We found the first ten PCs using hard call genotypes because those are currently supported by EIGENSTRAT. We did this once using all SNPs, and once restricting to SNPs measured on Illumina and imputed on Affy. Results were similar in the two approaches, and results from the latter are shown. The top three PCs are plotted against one another in Figure 2. We see that the second PC completely separates the cases (i.e., the Affy controls) and controls (the Illumina controls). Thus, when these PCs are included in a logistic regression predicting case-control status, we get a complete separation of data points, and the models cannot be fit.
