First, the PRS obtained using the full sample of GPpsy performed best at predicting MDD status in independent cohorts from PGC29-MDD (Nargelkerke’s r2=0.018, area uncer the curve (AUC) = 0.56 at a P-value threshold of 0.1; Fig. 7a and Extended Data Fig. 6). However, when equal sample sizes were used (randomly downsampled to 50,000 and case prevalence of 0.15; Methods), GPpsy no longer performed best at predicting MDD status in PGC29-MDD cohorts (Fig. 7b). Rather, the PRS from strictly defined CIDI-based MDD (LifetimeMDD) best predicted MDD disease status (Nargelkerke’s r2 = 0.0027, AUC = 0.52 at a P-value threshold of 0.1; Extended Data Fig. 6).
