Assessing the power of ascertained SNP datasets to reject incorrect admixture graphs.

A set of 100 topologically diverse poorly-fitting graphs was generated for each simulated random topology (see Methods) and fitted to both non-ascertained data and to all SNP sets ascertained using the HO one-panel scheme. (a) Boxplots summarizing distributions of worst f4-statistic residuals (WR) of incorrect graphs fitted to datasets of several types: non-ascertained, ascertained on the outgroup with an effective population size of 100,000 that was co-modelled with the other populations (“OG”), on the root population sample that was not co-modelled with the other populations (“true root”), on the root of all non-outgroup populations that was not co-modelled with the other populations (“non-OG root”), and on non-outgroup populations co-modelled with the other populations (“non-OG leaves”). Results for all the simulated topologies were pooled. (b) SNP sets generated by HO one-panel ascertainment on simulated data in the “bias” vs. “power” coordinates. WR of the true simulated graph serves as a measure of ascertainment bias (on the x-axis), and median difference in WRs of 100 poorly-fitting graphs fitted to ascertained and non-ascertained data (on the y-axis) serves as a measure of power to reject incorrect models. Results are shown separately for SNP sets ascertained on OG, true root, non-OG root, and non-OG groups. Each dot represents an ascertained SNP set. (c) This series of plots is similar to that in panel b, but another measure of statistical power is used: the proportion of graphs which fit non-ascertained data poorly (WR >3 SE) but fit ascertained data well (WR (PDF)