Filtering by the difference between the reference and the estimated minor allele frequency can effectively remove some genome wide significant SNPs. Of the 4016 genome wide falsely-significant SNPs, 3120 (77.7%) SNPs are removed by removing those SNPs whose minor allele frequency difference is greater than 0.01.However, there are still 832 (21% of the 4016 SNPs) that have passed the filter. Most of the 832 remaining SNPs share one character: they tend to have very low minor allele frequency (MAF median = 0.00096). Imputation tends to over-assign the major genotype to the imputed SNPs, resulting in different allele frequency and therefore inflating the P value. However, to filter by MAF difference at 0.01 is not an acceptable option. Most SNPs are correctly predicted even if the minor allele frequency is different. When we tried to remove all SNPs whose minor allele frequency difference was greater than 0.01, 583,456 of the total 788,944 available SNPs (74%) were removed. Most of these SNPs were correctly predicted even if minor allele frequency was different. This is because imputation does not assign predicted genotype based on minor allele frequency, but rather on haplotype modeling.
