Estimating the Effect of High-Risk Peers On Youth Alcohol Consumption in Chile Using Propensity Scores Stratification

Background: Does youth alcohol consumption proliferate when they socialize with peer-drinkers (positive selection hypothesis)? Or is peer-influence trivial among youth who are already at high risk for drinking (negative selection hypothesis)? Despite the centrality of these questions, estimating unbiased peer-effects is a challenge because youth's decision to engage in antisocial behavior is often related to the decision of with whom they associate. To overcome methodological challenges, we employed propensity score stratification to examine heterogeneous peer drinking effects across groups of youth with different tendencies of having drinking friends. A positive relationship between the magnitude of peer influence and the probability of having friends who drink would substantiate the positive selection hypothesis, while a negative relationship would suggest otherwise.

Methods: Data are from a NIDA-funded study of youth drug use in Santiago Chile (N=832). The dependent variable is number of drinks consumed in the past month. The independent variable is a binary variable of whether the youth associated with friends who consumed alcohol. Control variables include youth demographics, parent (SES, marital status, alcohol consumption), and school/neighborhood variables (school prevention programs, neighborhood crime, alcohol advertisement exposure).

Propensity score stratification establishes a framework similar to a randomized experiment, where the treatment is whether peers consume alcohol or not. By assuming balanced measured and unmeasured characteristics between the treatment and control groups within each stratum, we can estimate unbiased peer-effects (Rosenbaum and Rubin 1983). First, using logistic regression, we computed propensity scores by estimating an individual's predicted probability of associating with friends who drink based on observed information. Second, we stratified our sample into five groups based on the propensity scores. Third, for each stratum, we estimated the relationship between peer- and youth-drinking using negative binomial regression (Level-1). Fourth, we identified heterogeneous peer-drinking effects across strata using a hierarchical linear model (Level-2) (Brand and Xie 2010).

Results: Results support the positive selection hypothesis. When assuming heterogeneous treatment effects, peer effects increase with the predicted probability of having friends who drink (Level-2 slope=0.645). Level-1 slopes indicate that youths who are most likely to socialize with drinking friends are exposed to a greater peer-drinking effect relative to youths who are least likely by an additional 3.41 drinks/month. Under the assumption of homogeneity, however, the peer effect would have been 2.31 drinks per month. Computing a single estimate for peer-drinking would have underestimated the detrimental influence of peers particularly among at-risk youths. Finally, youth in the highest stratum scored significantly higher in rule-breaking and aggressive behavior in Achenbach's YSR problem score than those in the lowest stratum, which suggested that greater exposure to drinking peers increased the likelihood of youth's own risk-taking.

Conclusion: Findings indicate that the harmful effect of drinking peers is exacerbated among youth who have a high probability of socializing with them. These findings suggest that statistical models that do not account for selection underestimate peers' influence on adolescent drinking behaviors. Advanced methods are necessary to clearly understand the dynamics of youth socialization processes and alcohol consuming behaviors with U.S. and international populations.