Let’s say the two populations are Northern (orange) and Southern (red) Europeans and we are running a genetic association study (GWAS) of height, which tends to be greater in the North. With enough statistical power, the GWAS will identify all of the alleles that are slightly more common in the North (where people are taller) as “height increasing” and all of the alleles that are slightly more common in the South as “height decreasing”; whether they actually influence height or not. If we then use these “height” weights to build a genetic predictor of height for a completely new set of European individuals, the predictors will seem to show large genetic differences in height between the two groups. And these differences can grow very large as more variants are used in the predictor, since the stratification will always point the same way and accumulate. We thought we were training a predictor of height, but we actually trained a predictor of ancestry/environment that also happens to be directionally oriented with observed height. Not great. And because this is a predictor of environments, it will be correlated with all of the other environmental differences between Northern and Southern Europeans. So not only have we turned an environmental difference into one that looks like a much larger genetic difference, but we start to think that eating pasta or being a fan of Fellini movies or head size is also linked to a genetic propensity for lower height.
...
The worst part is that even though population structure itself is random, the GWAS orients all of that structure to match the phenotypes we actually observe, which makes the (false) genetic findings appear eerily plausible: genetically taller in the North and shorter in the South just like we see with our eyes! People sometimes ask why population stratification would just happen to line up so well with what we see phenotypically, but that is exactly what population stratification does: it lines up random genetic fluctuations with the observed phenotype in a way that then persists in independent samples.
...
This height example might seem far-fetched, but pretty much exactly what I described actually happened, and it led to a decade-long mess where the field was convinced that Europeans had undergone rapid natural selection on height (and other phenotypes correlated with height like … head circumference) only to learn in 2019 that it was all or nearly all explained by stratification (see Berg et al. and Sohail et al. eLife; or press coverage that concludes “this is a major wake up call … a game changer”). But prior to learning this error, the possibility of selection on head circumference got people speculating what else about the head could be under rapid recent selection. That speculation included an famous opinion piece by esteemed population geneticist David Reich raising concern that genetic analyses may soon reveal substantial biological differences among human populations on traits like intelligence; differences that we as a society were unprepared to grapple with. Naturally, in some circles, Reich’s cautious and circumscribed warnings that we may eventually find challenging genetic differences were read as a kind of Straussian message, a cryptic admission of precisely the “racist prejudices and agendas” Reich was attempting to head off (and, I should note, that he spent another two chapters in his book explicitly denouncing). Snippets from his editorial were further stripped of context, sometimes reworded entirely, and became meme fodder for open racists: Harvard’s superstar geneticist is secretly on our side, the truth about the inferior races will soon be revealed. And these memes continue to get passed around today, more than five years since the motivating height result was shown to be an artifact (in a paper on which Reich is a corresponding author no less). All of which is to say that poor control for population structure can have, well, some pretty big consequences.
From the article:
...
...