Fig. 4: Craving model comparison and parameter distribution: addictive condition.

a,b, As with decision models, ΔBIC was defined as the difference between each model’s BIC and the best-performing BIC in the alcohol group (a) and cannabis group (b). The joint outcome + EV model performed best across groups. c–f, In the alcohol group, for the best performing craving model, we calculated the correlations between model-predicted craving and true craving ratings. An example alcohol participant’s true (black line) versus predicted craving (blue line) is displayed. Two-sided one sample t test without correction was used for all significance testing (c). There was a high degree of correlation across participants in the alcohol group (mean r = 0.446), indicating strong model efficacy (d). Similarly, in the cannabis group, we visualize an example participant’s true (black line) versus predicted craving (yellow line) (e). There was a high degree of correlation across participants in cannabis group (mean r = 0.446; f). g,h, Distributions of parameters were extracted from the outcome + EV model. The left panel displays the joint distributions of outcome weight (\({w}_{\mathrm{outcome}}\)) and EV weight (\({w}_{\mathrm{EV}}\)), while craving baseline is visualized separately. In the alcohol group (g), \({w}_{\mathrm{outcome}}\) (t = 4.322, P < 0.001) was significantly positive and \({w}_{\mathrm{EV}}\) (t = −0.256, P = 0.798) was not significantly different from zero. In the cannabis group (h), neither parameter reached significance (\({w}_{\mathrm{outcome}}\), t = 1.603, P = 0.114; \({w}_{\mathrm{EV}}\), t = 1.298, P = 0.199). i–k, There were no significant group differences in baseline craving (t = 0.329, P = 0.776) (i), outcome weight (t = −0.911, P = 0.364) (j) or EV weight (t = 1.109, P = 0.255) (k). Two-sided independent-sample t test without correction was used for all significance testing in i–k.