Alternatively, both the observed strong genetic structure and deep divergence time estimates are consistent with the possibility that Ginsenoside-F4 populations of A. lyrata ssp. lyrata have persisted throughout the last glacial period. The existence of large, ecologically stable populations in A. lyrata makes it suitable for studying local adaptation. Indeed, a growing number of A. lyrata genes show evidence of local adaptation. Examples include the trypsin inhibitor ATTI2 gene in the Plech, Germany population, a centromeric region in the Russian population, a centromere specific histone gene, and a gene for trichome density in Swedish populations. Building on the idea that locally adapted loci Oxysophocarpine should have increased differentiation relative to neutral loci, we identified genes that exhibited extreme values of FST compared to expectations under the estimated demographic null model. This approach has the advantage of explicitly incorporating divergence and demographic processes, which can otherwise confound assessment of selection. Simple scans for loci of unusually low diversity would incorrectly suggest selection at many of the loci surveyed: in the Canadian population, for example, nearly half of the loci sequenced are devoid of variation. However, like tests of the site 
frequency spectrum, diversity, or LD, our approach can only reject a neutral null model �C it cannot provide evidence in favor of an alternative model with selection. It will also be limited by the accuracy and appropriateness of the demographic model used. However, our inferred model fits the data well, including a measure of the site frequency spectrum that was not used to fit the model. One difficulty with our model-based approach is that its statistical power is not well known. Because the initial demographic model is fitted to summaries of all the data from all loci, including variances of summary statistics across loci, it accounts for the full range of polymorphism among loci, including loci affected by recent selection. Simulating from the posterior distribution of parameter values rather than point estimates similarly broadens the range of summary statistics produced. The results of this analysis are thus likely to be conservative, sacrificing some power to detect selection but minimizing the potential for false positives. Finally, we have focused exclusively on FST as a measure of selection.