Anti-Racist Science Can Fail by the Same Statistical Move It Condemns
2026-08-23 · AI-generated insight
These two notes describe what look like separate controversies—Gould's Mismeasure of Man and Lewontin's argument against biological race—but the alleged errors share a common statistical shape: ignoring the structure that only appears when correlated variables are considered together.
Winegard and Carl's point about Lewontin is explicitly about aggregation: any single genetic locus varies more within populations than between them, just as nose size alone can't sort faces by sex. The signal lives in the correlations across many loci. Meanwhile, Carroll's critique of Gould (as Warne recounts it) is that Gould misunderstood factor analysis—which is precisely the tool for extracting a common signal from many correlated measurements. In both cases, a celebrated anti-racist argument allegedly worked by decomposing an aggregate into pieces, showing each piece was individually weak, and declaring the whole an illusion. Whether or not one accepts the critics' conclusions, the parallel suggests this isn't a coincidence of two careless scholars but a rhetorically attractive fallacy: disaggregation always makes patterns look smaller, so it's the natural move for anyone motivated to dissolve a category.
The Morton affair adds a darker recursive twist. Gould's central claim was that Morton's racist priors unconsciously bent his measurements—and the 2011 re-analysis accused Gould of exactly the same unconscious bending, in the opposite moral direction. Bias in measurement, it turns out, is not a property of villains; it's a property of having priors. Kaplan's position in the follow-up debate may be the wisest of the lot: Morton's interpretations were flawed, and Gould was wrong to think he could read Morton's motivations from the data. That's a general lesson for the whole "Politics Shaping Science" file—we're usually on firmer ground auditing the statistics than auditing the soul of the statistician. The statistics here suggest a symmetric failure mode: virtue is no vaccine against motivated aggregation errors; it just changes which direction the error points.