tests for SNPs imputed based on the union of genotyped SNPs available on either array revealed spurious results as indicated by inflated λgc values and SNPs having statistically significant P values across the MAF spectrum (Fig. 1b, e). Overall, 0.20 % of the imputed SNPs had P < 1 × 10−6 in both European Americans (71 false positives of 34,515 imputed SNPs) and African Americans (87 false positives of 42,963 imputed SNPs). This imputation strategy based on the union of genotyped SNPs incurred substantial deviation from expectation, as demonstrated in Figure S1. In contrast, conducting imputation based on the intersection of genotyped SNPs available on both arrays resulted in no spurious associations (Fig. 1c, f). To be sure that choice of software did not influence our observed pattern of results, we ran the Illumina 1M versus 550v3 comparisons in African Americans using IMPUTE2 and found parallel results to those obtained using MaCH (Figure S2).
