalculated these into standard units of alcohol consumption based on 10 grams of ethanol (SUs).Most RCTs measured alcohol consumption using time line follow-back (TLFB) approaches.For a few RCTs that did not report TLFB data, we estimated mean weekly SUs on the basis of the first two questions of The Alcohol Use Disorders Identification Test (brief, 3 items; AUDIT-C) scale [23] at post-intervention.Alcohol consumption at baseline was constructed identically.Most included participants were regular drinkers who were consuming more than the recommended low-risk weekly limits of 14 SUs (females) or 21 SUs (males) at baseline.Binge-only drinking is another problem drinking profile in which low-risk recommendations are exceeded.We defined binge-only drinkers by proxy as participants who drank more than 4 or 6 SUs (females/males) on at least one occasion per week, while still totalling less than 14/21 SUs weekly.
Secondary outcome measure.Treatment response (TR) was defined as an alcohol consumption level below 14/21 SUs per week for females/males at the first post-intervention follow-up.
Moderators.The following participant-level putative moderators were tested: gender (female/male); age (below 55/above 55); education (high/low, with 'high' referring to tertiary education and 'low' to primary or secondary schooling), employment (yes/no), and partner relationship (yes/no).Two dimensions of problem drinking were explored: regular drinking (>14/21 SUs female/male weekly) as contrasted with binge-only drinking (�4/6 SUs female/ male at least once a week but below 14/21 female/male SUs weekly); and heavy drinking (>35/ 50 SUs female/male weekly) as contrasted with non-heavy drinking (14-35 SUs weekly in females and 21-50 SUs weekly in males).Intervention-level putative moderators were therapeutic guidance (human-guided versus unguided interventions), intensity (single versus multiple sessions), therapeutic orientation (PNF-only versus integrated therapeutic principles), and intervention setting (in work, healthcare, or community populations).A study design moderator, type of control, was also included (waitlist control contrasted with assessment-only or minimal-intervention control).
One-stage and two-stage IPDMAs.Replications of individual study outcomes based on the raw data in comparison with the published results led to only one correction to the published tables [24].We next applied a one-stage individual patient data (IPD) model of analysis, as it is assumed to produce a more exact likelihood specification than a two-stage approach [25].In a one-stage IPDMA, the effect of iAIs is evaluated by fitting a single comprehensive model to the IPD from all trials, while simultaneously accounting for the nesting of participants within these trials.To account for the nesting structure, we assessed the summary effect of iAIs on the primary outcome using a linear mixed model (LMM).At the participant level, we used an ANCOVA model [26], regressing the post-intervention outcome score on the iAI intervention indicator, with the baseline alcohol consumption score used as a covariate.
To deal with missing baseline alcohol data, we used mean imputation to estimate scores [27].We subsequently analysed all available outcomes using complete cases-that is, including the full baseline outcomes (N = 14,198) but ignoring missing post-intervention outcomes.This analysis implicitly assumes that the missing data are missing at random (MAR) rather than missing completely at random (MCAR), allowing missingness of post-intervention scores to depend on the pre-intervention score.
To evaluate the effect of iAIs using an LMM, we regressed the post-intervention weekly SU level on the iAI intervention indicator, the baseline weekly SU level, and the comparison indicators (dummy variables contrasting the intervention arms with the control arms of the trials), assuming random effects (both intercepts and intervention slopes) for those comparisons and equal residual variances across trials.The estimates of the iAIs' effects are presented as unstandardised regression coefficients (b), which refer to the overall effect of the intervention on posttreatment drinking behaviour in terms of comparative SU levels.
