To investigate population structure, we use principal components analysis (PCA), essentially as described by Patterson et al. [2006]. The choice of which SNPs to use for principal components analysis is not obvious. Using all SNPs on a whole-genome array is computationally demanding, but feasible, and would seem to be the best approach in terms of utilizing all available information about genetic relationships. However, whole-genome arrays contain clusters of highly correlated SNPs and a single cluster may have a very strong influence on certain PCs, as noted previously [Novembre, et al. 2008; Tian, et al. 2008]. For example, in the Lung Cancer project (which consists entirely of European-ancestry subjects), when using all autosomal SNPs with missing call rate less than 5% (~545k SNPs), the first two PCs separate U.S. and Italian subjects, while the third PC separates both U.S. and Italian subjects into three distinct groups. These three groups correspond to the genotypes of a cluster of highly correlated SNPs in 8p23, a region that contains a polymorphic inversion. The same result was found previously in PCA of other European-ancestry populations
