In contrast to the cohort study design, it appears that there is no additional power to be gained by using Prentice-weighted Cox models for genetic associations in case-cohort studies. Indeed, there is a striking loss in power, which occurs because robust SEs are not necessary for the logistic regression model in addition to its effect estimates being greater in magintude. Furthermore, the computational cost of the Cox model was far greater than that of the logistic regression model. Hence, although we recognise the caveats of using the logistic regression model (inflated effect estimates, especially for associations of greater magnitude) in case-cohort studies, we propose that logistic regression could be used as a filter to detect SNPs below a pre-defined P-value threshold for GWAS (of a large number of SNPs (>500 000)) in case-cohort studies. This threshold should be set suitably high to ensure that all of the SNPs that would have been detected at the overall level of significance with the Cox model are contained within this subset, for example, 1 × 10−4 if the overall significance threshold is 5
