What Predicts Treatment Adherence and Low-Risk Drinking? An Exploratory Study of Internet Interventions for Alcohol Use Disorders

Introduction: Internet interventions for alcohol problems are effective, but not all participants are helped. Further, the importance of adherence has often been neglected in research on internet interventions for alcohol problems. Prediction analysis can help in prospectively assessing participants’ probability of success, and ideally, this information could be used to tailor internet interventions to individual needs. Methods: Data were obtained from a randomized controlled trial on internet interventions for alcohol use disorders. Twenty-nine candidate predictors were run in univariate logistic regressions with two dichotomous dependent outcomes: adherence (defined as completing at least 60% of the treatment modules) and low-risk drinking (defined as drinking within national public health guidelines) at two time points – immediately post-treatment and at the 6-month follow-up. Significant predictors were entered hierarchically into domain-specific logistic regressions. In the final models, predictors still showing significant effects were run in multiple logistic regressions. Results: One predictor significantly predicted adherence: treatment credibility (as in how logical the treatment is and how successful one perceives the treatment to be) assessed during the third week of the intervention. Four predictors significantly predicted low-risk drinking at the post-treatment follow-up: pre-treatment abstinence (i.e., not drinking during the 7 days before treatment started), being of the male gender, and two personality factors – a low degree of antagonism and a high degree of alexithymia. At the 6-month follow-up, pre-treatment abstinence was the only significant predictor. Conclusion: Adherence was not predictive of low-risk drinking. Personality variables may have predictive value and should be studied further. Those who abstain from alcohol during the week before treatment starts have a higher likelihood of achieving low-risk drinking than people who initially continue drinking.


Introduction
Identifying predictors of treatment outcome is well recognized as an essential part of clinical research. Predictors measured before treatment can have important value in everyday clinical work, for example, when deciding This article is licensed under the Creative Commons Attribution 4.0 International License (CC BY) (http://www.karger.com/Services/ OpenAccessLicense). Usage, derivative works and distribution are permitted provided that proper credit is given to the author and the original publisher.
what treatment a specific patient should be offered or which patient groups should be offered more resourceintensive interventions, and predictors measured during the early phase of treatment could be used in adaptive treatment strategies to guide clinical decisions on intensifying, adapting, or terminating the intervention [1]. When it comes to the treatment of alcohol use disorders (AUDs), identifying predictors has proven to be an elusive quest. Often, predictors that are significant in one study are not significant in another, and sometimes the direction of the predictor is even reversed [2]. To our knowledge, only two reviews have attempted to summarize the literature. The first literature review on predictors of positive alcohol treatment outcomes, published in 1977, concluded that although no predictors were consistently significant throughout all identified trials, two demographic factors (being married and being employed) and one treatment history factor (having had contact with Alcoholics Anonymous) consistently predicted a positive outcome across the majority of studies [3]. The only avowedly systematic review of predictors in alcohol treatment, published in 2009, suggested that a low degree of psychiatric comorbidity and four alcohol-related factors (low degree of dependence severity, high alcohol-related self-efficacy, high motivation, and having abstinence as a treatment goal) were the most consistent positive predictors across studies [4].
Although there are several effective treatment alternatives for alcohol problems in both mild and severe forms [5], the treatment gap for those with AUD remains large with only one in six receiving help within health care settings [6]. As stigma is believed to be an important deterrent from seeking help [7], internet interventions, by way of their anonymous nature, hold promise as a treatment alternative for the large group of individuals desiring help but who may hesitate to register as patients at a clinic. Internet interventions are effective for alcohol problems, generally rendering at least small effects [8]. Participants in these trials tend to differ somewhat from the population traditionally found in addition clinics, as the proportion of women is usually higher than in clinics, and participants tend to be younger and more highly educated [9]. Two studies and 1 individual patient data meta-analysis (IPDMA) have investigated demographic predictors and moderators of outcome in internet interventions for alcohol problems. Riper and colleagues (2008) found that being female gender and having a higher level of education modestly predicted positive treatment outcomes 12 months after randomization [10], while Blankers and colleagues (2013) found that having a shared living situation and high interpersonal sensitivity predicted a positive outcome 6 months after randomization [11]. In the mentioned IPDMA, however, being female, being younger, and having higher education predicted negative treatment outcomes [8]. Not only demographic characteristics have been investigated as predictors of outcomes; in a recent study, Ramos and colleagues used machine learning models on log-data collected during the first 3 days of access to a digital alcohol intervention and found, for example, that choosing a target goal to quit rather than reduce drinking, and drinking less during these first 3 days, predicted success (defined as completing all modules and achieving one's own drinking goal set at the start of treatment) [12]. Further, Witkiewitz and colleagues found that lower alcohol dependence severity and fewer drinks at the start of treatment predicted the achievement of low-risk drinking during treatment [13] Outcome has not been the only focus in prediction analyses. Since internet interventions often suffer from high attrition rates (defined as not providing follow-up data), several studies have investigated predictors of attrition (or its opposite -retention). Postel and colleagues (2011) found that higher treatment readiness, higher age, and lower baseline consumption predicted retention (treatment completion) [14]; Murray and colleagues (2013) found that higher age, being of female gender, having a university degree, and not having children were related to retention [15]; and Radtke and colleagues (2017) found that consuming more alcohol, reporting more severe alcohol problems, and being a student predicted attrition [16]. However, a more clinically relevant concept is adherence. In internet interventions, adherence has been defined as the extent to which a participant is exposed to, or actively works with, the content of the intervention [17]. There are several ways of measuring adherence, for example, the number of log-ins, time spent logged in, or patterns in treatment use. A fairly straightforward way to measure adherence is to calculate the mean number of actively completed treatment modules in a treatment. In relation to effectiveness, module completion has been found to be the adherence measure most related to outcomes in internet interventions [18]. Although a number of studies on psychiatric problems such as depression [19] and social anxiety disorder [20] have studied adherence measured as module completion, to our knowledge, no prior studies on alcohol internet interventions have explicitly investigated predictors of adherence defined as actively completing treatment modules. To address this gap in the literature, as well as to increase the knowledge base on predictors of outcomes in internet interventions for the AUD population, we used data from a randomized con-36 DOI: 10.1159/000527868 trolled trial [21] to investigate predictors of (1) adherence and (2) low-risk drinking according to national public health guidelines [22], at two time points: immediately post-treatment and at a 6-month follow-up.

