Fig. 4
Comparison of parametric P-values from different software packages for controlling heritable nuisance covariate in GWA (Simulation 5). Results shown for the distribution of null (βSNP = 0 and βnuisance = 0). P-values from score statistic based the simplified REML function using non-iterative random effect estimator (NINGA), FaST-LMM and EMMAX, for the GRM from 4000 unrelated individuals with 5000 realizations when nuisance covariate is included in the LMM (left panel) and residualised phenotypes are fitted to LMM (right panel). The nuisance covariate heritability is 60%, due to 10 SNPs25 using the additive model. There is no apparent differences between the packages when the nuisance covariate is included in the LMM. However, fitting the LMM to the residualised phenotypes produces invalid results in this setting of correlation between SNP and nuisance covariates. Confidence bounds created with the results of ref. 59, where ordered P-values follow beta distribution (see the Supplementary Note 1 for more details on simulation settings)
