been shown to be highly correlated with the IMPUTE-info score under the additive model.10 In both scenarios, we also filtered out SNPs with MAF <5%. Figure 1 illustrates the effects of altering post-imputation QC filters on the QCed data. On the basis of these results, we chose to use the IMPUTE-info score with a filtering threshold <0.8 and MAF <5%, which effectively eliminated ∼79% of the significant SNPs while retaining ∼85% of the nonsignificant ones (SNPTEST freq-add-proper-info <0.9 and MAF 5% would be roughly equivalent to this eliminating ∼73% of the significant SNPs while retaining ∼89% of the nonsignificant ones). We applied this post-imputation filter to each of our datasets and compared the results. We looked at the unQCed and QCed datasets first, as synopsised in Table 1. For each scenario, we examined frequency differences between the directly typed and the imputed genotypes as described above. In addition, we compared the imputed genotypes at imputed SNPs only for the unQCed and the fully QCed (QCed data with all poorly clustered markers removed) strategies.
