Globally, alcohol is one of the greatest risk factors for deaths and causes substantial health loss among the world’s population (1, 2). Even though there are several efficient ways to prevent and treat problematic alcohol use (3–5), only approximately 15% of individuals with alcohol use disorder seek and receive help (6). This constitutes one of the largest treatment gaps among mental disorders (7). Internet-based interventions for problematic alcohol use have emerged as an alternative to traditional face to face treatments (8–10). Availability and anonymity appear to make this option attractive to sufferers (11–14), creating a potential to attract those who would not otherwise seek help (15–17). Meta-analyses have revealed these interventions to be efficacious (8, 18), even comparable in effects to traditional face-to-face treatments (19).

However, as in traditional alcohol interventions (3–5), significant group-average decreases in drinking and symptom scores can mask substantial inter-individual variations in change (17). Past research has shown that the population of individuals with alcohol problems shows substantial heterogeneity with regards to many key characteristics (1, 20), which could reasonably be expected to moderate outcomes of an internet intervention. Indeed, matching individuals to different alcohol interventions has been a topic of some past research, dating back over twenty years (21–23).

One dimension well-known to moderate the presentation of many psychiatric disorders (24, 25), preferences for treatment (26) and even treatment outcomes (27), is sex. We recognize that it is often unclear whether studies within these field use sex (what sex one is assigned to at birth), or gender (which can be the same as what one is assigned to at birth, but also differ from it). Since this distinction is not made in most of the extant literature, we have opted to use sex consistently where the distinction is not clearly stated, since sex-specific guidelines on drinking are grounded in biological, pharmacokinetic sex-difference (28), but also recognize that this is a simplification. There is robust evidence that men, on average, drink more than women, and are over-represented in addiction care (1, 20), but whether sex is also a moderator of the effects of treatment intervention has received little attention in the extant literature. There are some studies suggesting that women with problematic alcohol use benefit from interventions that encompass recognition of gender expectations and the stigma of not living up to the results (29). A meta-analysis that investigated moderators of outcomes of internet-based alcohol interventions (30), concluded that data on sex were limited, particularly women, but that five studies that did investigate this failed to find that gender modified the difference in alcohol consumption between the trial arms. A later, individual-patient meta-analysis instead found that gender was a moderating factor, where females decreased their mean weekly standard units less than men (18). These results did not remain significant when imputing missing values, yet to what degree the imputation technique took gender into account, was not reported.

An often-overlooked dimension when examining sex-moderated outcomes of alcohol interventions, is that the outcome measure may in-itself be sex-moderated. As per clinical trial methodology, one needs to distinguish between change and final state. The former is captured by continuous measures such as reduction and response (typically denoting numeric decreases in symptom ratings, either absolute or relative, respectively), while the latter is captured by ratios of participants above or below a prespecified threshold, as in the case of remission. Importantly, the fact that men on average drink more than women, has important, but often neglected consequences for both types of outcomes. Many studies (31–33), including our own, have for example relied on national guidelines to threshold drinking into risk- and non-risky. In many countries, these guidelines are sex-specific (34), with the previous Swedish guidelines for example allowing men to drink 55% more standard units per week than women. Whether the average sex-difference in baseline drinking between men and women is equal, either in absolute or relative magnitude, to the sex-difference in remission thresholds, is not typically reported. Even if so, this assumes equidistance of change scores, i.e. that a 7-drink reduction from 22 drinks to 15, is the same as from 15 to 8. Importantly, it is mathematically impossible to match equidistance in both relative and absolute terms at the same time, assuming there is a baseline difference. This means that a baseline sex-difference may also confound numeric outcome measure such as reduction and response (henceforth used synonymously).

A baseline sex-difference in drinking may thus be a confounder in examining moderating effects of sex on treatment outcomes. Third-variable confounding (e.g. in psychiatric comorbidity) complicate the issue further. This highlights a potential concern: if studies investigate whether a treatment’s effectiveness differs by gender but rely solely on the number of drinks consumed as an outcome measure, they may mistakenly interpret a difference as treatment-related. However, this apparent difference might actually stem from baseline values or cofounders rather than the treatment itself. Therefore, the hypothesis of this study is that there will be a significant difference between genders, but this difference may be explained by baseline variations and/or other confounding factors.

