With regard to average r2hat, IMPUTE2 outperformed all other programs (Figure 3), but inclusion of more distantly related subjects led to reduced overall imputation quality (average r2hat = 0.68 for YRI+CEU+ASW, average r2hat = 0.62 for AFR+EUR, and average r2hat = 0.55 for ALL). The pattern of reduced overall imputation quality was entirely driven by low frequency SNPs (Figure 4), particularly SNPs with MAF≤2%, which represent 48%, 58%, and 64% of the imputed SNPs when using the YRI+CEU+ASW, AFR+EUR, and ALL panels, respectively. Inclusion of more distantly related subjects did not have the same effect on average r2hat when using the August 2010 release of 1000 Genomes (Figure S5), likely due to a smaller proportion of low frequency SNPs (e.g., 23% of imputed SNPs with MAF≤2% when based on the YRI+CEU+ASW panel and 25% when based on the AFR+EUR panel). Nonetheless, when using the February 2012 release of 1000 Genomes, imputation quality for low frequency SNPs was highest when using the most closely related reference panel (average r2hat = 0.51 when using YRI+CEU+ASW, 0.45 when using AFR+EUR, and 0.37 when
