The cosmopolitan approach of combining all available reference populations incurs more computational burden than other imputation approaches, but it has been advocated as the simplest and most practical approach without sacrificing performance [5], [6], [17], [23]. Studies focusing specifically on African American study populations have shown that inclusion of diverse reference panels is clearly advantageous over single ethnic panels [8], [28], but the optimal extent of diversity has not been fully evaluated. In European-derived study populations, Jostins et al. suggested that using diverse reference populations improved imputation of low frequency SNPs. However, their conclusion was drawn from HapMap phase III imputed SNPs that had already passed an r2 threshold of 0.9 [23]. Our study similarly showed that increasing the reference sample size by including more distantly related populations improved imputation quality, but this pattern pertained mostly to SNPs that were present in populations more closely related to African Americans. To efficiently remove the occurrence of low frequency SNPs that are most likely monomorphic in African Americans, investigators should consider filtering SNPs based on their MAF in subpopulations of interest. We
