Using social network analysis to examine alcohol use among adults: A systematic review

Data sources and literature search

A literature search of 4 databases (PubMed/MEDLINE, EMBASE, PsycINFO and Web of Science, Sociological) was conducted in March 2019 to identify studies that used social network analysis to evaluate the effects of social network characteristics on alcohol use. Keyword and terms used in search strategies varied based on the database to account for distinct indexing criteria and are described in detail in S2 Table. To identify additional studies not found in the literature search, the reference lists of relevant review articles of social network analysis [13, 44, 57] and included articles were reviewed. The literature search was conducted with guidance from an Education and Curriculum Librarian who serves as the Coordinator for Systematic Review Services at the New York University School of Medicine.

Screening

Duplicates were removed and screening of retrieved articles by title and abstract was conducted by two independent reviewers using Covidence systematic review software (Veritas Health Innovation, Melbourne, Australia. Available at www.covidence.org). Final inclusion was determined by two independent reviewers (JK & EG) screening the full-text of potentially eligible articles using the following criteria. Studies were eligible to be included in the review if they (1) were published in a peer-reviewed journal, (2) were written in English, (3) were conducted in human populations, (4) utilized a social network analysis design that included directed graphs (where there is directionality in the ties) and which data were collected beyond dyadic pairs (i.e., data linking participants were measured or inferred (e.g., known roommates were considered to be connected) and data were collected among alters for 3 or more connected individuals) and network measures were calculated based on this data, (5) evaluated an alcohol-related outcome, and 6) included a majority of participants that were adults (18 years or older). Studies were excluded if (1) the impact of social networks was assessed using simulations, (2) the study described qualitative research or was a case study. In the few cases on which opinions about inclusion differed, the reviewers met and reached consensus through discussion.

Quality assessment

We used a 12-item quality assessment tool (S3 Table) to evaluate study relevance and methodology. This tool was developed using modified sets of criteria from other quality assessment tools for assessing observational studies [65] and network studies [66]. The tool assessed description of: the research question, data collection procedures (data source(s), study setting, sample size, response rate, sample selection), measures used (exposure and outcomes), analysis of social network data, results and findings, strengths and limitations, and conclusions drawn.

Data extraction and synthesis

A data extraction form was developed to extract information on study objectives, study design and sampling approach, data collection method, setting and target population, participants, social network data collection procedures, outcome measurement, social network analysis methods used to analyze the data and calculated measures, statistical analyses used (as relevant to social network analysis), and key findings.

Identifying empirical studies

The literature search yielded 5,907 records for screening (5,477 after removing duplicates), and an additional 33 were included based on reference list review for a total of 5,510. The full text of 150 articles was reviewed to determine eligibility and 17 articles were judged potentially eligible and further assessed for quality and relevance (S4 Table). Scores on the quality assessment tool scale ranged from 78%-100% (mean = 95%). Based on these scores, all 17 articles were included in the review (Fig 1) [67–83]. Because the methodologies used in the included studies were heterogeneous, they did not lend themselves to meta-analysis. Therefore, a narrative synthesis of the studies will be conducted.

Population and setting

Among the included studies, 13 were conducted in the United States [67–70, 72, 73, 75, 76, 78, 79, 81–83], 3 were conducted in Europe [71, 77, 80] and one was conducted in Africa [74]. Most studies (n = 15) were conducted among young adults [67–74, 76–81, 83], the majority (n = 12) of which were among university students [67–71, 73, 76–81]. Five studies were conducted in community settings [72, 74, 75, 82, 83]. Mean age of the participants ranged from 18 years to 51 years. Proportion of male participants ranged from 25% [81] to 100% [71, 74]. Characteristics of the included studies are summarized in Table 1.

Sampling and study design

Sample selection procedures included the use of respondent-driven sampling, recruitment of peer groups, and the collection of complete sociometric social networks. The studies that used more robust sociometric social network sampling procedures [67–72, 77, 78, 81, 82] either collected data on complete (i.e. bounded) networks [77, 81] or used complete rosters but limited the number of peers that an individual was able to nominate [67–69, 71]. Two of the studies relied on peer nominations or recruitment without the availability of rosters [72, 74, 82]. Sample sizes ranged from 34 [81] to 12,067 [82]. Of the studies, 11 used cross-sectional data [67–70, 72–74, 77–80, 83], 5 used a longitudinal study design [71, 75, 76, 81, 82], and one was a naturalistic observation study in a ‘bar-lab’ [80]. Details of the individual studies, including objectives, study period and design, setting, participant characteristics, data sources, social network measures, analytic methods, and major findings are presented in Table 2.

