Despite numerous intervention and policy efforts, risky drinking (ie, hazardous levels of consumption resulting in increased risk for consequences) among youth in the United States remains a major public health issue. Although only 1.8% of youths aged 12 to 17 years and 10.0% of those aged 18 to 25 years met criteria for an alcohol use disorder in 2017 in the United States [1], risky drinking is common. For example, as one indicator of risky drinking, past-month binge (eg, ≥4 drinks for females and ≥5 drinks for males) drinking rates are 10.2% for ages 16 to 17 years, 26.2% for ages 18 to 20 years, and 45.4% for ages 21 to 25 years [2], although these may be underestimates because of possible underreporting in national surveys (eg, average past-year quantity or frequency questions) [3]. In fact, risky drinking among young people is associated with increased risk for other drug use, adverse health consequences (eg, injury and overdose), and development of substance use disorders [4-7]. Accordingly, late adolescence and emerging adulthood is a critical developmental juncture, distinct from childhood and adulthood, during which rates of alcohol and other drug use peak [8-10]. For example, 6.5% of adolescents and 22.1% of emerging adults report past-month cannabis use [1]. Thus, scalable, early interventions are urgently needed to address risky drinking and concomitant health risk behaviors (eg, other drug use and driving under the influence) among adolescents and emerging adults to disrupt risk trajectories. Here, we present the theoretical rationale and protocol for a randomized controlled trial (RCT) of social media–delivered interventions for risky drinking among adolescents aged 16 to 24 years recruited nationally.

The conceptual model guiding our intervention is rooted in social cognitive (eg, theory of planned behavior [11] and social learning [12]) and social ecological [13] theories, emphasizing the role of individual and social influences on alcohol use by adolescents and emerging adults. Furthermore, our intervention is implicitly grounded in a resiliency framework [14,15]. Across development, evolving interactions between individual and social risk and protective factors during the establishment of new roles (eg, relationships and employment) [16] can decrease or accelerate alcohol use trajectories [16-18]. Individual risk factors associated with alcohol use include low perceived risk of use, perceived norms, and mental health issues (ie, depression and anxiety), whereas disapproval of use, parenting practices, and protective behavioral strategies are protective [16,19-27]. Although parents are important during younger ages [28], peers comprise the most robust social influences on substance use among adolescents and emerging adults [16,29-38].

Over the past decade, social media has become increasingly prevalent in the day-to-day lives of young people, creating additional opportunities for exposure to positive and negative peer influences [39,40]. Social media content is user generated and constantly changing, providing frequent exposure to web-based peer influences, potentially resulting in reinforcing spirals of increasing exposure and involvement with alcohol use behaviors over time [41,42]. Although it is well known that offline peers can exert tremendous influence on alcohol use among youth [16,29-32], recent data suggest that online peers also influence alcohol use [43-45]. Emerging adults who consume more alcohol, for example, have more Facebook friends [46] and post more references to parties on Facebook than those who use less alcohol [47]. Among high school students, higher alcohol use is related to reports of friends posting alcohol content on social media [43], and in a laboratory study, researchers found that teens viewing Facebook profiles that contained positive references to alcohol had more positive attitudes and willingness to drink alcohol than teens who did not view these profiles [48]. As offline peer disapproval of risky substance use can be a protective factor [49,50], online peer disapproval of alcohol use may function similarly. Research shows that posting positive portrayals of alcohol use on social media is related to consumption among the youth [43,51]. Thus, social media provides an appealing platform for the delivery of alcohol interventions, wherein peer influences could be harnessed instead to promote harm reduction or reduced consumption. As described earlier, alcohol use is associated with other drug use; thus, social media interventions could be useful for targeting concomitant drug use, particularly because mentions of other drug use are also prevalent on social media [52,53]. For example, one study showed that more than one-third of a college student sample had seen a picture of a friend smoking cannabis posted on social media [54].

