Dudbridge et al.29 has noted that there are two purposes for GRS: association testing (that is, replication, reliant on significance/P-value) and prediction of phenotypic variance (for example, reliant on R2 estimates). Based on numerous simulations, he concluded that while most studies with approximately equally sized training (that is, COGA) and testing (for example, SAGE) samples are well-powered for association testing, current training samples are underpowered for prediction. Consistent with this observation, we were interested in the former rather than the latter and even though we report R2 values, the emphasis of our analysis was association testing. We confirm that selecting a testing sample of approximately the same size as the training cohort, as was achieved via our bootstrapping approach, yields significant association. Dudbridge29 notes that 10-fold cross-validation might be a more efficient approach to maximizing prediction. As two of our samples are family-based (COGA and OZ-ALC), such cross-validation approaches, which may necessitate disaggregating members of pedigrees or selecting subsets of pedigrees that may or may not be informative for the etiology of genetically transmitted AD, may not be applicable. Hence, our findings should be observed in the context of association (and not prediction) testing alone.
