The imputation accuracy was quite similar for the SNPs and the CNPs (Fig. 6b), given their similar haplotype properties. Accuracy depended on high SNP density; reducing the set of tag SNPs from the full HapMap 3 set to the subset found on an earlier generation of array (approximately a threefold reduction in density) reduced r2 by roughly a factor of two for low-frequency SNPs (Supplementary Fig. 8). Somewhat unexpectedly, the accuracy was consistently higher for YRI than for CEU for both classes of variant, despite the former’s greater haplotype diversity and the identical panel sizes and SNP frequencies. One possible explanation is that for less common variants, the relationship between frequency and age has been partly obscured by population bottlenecks in the history of European populations, so that minor allele frequency is less effective as a predictor of allele age than in samples from Africa.
