The HapMap 3 panel is still useful, however. Across the allele frequency spectrum, the difference in mean R2 between the GMB and HM3 panels was never larger than 2.5%, and the difference was smallest at low-frequency SNPs. In fact, the HM3 panel was more accurate than the GMB panel for SNPs in the 1–2% and 2–3% MAF bands (Figure 3B); this shows that an ancestrally inclusive, nonspecific reference panel can capture low-frequency alleles that are poorly represented in a Gambia-specific panel. A recent simulation study found that association power and mean imputation R2 have a roughly linear relationship with a slope near 1.0 (Zheng et al. 2011), which suggests that using the HM3 panel in place of the GMB panel would cause only a small loss of power in an imputation-based association scan. [We note that the results from Zheng et al. (2011) appear to conflict with those of a recent study by Huang et al. (2009b), which found that power drops quickly with mean R2. We believe that the Zheng et al. (2011) results more accurately reflect this relationship