Design
The study was based on data from a randomized controlled trial in which participants were randomized to one of two internet interventions (ePlus or eChange) that differed in intensity and therapist support. A more detailed description of the study can be found in the original trial publication [21]. See Figure 1 for the flow of participants in the original trial.

Participants and Procedure
Participants were recruited from all over Sweden through Google Adwords, information posts on Facebook, and the Remente health app. To establish a diagnosis of AUD, telephone interviews were conducted using a DSM-5-adapted version of the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) [23]. Further, the Mini International Neuropsychiatric Interview (MINI) was used to assess comorbid psychiatric disorders [24]. Inclusion criteria were: (1) being 18 years old or more; (2) having a score of at least 14 (women)/16 (men) on the Alcohol Use Disorders Identification Test (AUDIT); (3) having consumed alcohol during the past week at a level of at least 11 (women)/14 (men) standard drinks (with one standard drink representing 12 g of ethanol); and (4) having an AUD defined as at least 2 out of 11 possible criteria for AUD in the telephone interview. Exclusion criteria were insufficient skills in Swedish, significant reading and writing difficulties, concurrent psychological treatment with a content resembling the treatments in the study, severe depression, Registered for study (n = 508) Did not complete screening questionnaires (n = 508) Low AUDIT score (n = 47) Low TLFB (n = 26) High MADRS-S score (n = 10) High DUDIT score (n = 9) Was not reachable (n = 14) Declined to participate/consent (n = 28) Other reasons (n = 5) acute suicidal ideation, drug use problems, psychiatric or somatic conditions contraindicated for the treatment (for example, bipolar disorder, psychosis, PTSD), and recently initiated use of medication for alcohol or other psychiatric disorders.

