After imputation, investigators run association tests as usual, producing p-values for each SNP and looking for the most significant SNPs. However, the imputation has introduced differential measurement error: for example, some SNPs are measured almost perfectly (through actual genotyping) among the controls, but measured imperfectly (through imputation based on nearby measured SNPs) among the cases. Furthermore, the imputation itself may introduce bias. Many imputation programs base the imputation on a database of known genomes, such as the HapMap. If the minor allele frequency (MAF) of a SNP in the HapMap differs substantially from the MAF in study data, imputation in cases only or controls only can yield very different MAFs in cases and controls. This setting has been recognized as potentially problematic. For example, when discussing combining data from studies using different genotyping platforms, Li et al. (2010) recommends imputing and doing association tests within platform and then combining the results using a meta-analysis approach, which cannot be implemented unless each platform has at least some cases and controls.
