The study was planned as a randomized control trial to compare the Daybreak program plus online coaching via real-time chat messages (Intervention group) and the Daybreak program without online coaching (Control group) at one, three and six months. However, due to a programming error, some (n=48; 11.9%) people in the Control group were able to access online coaching and we could not tell if other members of the Control group found they could access the online coaching but did not opt to use it. Therefore, with the approval of the human research ethics committee, the trial was ended after the last person to consent reached their three-month follow-up rather than the planned six-month period.

We invited new registrants to Daybreak to join the study (see Multimedia Appendix 1). To be eligible, participants had to be 18 years or older, a resident in Australia, had to provide a valid email address and had to have access to the Internet. However, we excluded those who reported a history of treatment for cardiovascular disease, as this was likely to confer additional health risks during alcohol withdrawal [20]. Participants had to exceed the at-risk threshold (>7) on the 10-item AUDIT [18]. Scores on the AUDIT between 8-19 indicate a pattern of drinking that is likely to fulfil the International Classification of Diseases criteria for hazardous or harmful drinking, while higher scores equate to alcohol dependence [21], but the AUDIT is a screening tool and not a diagnostic indicator. Finally, as part of our duty of care, we assessed suicide risk with the P4 suicidality screening survey [22]. Those participants classified as above minimal risk were still eligible for the study, but they were also provided with the contact details for Lifeline (a 24-hour support service). Those people classified in the highest AUDIT group (≥20: probable dependence) had it recommended to them to speak to their general practitioner or other health professional prior to reducing their use of alcohol in order to minimize potentially serious complications during alcohol withdrawal. The study planned to use a simple, fully randomized allocation (eg, no blocking). Participants were blind to their condition and follow-up data were collected online by an automated survey and hence blind to condition. Those not meeting the eligibility criteria could still access Daybreak but were not part of the study cohort. The processes of screening, enrolment and randomization were fully automated.

The Daybreak program is available from the Google and Apple App Stores or https://www.daybreakprogram.org/, but it can also be accessed from sites such as healthdirect.gov or other health directories. Alternatively, visitors to the Hello Sunday Morning website (www.hellosundaymorning.org) are directed to the program. Daybreak is a self-guided program that can be accessed as frequently as required by the participant, and access is controlled via username (email address) and password. Those in the Intervention group also had access to an online health coach between 7:00 and 19:00 on weekdays, but as noted above, some of the Control group also accessed coaching. We commenced recruitment in February 2018 and closed it in November 2018. The Daybreak version used throughout the study was version 1.5.8, build 90. Since late 2016, Daybreak has reached more than 50,000 participants.

We emailed and sent a text to participants with a link to the relevant follow-up survey after one month and after three months. If required, we sent a reminder text message one week later. If there was no response, a research assistant who was blind to study group allocation telephoned the participant to ask them to complete the follow-up, with a maximum of three telephone reminders calls allowed to be made. At each follow-up, participants were eligible to enter a draw to win an Ipad2. The study received the required institutional ethics approval (Curtin University 2017-0855), and the trial procedure was registered (ACTRN12618000010291). In compliance with Australian ethical guidelines, as all participant were screened with at risk alcohol use, we used an active rather than placebo control group [23].

The sample size estimation was calculated for the planned outcomes at six months. For online alcohol interventions, typical effect size values are in the range of Cohen d=0.3-0.4 at 6 months [11,12,24]. We were not aware of previous investigations involving social networks as a means of reducing alcohol use; however, investigations with other behaviors (eg, diet, physical activity) typically report small but not significant changes [25]. Therefore, we based our sample calculation on a small effect (Cohen f=.10; equivalent to Cohen d=0.2) and assumed that the repeated measures would be correlated at r=0.5. To achieve a power of 0.80 with an alpha P<.05 would require 60 people per group, however, given the clinical interest in the effectiveness of brief alcohol interventions for those with more entrenched problems (probable dependence), and potential gender differences, we aimed to recruit 60 people to the smallest cell (sex by AUDIT risk level by study group). To achieve this, we projected recruitment of 300 participants in each group (600 in total), and assuming that 35% would be lost to follow-up, we targeted 467 per group (N=934).

