Single Loci vs. Correlation Structure: Where Lewontin's Argument Breaks
2026-09-18 · AI-generated insight
These two notes stage a genuine collision, and the fault line runs through a statistical subtlety rather than a moral one. Lewontin's famous finding — that most human genetic variation lies within populations rather than between them — is empirically true and has anchored the "race has no biological reality" position for fifty years. But notice how the note is filed: Correlation Across Many Variables. That filing names the exact weakness Edwards later called "Lewontin's fallacy." Any single randomly chosen locus tells you little about population membership; the aggregate correlation structure across thousands of loci classifies individuals into continental ancestry groups with near-perfect accuracy. Lewontin's argument is an argument about marginal distributions deployed against a question about joint distributions.
Warne's five lines of evidence live entirely in that joint-distribution world. Polygenic scores are nothing but weighted sums across many loci — they exist precisely because information that is negligible at each variant becomes substantial in aggregate. Admixture studies likewise depend on ancestry being genomically legible in a way Lewontin's framing implies it shouldn't be. So the debate isn't really "genetics says X" versus "genetics says Y"; it's a debate about whether you're allowed to sum.
There's a second asymmetry worth noting. Lewontin leaps from an empirical claim to a normative one in a single breath: racial classification is "positively destructive" and therefore no justification exists for it. Warne moves in the opposite direction, fencing his claim with caveats — >0% but of unknown magnitude, US-born native English speakers only, data "not fully conclusive." One can read this charitably (Warne is epistemically careful where Lewontin is motivated) or suspiciously (hedged framing is how contested claims get smuggled into respectability). Either way, the notes together illustrate a recurring pattern in politically charged science: the side claiming the moral high ground tends to argue from single variables and sweeping conclusions, while the side claiming the empirical high ground argues from aggregates and hedges. The honest synthesis is uncomfortable — Lewontin's statistic is true but doesn't support his conclusion, and Warne's methods are sound but can't yet deliver the effect size that would make his claim consequential.