We generated polygenic scores by summing genotypes weighed by the effect sizes estimated in the externalizing GWAS, among individuals of European ancestry in five hold-out cohorts: (1) Add Health71,72, (2) COGA73–75, (3) PNC76,77, (4) the UKB siblings hold-out cohort78, and (5) BioVU46 (Supplementary Information section 5). In each dataset, we generated three scores, of which two were adjusted for linkage disequilibrium (LD): (1) PRS-CS (version October 20, 2019; default Bayesian gamma-gamma prior of 1 and 0.5, and 1,000 Monte Carlo iterations with 500 burn-in iterations)79, (2) LDpred (version 0.9.09; infinitesimal Bayesian prior)80, and (3) unadjusted scores81, while using SNPs that overlapped the HapMap 3 Consortium consensus set82 (for comparability across methods and with previous work, and because PRS-CS imposes that restriction). We evaluated the incremental R2/pseudo-R2 (ΔR2) attained by adding the polygenic score to a regression model with baseline covariates, as in previous efforts17. The baseline model included covariates for sex, age, and genetic principal components (PCs), and genotyping array and batch. The choice of statistical model (e.g., least squares vs. logit) and adjustment of standard errors depended on (1)
