Our study is the first to highlight important considerations when using the cosmopolitan approach with 1000 Genomes to conduct imputation in an admixed study population, particularly an African American population. Imputation quality of all low frequency SNPs was relatively low, and imputation of low frequency, population-specific SNPs was especially prone to imputation error. Because imputation error leads to a loss of statistical power [37], [38], p-values of true association signals might be attenuated for the low frequency SNPs. Further, inclusion of more distantly related reference panels for imputation in African-derived study populations has been suggested to weaken association signals particularly near loci under strong selective pressure [39]. Regardless of the software program implemented, an optimal balance for African American studies may be provided by focusing GWAS analyses on SNPs imputed with reference to the ALL panel and then following up implicated regions based only on SNPs that are present in more closely related populations.
