Given this apparent minimal influence of input data QC on imputation outcome, we investigated further the small set of SNPs showing significant allele frequency differences for the presence of a common characteristic that could conceivably be used as a post-imputation filter. To rule out poor genotyping as the cause of these significant differences, we examined all cluster plots for the unfiltered significant SNPs (P<1 × 10−6, n=325). In all, 14 poorly clustered SNPs were removed and the data were re-imputed. After post-imputation QC, three additional SNPs were not significant and six were less significant. We then inspected the cluster plots for 10 SNPs on either side of the 61 SNPs remaining significantly different to rule out poor imputation due to flanking SNP poor clustering properties. We examined the cluster plots for 1008 SNPs and found that 36 of these were poor; these resided in the proximity of 35 of the significant SNPs. We subsequently removed these SNPs and re-imputed. We found that following post-imputation QC filtering, only 3 of the 61 SNPs were no longer significant, and the R2 remained
