|
Abstract:
|
Despite recent success of GWAS for time-to-event phenotypes, very few studies have been conducted to evaluate the time-dependent effect, i.e., non-proportional hazard (NPH), of genetic variants. Here, we hypothesize that a non-negligible proportion of GWAS signals for time-to-event phenotypes exhibit NPH. Applying proportional hazard models, e.g., Cox regression, to variants with NPH may suffer from severe power loss. We adapted a recently proposed omnibus test of change-point Cox regressions to GWAS settings. An efficient score test is developed to speed up computation and the Saddlepoint approximation (SPA) is adopted to ensure accurate p-value calculation for rare variants and low-frequent events. Simulation studies show that the proposed method provides remarkably higher power than existing methods, such as SPACox, when NPH is present. To validate our hypothesis on NPH of signals, we applied the proposed method to the UK Biobank inpatient data of 259,693 white British ancestry samples for the analyses of 12 time-to-event phenotypes. Out of 983 associated loci identified, 82 loci have p-values at least five times smaller than SPACox p-values and 44 loci would be missed by SPACox
|