Participants were screened with the 10-item AUDIT, which has been validated in Australia [18] and elsewhere. Scores range from 0-40 (0-7=low risk, 8-19=hazardous or harmful alcohol use, 20-40=probable dependence), and outcomes were assessed as change in the scores for the first three items (termed the AUDIT–C). Prior research has shown that the AUDIT–C can predict clinical outcomes at 12 months [26]. We also assessed secondary outcomes and other measures, which are listed in Textbox 1.

Engagement with the Daybreak program was recorded as: (1) engagement with coaching (defined as at least one message sent by the participant to the coach); (2) The number of experiments completed; (3) number of blog shares (posts on their own blog); and (4) number of blog comments (posts on another person’s blog).

Descriptive data and analyses (eg, t tests, Mann-Whitney U tests, χ² tests) were provided for each group. However, due to concerns over the randomization process, the main repeated measures analysis was for all participants rather than by randomized groups. In addition to the overall results, Multimedia Appendix 2 provides online descriptive data at 1-month and 3-months follow-up by randomized group for each outcome measure.

We used Weighted Generalized Estimating Equation (WGEE) analyses to investigate longitudinal changes in primary and secondary outcomes over the study period. WGEE is a repeated measures regression model that takes into account the correlation of repeated measures within each subject [43]. In contrast to repeated measures analysis of variance (ANOVA), WGEE has minimal assumptions about time dependence and uses all available longitudinal data, irrespective of single missing values at follow-up. WGEE is also more robust than unweighted generalized estimating equation, when the assumption of missing completely at random (MCAR) is violated. To control for attrition at 1-month and 3-months follow-up, we estimated weights as suggested by Salazar et al (2016). Attrition analysis revealed that 1-month assessments were more likely to be completed by participants who were older (t_721.92= 3.67; P<.001), not single (χ²₃=11.467; P=.009), who used less alcohol in the 7 days prior to baseline (t_786.919=–2.676; P=.008), who had higher sleep quality values (t_703.512=–1.953; P=.05), who had more blog shares (t_482.08=3.744; P<.001), blog comments (t_550.511=4.058; P<.001), experiments taken (χ²₁=15.65; P<.001), and finally, who stayed longer in the program (t_692.773=5.252; P<.001). In addition, the 3-month assessments were also more likely completed by participants who were female (χ²₂=20.366; P<.001), who had minimal initial suicide risk (χ²₂=8.706; P=.01), and who had lower psychological distress (t_657.61=–2.328; P=.02).

In the first round of analysis, WGEE models included only the time variable to examine significant changes in outcomes over the study course. In the second round, WGEE models included the following predictors: (1) time (baseline versus follow-up assessment); (2) study group; and (3) the interaction term for time by group. In the last round of analyses, WGEE models included: (1) time; (2) baseline variables; (3) program use variables; and (4) interaction terms for time by program use variables. We then used a hierarchical, backwards procedure where we removed predictors with the highest P-value one at a time until only significant predictors were retained within the model. In the WGEE parameters, the intercept shows the mean change in the outcome for all participants (and on a population level). Betas are added or subtracted to the intercept, which reveals the change in the outcome for the particular group.

Outcomes by AUDIT risk group (probably dependent versus hazardous or harmful) were assessed with a standard, repeated measures ANOVA, as we were interested in the subject level outcomes rather than the population interpretation from the WGEE. In addition, an alpha level of 0.05 (two tailed) was chosen for all statistical tests conducted in the study. All analyses were performed using the statistical tools SPSS version 22 and R version 3.3 via geepack [44].

There were 2616 people who viewed the invitation page. Of these, 1278 (49%) accessed additional information and 978 (77%) provided consent. However, not all were eligible for the study. Of everyone who provided consent, 61 indicated a previous cardiovascular disease diagnosis, 28 scored below the AUDIT cut-off score, and 36 did not complete the AUDIT (12 cases failed more than 1 eligibility criteria). Further, there were 50 people who attempted to enroll multiple times after failing an eligibility criterion. We excluded these people, plus an additional 22 people who were not randomized because of a programming error. Thus, the study cohort consisted of 793 people (81.1% of those who consented; see Figure 1).

There were 398 people randomized to the Intervention group (50.2%) and 395 to the Control group (49.8%). Of all the participants, 71% were female (one person selected other) with a mean age of 40.1 years (SD 10.0), and most participants (77%) had at least commenced tertiary education. The mean Kessler’s K-10 score was 24.8 (SD 7.0), with 19 participants (2.4%) reporting a serious suicide attempt, and nearly 30% of participants were in the highest risk category on the P4 suicide screen. A total of 69% (n=550) of participants were classified in the AUDIT highest risk category (probably dependent), with their mean alcohol consumption being 37.1 (SD 28.3) standard drinks per week. At baseline, the only significant difference between the groups was in terms of days completely out of role due to alcohol use (Table 1). Follow-up data were subsequently collected from 334 (42.1%) participants at 1 month, and from 293 (36.9%) at 3 months.