The Interventions
The programmes were both based on a CBT model of alcohol problems [25] and consisted of self-help texts divided into modules delivered over a 12-week period. The high-intensity intervention, ePlus, consisted of 13 interactive modules delivered weekly with about 3-5 pages of text per module, intended for use with therapist support (see publication for more details) [26]. The lowintensity intervention, eChange, consisted of 9 interactive modules with about 1-2 pages of text per module and was intended for use without therapist support (see publication for more details) [27]. Participants in eChange were automatically given access to one module each week, except for the last 3 weeks when no new modules were given. As for goal setting, in both ePlus and eChange, participants were able to choose in the work sheet accompanying module 2 whether they wanted to achieve abstinence or controlled drinking during the 12-week treatment period.

The Present Study
There were no significant differences between the two treatment groups on any measures at baseline, and there were no significant differences in adherence (module completion) between the groups (65% and 65%, respectively; χ 2 = 0.338; p = 0.56). There was one significant difference in drinking outcomes between the two treatment groups at post-treatment, with the high-intensity group reporting significantly fewer heavy drinking days (p = 0.05), but there were no significant differences in the number of drinks. At the 6-month follow-up, there were no significant differences in drinking outcomes. As the differences in drinking outcomes were low, data for the two groups were collapsed into one group (n = 143) for the analyses in the present study with participants in the wait list control group excluded. Of the participants in the two treatment groups, 123 (86%) provided data at post-treatment, while 111 (78%) did so at the 6-month follow-up. No data were imputed for the current study. Table 1 shows participant characteristics at baseline.

Outcomes
Outcomes were assessed at screening, pre-treatment (Time-Line Follow-Back (TLFB) only), post-treatment, and 6-month follow-up. Adherence as a predictor was defined as having completed at least 60% of the module worksheets, since this amount contained the components of the treatment believed to be the most important (i.e., goal setting, analysis of risk situations, coping with craving, and social situations). The primary outcome in the trial, alcohol consumption during the preceding week, was measured by TLFB [28], aggregated as (1) number of standard drinks and (2) number of heavy drinking days (HDD). For the purpose of this study, the variable "low-risk drinking" was created and defined as (1) ≤14 drinks for men/≤9 drinks for women and (2) no HDD (≤5 for men/≤4 drinks for women) during the preceding week. This is the recommended maximum level of drinking according to the Public Health Agency of Sweden guidelines [22]. Abstainers were included in the "low-risk drinking" category.

Selection of Potential Predictors
In total, 29 variables were chosen as potential predictors. The selection was mainly based on reviews and individual studies of predictors in alcohol treatment as well as recent literature on predictors in internet treatment for alcohol problems [3,4,10,11]. After the authors had selected which variables to include in an informal process of consensus, the variables were categorized into four domains: three consisting of baseline measures and one consisting of measures during treatment, as presented below. The numbers in parentheses refer to the variable number, from 1 to 29.

Treatment-Related Factors
Two treatment-related factors were investigated as predictors; (28) credibility of treatment assessed with the Treatment Credibility Scale (TCS) [39] (administered 3 weeks after initiation of treatment). TCS is an instrument that consists of 5 items assessing how logical the participant perceives the treatment to be and how successful they believe the treatment will be. Further, (29) treatment adherence defined as percent completed module worksheets, where 60% or more completed module worksheets were categorized as "adherent" and a lower percentage of completed modules were categorized as "non-adherent."

Statistical Analysis
A hierarchical logistic regression approach was used, where the first step for each of the predictors was to run univariate logistic regressions using adherence, low-risk drinking at post-treatment, and low-risk drinking at 6-month follow-up as the dependent variables. Second, significant predictor variables in the univariate logistic regressions were entered into domain-specific multiple logistic regression models. Significant predictor variables in the domain-specific logistic regressions were then entered into final models, one for each of the three outcomes. Implementing an initially higher p value is considered appropriate in prediction analysis as a standard p value (i.e., ≤0.05) often excludes important factors when the purpose of the study is exploratory [11,40]. Therefore, a p value of ≤0.15 was chosen in the univariate and domain-specific regressions. Predictor variables in the final models were assessed at the ≤0.05 level. All statistical analyses were performed using SPSS 24 (IBM Corp.).

