Machine learning reveals two key dimensions of developmental risk for substance use disorders in young adulthood

Prior work has identified a variety of developmental risk factors for substance use disorder (SUD) in adulthood, including externalizing behaviors, early substance use initiation, disinhibited personality traits, and poor neurocognitive abilities. However, many associations between specific risk factors and SUD could be accounted for by a smaller number of underlying dimensions of risk. We applied principal components regression (PCR) to a sample of 610 participants from the Michigan Longitudinal Study (MLS) to quantify the utility of a wide array of childhood and adolescent factors for predicting SUD in young adulthood and evaluated whether their associations with SUD could be accounted for by a smaller number of latent risk dimensions. PCR utilized nearly 300 variables collected from ages 3−19 across multiple domains (substance use, psychosocial factors, demographics, and neurocognitive testing) to predict SUD diagnoses between ages 20−26. PCR models involving all features classified individuals’ SUD diagnoses well above chance in held-out data, AUC = 0.736, CI = 0.696 – 0.776. Two latent dimensions were the minimal number necessary for reaching optimal performance relative to the full model. Remarkably, information drawn from hundreds of developmental features appears to reflect only two underlying SUD risk dimensions: an “externalizing/undercontrol” dimension characterized by impulsivity, inattention, worse neurocognitive abilities, and use of multiple substances, and a distinct “social adaptation” dimension characterized by behavioral approach, positive peer relationships, better neurocognitive abilities, and alcohol use. These dimensions were independent, had distinct correlates, and could provide a novel framework for investigating early person-level risk for SUD.