Overall, 106 participants engaged with coaching (eg, at least one message sent by the participant to the coach), including 68 from the Intervention and 38 from the Control groups (Table 2). A greater proportion of those in the Intervention group engaged with an online coach (17.1%; χ²₁=9.5; P=.002) than the Control group (9.6%). Blog posts were made by 526 people, with 258 in the Intervention group (65%) and 268 in the Control group (68%). In terms of shares, 230 people (58%) in the Intervention and 227 people (58%) in the Control group commented on the posts of others. There were no significant differences in the median number of experiments taken, shares, comments or days in the study.

Pre-post comparisons of primary and secondary outcomes are displayed in Table 3. The WGEE models revealed significant changes over the study period for all outcomes at P<.001, except for amount of exercise at baseline to 1 month (P=.15) and at baseline to 3 months (P=.65). Participants significantly reduced their alcohol use and their psychological distress, and also gained in sleep quality and quality of life. Further, WGEE revealed that neither study group, nor the interaction term between study group and time, were associated with changes in primary or secondary outcomes. WGEE accounted for missing patterns at T1 and T2, with weights for T1 accounting for age, marital status, alcohol use days (last 7 days), sleep quality, retention, blog shares, blog comments, and experiments seen or done, and weights for T2 accounting for age, gender, marital status, suicide risk, psychological distress (K-10), retention, blog shares, blog comments, and experiments seen or done. Effect size was calculated via Psychometrica [45].

Change in alcohol consumption was compared for those who were as classified on the AUDIT–10 at baseline as hazardous or harmful or as probably dependent. Retention was similar for the two groups at three months, with hazardous or harmful at 38.3% (χ²₁=0.3; P=.61) versus probably dependent at 36.4 % (χ²₁=0.3; P=.61). There were significant reductions in alcohol use for both groups (F₁₂₈₈= 24.7; P<.001), with estimated marginal means of 40.8 down to 20.1 drinks for probably dependent and 22.9 down to 11.9 drinks for hazardous or harmful, and there was also a significant time by group interaction (F₁₂₈₈= 8.3; P=.004) reduction for the probable dependence group. The time by sex and the time by group by sex interactions were not significant.

Factors that predicted significant changes in alcohol-related variables are displayed in Table 4. Online coaching was not significant in any of the models and was therefore not retained. Significant predictors for AUDIT–C were time (reduced at each follow-up), gender (males report higher scores than females), age (older participants report higher scores than younger participants), and the interaction between blog comments and time (more comments were associated with lower AUDIT–C scores at 1-month follow-up). Significant predictors for alcohol days out of role were time (reduced at each follow-up), gender (males report more days out of role than females), marital status (single participants report more days out of role than partnered participants), education (higher education was associated with lower scores), and the interaction between taking part in experiments and time (taking part in experiments was associated with fewer days out of role at 1-month and 3-months follow-up). Significant predictors for alcohol use in standard drinks were time (lower at each follow-up), age (increased with age), marital status (lower for partnered people), and the interaction term between blog shares and time (more shares were associated with lower alcohol use at 1-month and 3-months follow-up).

Factors that significantly predicted changes in psychological distress and quality of life are displayed in Table 5. Significant predictors for the K-10 were time (lower scores at both follow-ups), age (lower with increasing age), marital status (lower for partnered participants), and education level (lower for those with more education). None of the interaction terms for program use by time were retained in the model. Significant predictors for EUROHIS-QOL were time (higher scores at each follow-up), marital status (higher for partnered participants), education (higher for those with more education), and the interaction between blog comments and time (more comments were associated with higher quality of life scores at 1-month follow-up).

Use of the Daybreak program resulted in significantly improved outcomes in terms of alcohol measures, mental health and quality of life. An important finding was that the program was effective in those who were classified as probably dependent at baseline. Face to face brief interventions are recommended for those with at-risk alcohol use, but they are regarded as ineffective for those with alcohol use disorders [46,47]. It is therefore important to determine the range of alcohol use for which electronic health (eHealth) interventions are effective. Overall, these findings need to be tempered with the fact that the trial was not implemented as designed, and the longest period of follow-up was only three months.