For TR, a generalised LMM with participants nested within trials (a logistic model) was similarly used.TR at follow-up (yes/no) was the dichotomous dependent variable, and all fixed and random effects were identical to those in the LMM for the continuous primary outcome, except that fixed intercepts were removed for reasons of identification (convergence), resulting in a model with random intercepts (and slopes).We calculated odds ratios (ORs), representing the probability that an outcome will occur given a particular exposure as compared with the probability in the absence of that exposure [28].TR was additionally interpreted by transforming the OR to a number needed to treat (NNT) [29].
We subsequently tested whether participant, intervention, or study design characteristics moderated the effect of iAIs on either the primary or the secondary outcome or both.However, participant-level characteristics and study-and intervention-level characteristics were analysed differently.For participant-level characteristics, within-study and across-study interaction effects had to be separated to avoid ecological bias [25], whilst no such separation was needed for study-and intervention-level moderators.We additionally performed two-stage analyses to evaluate the sensitivity of the one-stage results, a recommendation by Burke and colleagues [25].The two-stage approach in our study derived aggregate data for effect estimates and their CIs for each study individually (step one), then combined these in a conventional meta-analysis model (step two).In two-stage analyses, participant-level moderators are estimated for each study separately and combined in the second stage, without risk of ecological bias, while intervention-and study-level characteristics are studied by comparing subgroups of trials in the second stage.
In the one-stage approach, we added a second sensitivity analysis to compare our results against those of a procedure applying multiple imputation to include all participants (an intention-to-treat [ITT] analysis).This analysis was conducted on the request of one of the reviewers.The multiple imputation procedure used chained equations to impute missing alcohol consumption scores-both before and after the intervention-together with missing values of the participant-level putative moderators.The ITT analysis employed logistic regression models for the dichotomous variables and predicted-mean matching for the continuous variables, with study indicators and intervention indicators (fully interacted) included as covariates.This second sensitivity analysis tested the main intervention effect and the moderator effects of all participant-level and study-level moderators on the primary outcome variable.
In the two-stage approach, we employed a third sensitivity analysis that evaluated the MAR assumption on the missing data mechanism, thereby answering an additional question of one of the reviewers.This additional, missing-not-at-random (MNAR) analysis is part of the ITT strategy suggested by White and colleagues [30]; it includes all randomised individuals in the analysis, taking baseline outcomes of dropouts into account.It evaluates a series of values of a sensitivity parameter δ, which equals the difference between the mean of the observed values of post-intervention SUs of alcohol and the mean of the unobserved values.Under MAR, δ (being the covariate-adjusted mean difference between missing and observed outcomes) is assumed to be zero.In our case, positive (or negative) values of δ correspond to the situation in which the dropouts, after adjustment for pre-intervention SUs, would have higher (or lower) mean values of post-intervention SUs than those who continued participation (see S1 Text for a detailed explanation of the evaluation of the MAR assumption).This third sensitivity analysis targeted the main intervention effect on the primary outcome variable only.
To examine heterogeneity, we calculated the I 2 statistic using the two-stage approach as well.This indicator is expressed as a percentage: an I 2 value of 0% is interpreted as no heterogeneity, 25% as low, 50% as moderate, and 75% as high heterogeneity [31].We calculated the 95% CIs around I 2 using the noncentral chi-squared-based approach within the heterogeneity module for Stata [32,33].All analyses were conducted with Stata 14.2.

