All approaches to PRS calculation involve parameter optimization in generating the final prediction model and are thus vulnerable to overfitting [14]. The best strategy to avoid overfitting is to evaluate performance in an independent validation sample, but such a sample is not always available. Alternatively, if the primary aim is to assess evidence for an association to test a hypothesis, then we can calculate an empirical P-value corresponding to the association of the optimized PRS, with the Type 1 error rate controlled [13].