With its multiple components, the Daybreak program does not fit within the typical definitions of a brief intervention, certainly not as described for face to face brief interventions [46,48]. In the context of eHealth interventions, the concept of a number of sessions of treatment or time to deliver treatment is ambiguous; however, Riper et al note a distinction is often drawn between “single session e-personalized normative feedback” and extended interventions that draw on a range of therapeutic techniques [11]. The Daybreak program lies within the more extensive group. Previously, the extent of engagement with the Hello Sunday Morning program (eg, in terms of blog posts, blog shares, following other participants) has been shown to correlate with improved outcomes (ie, reduced AUDIT scores) [19]. Despite the low overall level of engagement (see Table 2), these measures still predicted a range of improved outcomes (eg, in AUDIT–C score, in EUROHIS-QOL score with blog comments about alcohol use, and in blog shares at one month). These findings suggest that the community aspect of Daybreak has an important therapeutic role in supporting behavior change.

The outcomes for the study also need to be considered in light of the cohort who participated. Those who access eHealth interventions often have less extensive alcohol use, with a range of 9.1-43.6 standard drinks (mean 18.3) [12], than those in face to face interventions (mean 24.4 standard (10g) drinks) [49]. However, in the recent analysis by Riper et al, the mean consumption was 38.1 standard drinks per week [13]. Most of our participants were classified as probably dependent on alcohol, with mean consumption being 37.1 standard drinks, but the prevalence of severe alcohol problems was higher than typically found in eHealth trials. Thus, many of our participants (69%) were classified as probably dependent on the AUDIT whereas the respective figure reported by Riper et al was 22%, but they also reported that 34% were heavy drinkers (defined as >35/50 standard drinks for females or males) [13]. Also notable was the high prevalence of women in our trial, representing 71% of participants compared with typical values of about 50% [12,13] and even compared to the previous assessment of Hello Sunday Morning clients (64%) [19]. There are concerns that eHealth alcohol interventions may not be as effective with women as men [13]. Although men had higher AUDIT–C scores and more days out of role than women, gender was not significant as an interaction term in the current study.

Also notable was the high level of distress as measured with the K-10 and the suicide screening scores. Given the evident distress, it is important that eHealth interventions in this area provide appropriate support and referral pathways. In contrast, the days out of role were not as extensive as those reported for people with alcohol dependence [32], and the mean quality of life scores, while lower at baseline than the general population, were above those for a cohort with depression [34]. Further, by three months the mean score was above that for a general population sample [32].

One of the main aims of this trial was to evaluate if outcomes could be enhanced by giving participants access to online coaching. Few participants opted to use this facility, with 17.1% of the Intervention group accessing the service. Overall, online coaching was not associated with any significantly improved outcomes. The lack of an effect for online coaching is similar to previous studies where guidance has not improved alcohol-related outcomes [24], but remains in contrast to both eHealth interventions for other mental health problems where guidance appears to be beneficial (eg, standardized mean difference compared to unguided=–0.27) [50] and the most recent analysis of alcohol interventions [13]. The reason for the potential difference between those with alcohol use problems and other types of mental health issues is unclear but given that improved outcomes were obtained without the provision of this additional resource-intensive component, this means that Daybreak can be provided more widely.

Clearly the most significant limitation was that the trial was not implemented as designed, with some members of the Control group accessing the online coaching, so this necessitated a change in the planned analysis. In addition, we also stopped the trial at the completion of the three-month follow-up rather than at six months due to this error. Compared with other studies in the field, the level of attrition for this study was high (63%) compared to other alcohol studies of at least three months duration (range=8%-45%; mean 25.2% [12]) [24], thus limiting the generalizations that can be made from these data.

The current Australian guidelines do not provide a formal recommendation for the number of drinks per week that constitute low risk drinking, but they do recommend that average daily consumption should not exceed two drinks per day [3], equating to 14 per week. On this basis, the average reported here (17.5 drinks) still exceeds that figure, albeit from a high starting point. We do note that those with hazardous or harmful alcohol use did reduce their consumption to below the guideline (mean 11.9 at three months). The Daybreak program, due to its extensive reach, member engagement and clinical safety, has the potential to realize a population level impact for both at risk drinkers and for probably dependent drinkers [15].