Cross-population imputation can be less effective for low-frequency alleles both because the sets of alleles in the two samples do not overlap perfectly (see earlier), and because haplotype patterns differ between populations. To isolate the effect of differing haplotype patterns, imputation within a population (CEU or YRI) was compared with imputation into a closely related population (TSI or LWK), but restricting the analysis to SNPs that were polymorphic in both target and reference panels (Fig. 6a). Notably, the imputation worked well for low-frequency alleles when using the correct reference panel, with a mean r2 > 0.7 with only two copies of the minor allele in the reference panel, and a mean r2 > 0.6 when imputing from a single copy. Imputation accuracy into a closely related European population (CEU/TSI FST = 0.004) was almost indistinguishable from the accuracy within a single population. For the two African populations, where low-frequency diversity is greater and the populations more diverged (FST = 0.008), the difference between reference and target populations was more substantial, with mean r2 only rising above 0.7 when five copies
