0 otherwise; and where pseudo-counts of 1 are added to the numerator and denominator to avoid empirical P-values of 0 and reflecting (conservatively) counting the observed trait configuration as 1 potential null permutation [22]. While the empirical P-values for association will be controlled for the Type 1 error rate, because the same process of parameter optimization is performed explicitly under the null hypothesis, the observed phenotypic variance explained, R2, remains unadjusted and is affected by overfitting. Therefore, it is imperative to perform out-of-sample prediction, or cross-validation, to evaluate the predictive accuracy of PRS when using PRSice-2, and ideally the former given the problems of generalizability observed with PRS [14].