Comparison of IPDMA-included with non-included RCTs
The potential differences in treatment outcomes between the trials included and those that could not be included in preparing our IPDMA were assessed with a conventional meta-analysis (Comprehensive Meta-Analysis, version 3.3.070;S3 Data).

Results
Results are reported in accordance with the Preferred Reporting Items for Systematic Review and Meta-Analyses for IPD (S1 PRISMA Checklist) [34].

Selection of RCTs
Fig 1 illustrates the selection process for the trials included in our IPDMA.We identified 183 full papers, from which 24 eligible RCTS were found, five of which [35][36][37][38][39] could not be included (all involving unguided iAIs) because authors did not respond to our invitation.

Study characteristics
Table 1 shows the characteristics of the 19 included RCTs (26 comparisons).Most trials applied the full Alcohol Use Disorders Identification Test (AUDIT) (n = 9, cutoff �8) [40] or AUDIT-C scales (n = 4, cutoff �4 or �5) [23] as inclusion criteria.Four RCTs used cutoff thresholds based on daily or weekly low-risk drinking recommendations; the Fast Alcohol Screening Test (FAST) was applied in two trials [41].Participants were recruited either directly

Participants' characteristics at baseline
Of the total of 14,198 enrolled participants, 8,095 provided post-intervention outcome data (complete cases, Table 2).The mean age of the overall sample was 40.7 (SD = 13.2) and the sample was rather evenly divided by gender (47.6% women, 52.4% men).Some 51.9% of participants had tertiary education, 74.8% had paid employment, and 56.7% were in partner relationships.The mean weekly SU level at baseline was 38.1 (SD = 26.9).Most problem drinkers (80.1%,SUs 44.7, SD = 26.4)could be categorised as regular drinkers and 19.9% (SUs 11.9, SD = 4.1) as binge-only drinkers.Regular drinkers could be distinguished into heavy drinkers (34.2%, SUs 65.9, SD = 27.1) and non-heavy drinkers (65.8%,SUs 23.7, SD = 10.6).Heavy drinkers were found in both unguided and guided iAIs (34% and 30%, respectively).The mean full AUDIT score (n = 9 trials) was 15.0 (SD = 6.8), indicating hazardous or harmful alcohol use [40].Of the participants for which a full AUDIT score was available, 22.2% (n = 678) scored above 20, indicating a risk of alcohol dependence.Missing SU scores at baseline were virtually nil (0.4%).Missing data at the first post-intervention assessment for the primary outcome were considerable (43%), predominantly resulting from study dropout, which was not entirely random: participants under age 55 and those with baseline heavy-drinking profiles dropped out significantly more than others.

Risk of bias
The quality of the RCTs was relatively high (Fig 2 and S1 Table ).All but one scored high-risk on the blinding of participants, which was expected, as this criterion is difficult to meet for

Moderator analyses
Participant characteristics.Both men and women treated in iAIs decreased their mean weekly SU levels to a greater degree than controls, but women did so less than men (2.19 SUs, 95% CI 0.52-3.85,p = 0.013).Additional sensitivity analyses maintained this difference.In the �� These variables were not assessed in all studies, therefore the numerator and denominator are listed here separately. 1 Regular drinking denotes 14 or more SUs weekly for females or 21 or more for males (thus excluding binge-only drinking).
2 Binge-only drinking denotes more than 4 or 6 SUs (females/males) on at least one occasion per week, while still totalling less than 14/21 SUs weekly.
Abbreviations: AUDIT, The Alcohol Use Disorders Identification Test; SU, standard unit of alcohol consumption based on 10 grams of ethanol.
Study design characteristics.Treatment participants in waitlist-controlled (WLC) trials significantly reduced their mean weekly alcohol consumption by greater amounts in comparison to controls than those treated in otherwise-controlled trials (−9.27SUs, 95% CI −13.97 to −4.57, p < 0.001).Only the guided iAIs in WLC designs differed significantly from those in otherwisecontrolled trials (b = −10.51SUs, 95% CI −18.38 to −2.64, p = 0.009).iAI intervention participants in WLC trials were also significantly more likely to show favourable TR (OR = 3.74, 95% CI 2.13-6.53,p < 0.001) than those in other trials; significant differences were maintained for both Other control = MIC or AOC. 1 Studies 36-39 were regarded as outlier studies.
2 These were studies [52,56]. 3Regular drinking denotes 14 or more SUs weekly for females or 21 or more for males (thus excluding binge-only drinking).Binge-only drinking denotes more than 4 or 6 SUs (females/males) on at least one occasion per week while still totalling less than 14/21 SUs weekly. 4Heavy drinking denotes 35 or more SUs weekly for females and 50 or more for males; non-heavy drinking denotes 14/21 SUs or more, but less than 35/50 SUs, weekly for females/males. 5The number of persons refers to respondents for whom data were available on post-intervention drinking behaviour (complete cases) and, for the participant level, also on the moderator in question. 6Unstandardised regression coefficients (b) indicate the effect of the iAIs in terms of alcohol reduction in SUs.

