In our simulations and data analyses using PCA [Luca et al., 2008] we found that outliers can interfere with discovery of the major axes of ancestry and greatly increase the number of dimensions of ancestry discovered. To illustrate the effect of outliers we created a subsample from POPRES including 580 Europeans (all self-identified Italian and British subjects), 1 African American, 1 E. Asian, 1 Indian, and 1 Mexican. Smartpca removes the four outliers prior to analysis and discovers two significant dimensions of ancestry. If the outliers are retained, five dimensions are significant. The first two eigenvectors separate the Italian and British samples and highlight normal variability within these samples. Ancestry vectors 3–5 isolate outliers from the majority of the data, but otherwise convey little information concerning ancestry (Fig. 1). This example highlights our previous results [Luca, 2008], which show that outlier removal is an essential stage in PCA analysis to determine ancestry.
