To check if GEMMA and MTMM produce calibrated p values, we randomly selected 100,000 real genotypes in the HMDP data. We simulated 10,000 phenotypes under the null, based on the real relatedness matrix and the estimated genetic and environmental covariance matrices (for HDL and TG). We calculated p values for each SNP-phenotype pair in turn (one billion pairs). We did not perform comparisons based on the NFBC1966 data, partly because GEMMA and MTMM produce identical p values there, and partly because the sample size in NFBC1966 makes it computationally impractical to perform billions of association tests to check for the type I error at the genome-wide significance level.