Sensitivity analyses
In the first sensitivity analysis, we checked the extent to which the results would be different if we used a two-stage approach instead of a one-stage approach.The second sensitivity analysis involved the inclusion of all participants according to the ITT principle by use of a multiple imputation strategy.
The third sensitivity analysis concerned the MAR assumption that is commonly used to deal with missing outcome data.All three of our sensitivity analyses confirmed the results of our main analysis for the overall effect and for most of the moderating effects of participant-, intervention-, and study-level characteristics for the primary and secondary outcomes.This appears to verify the robustness of our findings (see Tables 3 and 4 for the results of the twostage approach and S2 Table and S3 Table, in which the results of the multiple imputation analyses are presented).Some minimal differences for moderators were seen in the multiple imputation analyses.The moderating role of gender and education for the primary outcome lost significance after multiple imputation.For the secondary outcome, the moderating role of single versus multiple sessions became significantly different in favour of multiple sessions, while intervention in the work setting became effective (p = 0.041), as was the case for assessment-only interventions.The contrast between PNF versus integrated iAIs became nonsignificant, as was the case for the contrast between unguided iAIs with WLCs versus other types of control conditions.Thus, in some cases these made our moderator analysis appear more conservative, while in some other cases the MI was more conservative.
Fig 3 depicts the results of the third, MNAR sensitivity analysis, which assessed departure from the MAR assumption.The figure shows estimates (and 95% CIs) of the overall intervention effect on our primary outcome variable, SU, for differing values of δ.The value of δ = 0 corresponds to the MAR assumption, on which the results displayed in Table 3 are based.Positive (or negative) values of δ correspond to situations in which-in each study included in the IPDMA and in both the intervention and the control arms-the mean of unobserved scores for post-intervention SUs would be higher (or lower) than the observed post-intervention SUs, Other control = MIC or AOC. 1 Studies 36-39 were regarded as outlier studies.
2 These were studies [52,56]. 3Heavy drinking denotes 35 or more SUs weekly for females and 50 or more SUs for males; non-heavy drinking denotes 14/21 SUs or more, but less than 35/50 SUs, weekly for females/males. 4The number of persons refers to respondents within the subsample of baseline regular drinkers for whom data were available on post-intervention drinking behaviour (complete cases) and, for the participant level, also on the moderator in question.Binge-only drinkers are excluded in this after adjustment for pre-intervention SUs.If MAR holds, the overall effect is estimated in the two-stage method at −4.80 SU.Fig 3 shows that if the post-intervention SUs of dropouts, adjusted for the pre-intervention SUs, were to be 35 SUs higher on average than the post-intervention SUs of participants (being about 1.4 SD above the pre-intervention SUs shown in Table 2), then the estimate of the overall effect would be −4.06SUs (95% CI −6.25-1.87).If the mean post-intervention SU level of dropouts were to be lower than those of participants (negative value of δ), then the overall effect would be stronger; for instance, if δ = −20, then the estimated overall effect would be −5.32SUs (95% CI −7.64 to −3.01).This sensitivity analysis leads us to conclude that our results would remain rather stable, even in the event of substantial deviations from the MAR assumption (see S1 Text).Heterogeneity for the overall primary outcome was high and significant (I 2 = 89.6%,CI 78.4%-95.2%,p < 0.001) and for the secondary outcome as well (I 2 = 78.2%,CI 56.3%-89.9%,p < 0.001).It could be partly explained by the identified outliers, as it dropped from high to moderate for the primary outcome (I 2 = 55.5%,CI 16.2%-80.3%,p < 0.001) and from high to small for the secondary outcome (I 2 = 30%, CI 0%-69.1%,p < 0.001) after removal of the outliers from the analyses (Tables 3 and 4).

Conventional meta-analysis comparing included with non-included RCTs
The conventional meta-analysis (24 trials, 34 comparisons) was based on our search up to 31 December 2016 and included additional data from two RCTs published in 2017 [20,47].It revealed a small significant difference in mean weekly SUs at the first follow-up in favour of iAI participants as compared with controls (Hedges' g = 0.26, 95% CI 0.17-0.34,p < 0.001; Fig 4, forest plot of results of conventional meta-analysis).There was significant, moderate heterogeneity, indicating that the effect was greater in some trials than in others (I 2 = 65%, p < 0.001; We could not conduct a conventional meta-analysis for our secondary outcome, as only a limited number of studies reported on it.In S3 Data, this conventional meta-analysis has been expanded with two further eligible studies published between 1 January 2017 and 30 May 2018 that could not be included in our IPDMA.Our aim here was to explore whether more recent studies could potentially alter our IPDMA results; as they did not significantly alter the effect size in our conventional analysis, we believe this confirms the robustness of our analysis.

