One approach to avoiding the strong influence of SNP clusters is to prune the full genome-wide SNP set before PCA to obtain a subset of SNPs in which all pairs have low correlations. In the T2D project, we compared two such SNP sets, one reported by Yu et al. [2008] selected to have pair-wise linkage disequilibrium (LD) of r2<0.004 and minor allele frequencies (MAF)>0.05 in a European-ancestry population. The other SNP set we selected, from among the 870,000 autosomal SNPs assayed on the NHS subjects, to have LD r2<0.04 and MAF>0.05. Both sets contained about 12,000 SNPs and the overlap is 445 SNPs. The first two eigenvectors obtained from the two SNP sets are very similar, whereas the third, fourth and fifth have much lower correlations (Figures S7 and S8). Similar sensitivity to SNP selection has been observed in other projects. These results suggest that, beyond the first one or two components, eigenvectors ordered by sample eigenvalues may not be robust indicators of population structure. However, we note that this does not rule out the eigensystem as a whole being