A common feature of social media use among adolescents and emerging adults is the frequent use of more than one platform, which reflects increased smartphone ownership and Wi-Fi access. As of 2018, 95% of teens reported having a smartphone or access to one, of which 45% reported they are on the Web almost constantly [55]. Among emerging adults (aged 18-25 years), 88% use Facebook, compared with 68% for Snapchat, 59% for Instagram, and 36% for Twitter [56]. In addition, engagement is more frequent for Facebook, with 74% of all users checking it daily, compared with 61% for Snapchat, 63% for Instagram, and 42% for Twitter [56], and 51% of Facebook users log in several times per day [57]. In contrast, among adolescents (aged 13-17 years), 51% use Facebook (which has declined in recent years) [58], compared with 69% for Snapchat, 72% for Instagram, and 32% for Twitter [56]. It may be that adolescents and emerging adults who use Facebook regularly differ from those who do not, which could affect the utility of interventions. For example, data suggest that a larger proportion of teens from lower income households use Facebook than those from higher income households [59]. Thus, when choosing a social media intervention platform to reach both adolescents and emerging adults, there is no clear single best choice. Furthermore, as trends in social media use shift over time and/or within demographic groups, there may be unique opportunities to leverage content for delivery across various emerging platforms with shared features (eg, ability to post personal content, videos, articles, etc) to reach certain at-risk groups.

To date, there a very few social media interventions to reduce risky drinking (and/or other illicit or prescription drug misuse) among young people [60], with a recent publication describing the development of a tobacco and binge drinking intervention [61]. Most prior research testing early interventions for alcohol (and other drug use) has examined interventions delivered by therapists and/or static computer programs, with demonstrated efficacy in medical and university settings [62-74]. Overall, effect sizes are modest [75], with newer studies in the substance use field and other health fields testing technology-driven methods to extend delivery [76-79]. An advantage of social media interventions is that they can be designed to blend therapist and computerized interventions to deliver dynamic, evolving content; harness online peer influences; and provide access to electronic coaches (e-coaches), which can increase exposure to content at the time the person chooses to engage. Social media interventions (typically delivered over 8-12 weeks), addressing other health outcomes (eg, exercise/weight, HIV risk reduction/sexual health, and smoking cessation) among varied samples (eg, postpartum women, college students, and general community), have demonstrated promising effects [80-88], supporting the potential of this approach to address alcohol and other drug use.

Prior social media interventions have used Facebook for delivery, likely because it remains the most popular social media site among emerging adults and it has unique features that support intervention delivery. For example, Facebook allows private, secret groups to address privacy and confidentiality concerns (which are not searchable or viewable by others and can be joined by invitation only). In addition, the content is sorted into threads, promoting group interaction, with active conversations bumped to the top of the group or one’s newsfeed. Moreover, Facebook content does not disappear (eg, as in Snapchat), so it can be viewed an unlimited number of times and discussions can be revisited, as group members post new comments. Finally, Facebook does not restrict the character count of posts, which is a limitation of other platforms.

Critical issues related to designing social media interventions are exposure, dose (engagement or response showing the degree to which content may be processed), and diffusion (reach or interaction among online peers via shares, comments, etc) [89]. Exposure can be measured using metrics of reactions (eg, likes), comments, replies, and posts to Facebook groups. For instance, researchers found more than half (approximately 63%) of participants in a physical activity intervention condition reported visiting the Facebook group 2 to 3 times per month during a 12-week intervention period; among those who posted in the Facebook group more than once, they averaged 8 interactions each over 12 weeks [80]. Thus, an important methodological question is related to how to encourage engagement, increasing dose and diffusion [89]. Our study sought to accomplish this in 2 ways. First, content was informed by social marketing research tips regarding characteristics of posts that increase interaction: (1) give (eg, photo/video contests), (2) advise (useful tip for concerns, eg, coping strategies), (3) warn (dangers could affect anyone, eg, overdose and impaired driving), (4) amuse (amusing photos/videos), (5) inspire (moving quotes or stories), (6) amaze (amazing pictures or facts, eg, norms), and (7) unite (brag about group membership and social support) [90]. Second, to our knowledge, no researchers have compared intervention conditions that vary incentives for engagement. Increased interaction via incentives among peers could theoretically reinforce group interactions, increasing dose, which is thought to result in behavior change. Thus, we sought to compare an intervention condition that provided modest financial incentives for engagement as measured by daily interactions (ie, posts with status updates or comments to another’s post) with a condition that did not provide incentives. Thus, in addition to comparing the interventions to an attention control condition to determine efficacy, our goal was to examine whether externally incentivized interaction produced greater engagement, and if so, whether that enhances intervention efficacy, relative to the nonincentivized intervention condition containing organic, individually motivated interactions.

