An important component in the analysis of genome-wide association studies involves the imputation of genotypes that have not been measured directly in the studied samples. The imputation procedure uses the linkage disequilibrium (LD) structure in the population to infer the genotype of an unobserved single nucleotide polymorphism. The LD structure is normally learned from a dense genotype map of a reference population that matches the studied population. In many instances there is no reference population that exactly matches the studied population, and a natural question arises as to how to choose the reference population for the imputation. Here we present a Coalescent-based method that addresses this issue. In contrast to the current paradigm of imputation methods, our method assigns a different reference dataset for each sample in the studied population, and for each region in the genome. This allows the flexibility to account for the diversity within populations, as well as across populations. Furthermore, because our approach treats each region in the genome separately, our method is suitable for the imputation of recently admixed populations. We evaluated our method across a large set of populations and found that our choice of reference data set considerably improves the accuracy of imputation, especially for regions with low LD and for populations without a reference population available as well as for admixed populations such as the Hispanic population. Our method is generic and can potentially be incorporated in any of the available imputation methods as an add-on. Genet. Epidemiol 34:773-782, 2010. © 2010 Wiley-Liss, Inc.