Hierarchical density-based clustering (R dbscan package) (55) was used to detect and remove temporal outlier responses to the follow-up questionnaire (Supplement 1, Figure S5 in Supplement 1). The R nlme package (56) was used to fit a longitudinal multilevel linear model to examine whether GMV predicted AUDIT-C at follow-ups. The model included both random intercept and random slope components with a continuous autoregressive correlation structure. Time was coded as both linear and quadratic age at the date of response (baseline or follow-up). Models tested the interaction between brain volume and age (i.e., does baseline ROI volume predict a different slope of change in drinking behavior as participants age?). Covariates were z-scored, and they were identical to those used in neuroimaging analyses, with the addition of second-order interactions between covariates and primary variables (57,58). Each of the 2 ROIs was tested in a separate model, and p values were false discovery rate (FDR) corrected (i.e., 4 tests: middle × linear age, superior × linear age, middle × quadratic age, superior × quadratic age).
