This was a 2-arm participant-blinded (single-blinded) RCT (1:1) to compare a mobile app that provided government guidance on alcohol consumption only (control arm) with the mobile app Drinks:Ration, a personalized app based on BCT principles. We hypothesized that the intervention arm would be efficacious in reducing alcohol consumption when compared to the control arm.

Both the control and intervention arms were delivered via the Drinks:Ration app. Participants in the control arm were given access only to the alcohol consumption guidance and a unit calculator based on guidance issued by the UK Chief Medical Officer. Those in the intervention arm were given access to the full version of the app, which included theoretically driven components and personalized messaging. This included individualized normative guidance alongside features designed to enhance participants’ motivation through interactive guidance focused on self-efficacy to help modify alcohol consumption. Participants in both arms were asked to use the app for a minimum of 28 days. After this, they could continue to use the app, but they did not receive personalized messaging. This was undertaken to assess the long-term effectiveness of the app.

This study was designed such that the control arm structurally resembled the intervention arm but excluded active intervention techniques, such as a drinks diary, drinks in pixels, and drinking zones (based on the Global Positioning System [GPS] location). This approach increased uniformity across the arms (eg, ensuring both arms received a digital intervention) and maintained treatment allocation concealment.

Data were collected at baseline and follow-up assessments at 84 and 168 days postbaseline. Additional questionnaires were collected on days 7, 14, and 21. This information was used to personalize the Drinks:Ration app for the intervention arm only. Please refer to the published trial protocol for further details [19].

This trial was approved by the local ethics committee of King’s College London (HR-19/20-17438).

Participants were recruited between October 2020 and April 2021 in succession via a clinical group, an existing research cohort [2], and social media [20]. The clinical group was derived from Combat Stress, a third-sector charitable organization that provides mental health services, including substance misuse management, to UK veterans. The research cohort was the King’s Centre for Military Health Research health and well-being longitudinal cohort study [2], where a sample of self-reported help seekers were identified and extracted. Social media platforms, such as Facebook and Twitter, were also used to promote the RCT via free and paid promotional advertisements with a link enabling potential users to express an interest in taking part (for further information, please see [20]).

Potential participants were invited to take part via email with an explanation of the study, a link to the participant information sheet, and instructions on how to download Drinks:Ration using a unique quick response (QR) code. Once participants had downloaded the app, they were asked to report alcohol consumption using the validated Timeline Follow Back for Alcohol Consumption (TLFB) [21] for the prior 7 days and confirm their military serving status. Those meeting the study eligibility criteria were allowed to proceed and complete the baseline questionnaire.

Eligibility was assessed at baseline. To be included in this RCT, participants needed to download the Drinks:Ration app onto an iOS or Android device; be aged 18 years or older; currently reside in the United Kingdom; consume at least 14 UK units (approximately 112 g of ethanol) of alcohol or more per week at baseline (hazardous or harmful levels of alcohol consumption); confirm that they currently receive, or had received, support for mental health symptoms in a clinical setting; provide a mobile phone number; and be a veteran of the UK AF.

A power calculation was performed based on Alcohol Use Disorder Identification Test (AUDIT) data previously reported from Combat Stress [5]. To detect a difference in alcohol consumption of 4 UK units (approximately 40 g of alcohol per week) between the control and intervention arms at day 84, with a 2-sided 5% significance level and a power of 80%, we needed a sample of 37 participants per arm with complete primary outcome assessments. We selected 4 UK units based on reductions observed in similar studies [8,22,23] and reductions observed in the feasibility trial of Drinks:Ration, which found a 7-unit decrease at week 4. To allow for attrition of 40%, we aimed to recruit a total of 124 participants (62, 50%, per arm). To account for an expected response rate of 30%, we estimated that we would need to invite 620 veterans to participate in the study.

Randomization was carried out automatically as part of the Drinks:Ration platform. When participants registered for the app, they were assigned a unique identifier and asked to provide their gender. They were then randomly allocated [1:1] to receive either the control or the intervention arm. Stratification was used to ensure equal gender distribution across arms. This is because those who identify as female only represent approximately 10% of the UK AF.

The randomization procedure was based on a list of random numbers computer-generated by the Drinks:Ration platform. All members of the research team were blind to participant treatment allocations except for authors DL and CW. CW conducted participant management, and DL led the development of the Drinks:Ration app, had access to raw study data, and conducted the primary analyses. Except for automated weekly backups, access to the data was disabled.

Participants were not informed of their treatment allocation. However, they may have been able to deduce their allocation condition based on app content. As the intervention was automated and delivered via an app, there was no contact between researchers and participants during the intervention unless a risk to health (adverse event) had been detected or if technical problems arose [19]. An adverse event in this study was defined as participants reporting (via the drinks diary or during contact with the research team) that they had consumed more than 25 UK units of alcohol within a 24-hour period. Once detected, a clinician would contact the participant over the telephone to check their health and provide signposting to other services (which are listed in the app). The clinician would not disclosure the treatment arm allocation.