Principal findings and their interpretation
This study found that participants treated in iAIs showed a higher mean weekly decrease of 5.02 SUs of alcohol consumption and a greater likelihood of favourable TR (OR 2.20) than controls.Women decreased their mean weekly alcohol consumption significantly less than men (around 2 SUs).Our sensitivity analysis confirmed our assumption that this difference was not an artefact of the higher cutoff thresholds for men than for women at study inclusion (leaving women less space for alcohol reduction) [60].More highly educated participants reduced their mean weekly consumption significantly less than lesser educated ones (around 2 SUs).This result differs from the few studies that have reported on education as a moderator of iAI treatment outcomes; these showed either improved outcomes for more educated participants [61] or no such impact [62].For gender and education as moderators of the primary outcome, our sensitivity analyses pointed in similar directions to the outcomes of our main analysis, although the results were no longer significant.In our study, age was found to have moderated TR, with participants above 55 showing greater likelihood of post-intervention adherence to low-risk drinking recommendations than younger people.None of the other participant characteristics moderated treatment outcomes.Internet interventions appear effective when applied in community and healthcare settings, but effectiveness in work settings is still inconclusive.
Guided iAIs yielded significantly better results than unguided ones for both treatment outcomes.iAIs based solely on PNF showed a lower likelihood of TR than iAIs based on integrated therapeutic principles.Waitlist control moderated both types of treatment outcomes, with iAIs in WLC studies showing significantly better outcomes in terms of both SU reduction and TR than those in otherwise-controlled studies.It thus appears that iAI treatment outcomes could have been overestimated in studies in which WLC groups were applied as comparators.One possible explanation for such higher effect sizes in WLC studies would be that problem drinkers allocated to waiting lists might delay their alcohol reduction because they anticipate treatment soon.In contrast, people in other types of control groups might have already found alternative support by the time of the follow-up assessment, thus potentially reducing their alcohol consumption more than WLC controls.By the same token, such tendencies could deflate effect sizes in non-WLC studies [63].
The overall greater reduction of 5.02 SUs of alcohol consumption seen here in iAI treatment participants as compared with controls was higher than the 2.2 SUs we found in our earlier, conventional meta-analysis [14].One potential explanation for that difference is the higher number of guided iAI studies included in the present IPDMA; these showed higher treatment outcomes than unguided ones.Our current finding is comparable to the 5.61-SU reduction by adult iAI participants over controls reported in the conventional meta-analysis by Kaner and colleagues [16].Our results compare quite favourably with outcomes of patients treated in primary care settings with brief guided face-to-face interventions, who showed decreases from 2 to 4 SUs [64,65].We were also able to assess TR in terms of NNT (4.15).Due to data limitations, conventional meta-analyses have not been able to report on NNTs or on potential moderators of iAI treatment such as gender, age, and drinking profiles [16].

Methodological considerations
To the best of our knowledge, this is the first IPDMA to test the impact of iAIs and their moderators on treatment outcomes with adequate statistical power.The included RCTs had a low overall risk of methodological bias.Our results appear robust after comparison with our two-stage IPDMA results, as well as with those from our multiple imputation analysis and those from our conventional meta-analysis.The ANCOVA model that underlies our IPDMA implicitly relies on the MAR assumption, allowing dropout, which was 43% in our study, to depend on baseline consumption level.Although it cannot be ruled out that dropout was actually attributable to characteristics not included in the model, our MNAR sensitivity analysis suggested that the estimate of the overall effect would be reasonably stable against moderate deviations from the MAR assumption.The generalisability of our results to people in real-life settings might be hampered by poor assessment of ethnicity and by the focus on studies from high-income countries.In addition, only a small number of studies addressed effects of iAIs administered in care settings other than the community (such as in primary care practices, emergency departments, or workplaces).Another limitation is that all studies applied selfreported alcohol consumption measures, which is possibly a source of social desirability bias [66].We also observed high heterogeneity in our analyses, and it could be explained only partly by excluding outliers or by some of the subgroup analyses that we conducted.Hence, the moderating factors we identified offer only partial clarification of moderating influences on treatment outcome.We were bound, of course, by the available data.Other moderators, such as self-efficacy or participants' preference for iAIs over other types of interventions, cannot be ruled out [67].Longer-term outcomes of iAIs could not be assessed, as few studies addressed them.

