Logistic mixed effect models (logME) can be used to test for associations between binary traits and genetic variants while accounting for population structure and the relationships among individuals using a random effect defined by an empirical genetic relationship matrix (GRM). However, the optimal choice of variants to include when generating the GRM is not well understood. Here, we investigate the performance of logME models under a range of GRM choices using meta and joint analysis when the study sample includes samples from two distinct ancestries.
Genotypes from 4177 adults of European ancestry (Framingham Heart Study, FHS), and 3417 adults of African American ancestry (Jackson Heart study, JHS) were measured from the Trans-Omics for Precision Medicine program freeze4 whole genome deep sequencing data. We computed GRMs in FHS and JHS and in the combined FHS+JHS sample using 1) MAF≥0.1% and 2) MAF≥5% variants after linkage-disequilibrium pruning. A binary trait with 50% heritability and a difference in prevalence of 10% versus 5% in the two samples was simulated based on a mix of variants with large and small differences in frequency between the FHS and JHS samples. On average, power was higher for combined than for meta-analysis. The ≥5% MAF and ≥0.1% MAF GRMs produced equivalent type I error and power with meta-analysis, but for combined analysis the power was lower for the ≥0.1% MAF GRM than for the ≥5% MAF GRM. We recommend combined analysis and GRMs based on common variants for logME association analyses with binary phenotypes.