Peer Influence and Context: The Interdependence of Friendship Groups, Schoolmates and Network Density in Predicting Substance Use

In comparing those adolescents who did not have enough data to be part of the final samples to the youth who comprise our sample, excluded subjects, on average, were the same age (15 years old), but were less likely to be White (43 vs. 60 %), less likely to be female (43 vs. 55 %) and more likely to have been on public assistance (11 vs. 8 %).

Imputation procedures were used with Stata’s Imputation with Chained Equations (ICE). This method predicts values (across several iterations) using a procedure where (a) initially all missing values on a particular variable are filled in at random (x), (b) that variable is then regressed on others in the set, (c) the values on that initial variable are then replaced by draws from the posterior distribution of predicted values (x), and (d) the process is replicated across a number of cycles so that stable imputation results can be achieved (Royston and White 2011).

Our focus is on schoolmates, but some readers may wonder about grademates. Students are organized by grades within schools, but these are not fixed boundaries. Sports teams and clubs are school-based, not grade based, and some academic classes contain students at different grade levels. Indeed, research has suggested that the peers from sports teams and other organizations (which cut across grades) have an important influence on adolescent behavior (Fujimoto and Valente 2013; Fujimoto et al. 2012; Kreager 2007), and that the culture of the overall school is important for understanding individual substance use (Bisset et al. 2007). We also believe our data underscore the focus on schoolmates rather than grademates. 25 % of identified friends were not in the same grade as the subject. Importantly, this carries over even to best friends—nearly 25 % were likewise in a different grade than the subject. Still, we completed supplementary analyses looking at the conditioning effect of grademates. The results for smoking are similar to those reported in the main text. For example, the interaction term in the subsample for adolescents whose friendship group has a density score of <.80 is negative and achieves statistical significance (b = − 1.559, SE = .635, p = .016). This suggests that friends become more dissimilar from grademates with regard to smoking, the friends’ influence on the smoking behavior of the subject declines. For the subsample of adolescents whose friendship group has a density coefficient of .8 or greater, the interaction term is positive and non-significant (b = 2.718, SE = 2.773, p = .330). With regard to drinking, the pattern of results is similar, though weaker.

Hierarchical models that specify schoolmates as a second-level variable would not be appropriate for the questions under consideration. This would fix the schoolmate variable to be the same for all members of the same school, and thus the cross-level interaction between peers and schoolmates would simply reflect whether delinquent peers mattered more in schools with higher or lower levels of such behavior, on average. Still, because there is school-level variability in the smoking and drinking outcomes, we conducted sensitivity analyses in which we ran random-intercept logistic regression models (estimated in HLM), which does not force the intercept to be constant across all schools. The results from these models are consistent with those presented in the main text.

Some readers may be interested in whether there is a direct effect of schoolmates’ substance use on subjects’ own use. When accounting for friends’ substance use along with demographics, parental attachment, school attachment, parental supervision, impulsivity and (the respective) substance use at Wave 1, we found that the portion of schoolmates who smoke is not a statistically significant predictor of later smoking ((b = −.025, SE = .473, p = .957). Likewise, the portion of schoolmates who drink is not a statistically significant predictor of later drinking (b = .224, SE = .371, p = .546). Such results are interesting, as they may prompt scholars to prematurely assume that schoolmates “don’t matter” for substance use. But, our focus and analysis suggests they do matter, in an indirect way, by conditioning the influence of friends’ substance use.

Note that if the mother took the respondent to school or picked him/her up from school, this was coded as “always.”.

Some readers may be interested in age effects with regard to the relationships under study. First, because smoking and alcohol are perceived differently according to age, we investigated whether our results held for smoking with a subsample of adolescents younger than 18 (at which point individuals can legally purchase cigarettes) and for drinking with a subsample of adolescents younger than 21 (at which point individuals can legally purchase and consume alcohol). The results are substantively the same as those for the whole sample. Next, some may wonder whether our results vary across developmental periods. We did complete supplemental analyses for different subsamples clustered by age—specifically, age 15 and younger and between ages 16 and 18 (the ability to estimate models for a highly dense subgroup within an older subsample was not possible due to low sample size). It is important to note that, although participants in our sample range from 10–22 years of age, over 75 % of the sample is between ages 14 and 18. The results for the younger sub-sample were in the same direction as those presented in the main text, but the relationships are stronger (and statistically significant) for the 16–18 subsample.

Recent research on peer effects increasingly relies on stochastic actor-based (SIENA) models (Snijders 2001), which allow researchers to simultaneously model dynamic changes in network characteristics and changes in behavior to disentangle selection and socialization effects (Weerman 2011). Though we use social network data, the hypotheses posed in this inquiry are not yet translatable to a SIENA framework.

We use the following Stata command when calculating cross derivatives: margins, dydx(portion of peer group that is deviant) at (absolute different between peer group and schoolmates = (0 .1 .2 .3 .4 .5 .6 .7)) post (see Karaca-Mandic et al. 2012).