Tootfinder

Eyeballing Figure 1 of their response actually seems to support this: the three subregions in the Americas contain nearly 80 % of all polysynthetic languages. In each of them, the median population size lies below the global median. However, if we compare within each of these three regions, polysynthetic languages have a higher median L1_population size than non-polysynthetic ones. Might this pattern point towards a classic Simpson's paradox?
A negative global association arises because polysynth lang are concentrated in regions with smaller overall populations, even though within regions the relationsh is positive. Once we account for that structure—as our mixed logit models do—the supposed "global" negative effect reverses direction.

Finally, what Xia & Lindell call a "separation problem" is, in our view, a feature of our approach and not a bug.
If, e.g., all languages in a family are polysynthetic (or none are), that’s not a statistical artefact – it’s the signal. The outcome is well associated with genealogy, showing that family membership captures someth genuinely informative about the process. When the model finds that family explains a large share of the variance, that's not a failure–it's evidence that phylogenetic structure dominates the pattern.
So while Xia & Lindell insist that "autocorrelation due to relationships and distance cannot be captured in family or regional-level analyses", we see that as an empirical question – and we treated it as one.
The real test is whether a mixed model that explicitly represents phylogeny and geography performs worse than their alternative, where the entire shared history of languages and environments is effectively collapsed into a single dimension (an eigenvector).
In other words: we model relationships – Xia & Lindell summarise them into one number per language.