To evaluate the bias induced by imputation across highly different arrays, we combined SAGE subjects genotyped on Illumina 1M with subjects genotyped on Affymetrix 6.0 from the Genetic Association Information Network (GAIN) GWAS of Schizophrenia (Manolio et al. 2007). European American and African American subjects from these studies were analyzed separately. A single false positive SNP association was observed among the genotype SNPs shared on both arrays in European Americans (Fig. 2a), but no false positive SNP associations were observed in African Americans (Fig. 2d). Compared to the analyses across the Illumina family of arrays, substantially more false positive associations were observed when using imputed SNPs based on the union of SNPs available on the Illumina or Affymetrix array (Fig. 2b, e): 184 (0.53 %) false positives of 34,503 imputed SNPs in European Americans and 271 (0.63 %) false positives of 43,035 imputed SNPs in African Americans. The deviations from expectation were substantial (Figure S4). Our strategy of using the intersection of SNPs as the basis for imputation was able to eliminate these biases, even when combining these highly different arrays (Fig. 2c, f).
