the matrix of genetic relationships between individuals). Using the software GCTA (Yang et al. 2011), we computed the genetic relationship matrix (GRM) for our LD-pruned genotype data, combining the COGA and SAGE samples for the EA (n = 2,763) and AA (n = 1,167) study populations, with the exclusion of individuals with estimated relatedness greater than 0.025 (i.e., corresponding to third cousins or closer). The GRMs were then fitted to the linear models for AD status, parameterized on an unobserved continuous liability scale via a probit transformation (Lee et al. 2011), using a restricted maximum likelihood (REML) approach, with the covariates age and sex. The estimates of AD variation explained by the GRMs were corrected for ascertainment bias using population-specific prevalence rates (0.038 and 0.036 for EAs and AAs, respectively) (Grant et al. 2004).
