data could result in many SNPs being excluded from the data, and thus potential true association signals could be missed. Some of the inaccurately imputed variants were due to poor clustering properties. It is plausible that the handful of variants that still remained inaccurately imputed could be because of the differences in ethnicity between our data and the HapMap CEU reference panel from which the genotypes were predicted. We have used IMPUTE, but do not expect our results and conclusions to qualitatively differ with different imputation methods, for example, BEAGLE and MACH exhibit similar imputation accuracy to IMPUTE.11 Differences in population structure between the reference panel and target dataset can be a source of imputation inaccuracy. Imputation accuracy for common SNPs may be further increased by using larger reference panels with data on denser sets of variants. Our results show that GWAS QC is not of paramount importance for the imputation of common variants. This may be different for the imputation of low frequency and rare variants based on emerging reference panels such as the 1000 genomes (http://www.1000genomes.org) and UK10k (http://www.uk10k.org) projects. In summary, our study demonstrates that imputation of common variants is generally very accurate and robust to GWAS
