A Bayesian learning model to predict the risk for cannabis use disorder.

Area of Science:

• Psychiatry and Behavioral Sciences
• Public Health
• Data Science

Background:

• Cannabis use disorder (CUD) prevalence is rising, exacerbated by increasing cannabis legalization.
• A predictive model for CUD risk in adolescent and young adult cannabis users is needed.
• Existing models lack validation using nationally representative longitudinal data.

Purpose of the Study:

• To develop and validate a predictive model for identifying adolescent and young adult cannabis users at high risk of developing CUD in adulthood.
• To utilize a novel Bayesian learning approach with nationally representative longitudinal data.

Main Methods:

• Employed a Bayesian learning approach with logistic regression models and four regularization priors (lasso, ridge, horseshoe, t).
• Utilized data from the Add Health study (n=8712) for model development and external validation (n=570).
• Compared models using 5-fold cross-validation, assessing discrimination (AUC) and calibration (E/O ratio).

Main Results:

• The final model, based on a lasso prior, includes seven predictors: biological sex, neuroticism, openness, conscientiousness, adverse childhood experiences, delinquency, and peer cannabis use.
• The model demonstrated good discrimination (AUC=0.69) and calibration (E/O=0.95) via cross-validation.
• External validation showed an AUC of 0.71 and an E/O of 1.10.

Conclusions:

• An externally validated model can identify young cannabis users at high risk for developing CUD.
• This model offers a valuable tool for public health interventions and personalized risk assessment.