Identifying which adolescents will develop substance use disorders before those disorders take hold could transform prevention programs — shifting resources from broad campaigns to high-risk individuals. A new predictive model takes a meaningful step toward that goal by jointly forecasting adult risk for both alcohol use disorder and cannabis use disorder using data collected during youth.
Built on longitudinal data from the National Longitudinal Study of Adolescent to Adult Health (Add Health, n = 12,503), the model uses a Bayesian framework that treats alcohol use disorder and cannabis use disorder as correlated outcomes rather than independent events. It distinguishes three user groups — alcohol-only, cannabis-only, and dual-substance users — and applies ten predictors capturing both shared and substance-specific risk factors. Predictive performance, measured by the area under the receiver operating characteristic curve, ranged from 0.65 to 0.75 across two external validation datasets, indicating moderate-to-good discrimination. The joint modeling approach outperformed separate univariate models in simulation testing, suggesting that capturing the co-occurrence structure of these disorders adds predictive value.
This work sits at the intersection of psychiatric epidemiology and machine learning, and its methodological contribution may ultimately matter more than the specific AUC figures. AUC values in the 0.65–0.75 range are useful for population-level screening tools but fall short of clinical decision thresholds, meaning the model is best viewed as a risk-stratification aid rather than a diagnostic instrument. The reliance on Add Health data — collected across multiple decades — introduces questions about generalizability to contemporary adolescents, particularly given the sharp rise in high-potency cannabis products since those cohorts were surveyed. The external validation on two independent datasets is a genuine strength, rare in this literature. Practically, a tool of this kind could help school-based or primary-care settings flag adolescents warranting early counseling, though prospective real-world testing will be necessary before clinical deployment. The finding that joint modeling outperforms univariate approaches also has broader implications for comorbid disorder research, where correlated outcomes are the rule rather than the exception.