Objectives

An objective of all the included studies was to assess how a social network characteristic was associated with alcohol consumption. Specifically, 12 studies assessed the association between characteristics of network members (e.g. peers’ weekly alcohol consumption or peers’ beliefs about alcohol) and alcohol consumption [68, 70, 71, 73–76, 78–80, 82, 83]. Eight studies assessed the association between characteristics of network structure, such as network size, network shape, network attributes, or position within a network, and alcohol consumption [67–69, 72, 77, 78, 81, 82].

Measures

The measures used in the included studies were heterogeneous. Table 3 provides definitions for the social network characteristics that the studies examined, which included: attributes of network members (e.g. peers’ weekly alcohol consumption or peers’ beliefs about alcohol), indegree/indegree centrality, betweenness centrality, outdegree/outdegree centrality, reciprocity, prestige, group integration, density, transitivity, closeness, gender heterophily, effective size, cluster identification, and homophily.

All the studies measured some form of alcohol consumption over a specific time period as the outcome. These included frequency of alcohol consumption, quantity of alcohol consumption, and frequency of binge drinking. Some studies used a single item to measure alcohol consumption, while others measured multiple forms of alcohol consumption using previously validated scales. One study looked at frequency of drug and/or alcohol use with a social network member in the past 6 months without distinguishing between drug and alcohol use [72]. The naturalistic observation study relied on observed counts of alcoholic drinks consumed [80], all the other studies relied on self-reported data, including one that compared perceived levels of alcohol consumption by peers to self-reported alcohol consumption by participants [79].

Statistical analyses

The statistical analyses used to assess social network characteristics were also heterogeneous, although the majority used some form of regression modeling correcting for nonindependence of observations/autocorrelation among network members. The longitudinal studies used Generalized Estimating Equations (GEE) to account for repeated measures. Overall, the methodologies described by the included studies were rigorous enough that the studies met inclusion criteria based on the quality assessment tool (mean = 95%).

Findings

All the included studies reported a statistically significant association between a social network characteristic and an alcohol-related outcome. The different types of social network members whose alcohol consumption (or in one study, their beliefs about alcohol [76]) was associated with participants’ alcohol consumption, included: peers/friends [69, 70, 74, 76, 78, 80, 82, 83], dorm mates [68, 73, 79], drug network members [75], and parents [76]. In one study, participants’ alcohol consumption was associated with alcohol consumption of relatives’ and friends’ but not that of immediate neighbors or co-workers [82].

Social network members’ drinking behaviors were associated with participants’ drinking behaviors in multiple ways. For example, among first-year university students in Germany, participants’ quantity of drinking was associated with friends’ quantity of drinking but participants’ frequency of drinking was only associated with friends who also identified the participant as a friend [71]. At a medium-sized university in the Northeast US, weekly volume of alcohol consumed among nominated peers was significantly associated with that of participants but alcohol problems (measured using the 24-item Brief-Young Adult Alcohol Consequences Questionnaire [84]) were not [68]. In a subsequent study that sampled the entire first-year class at the same university, participant's binge drinking frequency was positively associated with both perceived and actual binge drinking frequency of important peers, and resistance to peer influence weakened the effect of perceived peer binge drinking on participant binge drinking, but did not interact with actual norms [70]. In a sub-sample of past-month drinkers from the previous study, the number of drinks consumed by your peers on their heaviest drinking occasion was associated with greater drinking quantities on one's own heaviest drinking occasion [78]. Among black men who have sex with men (MSM) in a community setting in South Africa, individuals whose social networks included a higher proportion of hazardous drinkers were more likely to be hazardous drinkers themselves [74]. In the naturalistic observation study, peer group alcohol consumption was the strongest predictor of participants’ alcohol consumption [80]. Data from the Framingham Heart Study was used to show how the impact of peers drinking habits on one’s own drinking behaviors diminished across degrees of separation, i.e. one’s own friends influence drinking more than the friends of one’s friends [82].

A few studies looked at peer effects in social networks by examining homophily [68, 81, 82]. One study showed that quantity of alcohol consumption clustered within nominated peers of a university residence hall, then used regression to show that participants’ quantity of alcohol consumption was associated with nominated peers’ quantity of alcohol consumption [68]. Two studies observed that homophily increased over time (i.e. social networks became more homophilous), and provided support for induction because the directionality of friendship nominations mattered after controlling for participants’ previous alcohol consumption [81, 82].

