More than 1 in 10 US adults meet diagnostic criteria for alcohol use disorder [1], and 25% engage in binge drinking (4+ drinks for women and 5+ for men in 2 hours) [2]. In addition, 3 in 10 adults use other drugs, like cannabis, stimulants, and opioids [2], and 4% meet diagnostic criteria for a drug use disorder [3]. Together, alcohol and other drug use (ie, substance use) are major public health burdens [4,5] and cost the United States $500 billion annually [6,7]. Only 10% of those with substance use disorder (SUD) seek specialty treatment, and of those who do not seek SUD treatment, 95% do not perceive a treatment need [8]. The reach of public health approaches to boost clinical screening for substance use problems [9] may be enhanced by developing and testing novel, technology-based strategies to identify individuals who use substances, including but not limited to those engaging in harmful use [10]. If effective at identifying such individuals, technology-based strategies could then be paired with alternative and innovative interventions, which may be scaled up and deployed to help individuals with milder SUD variants who may be less likely to seek formal care [11], as well as individuals who might benefit from harm reduction–based psychoeducation. One such technology-based strategy leverages data from commonly used social network sites (SNSs) such as Facebook, Twitter, and Instagram. More than 70% of US adults have at least one SNS account, including nearly 70% who use Facebook, 37% who use Instagram, and 22% who use Twitter [12]. Among young adults specifically, who are disproportionately represented among individuals with SUD and hazardous drinking [2], Instagram is nearly as popular as Facebook, with 67% versus 79%, respectively, having ever used these platforms [12].

As detailed below, emerging research suggests that individuals’ SNS posts can be leveraged to detect substance use [10,13,14]. However, these computational strategies function more efficiently and reliably with increasing amounts of data [15]. Thus, if certain types of individuals are less likely to participate or have lower levels of participation on certain SNSs, then it may be more difficult to detect substance use in these individuals, even if models are ostensibly well designed. As such, a greater understanding of the associations among individuals’ demographic characteristics, SNS participation, and substance use can help inform the utility of these computational screening tools. In the current study, we examined associations between substance use, including at-risk drinking and drug use, and Instagram posts and tested whether such associations differ as a function of age, gender, and race/ethnicity.

In a prior study, our team [10] employed a novel deep neural network (ie, deep learning) framework [16] capable of processing data points with varying dimensions, including text and pictures of various sizes, to detect at-risk drinking and drug use from Instagram images, captions, and comments among a community sample of adults aged 18 and older, recruited primarily through Clickworker (Clickworker GmbH). Findings showed the deep learning model significantly predicted at-risk drinking, defined as having “had more than a few alcoholic drinks a day” at least once in the past year. In this model, the area under the receiver operating characteristic curve (AUROC) was 0.65, meaning there was a 65% chance that this novel classification model would correctly assign a higher score to a random positive example (participant reported at-risk drinking at least once in the past year) than a random negative example (participant reported no at-risk drinking in the past year). While the consensus is that AUROCs less than 0.7 provide low discrimination [17], the proof-of-concept study showed that a combination of visual and text-based SNS data could be used to detect self-reported at-risk drinking. The deep learning model was unable to detect other drug use any better than chance. Of note, the combination of images, captions, and comments evidenced superior detection of at-risk drinking compared with other combinations of Instagram content. To our knowledge, this study represents the only published research to date that has leveraged machine learning to detect individuals’ substance use from their SNS data in a nonclinical sample.

Studies showing that online forum content can be used to detect risk in clinical SUD samples are also instructive. Kornfield et al [13,14] used natural language processing in 2 studies, one of which also employed machine learning, to predict risk for negative substance use outcomes among individuals with both alcohol and other drug use disorders participating in a smartphone app online recovery forum. In the first study [14], they examined the utility of natural language processing with a Linguistic Inquiry and Word Count approach to predict binge drinking (5+ drinks for men and 4+ for women in 2 hours) in participants with alcohol use disorder. Individuals in this secondary analysis of a randomized controlled trial testing a smartphone-based intervention after residential treatment [18] were both randomized to receive the smartphone app intervention and participated in the application’s online forum between baseline (ie, residential treatment discharge) and the 4-month follow-up. Controlling for individual and system use characteristics (eg, number of messages posted), a greater percentage of words capturing swearing, negative affect, inhibition/control, and love and a lower percentage of words capturing higher-order cognitive processes (eg, insight) and achievement predicted past-month binge drinking at the 12-month follow-up [14]. In a related study including both participants from Gustafson et al [18] and participants with a range of SUDs recruited from primary care in Quanbeck et al [19], Kornfield et al [13] showed that decision tree machine learning algorithms can be used to detect recovery problems from online forum content. Algorithms included both Linguistic Inquiry and Word Count as well as Bag of Words natural language processing approaches. Recovery problems were determined by the research team based on a codebook informed by forum moderators’ perspectives regarding which posts warranted concern or intervention.

