We have no complete solution to these issues, though there is no reason to think that our eigenvector-based methods are more sensitive to the problems than other techniques [9]. One check we do recommend is to generate a test matrix by taking the initial counts C(i,j) to be 0 if the corresponding data is present; otherwise, set C(i,j) = 1. This is equivalent to only focusing on the pattern of missing data. The eigenanalysis on this test matrix will show significant TW statistics if the missing data by itself is showing evidence of population structure. If so, the results should be regarded with some suspicion, especially if the eigenvectors show high correlation to the eigenvectors of the main analysis. We here echo [38] and recommend that the analyst should “control all aspects of source, preparation and genotyping, using the paradigms of blindness and randomization,” but, as the reference states, this will not always be possible.
