Fig. 5: Clinical risk prediction analysis.

Regression analyses were conducted to determine the efficacy of Comp-only, Agnostic-only, Demo-only, Demo + comp, Demo + agnostic and Demo + comp + agnostic in predicting clinical risk. Note that ASSIST scores were normalized before regression. Pearson correlation between true and predicted ASSIST scores was used to perform significance testing. a, In the alcohol group, Demo + comp + agnostic outperformed all alternatives (elpd_waic = −92.99). b, In cannabis users, Demo-only was the best-performing model (elpd_waic = −94.032). c,d, The best-performing models in both groups generated predictions that were highly correlated with true ASSIST scores (alcohol (c), r = 0.545, P < 0.001; cannabis (d), r = 0.372, P = 0.003; the shaded region represents the 95% CI). e,f, Parameters were extracted from the samples from the posterior distribution, and the 89% highest density intervals are plotted for each significant variable. Weight distributions for race and sex were eliminated due to a high degree of variance and low contribution to prediction. In alcohol users (e), α (learning rate), craving baseline and \({w}_{\mathrm{outcome}}\) were found to be significantly positively associated, while β (inverse temperature) and standard deviation of craving were found to be significantly negatively associated, with ASSIST scores. In the cannabis group (f), income was negatively associated, and there was a marginally negative association for age and education as well.