Adherence
Univariate Logistic Regressions Results from the univariate regressions between individual predictors and adherence are presented in Table 2.
Multiple Logistic Regressions Two demographic factors (age and good economic situation) were significant in the univariate analyses and were entered into a domain-specific regression. Both of them remained significant and were therefore entered into a final model. Three other factors (drinks preceding the week, hedonic capacity, and treatment credibility), one from each of the other three domains, were signifi-cant in the univariate analyses and were thus directly added to the final model. In the final model, treatment credibility emerged as the only significant predictor of adherence (Table 3).

Low-Risk Drinking
Univariate Logistic Regressions Results from the univariate regressions between individual predictors and low-risk drinking at post-treatment and at the 6-month follow-up are presented in Table 2. Table 4 displays the two final models, with the predictors that were significant through the domain-specific models. Being of male gender, showing a lower degree of antagonism, a higher degree of alexithymia and reporting pre-treatment abstinence emerged as significant predictors of low-risk drinking at post-treatment. Reporting pre-treatment abstinence was the only significant predictor of low-risk drinking at the 6-month follow-up.

Sensitivity Analysis
Since gender influences the definition of low-risk drinking, as this outcome is calculated based on two measures (drinks and HDD in the preceding week) with different cut-off scores for men and women [22], the robustness of our finding of male gender as a predictor of lowrisk drinking was assessed; we conducted a sensitivity analysis where change scores between screening and post-treatment in number of drinks were used as an outcome instead of low-risk drinking. We found no significant differences in change scores between men and women (t = −1.365, p = 0.176).

Discussion
In this study, data from a randomized controlled trial was used to investigate predictors of adherence to an intervention and low-risk drinking after the intervention. One significant predictor of adherence was identified: treatment credibility measured 3 weeks into the treatment. Significant predictors of low-risk drinking at posttreatment were pre-treatment abstinence, a lower level of antagonism, a higher level of alexithymia, and being of the male gender (although our sensitivity analysis presented conflicting results regarding the finding on gender). At the 6-month follow-up, pre-treatment abstinence was the only remaining significant predictor. The only factor predicting adherence in this study was treatment credibility. Although participants perceiving the treatment as credible early in treatment were more likely to adhere to the treatment in the form of completing more than 60% of the treatment modules, neither treatment credibility nor adherence were significantly predictive of low-risk drinking. Thus, we did not find that module completion was associated with outcome, as has been found to be the case in internet interventions for depression [41] and social anxiety disorder [42]. When it comes to predictors of low-risk drinking, the results of our study differed completely from the findings in other predictor studies. For example, of the factors that we investigated, none of the ones that were considered consistent predictors in the systematic review from 2009 (i.e., self-efficacy, baseline alcohol consumption, dependence severity, psychiatric comorbidity) [4] were significant in our study, and neither were the significant predictors in the two studies on internet interventions for alcohol problems (i.e., female gender, high level of education, shared living situation) [10,11]. Pre-treatment abstinence was the only predictor found to be significant both at post-treatment and at the 6-month follow-up. Of our findings, pre-treatment abstinence is also the one most supported by prior research. Several studies have shown that the greatest change in alcohol treatment occurs early in treatment, sometimes even before treatment starts, and that this is often related to treatment outcome [13,32,33,43]. Indeed, a recent systematic review showed that the placebo response in pharmacological trials of AUD was higher among early abstainers, suggesting that this subpopulation may fare better in treatment than non-early abstainers [44]. As none of the above-mentioned trials randomized participants to pre-treatment abstinence or not, it is unclear whether it is pre-abstinence/reductions in itself that is beneficial or whether the success of those with pre-treatment abstinence/reductions simply reflects a highly motivated sub-population among participants in these trials. Future trials where participants are, for example, randomized to either immediate abstinence or moderation could help elucidate the significance of this predictor further. A second significant predictor at posttreatment was being of male gender. This finding was in line with the only individual patient data meta-analysis on alcohol internet interventions [8]. However, it contrasts with previous findings on predictors in internet interventions for alcohol problems [10], as well as with general findings in alcohol treatment, where female gender has mostly been found to positively predict outcome [4]. It has been argued that applying different cut-offs for men and women in alcohol trials may lead to underestimation of effects on women [45]. For example, a consequence for trial eligibility when using lower cut-offs for women is that included female participants will, on average, consume less alcohol at baseline than included male participants, leaving less room for them to reduce consumption thereby diluting treatment effects [45]. Using different cut-offs to produce a dichotomous treatment response variable, as we have done in this study, may have the same consequence, as it will be harder for women to reduce their alcohol consumption to below the cut-offs than for men. Our sensitivity analysis showed that men and women had made comparable quantitative reductions, implying that the treatment effect was similar among men and women. Although it is a matter of debate which of these two analyses is preferable, it can at least be argued that in future trials where different gender cut-offs are used to assess eligibility and/or to generate treatment outcome, gender differences should be explored further with appropriate sensitivity analyses to assess robustness of findings. Personality as a predictor of outcome has been sparsely investigated in alcohol treatment research. In relation to the five factors of the FFM, people with alcohol-related problems tend to score higher on neuroticism and lower on agreeableness and conscientiousness, while research on extraversion has provided mixed findings and research on openness has shown no associations at all [46]. When it comes to treatment outcomes, several studies have shown that a high degree of neuroticism and a low degree of conscientiousness are associated with relapse after treatment [47,48]. In our study, neither negative affectivity (corresponding to the FFM factor neuroticism) nor impulsivity (corresponding to the FFM factor conscientiousness) were significant predictors of low-risk drinking post-treatment. Instead, the personality factors associated with low-risk drinking were alexithymia (negatively corresponding to the FFM factor openness) and antagonism (negatively corresponding to the FFM factor agreeableness). Alexithymia, which has been described as difficulty in identifying and communicating feelings, has previously been linked to AUD factors [49]. In relation to treatment, it has been found to have negative predictive value for outcomes following psychodynamic psychotherapy, but not following CBT [50]. A possible explanation for the finding that alexithymia was positively predictive of low-risk drinking is that internet interventions, by way of their anonymous nature, suit alexithymic individuals well, as these individuals generally feel uncomfortable in social situations and tend to want to minimize human interaction [51]. The second factor associated with outcome, a lower degree of antagonism, was surprising, as this domain has not previously been found to be associated with either alcohol problems per se [52] or alcohol treatment outcomes [47,48]. Lastly, although both alexithymia and antagonism were predictive of low-risk drinking, their predictive value occurred in opposite directions, which is also surprising as they seem theoretically similar and were also highly correlated in our study (r = 0.414, p=<0.0001). More research on the correlation between FFM factors and alcohol treatment outcomes should be conducted to further assess their clinical value in this context.

