Social Selection and Social Influence Theories

There are two competing theories which contend different directions of influence between peers and substance use. Social influence theory [16] suggests that observing substance use by network members leads to changes in the substance use among others in the network, such that the other member's substance use becomes more like their friends' behavior [17,18]. In contrast, social selection theory [16] suggests that a person who is using drugs would change networks in order to spend more time with others who are also using drugs [17]. The key difference between these two theoretical perspectives is the direction of influence.

Several studies have examined social selection and social influence processes in relation to substance use among adolescents. Kandel and colleagues [5] found that having peers who use marijuana was a strong predictor of initiating use of marijuana or other illicit drugs among adolescents. Various aspects of the peer network explained 68% and 55% of the variance in marijuana and alcohol use, respectively [10]. Another study of adolescents found that youth who reported being influenced by their friends' behavior were more likely to report drug use [11]. Cleveland and Wiebe [12] found that the similarity between adolescent friends in tobacco and alcohol use was moderated by the level of substance use by others at the school. In addition, Johnson [14] found stronger evidence for the influence of peers on alcohol use compared to marijuana use. This relationship did not vary by gender.

A Social Network Approach

Social networks are an approach to conceptualize and measure the people with whom an individual (index) has social interactions [7,19] and has proven to be particularly instructive in examining these two theories in the context of adult alcohol use [17,20]. A number of studies have used social network data to examine drug use [7,21-24] and have confirmed that the behavior of the social network is associated with the drug behavior of the index member. For example, Latkin and colleagues [7] found that the number of network members the index member shares drugs with was associated with the index's frequency of injection. Among injection drug users (IDUs), turnover in the social network is associated with more HIV risk behavior [21]. Exposure to network sharing of filters has also been linked with index filter sharing among IDUs [22]. Furthermore, a preliminary analysis for the present study found that quitting drugs was associated with a subsequent decrease in social network drug use [23].

Two previous longitudinal studies have used network data to examine the association between peers and alcohol consumption among adults [17,20]. In a study of college students, Reifman and colleagues [17] used structural equation modeling (SEM) with an autoregressive cross-lagged panel design to test bidirectional influences between alcohol use and network drinking. In their model, both alcohol use and network drinking were modeled simultaneously over time. Each was a function of a stability path from the same variable at the previous assessment, the cross-lagged path of the other variable at the previous assessment, and a correlation of the residuals for variables measured at the same assessment.

Reifman et al. found that network drinking predicted later respondent drinking, but respondent drinking only predicted network drinking for one of the two cross-lagged paths examined. The authors also found that drinking by key network members (i.e., those that the respondent rated the highest level of closeness and those who the respondent spent the most time with) was correlated with respondent alcohol misuse. They further examined the proportion of change in network member drinking that was accounted for by change in the behavior of network members of time and changes in the composition of the network itself in a microsocial analysis. Consistent with social selection, changes in network behavior were mostly accounted for by the gaining and dropping of unique network members (i.e., turnover due to seeking new friends). Also using cross-lagged autoregressive SEM, Bullers and colleagues [20] found evidence of a bidirectional relationship between individual and network drinking among adults.

Despite the increasing number of studies examining the association between drug use and network drug use, there are still a number of gaps in the available research. For example, much of the research examining social influence and social selection theories has focused on adolescents rather than adults [5,10-15]. Previous research on social networks and illegal drug use has largely focused on initiation rather than maintenance or cessation [25]; additional research is needed to determine the direction of the association between illicit drug use and peer drug use at stages of use beyond initiation. Furthermore, it is unclear if findings regarding the association between use and peer alcohol use previously observed in young adults [17,20] apply to illegal substances..

