Linear and logistic regression models were used to examine whether AUDIT-C and AUDIT-P PRS were associated with each of the alcohol-related phenotypes (see Table 1), and to determine which PRS threshold (i.e., pT) was most predictive of each measure based on the p-value and observed (linear) or pseudo (logistic) R2. For the cohorts of unrelated individuals (ALSPAC and UKB), the partial R2 was extracted from linear regression models for continuous traits, while Nagelkerke’s pseudo-R2 was extracted from logistic regression models for binary traits. For the cohorts that employed mixed-effect models to account for within-sample relatedness (COGA and GS), variance explained by the PRS in the continuous outcomes (e.g. MaxDrinks, CAGE) was calculated by multiplying the PRS by its regression coefficient and dividing the variance of that value by the variance of the outcome to derive a coefficient of determination between 0 and 1 (Nakagawa & Schielzeth, 2012); in COGA, the ‘MuMIn’ package in R was used to calculate marginal R2 for the logistic mixed-effect models for the binary outcomes (Barton, 2011). Once the most predictive AUDIT-C and AUDIT-P PRS thresholds