Limitations
As this was an exploratory study with almost 30 potential predictors and three outcome variables, a large number of significance tests were performed, raising the possibility of chance findings. Furthermore, although we collapsed the two treatment groups into one, thus increasing power, the sample is still relatively small.

Conclusions
Adherence, defined as the completion of treatment modules, was not predictive of low-risk drinking. People with AUD who abstain from alcohol shortly after signing up for internet treatment for AUDs may have a higher likelihood of achieving low-risk drinking than people who initially continue drinking. Personality variables may have predictive value and should be studied further.

Acknowledgments
Claudia Fahlke passed away during the writing of the manuscript but contributed significantly to the design and analyses of the current study.

Statement of Ethics
This trial on which this study is based was reviewed and approved by approval number the Regional Ethics Vetting Board in Stockholm (2015/2014-31; amendment 2016/295-32). Digital informed consent was collected from all participants included in the trial.

Conflict of Interest Statement
Christopher Sundström has been on the Clinical Review Board of the digital mental health company Alavida. Niels Eék is cofounder of the Remente mental wellbeing app. Anne H Berman is an Editorial Board Member of "European Addiction Research." The other authors report no conflicts of interest.

Author Contributions
Christopher Sundström and Anne H. Berman designed the study, analysed the data, prepared the first version of the manuscript, revised the manuscript, and approved the final manuscript. Niels Eék, Martin Kraepelien, and Viktor Kaldo designed the study, revised the manuscript, and approved the final manuscript.

Data Availability Statement
All data generated or analysed during this study are included in this article. Further enquiries can be directed to the corresponding author.