Certain studies assessed associations between other aspects of social network structure than homophily (sometimes in addition to looking at the association between social network members drinking behaviors and participants’ drinking behaviors), such as network size, network shape, network attributes or position within a network, and alcohol-related outcomes [67, 72, 77]. For example, among university students in the US, outdegree was positively associated with number of heavy drinking days [67]. In the same study, betweenness centrality was positively associated with alcohol-related problems, with a stronger association among women [67]. In a subsequent study that sampled the entire first-year class at the same university, popularity (indegree) was positively associated with participants’ alcohol consumption, binge drinking frequency, and alcohol-related problems [69]. In a sub-sample of past-month drinkers from the previous study, being present at other peers' heaviest drinking occasions was associated with greater drinking quantities on one's own heaviest drinking occasion [78]. Among university students in Belgium, indegree was positively associated with binge drinking, while gender heterophily and effective size were negatively associated [77]. Among young, MSM in a community setting in the US, transitivity was positively associated with frequency of alcohol and/or drug use [72]. Taken together, these findings suggest that characteristics of social networks and one’s position in a social network also are associated with alcohol-related outcomes.

Critical evaluation

Almost all (15 out of 17) of the included studies were conducted in young adults [67–74, 76–81, 83], mostly in university settings [67–71, 73, 76–81], with just 2 in community settings [72, 74]. Two studies were conducted among adults whose mean age was greater than 30 [75, 82]. Considering the ways that social networks are formed, it is possible that school settings, in which individuals live in close proximity to similar-aged peers who share a large number of commonalities, have a qualitatively different effect on the ways that social networks are formed and how information is transmitted through these networks than networks of older adults or those in community settings.

The variation in measures used, both of social network characteristics and alcohol consumption, makes it challenging to summarize the body of evidence for how social networks influence alcohol consumption in young adults, and vice versa. Also as a limitation, many of the studies were cross-sectional [67–70, 72–74, 77–79, 83]. Cross-sectional designs allow researchers to test for correlations between an individuals’ alcohol consumption and that of their peers, but not to distinguish whether the correlations result from selection, confounding or induction (i.e. that the observed homophily with regards to alcohol consumption observed among social networks of adults is due to the tendency for people to befriend those similar to themselves, the effects of shared environments, or the spread of drinking behaviors within networks). Two studies examined directionality and found support for induction [81, 82].

Some studies collected data on complete networks [77, 81] or used complete lists of members in a social network (i.e. rosters) but limited the number of peers that participants were able to nominate [67–71, 73, 78, 79]. Other studies relied on peer nominations or recruitment without the availability of rosters [72, 74–76, 80, 82, 83]. The studies with less complete networks suffer from missing data that raise doubt about whether the results apply to the entire network, and whether the network metrics are accurate. Dyadic index-peer data can provide useful information, but is incomplete social network data because it limits researchers to investigation of the effects of peer consumption rather than social network position or other characteristics of social network structure. Such studies contribute only limited information about how social networks influence adult alcohol use. Furthermore, results that rely on incomplete social network data are likely to suffer from selection bias because peers who are closer to the participant (and thus, more likely to share similar characteristics) are more likely to be selected into the study. While many of the studies in adults were limited in the completeness of social network data collected; collectively, they featured a wide variety of types of social network members studied (e.g. peers/friends, classmates/dormmates, drug network members, relatives/parents, neighbors, co-workers). In a few of the studies [72, 74, 83], the nature of the relationships among participants was not even known, just that there was some sort of connection between them because one participant’s recruitment was attributed to the other. This is reflective of how peer groups are more expansive and diverse in adulthood [60], but also more difficult to capture completely. In nearly all cases, these social network members were found to influence participant alcohol consumption. Lastly, many of the included studies were conducted among specific populations (e.g. fraternity brothers [81], injection drug users [75], men who have sex with men [72, 74]), raising questions about the generalizability of the findings.

Implications for research and intervention

There are a few important take-away messages from this first review of the growing number of empirical studies that have used social network analysis to explore alcohol use in adults. First, peer alcohol use and other social network characteristics (e.g. network attributes or position within a network) were associated with adults’ alcohol consumption across studies. Second, we identified a lack of research on the impact of social networks on alcohol use in middle-age and older adults, especially those residing in community settings (i.e. without explicit boundaries). Future empirical research should work to address these gaps in our understanding. Efforts should also be made to reduce heterogeneity in social network analysis studies (e.g. agreeing on standardized definitions of social network measures, assessing and reporting the results of them consistently) to facilitate generalizability. This review informs alcohol researchers, health service providers, and policymakers about how social networks have been studied, thus far, to better understand alcohol consumption. As social networks appear to play a role in the consumption of alcohol in young adulthood, this suggests potential for interventions that utilize social networks to help reduce the burden of harmful alcohol consumption [92]. Interventions that utilize social networks to promote behavior change are increasingly available [44], calling for studies of their feasibility and efficacy in reducing alcohol consumption in young adult populations.