Drinks:Ration (formerly called InDEx [15-17]) was developed by the King’s Centre for Military Health Research (at King’s College London) and the University of Liverpool following the Medical Research Council Complex Intervention Guidelines using co-design methodology in collaboration with end users. The app was designed to support veterans drinking at hazardous or harmful levels by providing bespoke advice and support over a minimum of 28 days. The app was developed without any organizational branding to promote use. The iterative development process, theoretical framework, and feasibility trial are reported elsewhere [15-17,19]. Briefly, Drinks:Ration was developed and tested with 5 core modules (see Figure 1):

Participants in the intervention arm completed additional questionnaires on their mood and general mental health each week. These responses were used to personalize app content, push notifications, and SMS text messages.

A summary of measures and data collection schedule is provided in Table 1 (and see [19]). Measures were the same for the intervention and control arms.

The statistical analysis plan was prospectively registered on the Open Science Framework [29]. Data were analyzed using Stata 16.1 MP (StataCorp).

Descriptive statistics were reported either as unweighted frequencies and percentages, the mean with the 95% CI, or the median with the IQR.

The primary and secondary outcomes were modeled using linear mixed effects models. Each outcome was tested in a separate model. Each model included up to 3 repeated outcome assessments, collected at days 28, 84, and 168. Repeated measures were clustered within individual participants, represented with a random intercept. The mixed model used all available information (ie, participants with at least 1 follow-up assessment were analyzed), leading to more precise estimates of the treatment effect.

We used multivariable binary logistic regression to assess whether baseline variables were associated with missingness in the primary outcome variable (1=missing primary outcome at day 84, 0=nonmissing). Each model included as covariates (1) time (measured as days since baseline), (2) a dummy variable to represent treatment allocation (0=control, 1=intervention), (3) a time × arm interaction term, (4) the baseline measurement of the outcome, (5) relevant covariates (age and gender), and (6) baseline variables associated with missingness (number of days off work due to alcohol consumption). Treatment effects were estimated as the difference between baseline and follow-up assessment (day 28, 84, or 168) and reported as the absolute alcohol unit difference between the arms.

Between-group effect sizes (Cohen d) were calculated by subtracting baseline total units consumed/AUDIT score from the day-of-assessment margin mean values (d=0.2, small effect; d=0.5, intermediate effect; and d=0.8, strong effect). The threshold for statistical significance reported in these analyses was P=.05.

The intention-to-treat analyses included all participants who completed at least 1 follow-up assessment (day 28, 84, or 168).

Between October 2020 and April 2021, 2708 individuals were invited to take part, of whom 2531 (93.5%) did not respond to the invite or declined to take part (n=150, 5.5%). In total, 177 (6.5%) participants were invited to complete a baseline assessment, of whom 54 (30.5%) were found to be ineligible based on study criteria (Figure 2). Therefore, a total of 123 (4.5%) participants completed the baseline assessment and were randomized into the study. Of these, 78 (63.4%) completed outcome assessments at day 28, 79 (64.2%) completed outcome assessments at day 84, and 27 (22.0%) completed outcome assessments at day 168. A total of 19 (15.4%) participants withdrew from the study by day 84. This included 7 (36.8%) participants who withdrew due to the limited functionality of the control version of the app.

Of the 123 participants, 62 (50.4%) participants were randomized to the intervention arm and 61 (49.6%) to the control arm. Baseline characteristics are shown in Table 2. The overall mean age was 47.6 years (95% CI 45.8-49.3), 117 (95.1%) participants were male, and 95 (77.2%) were married or in a long-term relationship. In addition, 87 (70.7%) had served in the army, and on average, the participants had served 14.4 years (95% CI 12.9-15.9) in the UK AF. The participants had a median AUDIT score of 16 (IQR 10-22) at baseline, and 66 (53.7%) were identified as having no probable PTSD. A total of 65 (52.9%) participants reported probable depression. Most participants entered the study with an Android device (n=67, 54.5%), and 79 (64.3%) participants completed the primary outcome assessment at day 84, with 76 (61.8%) completing the secondary outcome assessment.

For the primary outcome of the TLFB (units of alcohol per week) at day 84 (Table 3), we found that participants in the intervention arm had significantly larger reductions in self-reported alcohol unit consumption from baseline (marginal unit mean 56.3, 95% CI 50.6-62.0) to day 84 (marginal unit mean 28.1, 95% CI 21.1-35.1) compared with those in the control arm (marginal unit mean from 54.0, 95% CI 48.2-59.8, to 43.5, 95% CI 36.3-50.8; interaction P=.01). The effect size for the difference between the intervention and control arms in the mean change of units between baseline and day 84 was Cohen d=0.35, which is consistent with a moderate effect size.