The current study aimed to address these gaps in the research by using social network inventory data to examine the direction of the association between index heroin and/or cocaine use and social network heroin and/or cocaine use among drug-experienced adults. Specifically, we tested the following hypotheses: 1) Substance use by social network members will be associated with index substance use at the subsequent study assessment, which would support social influence theory. 2) Index substance use will be associated with substance use by social network members at the subsequent study assessment, which would support social selection theory. 3) A greater amount of the change in network drug use between assessments will be accounted for by turnover than by changes in the behavior of network members over time. This tested the mechanism of social selection that an individual chooses friends whose behavior matches their own. 4) Drug use by close network members will have a positive association with index drug use at the following assessment. This hypothesis further explored social influence by testing if drug use by closest peers exerts influence on index drug use.

Sample

Data come from waves one, two, and four of the SHIELD (Self-Help In Eliminating Life-Threatening Diseases) Study [26]. Participants were recruited from neighborhoods in Baltimore, Maryland identified via ethnographic research and police records to have high levels of drug use. Eligibility criteria included reporting weekly contact with drug users, being 18 years of age or older, and willingness to engage in a HIV risk reduction intervention. A number of respondents (n = 567) were recruited from among the social networks of other participants. Only 154 participated in the intervention.

Wave one (1997-1999) included 1637 interviews. Due to funding constraints, participants were prioritized for follow-up at wave two, but all baseline participants were eligible for subsequent interviews. Wave two (1998-2001) included 896 interviews and wave four (2001-2003) totaled 905 interviews. Waves three and five did not include social network data. The median number of days was 245 and 940.5 from waves one to two and two to four respectively. All participants provided written informed consent and were financially compensated. The study was approved by the local Institutional Review Board.

Of the baseline sample, 1,158 had network data for wave two and/or four. An additional 50 individuals had a negative or missing lifetime history of heroin or cocaine use and were not included. Compared to participants lost to follow-up, the retained sample (n = 1,108; see Table 1) was less likely to have used heroin and/or cocaine in the prior six months (76.6% vs. 97.3%, p <.001) and reported fewer drug-using network members (3.5 vs. 4.0, p <.001) at baseline.

Measurement

Participants self-reported the frequency of their drug use in the prior six months for each substance via each route (e.g., snorting, injecting, smoking) on a 8-point Likert scale. Consistent with previous studies of urban adults [27,28], use of cocaine and heroin (though not necessarily together) was more common than use of only cocaine or heroin in this sample [29]. Consequently, we coded whether participants had used heroin and/or cocaine via any route in the past six months [23,30].

The social network was measured using an adapted version of the Arizona Social Support Inventory [31], during which respondents listed all individuals they considered among their social ties. Index report in the inventory was used to determine if each nominated network member was a current heroin/cocaine user, and total count of active drug users in the network indicated the index's exposure to network drug use [23,30,32].

Analyses

Preliminary Analyses to Explore Clustering

Due to the clustered nature of the recruitment techniques, we calculated the Design Effect (DEFF = 1 + [average cluster size -1] * ICC), where a DEFF > 2 indicates that multilevel modeling is preferable [37]. The Design Effect for clusters by primary index and by neighborhood was estimated to be 1.04 and 1.01 respectively, using the ICC from baseline number of network drug users. Consequently, we did not use multilevel modeling.

Structural Equation Modeling

Figure 1 reports the standardized estimates of the primary SEM. Estimates of both cross-lagged associations which supported social selection (index drug use predicted later network drug use) were statistically significant. Only one of the two estimates of the cross-lagged associations which supported social influence (network drug use predicted later index drug use) was statistically significant. Inspection of the fit indices indicated good model fit. Baseline demographic characteristics (sex, age, and education) were entered into the model regressed on baseline drug use and network drug use, but were not significant and not included in the final model. Intervention status was not significantly associated with the study outcomes, and inclusion of the intervention variable in the model resulted in little change to the estimates. No moderating effect for gender was found.

