Applying machine learning methods to model social interactions in alcohol consumption among adolescents

Background: Existing research using machine learning to investigate alcohol use among adolescents has largely neglected peer influences and tended to rely on models which selected predictors based on data availability, rather than being guided by a unifying theoretical framework. In addition, previous models of peer influence were typically estimated by using traditional regression techniques, which are known to have worse fit compared to the models estimated using machine learning methods. Methods: Addressing these limitations, we use three machine-learning algorithms to fit a theoretical model of social interactions in alcohol consumption. The model is fit to a large, nationally representative sample of U.S. school-aged adolescents and accounts for various channels of peer influence. Results: We find that extreme gradient boosting is the best performing algorithm in predicting alcohol consumption. After the algorithm ranks, the explanatory variables by their importance in classification, previous year drinking status, misperception about friends' drinking, and average actual drinking among friends are the most important predictors of adolescent drinking. Conclusions: Our findings suggest that an effective intervention should focus on school peers and adolescents' perceptions about drinking norms, in addition to the history of alcohol use. Our study may also increase interest in theory-driven selection of covariates for machine-learning models.