Genotype Imputation

Abstract A missing data problem arises in genetic epidemiological studies when genotypes of particular markers are unavailable for analysis for reasons of data quality, cost efficiency or technical design. In such instances, imputation methods can be used to extend the process of scientific inference making from the typed to the un‐typed markers. The information required to infer unobserved genotypes from observed genotypes is provided by the so‐called imputation base, an internal or external set of comprehensively typed individuals (often taken from HapMap or the 1000 Genomes Project) that is representative of the study population as a whole. The most popular genotype imputation methods, including IMPUTE, fastPHASE, MaCH and BEAGLE, employ a Markov chain model of the haplotype distribution in the population of interest. Although these frameworks have been shown to provide accurate and efficient tools of ‘ in silico genotyping’ under certain conditions, their uncritical use nevertheless must be cautioned against. Key Concepts Relevant genotype data may be missing in genetic epidemiological studies for technical or efficiency reasons. Scientific inference that takes missing genotype data properly into account can be made using data imputation methods. Genotype imputation requires an imputation base, that is, a population‐representative set of individuals (such as the HapMap or the 1000 Genomes Project) who are genotyped for all markers of interest. Established genotype imputation methods employ a Markov chain model of the haplotype distribution in the population under study. Genotype imputation may achieve 90% accuracy for highly polymorphic markers but performs less well for rare variants. While genotype imputation may provide valid statistical tests of genotype–phenotype association, their use for effect size estimation and significance assessment must proceed with caution. Genotype imputation needs to follow the same rules of good scientific practice as laboratory‐based data generation.