I wrote about how population stratification in genetic analyses led to a decade of false findings and almost certainly continues to bias emerging results. But we are starting to have statistical tools to sniff it out. A 🧵:
Mar 28, 2025 · 9:49 PM UTC
9
154
749
102,200
First, stratification = genetic structure + environmental structure. If two populations have some genetic variation (e.g. due to drift) and differing environmental influences on a trait, that will induce a false/non-causal correlation between genes and the trait.
2
5
82
6,232
When such false correlations are further aggregated into polygenic scores, they can accumulate into very large *apparent* genetic differences between even closely related populations. And these false differences will mirror the environment: environment looking like genes.
1
3
66
6,072
The backstory here is that population stratification led to a decade of thinking that very recent natural selection was acting on height. All shown to be false in co-published analyses in 2019. There's even a good press article on this specific fiasco: quantamagazine.org/new-turmo…
1
5
80
6,510
How common is this? Tan et al. (medrxiv.org/content/10.1101/…) derived an estimator of GWAS confounding by contrasting population and within-family effects; the latter ~immune to pop strat. And the confounding is substantial, >50% for socially stratified traits like income and IQ.
1
4
80
42,316
More recently, Smith et al. (biorxiv.org/content/10.1101/…) proposed an estimator of polygenic score confounding that can be directly attributable to stratification/ancestry. They show that it is present in target data from Europe, other continents, even ancient DNA!
1
1
56
5,257
To get a sense of the magnitude, we can build polygenic scores from the population and family GWAS data from Tan et al. for ADHD: a trait that showed 79% confounding and ~zero heritability -- a nice null. Then compute polygenic score means in different continental populations.
2
1
45
5,220
There are many reasons this leads to biased estimates, w/ multiple papers warning against it. And by comparing the population vs. family-based estimates, we can see that the biases are massive: significant differences that flip around arbitrarily for this ~0 heritability trait.
1
1
43
7,335
Large biases are also observed for heritable traits. When we play the same game for IQ GWAS, we see significantly different estimates for population vs. family-based weights, with African populations exhibiting a 1SD higher than average mean for the latter.
3
5
71
49,616
And all of the results are highly sensitive to parameter choice. An unscrupulous researcher could spin out all sorts of evo psych theories from these results -- maybe hot climates improve brain development! In reality, this is just a bundle of bias, stratification, and noise.
3
3
73
5,701
This problem is not restricted to global populations. One popular approach is to correlate the education polygenic score with number of offspring, as a measure of extremely recent natural selection. But this too is highly susceptible to stratification biases.
1
46
4,677
Indeed, when family-based statistics are used to estimate genetic correlation, with better control for stratification, the relationship is null or even reversed! Having more children is genetically correlated with higher IQ (green):
2
3
56
6,376
Yet another genetic finding that's been hanging around for ten years -- even leading to fringe concerns about "dysgenic fertility" -- but appears to be substantially or even entirely due to stratification.
1
50
3,833
And this may keep happening. Another recent study of polygenic score/offspring correlations made a big splash finding the most significant association with ADHD. What a coincidence that ADHD is the most confounded and least heritable trait in family-based analyses.
3
4
63
6,136