In sum, there are inconsistent findings in the extant literature as to whether internet-based interventions for alcohol use disorder have different effects for men and women and to our knowledge, no previous study examining sex-moderation of outcomes in interventions for a problematic alcohol use has systematically examined confounding. It remains unknown whether previous positive findings were due to confounding. To examine this important question, we performed secondary analyses of a randomized controlled trial.

The current study replicated past research in showing differential treatment outcomes between men and women when examining response (32), but not remission (33). However, after taking a baseline difference in drinking into account through a subsampling strategy, the difference in response was no longer significant, suggesting that this apparent sex-difference was confounded by intercept-slope correlation – i.e. since men on average drank more at baseline, this offered larger room for decreases.

Our findings stress the importance of carefully considering which outcome that best captures the desired change after treatment, as well as how other study characteristics impact outcome modeling. First, the definitions of both low-risk drinking and hazardous drinking differ greatly between countries – and in most countries, the definition of low-risk drinking also differs by sex (34). A direct consequence of this sex-based target differentiation is that women must reduce their drinking more to achieve low-risk drinking, assuming they start from the same baseline level. This, in turn, is however seldom the case (including in the current study). In several studies similar to ours, sex-specific definitions are also used as outcome, with inconsistent results. For instance, some studies on predictors of change in internet interventions have not found any significant difference among women (47) or somewhat better results among women (48). But in a study that investigated predictors of change in a similar intervention as the current study, women were found to be less likely to have low-risk consumption at follow-up compared to men when previous Swedish sex-specific guidelines were used as outcome (32). These results were replicated a few years later (49). A more recent study, exploring the effects of a web-based intervention for alcohol and PTSD symptoms among veterans, also showed that significantly fewer women achieved low-risk drinking after one-month, but also that women did not reduce their weekly drinking as much as men after six months (33). These results on continuous drinking outcome are similar to the findings in our study and the findings from the previously mentioned individual data meta-analysis by Riper et al. (18). In a British study investigating the predictors of outcomes of a mobile app targeting harmful alcohol use, the only predictor associated with the extent of alcohol reduction was how much the participant drank at baseline (50), similar to the findings in the current study.

Multiple studies have revealed that both the sensitivity and specificity of the sex-specific definitions have had large variation (51, 52). Also, there are ethical aspects in using assigned sex rather than individuals identified gender. Gender is not necessarily binary, and using uniform measures could result in more inclusive standards (53). Further, there has also been an ongoing discussion about using categorical outcomes for alcohol interventions, such as cut-off scores for heavy, or hazardous drinking (54). Unless the explicit target of an intervention is to decrease drinking to a specific, sex-indifferent level [e.g. before planned surgery (55)], capturing change after treatment with an absolute or relative response metric will circumvent this issue; should sex-differences in outcomes be of special interest, analyses should then preferably be adjusted similar as to in the current study. Of note, this applies only when considering any change in drinking as clinically meaningful: if total abstinence (i.e. a naturally occurring zero) is the only intended outcome, the entire issue of sex-specific outcomes is largely rendered irrelevant.

Strengths of the current study are that the sample is both large and inclusive. Another strength is that a multitude of confounders were considered. There are also several limitations to the study. First, we opted to focus on total number of standard units per week, since this is the most common outcome in the field of digital interventions for alcohol problems, and also the main metric (along with daily drinks) on which national drinking guidelines are based. Similar analyses could also be performed for other TLFB-derived metrics like drinking days, average number of drinks per drinking day, days with binge drinking, maximum drinks on any given day, and other clinically pertinent metrics. Second, it was deemed out of scope in the current study to examine whether popular imputation techniques for missing data should be performed separately by sex. Third, the current study did not attempt to associate change in drinking to treatment adherence; such analyses would however need to account for the non-randomized nature of this variable, which has shown to be associated with baseline severity in at least one other study on internet interventions for addictive disorders (56).

Considering the magnitude of the alcohol problem, and that iCBT already has a proven track record of reaching and attracting large samples, there are excellent reasons to continue developing and evaluating the effects such similar interventions not only on a group-level, but also subgroup-level. In choosing which potential moderators to examine, it is important that these are anchored in evidence and proper deductions that show why these may indeed moderate outcomes, as to avoid Type 1 errors through involuntary hypothesizing after the results are known (57). Findings of the current study highlight the importance of carefully consider which outcomes to specify when conducting studies on internet interventions for addictive disorders which accept both sexes.