Finally, sentiment analysis (eg, examining the relative positive or negative valence [tone] and arousal [activation] in text) [91,92] is a potentially useful tool to understand characteristics of engagement in social media interventions [93]. Using state-of-the-art software, natural language processing can evaluate slang and common misspellings, with 85% accuracy [94,95]. To date, sentiment analysis has been applied to smoking cessation interventions but has not been applied to alcohol interventions, although researchers are coding content of social media related to alcohol use [96,97]. Thus, because our interventions sought to encourage interaction within secret groups, sentiment analysis is an innovative approach to understanding the characteristics of engagement (eg, valence and arousal) and alcohol use outcomes.

Goal of This Study

We recruited adolescents and emerging adults using Facebook and Instagram advertisements and conducted web-based screening to enroll risky drinkers in an RCT comparing 3 conditions: (1) social media intervention (SMI) with incentives (SMI+I), (2) SMI only, and (3) attention control condition, with follow-up assessments at 3, 6, and 12 months. Interventions comprised access for 8 weeks to unique, private secret Facebook groups facilitated by e-coaches (supervised by licensed therapists), with dynamic content addressing motives for risky drinking and reducing consumption as well as concomitant risk behaviors (other drug use). The attention control condition included access for 8 weeks to entertaining content (eg, sports, lifestyle, fun, etc). As described earlier, we used 2 intervention approaches, with and without financial incentives for participation, and we will measure engagement within the intervention groups. By providing incentives for participation in one condition, we attempted to harness participants to provide group support, thereby delivering intervention content, facilitated by e-coaches.

The specific aims are to (1) test the efficacy of the 2 intervention conditions compared with the control, in reducing alcohol consumption and alcohol-related consequences at 3-, 6-, and 12-month follow-ups; (2) compare the intervention conditions in participant engagement and efficacy in reducing alcohol consumption and alcohol-related consequences at 3-, 6-, and 12-month follow-ups; and (3) examine how level of engagement in intervention conditions (eg, engagement metrics) and characteristics of intervention engagement (sentiment analysis) relate to alcohol use outcomes in the 2 intervention conditions. The secondary aims include examining the efficacy of the interventions on other drug use, moderators of outcome, and conducting cost analyses. This paper describes the study protocol in relation to the primary aims.

The study procedures were approved by the University of Michigan Institutional Review Board (IRB), and the study was registered in Clinicaltrials.gov (#AA024175). We received a waiver of parental consent for all aspects of the study for youths aged 16 to 20 years (the age of majority varied based on state residence). The rationale for this waiver was based on (1) the determination of teenage participants as mature minors (ie, they can understand the study risks), with decisional capacity to promote health-seeking behavior including substance use treatment [98]; (2) the fact that disclosure of high-risk behaviors may increase the risk of adverse effects on participants’ well-being because of potential reactions from parents (eg, rejection and abuse); and (3) the study could not practicably be carried out without this waiver, given potential bias in participation because of fear of disclosure of risky drinking to parents [99]. Furthermore, our study involved a two-phase consent process, with separate web-based consent obtained for the screening and RCT phases. Confidentiality and privacy were also enhanced by requiring participants to agree to abide by our own User Safety Agreement (see the Interventions section). We obtained a Certificate of Confidentiality from the National Institutes of Health.

