We also note that the boost in power is more substantial than that estimated in another recent study [11]. A close look at the details of this other study shows that the only imputed SNPs used were those (a) which had real genotype data from one of the other chips, and (b) the imputed and real data at the SNP agreed with an r 2>0.8. So for example, for the Affy 500 k chip only genotypes at 427,838 imputed SNPs were used, rather than all those available from HapMap (approximately 2.5 milion SNPs), as normal practice when carrying out imputation. Using such a filter clearly creates a bias towards imputed SNPs that are almost perfect tags for SNPs on the chip so it is not surprising that this study shows such small increases in power when using imputation.