It is worth acknowledging the large and growing body of literature on the use of aggregated SNS big data to help surveil public health trends in drug use [20,21]. This big data surveillance literature and the use of SNS data to detect substance use risk share the use of SNS data as a proxy or marker of human behavior, which can then be downloaded for analysis. Aims of the current study, however, are intended to inform strategies that detect substance use at the individual level. Such a risk detection approach, when determined to be sufficiently reliable, would ultimately then be paired with interventions as mentioned above. When using SNS data in the context of drug use surveillance, data are analyzed in aggregate to identify macro-level trends and are thus outside the scope of this study of individual behaviors.

Following from studies of computational methods using SNS data to detect substance use, it is also important to examine whether, and in what contexts, individuals who use substances also engage with SNSs to inform the utility of these technology-based screening tools. In a meta-analysis of 17 studies targeting adolescents and young adults, Curtis et al [22] found that self-reported alcohol consumption, including both general consumption and measures of risky drinking, is moderately associated with self-reported and hand-coded SNS engagement, including both alcohol-related SNS posts as well as exposure to alcohol-related SNS posts of others (overall r=0.36). In a related meta-analysis of 7 studies [22] authors similarly found that alcohol-related problems, as measured, for example, by the Alcohol Use Disorders Identification Test [23], were also moderately associated with SNS engagement (overall r=0.37). Thus, younger individuals may provide enough SNS data for these computational strategies to detect at-risk drinking.

Most studies examining the association between SNS participation and substance use to date have focused on the association between Facebook engagement and drinking. In an exception, Instagram participation was negatively related to past-month days of cannabis use but positively related to alcohol use in emerging adult (ie, aged 18-29 years) Instagram users recruited through Amazon mTurk [24]. Given the literature’s focus on Facebook and drinking among youth, studies that include individuals across age groups and that examine the association between substance use, including both alcohol and drug use, and popular platforms other than Facebook, such as Instagram, may help build on this emerging scientific literature.

Knowledge about whether demographic characteristics moderate associations between SNS participation and substance use might further improve research on computational detection strategies. If, for example, there was an association between at-risk drinking and SNS participation only for younger individuals, this might reduce the utility of an SNS data–based model in detecting at-risk drinking for older individuals. To date, however, there are no existing studies that examine whether demographic characteristics moderate the association between substance use and SNS participation in community (ie, nonclinical) samples.

In the absence of prior work that might inform whether demographic characteristics moderate SNS–substance use relationships in the current study, it is worth mentioning that, consistent with the general population [12], emerging adults (aged 18-29 years) in SUD treatment report greater SNS participation compared with middle- and older-aged adults [25]. Similarly, in a nationally representative sample of US adults who resolved a substance use problem, emerging adults were more likely than their older counterparts to have used online resources, including but not limited to SNS platforms, to address their substance use or enhance their SUD recovery [26]. Also, in this same nationally representative recovery sample, Hispanic race/ethnicity relative to White race was related to a greater likelihood of recovery-related use of online resources, though men and women reported similar rates of this online help-seeking behavior. While these data are based on treatment and recovery samples, whereas the current study focuses on substance use in nonclinical samples, they suggest that any observed relationships between substance use and SNS participation might differ by age and race/ethnicity.

Emerging technology-based strategies to detect substance use with SNS data hold promise as scalable health risk screening tools. These computational strategies, of course, can only be effective among individuals who participate on SNSs. Even among those who participate, these screening tools are more powerful and, therefore, more useful, with increasing amounts of data. Thus, the real-world utility of these tools can be informed by research examining the associations between substance use, SNS participation, and demographic factors that moderate these relationships. Existing research suggests that substance use, including but not limited to hazardous use, is associated with greater SNS participation, but studies have focused primarily on drinking among young people on Facebook. To expand on this work, the current study targeted Instagram participation among adults of all ages and had the following aims: (1) to examine whether at-risk drinking and drug use is related to Instagram posts and (2) to examine whether these relationships between substance use and Instagram posts are moderated by age, gender, and race/ethnicity. We hypothesized that at-risk drinking would be related to more Instagram posts; we made no a priori hypotheses about drug use. We also hypothesized that the relationship between at-risk drinking and Instagram posts would be greater among younger individuals; we made no a priori hypotheses about other potential moderating demographic characteristics.

