Contents

Estimating admixture proportions when most individuals are admixed.

Estimating admixture proportions can be particularly challenging if there are very few representatives of the parental populations. There is an example of this for simulated data in Pritchard et al. 2000b. The data were supposed to approximate a sample from an African American population in which most individuals had some degree of European ancestry. For those data, the estimated ancestry proportions were highly correlated with the true (simulated) values, but the actual proportions of ancestry were biased. That example is also representative of our more recent experience with real data. This occurs because in the absence of any non-admixed individuals, there may be some non-identifiability where it is possible to push the allele frequencies further apart, and squeeze the admixture proportions together (or vice-versa), and obtain much the same degree of model fit. For this reason, the choice of prior model for the allele frequencies can have a big impact on the magnitude of estimated admixture proportions in this type of situation, and care is needed not to over-interpret the results.