Leveraging Machine Learning to Understand the Link Between School Climate and Youth Substance Use: a Focus on...

This paper focuses on school climate indicators, which have been previously linked with aspects of students' well-being and school-related success, to explore how they relate to alcohol and cannabis use. We used machine learning (ML) approaches and leveraged data from a diverse sample of 69,513 students (45.4% White, 23.9% Black, 8.9% Latine) across 111 middle and high schools, with 12% (n = 7783) reporting cannabis use and 18.8% (n = 12,220) reporting alcohol use in the past 30 days. We focused on 154 items related to school climate, student attitudes and behaviors, and demographics. We employed a two-stage feature selection method, initially reducing the 154 features to 31, and subsequently to 20, for both alcohol and cannabis use. Alcohol and cannabis use shared 15 common features and 5 distinct features, though some variation occurred across these two outcome variables. We identified both unique and shared factors that best classified current users vs. non-users. Specifically, gender, sense of pride in the school, weapon carrying, and bullying others were unique indicators that best classified alcohol use. In contrast, difficulties overcoming challenges, problems controlling temper, and becoming angry easily were more strongly associated with cannabis use. Shared indicators associated with both substances included gang membership, skipping school, violent behavior, school-parent and school-student engagement, and gambling. The inclusion of diverse classification factors underscored ML's ability to capture complex social and environmental factors that may be associated with substance use differently across student demographics. These features were tested in 12 classification models for both substances, achieving ROC-AUC scores up to 86% with fine-tuning of the best-performing models. The results highlight the utility of ML for examining complex, multidimensional indicators associated with substance use that complement traditional models.