Comparing five methods of admixture estimation

In my comments on the Moorjani et al. (2011) I argued that admixture proportions presented in the paper are inaccurate, and gave my reasoning behind this claim. Moorjani et al. (2011) also present STRUCTURE 2.2 results.

Naturally, I wanted to see whether independent admixture estimates on some of the same populations had been estimated in the literature. This brought me to Pugach et al. (2011) which introduced a wavelet-based admixture estimation method called StepPCO. In that paper, the authors presented estimates of the extent and timing of admixture for some populations also included in Moorjani et al. (2011). They also compared with HAPMIX, a well-known method using a completely different methodology, and presented their comparative data in this table and in the body of their paper.

Hence, we now have 4 different estimates of admixture for some populations. To these, I decided to add supervised ADMIXTURE 1.1 results. I used CEU and YRI as "West Eurasian" and "Sub-Saharan" references so that I would be in accordance with these other methods.

(The ADMIXTURE results were obtained by merging the datasets in PLINK with the --geno 0.001 option, then pruning the combined set for LD with --indep-pairwise 50 5 0.3)

The following table summarizes the estimates.

Note that these are estimates of Sub-Saharan admixture assuming two parental populations; also, Moorjani et al. break up the Bedouin sample into the two distinct groups it is composed of, so I have taken a weighted average of the figures in their paper.

It is difficult to make meaningful statistical inferences on only a few comparison points, but I do observe that STRUCTURE 2.2 on 13,900 markers gives the higher estimates, followed by Moorjani et al.

The other three methods cannot be ordered, giving higher estimates in some populations and lower in others. They all give, however, lower estimates than both Moorjani et al. (2011) and STRUCTURE.

As I've explained in my earlier post, Moorjani et al. (2011) have higher estimates of admixture because they measure it by comparing populations' shift on the East Eurasian-African axis, ignoring the Asian-shift of North Europeans and adding it to the African-shift of southern Caucasoids. This leads them to conclude a few percentage points of African admixture in populations that have virtually none (such as Sardinians and North Italians, even Swiss French). For populations that do have noticeable African admixture (such as those on the table) their overestimates amount to a a few percentage points.

Three of the methods also provide an estimate of time since admixture:

There is no simple relationship between these times, but an obvious pattern is that the dates of Moorjani et al. are younger, perhaps less than 50% of the other two methods.

Conclusion

Clearly, the art of admixture estimation is still in its infancy, and different methods provide different results even with a simple 2-population model. I've argued how the results of one method can be harmonized with those of the others, but I don't have a ready explanation about the substantial age differences. Pugach et al. argue that their method is better than HAPMIX, but the differences between the two seem small compared to the differences of both methods to ROLLOFF (the method of Moorjani et al.'s paper).

The discrepancy is even more interesting if one takes into account the fact that HAPMIX and ROLLOFF were done by many of the same people. Hopefully, someone will be able to figure out the cause behind the discrepancy. A commenter in my earlier post suggested that ROLLOFF produces younger ages because its age estimation is tied to its inflated admixture proportions; this could be true, however, the discrepancy exists even in populations where the relative difference is small.

A speculative historical coda

Historical explanations about the circumstances of this admixture need to be made with some caution, due to the uncertainty about admixture times.

For example, a doubling of the Moorjani et al. age estimates would disentangle the Sub-Saharan element in Levantine Arabs from the Islamic epoch. A doubling of the admixture date for Jewish populations, as presented by Moorjani et al. would bring that admixture's age to the middle of the 2nd millennium BC, a period in which the Hebrews were said to be in Egypt, where potentially they may have collectively acquired a small African admixture.

Hopefully, with time and full genome sequencing, we will get a better idea of what these African signals in some West Eurasian populations represent.