Post date: Sep 13, 2013 9:25:39 PM
Patterson et al. (2012; Ancient admixture in human history) describe a 3-population test for admixture. Inferences is based on the expectation of the produce to the allele frequency difference between the putative admixed population and each of two putative source populations. This statistic should be >= 0 unless the putative admixed population is in fact admixed. Thus negative values are evidence of admixture (but see caveats in the paper about the 'outgroup issue'). One might fail to detect admixture if there has been substantial time for evolutionary change in the admixed population.
Unfortunately this software requires integer-valued genotypes. I could of course re-implement the methods in my own software, but I don't want to do this (at least not right now). So I converted the common variant mean genotypes to integers (0, 1, or 2) and replacing any mean genotype that was not within +- 0.1 of an integer value with 9 (the value for missing data). I wrote the resulting genotype matrix to a file (lycaeides.geno in projects/lyacaeides_admixture/admix3poptest/; the file is in eigenstrat format and is one of the three files needed).
I formatted the data to calculate f-3 statistics with admixtools (qp3Pop is the actual program). There were a few weird things I had to deal with. First, you have to supply chromosome numbers. I tried using the contig numbers but you are only allowed 23 (humans). The program actually uses this information for determining the independence of loci for a jacknife procedure. Because we only have one SNP per contig, I just arbitrarily assigned them numbers 1-20. Here is a document the describes how I binned sites into 'populations' (this is worth thinking about more), and here are the tests I ran. The results are here, right now we have evidence of admixture in dubois, carson pass, the warners, and most idas populations (including longinus or jacksonhole). More thoughts to come.