We evaluated the predictive accuracy of PRS based on PTSD Freeze 3 in a set of MVP holdout samples (Fig. 5). In EA holdouts, risk was significantly different across the range of PTSD PRS. For example, individuals in the highest quintile of PTSD PRS had 2.4 times the relative risk of PTSD (log relative risk s.e. = 0.032; 95%CI = [2.25, 2.56]; P = 1.16 × 10−167) than individuals in the lowest quintile. PRS explained 6.6% of the phenotypic variation in PTSD (Nagelkerke’s R2 transformed to the liability scale at 15% population and sample prevalence), representing a major improvement over PRS based on Freeze 2. In contrast, among AA holdout samples, PRS explained only 0.9% (liability scale) of the variation in PTSD, consistent with previous work suggesting that AA PRS based on EA data lag behind in prediction64.