Using recruitment via social media advertisements, we enrolled and randomized 955 adolescents and emerging adults (aged 16-24 years) in an RCT comprising the 2 intervention conditions and the control condition. Participants were assigned their conditions for a period of 8 weeks and were prompted to self-administer follow-up assessments at 3, 6, and 12 months postinitiation of groups. All assessments and interventions occurred on the Web, with surveys administered through Qualtrics [100].

Potential participants were recruited in 10 waves, separated by age (16-20 and 21-24 years) via paid advertisements on Facebook and Instagram. Each wave contained an average of 95.5 participants, which helped ensure that the 3 groups in each wave contained approximately 30 participants (mean=31.8 participants per group) to allow for sufficient online group interaction.

On the basis of prior work [101], Facebook/Instagram advertisements were initially placed by setting the audience location to include users in the United States. Advertisements were also specified to be displayed to users with certain demographic characteristics (ie, age groups 16-17, 18-20, 21-22, and 23-24 years and English-speaking users) and detailed targeting displayed advertisements to users who liked Facebook pages related to alcohol (eg, popular brands, drinking games, etc). Starting in wave 5, we added user characteristics to increase the recruitment of racial and ethnic minorities using affinity targeting within Facebook Ads Manager. After wave 6, ethnic affinity targeting was temporarily removed from the Facebook Ads Manager but was available again and used for waves 9 and 10. Each advertisement featured headlines to encourage potential participants to take the survey (eg, “Drink alcohol? Participate in a research study; earn $$$ for your time.”). We used 3 images (from Facebook Ads Manager and stock photos) of alcohol/individuals with alcohol and one image of the study logo. To encourage minority representation in the sample, advertisements pictured individuals of varying races/ethnicities. Also starting in wave 5, white individuals and females were informally excluded after preset quotas were filled. Advertisements initially directed participants to the study website, but starting in wave 2, advertisements led participants directly to the consent page and screening survey. The study’s website URL was provided to participants throughout the study (eg, in Facebook secret groups and texting/email communications).

Among screening-eligible participants, the 21-item web-based screening survey was used to determine RCT eligibility using a past 3-month version of the Alcohol Use Disorders Identification Test-Consumption (AUDIT-C [102-104]; where binge drinking was defined at ≥4 drinks for women and ≥5 drinks for men) embedded among other standard items querying demographics and other substance use. To ensure real people completed the survey (as opposed to bots), the web-based screening consent page included a Completely Automated Public Turing test to tell Computers and Humans Apart. Participants who reported risky drinking the past 3 months (AUDIT-C score: ages 16-17 years: ≥3 females and ≥4 males and ages 18-24 years: ≥4 females and ≥5 males [105,106]) were eligible for the study.

Before enrolling eligible participants, we reviewed their screening data as a second step to ensure data integrity and to deter fraudulent participation. Procedures included checking data for duplicate internet protocol addresses, survey completion times >60 seconds, and the existence and legitimacy of the participant’s Facebook profile based on published recommendations [107,108]. Once initial identity verification procedures were passed, eligible youths were sent an email invitation to participate in the study, with a link that automatically directed them to the RCT consent form followed by a web-based baseline survey and a contact information form. Participants were informed in the consent form that “the purpose of the study is to develop and test social media interventions to help young people reduce risky behaviors, such as alcohol use.” To help ensure identity and age, as part of the baseline procedures, participants were required to upload a selfie containing a handwritten sign with the date and time that included their head and shoulders. Study staff compared the selfie with the participant’s Facebook profile for verification before randomization and group assignment. In rare cases where a participant’s Facebook profile did not already contain a photo of themselves, we asked them to temporarily upload a second photo (different than the selfie) for real-time, immediate verification against their time- and date-stamped photo.

