rolloff analysis of North European admixture in Greeks.

One of the signals of admixture in my recent post on the Greeks on the crossroads of Eurasia was between north European and Near Eastern populations, with several pairs of such populations showing a significant negative f3(Greek_D; North European, Near Eastern) statistic. I used rolloff to estimate the date of this admixture.

Note that rolloff assumes a pulse model of admixture, whereby the two populations mix at a point in time, rather than experience gene flow over a protracted period. This may not be applicable in the case of Greeks, since gene flow may have occurred repeatedly throughout history. Also, rolloff estimates admixture times in the absence of very accurate ancestral populations, by exploiting allele frequency differences between them. So, for the first example below, with Finnish_D and Yemen_Jews as reference populations, which showed the most negative f3 statistic, this does not mean that Greeks are the product of admixture between Finns and Yemen Jews, but rather that allele frequency differences between these two populations reflect a contrast between North Europeans and Near Eastern populations, which may, presumably map, to the west Eurasian cline of diminishing Near Eastern "Neolithic" ancestry.

This experiment was performed on a set of 292,223 SNPs, and using the Rutgers map for Illumina chips. The first plot is using Finnish_D and Yemen_Jews as references. The fit does not visually appear extremely convincing, perhaps due to a smaller number of SNPs, or to the aforementioned deviation from the pulse admixture model.

(NB: I used the expfit.sh script with default parameters to do the exponential fit/plotting; note that negative weighted correlation points are not visible in the produced output)

The jackknife estimate of this admixture is 87.849 +/- 20.254 generations, or, assuming a generation length of 29 years, into 2,550 +/- 590 years.

The second plot uses Polish_D and Saudis as references:

The jackknife gives 67.235 +/- 22.148 generations, or 1950 +/- 640 years. The fit seems to capture the rise of the exponential for smaller cM genetic distances reasonably well.

Recently, Graham and Coop used fastIBD to identify a signal of possible Slavic admixture in the Balkans dating to the medieval period, using a similar generation time of 30 years. That method uses shared IBD segments between populations, so it may be limited to uncovering the most recent signal of admixture. Another piece of evidence comes from an abstract in ASHG 2012, according to which an Iron Age individual from Bulgaria was Sardinian-like. Since the Iron Age starts at the conclusion of the 2nd millennium BC, it might seem that a northern European element -whether present or not- had not admixed yet with the people who lived in the Balkans at the time. This seems to parallel the situation in two other earlier locations (c. 5ka in the Tyrolean Iceman and Gok4 Swedish TRB farmer), in which the North_European component was absent, although we cannot yet exclude its absence from the Iron Age Bulgarian, since a little such admixture might still leave an individual mostly Sardinian-like. Finally, levels of the North_European component in Greek individuals seem fairly variable, and this might indicate that levels of this element of ancestry had not had sufficient time to even out in the population.

(A different possibility is that the admixture signal reflects admixture of a Near Eastern kind. I consider this less likely, since there is evidence of "Southern" and "Southwest Asian" ancestry (in the K7b/K12b sense) already in Neolithic Europe.)

More research on the issue is certainly needed, but a first reading of the evidence suggests that this type of admixture may reflect events that took place during the historical period of Greek history.

Each of these experiments took about 1.5 days to complete. I am currently running another set of experiments with ~2-fold more SNPs, and assuming that finishes in good time, I may re-visit the question addressed in this post, to see if standard errors decrease and/or time estimates change with denser coverage.