When we did in fact observe inflated Type I error after applying standard imputation quality filters, we explored a number of ways to lessen the inflation. We first considered controlling for platform effect as we would control for population stratification: by using principal components (PCs) as covariates in logistic regression. However, the platform effect was so strong and confounded with case-control status that we could not fit the models. Then we considered restricting to SNPs imputed with good accuracy. This approach yields excellent results, but reduces power by reducing the number of SNPs we can test. Finally, we considered the possibility of genotyping a small number of additional controls alongside cases on the new platform, who could be compared to the original controls in a preliminary analysis to identify aberrant SNPs. This approach yields good results, but requires the additional expense of genotyping more subjects.