Conclusions and clinical implications
Both men and women from different age groups and with different drinking profiles, including heavy drinking and binge-only drinking, can benefit from iAIs, and in particular from the therapeutically integrated ones as opposed to PNF-only interventions.Participants in iAIs reduced their mean alcohol consumption from 38.1 to 32.9 SUs per week, and they had a substantially higher probability of posttreatment adherence to low-risk drinking recommendations.The fact that heavy drinkers decreased their alcohol consumption by amounts similar to those of non-heavy drinkers has favourable implications, as the health impact of a given reduction is greater at higher levels of alcohol consumption [68].Despite the finding that many participants were still consuming beyond low-risk limits at posttreatment, the population health gains could nevertheless be substantial, in view of the high number of participants that can be reached with iAIs and the positive relationship between decreased alcohol consumption and the lower risks of physical and mental health disorders in the long term.These include earlieronset dementia [69], several types of cancer [70], cardiovascular diseases [3] (Wood 2018), and depression and anxiety [68,69,71].iAIs have great scaling potential, partly by virtue of their swift entry procedures for patients and the relatively low cost of repeated reuse, especially if unguided.For many people, iAIs could serve as a first step towards changing their problemdrinking behaviours and towards more intensive treatment, if needed.
In view of the constraints experienced with face-to-face BAIs in primary care settings, future studies should also explore various types of brief interventions, in order to gauge how problem drinkers in such settings can best be targeted.Those could be either face-to-face BAIs or iAIs, and the latter could be guided by general practitioners (GPs) or other professionals.For some patient populations, referral to unguided forms could be more beneficial [72].More primary care studies are needed, however, including head-to-head comparisons of unguided versus guided versus face-to-face interventions.The same applies to the optimum treatment orientations and levels of intensity and duration [73,74].As we have seen, not all treatment participants benefited from iAIs.We therefore need to better understand for which people such interventions work, how they work, and in what contexts (an approach also highlighted by Babor in 2008) [75,76].A final observation is that some countries have now substantially lowered the advised limits for daily and weekly alcohol consumption, in response to mounting epidemiological evidence of health risks inherent in the conventional limits [77].A threshold not exceeding 10 SUs of weekly alcohol consumption for both men and women has been proposed [3].Future studies should correspondingly adjust their sample inclusion criteria based on units of alcohol consumption.

Table 1 . Characteristics of studies analysed in IPDMA (19 studies, 27 comparisons). Study Target group/Screener Setting/Recruitment Intervention Mode of delivery N Control Timing of FPTA
from the community (n = 12 trials), from healthcare settings (n = 4), or from work settings (n = 3).Eight trials employed a minimal-intervention control design, six trials applied assessment-only control, and five included a waitlist-control comparator.Eleven trials estimated the effects of multiple-session iAIs, seven studied single-session iAIs, and one study included both types.Twelve investigated effects of therapeutically integrated iAIs and seven studied PNFonly interventions.Most comparisons (n = 19) involved unguided iAIs; eight involved human-guided interventions.The first post-intervention assessment occurred in most trials (n = 15) between 1 and 3 months after treatment, in three trials at 6 months, and in one study at 12 months.A total of N = 14,198 participants was included, out of the 17,545 participants in the 24 identified trials (a 79.77% inclusion rate).
table, as these would have unjustifiably satisfied our mean weekly SU criterion for favourable TR.Abbreviations: AOC, assessment-only control; AUDIT, The Alcohol Use Disorders Identification Test; iAI, internet-based alcohol intervention; MIC, minimalintervention control (e.g., information brochure); PNF, personalised normative feedback; SU, standard unit of alcohol consumption based on 10 grams of ethanol; TR, treatment response; WLC, waitlist controlled; 95% CI, 95% confidence interval.