Following the web-based baseline assessment and selfie verification, participants were randomized to 1 of the 3 conditions. Given differences in severity of drinking by age and sex, which could affect response to the intervention [2], computerized, stratified random assignment by sex and age group (16-20 and 21-24 years) took place within condition, in blocks of 20 within cells to equalize randomization over time. Randomization occurred by a computer algorithm generated with supervision by the data manager; thus, research staff were not able to manipulate condition assignment. Given e-coach interaction in groups, it was not possible to blind staff to condition assignment; regarding participant blinding, although participants were not told whether they were assigned to an intervention or control group, the control group did not receive alcohol content; thus, as with most behavioral trials, it is possible that participants discerned their condition assignment. Specifically, the consent form described the 2 intervention conditions, “You will have access to the secret Facebook page that will deliver health information focused on reducing risky behaviors, including alcohol use.” The consent form described the control condition as, “You will have access to the secret Facebook page that will share news information about things like entertainment, sports, weather and world news.” In addition, for the SMI+I group, participants were informed that they would earn points for interacting on the group page and be paid for the points earned, so participants were not blinded to being assigned to the payment condition. After randomization, participants were sent a friend request from an e-coach; once participants accepted the request, they were added to their corresponding secret group where the 8-week condition was delivered.

Consecutive web-based assessments, mirroring the baseline survey measures, were distributed by a research assistant using a generic study email address at 3, 6, and 12 months after group initiation. Participants were assured that e-coaches would not view their individual outcomes on their self-administered follow-up surveys.

Each participant received a US $30 Amazon gift card code for completing the web-based baseline survey and providing a selfie, which took approximately 30 min. Compensation for follow-up assessments was US $35 for the 3-month assessment, US $45 for the 6-month assessment, and US $55 for the 12-month assessment. Participants in the SMI+I condition received incentives to encourage interaction, earning US $1.00 for each day they posted text and/or images in the secret group (ie, status update, comment, reply, or share) for a maximum of US $56 per participant over 8 weeks. Note that likes or reactions (eg, heart and sad face) were not incentivized. Incentives were paid weekly via an electronic Amazon gift card by study staff (student research assistants, e-coaches, and/or supervisors) who reviewed posting data to confirm the number of days on which participants posted.

Similar to prior work [86], participants in the control group were given access to an 8-week attention placebo entertainment condition using private secret Facebook groups. Weekly topics included posts related to nonalcohol or drug-related topics that involved entertaining content (eg, sports, lifestyle, fun, etc). E-coaches posted content within the groups daily, at the same intervals as the intervention group posts. As in the intervention groups, the User Safety Agreement was enforced, the crisis line information was displayed in the group cover photo 24/7, and the downloadable files section included national resources (eg, suicide hotlines, mental health, and substance use treatment), the User Safety Agreement, and a copy of the consent form.

We will use generalized linear mixed models (GLMMs) to examine treatment effects and changes in the dependent measures (for both primary and secondary outcomes). We chose GLMMs for 2 primary reasons: (1) GLMMs adjust for correlations between data points (eg, repeated measurements on individuals); and (2) within GLMMs, one can retain participants who do not complete all follow-up assessments in analyses. As the primary outcome variable, alcohol consumption, is unlikely to have normally distributed errors and is effectively integer valued, the Poisson distribution, allowing for overdispersion [125], is a natural choice. This assumption will be scrutinized, and, as needed, modifications (eg, zero inflation) [126] and alternative families of distributions (eg, negative binomial) will be considered. For models treating level of interaction, quantified by engagement metrics, as the dependent variable (aim 2), we do not have any a priori judgments about the appropriate distributional family, and this will be assessed based on the observed distribution. Our initial choice will be the Gaussian (normal) distribution. In all cases, we will implement an intent-to-treat analysis [127]. Analyses pertaining to secondary outcomes of other drug use will be conducted in parallel manner to the primary outcomes analysis.