In this work, we have highlighted the fact that combining reference data from different populations can improve imputation accuracy at low-frequency variants. This finding reflects both the limited sample sizes of existing reference panels and the shared ancestry of human populations: an allele that occurs at low frequency in a study population may be poorly represented in a well-matched reference panel due to sampling effects; however, that same allele may be found in reference sets from other populations due to genetic drift or introduction by recent migrants. While a multipopulation reference panel can improve accuracy at this kind of variant, there are other situations in which accuracy might be harmed by such panels. Possible mechanisms for decreased accuracy include (i) the imputation of variant alleles at sites that do not segregate in a study dataset, (ii) signal dilution from reference haplotypes that are similar to those in a study population but do not carry a variant allele that segregates in that population, and (iii) misleading results from reference haplotypes that carry recurrent mutations. We discuss these issues in File S5;
