Researchers have built a model that predicts which teenagers will start drinking or using marijuana, using data from nearly 12,000 kids tracked over several years.
Using data from about 11,860 participants in the long-running Adolescent Brain Cognitive Development (ABCD) Study, the team compared tree-based models, recurrent neural networks, and graph-based networks built from family, school, and similarity connections. All were tasked with predicting alcohol sipping, alcohol use, marijuana use, and combined alcohol/marijuana use. A method called temporal XGBoost, which tracks how risk factors change over time, turned out to be the strongest single approach. Graph-based models mapping a teen's social and family ties did not beat that on their own, but stacking the two together produced the best results across every outcome, with AUC-ROC scores above 0.79. The most predictive factors were peer deviance, age, externalizing symptoms, parental monitoring, cultural norms, and neighborhood context.
That combination matters because most substance-use screening leans on a single snapshot, a form filled out once, when what actually signals risk is how a kid's behavior and surroundings shift over years. Building peer networks and neighborhood context into the model, not just individual traits, points toward prevention efforts that account for who is around a kid, not only what is going on inside their head.
An AUC-ROC above 0.79 is a solid research result, not a crystal ball. Predicting risk across a population of nearly 12,000 teenagers is a different exercise than knowing whether any one specific kid is about to pick up a drink.