Standardized estimates from the recursive models are reported in Table 2. Because dependent variables in the SEM were independent variables in these models, some missing data were not imputed. Consequently, these models included only those participants with full data for the variables included in each model (814 for Model 1, 826 for Model 2, and 631 for Model 3). For the cross-lagged path supporting social influence (Model 1), the estimate remained significant, indicating that the influence of the network on an individual functions through long-term processes. For the two cross-lagged paths supporting social selection (Models 2 and 3), the estimates for the longitudinal path became non-significant (Model 2) or reversed directions (Model 3), indicating shorter-term processes.

Microsocial Analyses

For the 40 observations with the greatest decreases in drug use in the network, the mean ΔU was -6.08 (S.D. = 0.32) and the mean ΔC was -0.83 (S.D. = 0.16), with a statistically significant greater decrease in drug use through entrance and exit of unique members than change in behavior of carryover members (t(39) = -12.32, p <.001). Similarly, among the 40 observations with the greatest increases in drug use in the network, the mean ΔU was 5.20 (S.D. = 0.32) and the mean ΔC was 0.43 (S.D. = 0.15). Again the increase in drug use through entrance and exit of unique members was statistically greater than the increase in drug use by behavior change in carryover members (t(39) = 11.33, p < .001).

Key Peer Analysis

The model in Figure 2 includes the influence of drug use by network members at the highest level of trust and Figure 3 the influence of drug use by network members who the index sees daily. The association between key peer drug use with subsequent index drug use was not statistically significant in the expected direction in either model when total network drug use was also included. Both models exhibited good fit. Gender was not a modifier of the effect of key peer drug use and index drug use in either model.

Limitations

Although the present data have the advantage of being longitudinal, inferences are limited by the complexity of causality in social, behavioral, and psychological dimensions. Given the focus on heroin/cocaine-experienced individuals, the structure and influence of the networks observed over the course of this study may have been influenced by unmeasured prior attempts to cease drug use. The data were collected in one metropolitan area, and thus may not be representative of drug users in other locations. It is likely that the participants, drawn from inner-city neighborhoods with high rates of drug use, have a greater probability of having drug users in their network than other populations. Additionally, participants may have larger social networks compared to the broader population and may be individuals who are less easily influenced, as opinion leaders are better suited for network-based interventions [42] such as the SHIELD study. The relative poverty of the sample may limit the ability of respondents to cut network ties, in comparison to more affluent samples.

Despite its limitations, self-reports of drug use have been shown to have moderate to strong reliability and validity [43-45]. Social network members' behavior was reported by the index participant. Participants not using drugs do not have drug interactions with network members, and thus may have less knowledge of their friends' drug use than indexes who use drugs.

Implications and Future Directions

These findings suggest that the social network may be an important avenue for intervention with adult drug users. Consistent with the evidence in support of social influence, a possible intervention to reduce substance use could include training individuals to conduct peer outreach to their network [26] via education on drug use risks and disseminating treatment information. The focus of the present study naturally poses the questions of whether eliminating drug users from the social network would be an effective way to reduce drug use, as recommended by many treatment programs. These findings do suggest that friends influence drug use, and thus, there may be benefit in reducing the number of people one spends time with when attempting to quit drug use. However, a prior study examining cessation found that many individuals who quit continue to associate with drug users, albeit fewer [23]. Nonetheless, interventions to reduce drug use could engage drug users in discussing how they seek out or avoid spending time with social network members who use drugs, and promote more awareness of seeking out friends who use drugs as a conscious or unconscious expression of a desire to use drugs and a risk for relapse.

Social selection is a complex and dynamic process which appears to play an important role in either perpetuating or terminating an IDU's drug use career. Given the bidirectional association between a person's drug use and the drug use of social network members, future studies should use an individual-centered approach to determine if some individuals are more susceptible than others to peer influence on drug use [15]. Additional research should also explore drug availability through social networks, the relationship of network use with addiction severity, and characteristics of influential network members as moderators of the observed associations [46]. An enhanced understanding of these factors may elucidate potential intervention targets.