A random-forest approach for identifying classifiers for concurrent cannabis and tobacco use in US youth

Nicotine and tobacco products (NTP) and cannabis co-use entail unique risks. However, research on the correlates of co-use is limited, focusing on small samples, subpopulations, or specific correlates, without holistically examining multiple risk behaviors. Using an exploratory, hypothesis-generating approach, this study aimed to identify classifiers of past 30-day co-use of cannabis and NTP using random forest regression in a nationally representative adolescent sample.

Data were drawn from the 2023 Youth Risk Behavior Survey (YRBS). Participants were 20,103 US adolescents in grades 9-12. Complete-case analyses included 8263 participants. The sample was 51.9% male and racially and ethnically diverse. The primary outcome was a binary indicator of past 30-day co-use of cannabis and NTPs. Classifiers included demographic, behavioral, and psychosocial factors. Random forest regression, including hyperparameter tuning, importance ranking of classifiers, and model robustness measures, were conducted in complete-case and multiply imputed data.

Ten variables emerged as consistent and strongest classifiers of co-use, including past 30-day alcohol use (M_rank=1.00; SD=0.00), ever having had sexual intercourse (M_rank=2.36; SD=0.50), past 30-day binge drinking (M_rank=2.64; SD=0.50), and alcohol or drug use during last sexual intercourse (M_rank=4.00; SD=0.00). Models using imputed data achieved lower OOB error rates compared to the complete-case model.

Among US adolescents, alcohol use and sexual behavior were the most salient classifiers of cannabis and NTP co-use. These findings highlight the importance of holistic examinations of multiple health risk behaviors.