Overall, we found that between baseline and day 84, weekly alcohol consumption reduced by –10.5 (95% CI –19.5 to –1.5) units in the control arm and –28.2 (95% CI –36.9 to –19.5) units in the intervention arm (P-value for the difference between arms at day 84=.003) at the primary outcome measure. The difference in unit marginal means was –15.4 (95% CI –25.5 to –5.4) units of alcohol in favor of the intervention arm.

There was evidence of a strong effect between the 2 arms by day 28 (Cohen d=0.50) but no evidence of a difference between the 2 arms by day 168 (Cohen d=0.11) for self-reported alcohol consumption (Figure 3).

For the secondary outcome change in the AUDIT score by day 84, we again found that participants in the intervention arm had significantly larger reductions in the AUDIT score from baseline (score marginal mean 16.3, 95% CI 15.0-17.5) to day 84 (score marginal mean 10.1, 95% CI 8.5-11.8) compared with those in the control arm (score marginal mean from 16.0, 95% CI 14.6-17.3, to 14.1, 95% CI 12.4-15.7; interaction P=.003). The difference was –3.9 (95% CI –6.2 to –1.6) points on the AUDIT score in favor of the intervention arm. The effect size for the difference between the intervention and control arms in the mean change in the AUDIT score between baseline and day 84 was Cohen d=0.48.

There was evidence of a strong effect between the 2 arms by day 28 (Cohen d=0.53) but no evidence of an effect between the 2 arms by day 168 (Cohen d=0.06) for the AUDIT score (Multimedia Appendix 3).

Sensitivity analyses of primary and secondary outcomes using complete case analysis produced the same patterns as those identified in the main analysis (Multimedia Appendix 4).

Over the entire study period (168 days), participants in the control arm used the app for a median of 1 week (IQR 1-2), initialized the app a median of 3 times (IQR 2-9), and had a median session duration of 60.9 seconds (IQR 35.7-75.6). Participants in the intervention arm used the app for a median of 3.5 weeks (IQR 2-6), initialized the app a median of 13.5 times (IQR 4-27), and had a median session duration of 43.8 seconds (IQR 32.3-67.9); see Table 4.

Participants in the intervention arm reported a median of 7 drinking days (IQR 4-11) during the first 28-day period, a median of 3.5 drink-free days (IQR 2-7), and a median of 12.8 units of alcohol per drinking day (IQR 4.4-16.5). A median of 18 push notifications (IQR 9-19) were sent to participants in the intervention arm, along with a median of 12 SMS text messages (IQR 10-14).

App use of participants in the intervention arm is shown in Table 5. Participants engaged with all modules of the app, but most of the app engagement was spent using the screening module (mean 201.0, SD 994.6) and the normative guidance module (mean 510.4, SD 1012.7).

The participants completed the MAUQ at day 28 (Table 6). They responded to a set of usability questions on a scale of 1-7, with a higher value indicating improved usability. Participants in the control arm reported a mean overall app usability score of 4.1 (SD 1.5), a mean ease-of-use score of 4.4 (SD 1.6), a mean interface and satisfaction score of 4.1 (SD 1.6), and a mean usefulness score of 3.6 (SD 1.7). These scores were lower than those of the intervention arm, which reported a mean overall app usability score of 5.9 (SD 1.1), a mean ease-of-use score of 5.9 (SD 1.2), a mean interface and satisfaction score of 5.9 (SD 1.1), and a mean usefulness score of 5.7 (SD 1.1).

In total, 2 (1.6%) of 123 participants were identified as having a single adverse event of consuming more than 25 units of alcohol within 24 hours during the study period. Following our risk protocol [19], a signposting booklet to relevant charities was provided, as well as a call with the study clinical lead. After a clinical interview, both participants were allowed to continue in the study. Their treatment allocation was not disclosed to the participants. No other adverse events were identified. No technical issues occurred during the trial.

This study is 1 of the few RCTs to date to examine the efficacy of a fully automated 28-day brief alcohol intervention delivered via a mobile app in a help-seeking sample of UK veterans with at least hazardous alcohol consumption. Help-seeking veterans were consuming on average 55 units of alcohol per week at the outset of the study, well above the 14 units of alcohol per week recommended as the maximum by the UK Chief Medical Officer. At the primary outcome (day 84), the difference between the estimated marginal means for the intervention and control arms was 15.4 units of alcohol lower in the intervention arm than in the control arm. A similar pattern was also observed for the AUDIT score, where the difference between the estimated marginal mean between the arms was 4.0 points on the AUDIT scale, also lower than that of the control arm. Overall, the intervention arm achieved significantly better reductions in alcohol consumption and AUDIT score. These effects disappeared at day 168.

