Allowing for genotyping error in analysis of unmatched case-control studies

A commonly-used method for testing for association between disease and a single-nucleotide polymorphism (SNP) is to compare the frequencies of the SNP genotypes in a sample of unrelated cases to those in a sample of unrelated controls drawn from the same population (an unmatched case-control study). A drawback of such a study is that it is impossible to detect genotyping errors, and few methods have been developed to allow for the presence of undetected genotyping errors. In this paper, we obtain analytic formulae for estimates of genotypic relative risks in terms of error probability (e). In general, e will be unknown. We investigate the effect of assuming both correct and incorrect values of e on power and type I error, and also on the genotypic relative risk estimates. The choice of e was found to have no effect on power or Type I error probability (provided a 2df test was used, allowing relative risks of homozygotes and heterozygotes to differ). However, overestimating e in the presence of a true association was found in general to bias relative risk estimates away from the null, with underestimates of e having the opposite effect. Although e is unknown, it may be estimated from an external "validation" study, such as genotyping a sample of unrelated individuals twice and counting the discrepancies. Simulation results suggest that, for such a study, 25 individuals would be sufficient to give approximately unbiased estimates of relative risks.