As detailed in Hassanpour et al [10], study participants were recruited in the winter of 2016, primarily through the Clickworker crowdsourcing platform, which compensates individuals directly for study participation. Participants were also recruited via word of mouth and SNS advertisements on social media. Each recruitment avenue directed participants to the study website, where, following online consent, they completed an online survey consisting of demographic information and the National Institute on Drug Abuse’s (NIDA’s) Quick Screen substance use screener (see the “Measures” section below for more details of the NIDA Quick Screen) [27].

Instagram permitted the use of their application program interface (API) to collect participants’ data with individuals’ permission. Specifically, upon completion of the survey, the online study site linked participants to the Instagram gateway where they could give their permission to allow an application developed by our team to communicate with the Instagram API. If participants granted permission, their posts were downloaded onto a secure server and stored under an anonymized unique identifier for restricted use in this study. The team piloted the procedure and application, facilitating download from Instagram’s API, on 81 individuals, whose data were not included in the final, analyzed sample.

Out of a total of 3117 participants, 861 (27.62%) had an account with 0 posts and 2256 (72.37%) had an account with one or more posts. For those with one or more posts, the average number of posts was 201.9 (SD 394.7), ranging from 1 to 4999 posts. For substance use, in the past year, 55.34% (1725/3117) engaged in at-risk drinking, 22.30% (695/3117) used drugs, and 15.27% (476/3117) used prescription drugs.

With respect to demographic characteristics, participants were aged 29.9 years on average (SD 9.86; range 18-73 years), with 40.52% (1263/3117) aged between 18 and 25 years, 42.38% (1321/3117) between 26 and 38 years, and 17.10% (533/3117) 39 years and older. For gender, 37.12% (1157/3117) identified as male, 62.14% (1937/3117) as female, 0.38% (12/3117) as transgender, and 0.35% (11/3117) as other gender. For race/ethnicity, 6.48% (202/3117) identified as Asian, 16.27% (507/3117) as Black, 10.27% (320/3117) as Hispanic/Latino, 62.59% (1951/3117) as White, and 4.40% (137/3117) as other (including Alaskan Native/Native American or Hawaiian Native/Pacific Islander).

For aim 1, we used logistic regression to examine the main effects of demographic characteristics and substance use variables on having an account with 0 versus any Instagram posts. We used negative binomial regression to examine the main effects of demographic and substance use variables on the number of Instagram posts for individuals. Since Instagram posts are count data and its distribution is overdispersed, we considered both Poisson regression, which assumes that the variance and mean are equal in the dependent variable, and negative binomial regression. Upon examination of the quantile-quantile plot (Multimedia Appendix 1), the distribution of Instagram posts was similar to a negative binomial distribution, supporting the use of a negative binomial regression to examine those factors associated with the number of Instagram posts [29].

For aim 2, we used the same logistic regression and negative binomial regression models as in aim 1 but examined the interaction effects between each of the demographic characteristics and at-risk drinking, drug use, and prescription drug use in the prediction of Instagram posts.

Of note, negative binomial regression models can yield inflated observed significance levels when there are high numbers of 0-count variables. In addition, the negative binomial regression models were unable to discriminate “true zeros” (ie, individuals who had Instagram accounts but 0 posts) from “artificial zeros” (ie, individuals who had 0 posts because they created accounts to participate in the study). In order to test whether removing those with 0 Instagram posts from the model would alter the pattern of findings, we conducted sensitivity analyses with zero-truncated negative binomial models, as in aims 1 and 2, with only individuals who had accounts with one or more Instagram posts (n=2256). The pattern of results for these sensitivity analyses was nearly identical, with the same significance testing results for all effects. Thus, we present only the primary analyses but include these tabulated sensitivity analysis results in Multimedia Appendix 2.

We used R 3.5.3 (The R Foundation) to conduct study analyses. All analyses tested significance at the .05 level. The institutional review board at Dartmouth College approved all study procedures. Data collection for this research project was conducted with informed consent from all participants and complied with the World Medical Association Declaration of Helsinki on Ethical Principles for Medical Research Involving Human Subjects.