Imputation of missing genotypes: an empirical evaluation of IMPUTE

Imputation of missing genotypes is becoming a very popular solution for synchronizing genotype data collected with different microarray platforms but the effect of ethnic background, subject ascertainment, and amount of missing data on the accuracy of imputation are not well understood.

We evaluated the accuracy of the program IMPUTE to generate the genotype data of partially or fully untyped single nucleotide polymorphisms (SNPs). The program uses a model-based approach to imputation that reconstructs the genotype distribution given a set of referent haplotypes and the observed data, and uses this distribution to compute the marginal probability of each missing genotype for each individual subject that is used to impute the missing data. We assembled genome-wide data from five different studies and three different ethnic groups comprising Caucasians, African Americans and Asians. We randomly removed genotype data and then compared the observed genotypes with those generated by IMPUTE. Our analysis shows 97% median accuracy in Caucasian subjects when less than 10% of the SNPs are untyped and missing genotypes are accepted regardless of their posterior probability. The median accuracy increases to 99% when we require 0.95 minimum posterior probability for an imputed genotype to be acceptable. The accuracy decreases to 86% or 94% when subjects are African Americans or Asians. We propose a strategy to improve the accuracy by leveraging the level of admixture in African Americans.

Our analysis suggests that IMPUTE is very accurate in samples of Caucasians origin, it is slightly less accurate in samples of Asians background, but substantially less accurate in samples of admixed background such as African Americans. Sample size and ascertainment do not seem to affect the accuracy of imputation.