The findings of this RCT demonstrate the efficacy of an automated brief alcohol intervention with personalized messaging for those who consume alcohol at least at hazardous levels and have sought help for a mental health difficulty [15]. The findings also mirror those obtained in other studies [7,12]. In particular, the findings demonstrate the efficacy of Drinks:Ration within a group that has been shown to be at increased risk of dual diagnosis. The between-arm difference compares favorably to stand-alone self-help interventions [33], and the differences are greater than those typically found with face-to-face therapies and the most successful therapist-guided interventions [34].

This study found significant differences between the arms for the primary and secondary outcomes at the primary data collection point (day 84), and the differences were even larger at day 28 when the first assessment took place. At follow-up (day 168), only 22% of participants responded, which limited our ability to discern differences at the timepoint. An alternative explanation is that the effect of the intervention reduced over time, so the long-term benefits of the Drinks:Ration app may need reinforcing beyond the intervention month (ie, the first 28 days).

The results of this RCT should be placed in the context of the wider literature. A recent literature review exploring the effectiveness of alcohol reduction apps and the availability of evidenced-based apps on top commercial app stores identified 21 articles representing 19 unique mobile apps [35]. Of these, 7 (36.8%) apps were targeted at adolescent drinkers, and the remainder on the general population. No studies that targeted the AF were identified. The overall effectiveness of the included interventions was mixed, with standards of reporting making direct comparisons difficult (eg, AUDIT-C, binge-drinking days, alcohol unit consumption). Drinks:Ration results compare favorably to all included studies, in so far as we identified the largest reductions in alcohol consumption; however, our base starting point was significantly higher than the general population.

Although there are no reported data on waiting times between referral and treatment in the United Kingdom, National Health Service (NHS) Scotland national drug and alcohol treatment waiting times for alcohol treatment were reported to be around 3 weeks or less between January 2021 and March 2021 [36]. The delay between referral and treatment may be an opportunity to deploy the Drinks:Ration app to support help-seeking veterans while they wait for formal treatment. Help-seeking veterans misusing alcohol attend fewer mental health appointments [6], probably because many are prevented from receiving treatment for mood disorders and PTSD until they have reduced their excessive drinking. The use of the Drinks:Ration app to support reduction in alcohol consumption could enable more help-seeking veterans to access services. In addition to the Drinks:Ration app, the use of personalized messages sent via SMS and push notification may have contributed to improved performance of the intervention arm.

The efficacy of the app in relation to our primary outcome is encouraging, but there are some issues worth considering. The recruitment for this study was lower than anticipated. We expected that 30% of eligible veterans would enter the study, but though it was difficult to estimate the true total number of eligible individuals, the percentage who consented to participate was less than 10%. The second potential problem is that the effect seen at 84 days may need a reinforcement to encourage persistent changes in behavior over the long term. This could be achieved by enabling personalized messaging over the entire life course of app usage.

There were 2 adverse events (involving 2 participants), which were unlikely to be caused by our app and more likely to be caused by the ongoing alcohol consumption of the individuals involved. This highlights the challenges of monitoring adverse events in remote/automated interventions for which the implementation of the app needs constant monitoring while being used. Finally, it is important to consider how Drinks:Ration can be integrated into the treatment pathway to support veterans prior and during treatment, while also monitoring the degree to which its efficacy transfers to a clinical context.

Several limitations of this trial should be noted. First, as already acknowledged, the majority of those invited to participate in the study did not take part. It is not possible to ascertain why these individuals chose not to take part, but it may be due to digital fatigue because of the COVID-19 pandemic. Therefore, we consider that our study assessed the efficacy of the intervention in those willing to engage with the app rather than effectiveness in the target population. Second, participants self-identified their military and help-seeking status among those recruited through social media and the status was not verified. Third, we only used self-reported data provided via outcome assessments and did not use data collected via the drinks diary. This decision to use the TLFB for assessing our primary outcome was to ensure comparability with other studies and to also ensure the control arm did not complete the drinks diary. This resulted in duplicating participant input, which could have created user frustration and negatively impacted usability and participation. Finally, this RCT was conducted during the COVID-19 pandemic. This period resulted in meaningful behavioral changes to how UK military veterans consumed alcohol due to lockdown. In a recent study of UK veterans, they were found to be drinking less alcohol during the first phase of the pandemic, reducing their hazardous drinking from 49% to 28% [37]. This may have reduced the available population that consumes alcohol at a harmful-to-hazardous level.

Our findings suggest that Drinks:Ration is efficacious in reducing alcohol consumption in help-seeking veterans and that wider uptake of Drinks:Ration in this population would be beneficial. However, strategies to increase use of the app and ensure that the gains in decreasing alcohol consumption persist over time need to be well thought out. This could be achieved by promoting app use and continuation of messaging and more personalized goal setting.