The best use of the full posterior probability distribution for the three genotypes at imputed SNPs when testing association is a matter of ongoing investigation. Three possibilities are to call only those genotypes for which the posterior probability exceeds some threshold (the posterior mode approach), to summarize the distribution by using the posterior mean (the posterior mean approach), and to use the full distribution in the likelihood framework (the full data approach) [Marchini et al., 2007; Guan and Stephens, 2008]. The full data approach makes full use of the uncertainty in the imputed genotype calls whereas the posterior mode and posterior mean approaches are quicker and work best if there is high certainty about the imputed genotype calls at a given SNP [Marchini et al., 2007]. Depending on the downstream application of imputed genotype calls, discrete calls may be useful. We chose to adopt the posterior mode approach and make discrete imputed genotype calls. By setting a suitably stringent threshold, we were able to reduce the effect of genotype uncertainty. Under this approach, it is straightforward to test for Hardy-Weinberg
