Murray and most of race twitter has apparently been fooled by this...

145 views Oct 14, 2025 ~5 min read

Murray and most of race twitter has apparently been fooled by this completely fabricated analysis purporting to show African ancestry is associated with IQ. People lie on twitter all the time, but this is both more revealing and more disturbing than usual. A 🧵

Charles Murray@charlesmurray

Proposition: Everything that has been learned about the nature of the B/W difference in the subsequent 56 years of research is consistent with Jensen’s expectations in 1969. I stand ready to be corrected.

                            2
                        
Revealing in that it shows how quantitative racism is a just an exercise in manipulating data to fit the preconceived conclusion. Disturbing because this time private data is being used and the results, which cannot be easily verified, are just flatly invented.

What's actually going on? Some guy claims to have an analysis showing that African ancestry differences between siblings are associated with IQ differences in the UK Biobank. Implying an ancestry difference in the within-family influences.

But there's an obvious glaring error: in the same analysis, African ancestry was *not* associated with darker skin! Which would mean ancestry-specific alleles DO NOT have a direct influence on skin color. Pigment is, according to these guys, a social construct.

                            5
                        
We, of course, know that alleles influencing pigment differ across populations, and that more African admixture is associated with darker pigment. Anyone honestly looking at this result would realize something went wrong and try to debug it. But not this crack squad.

Instead, the result is forced to fit the narrative. We get a whole wall of text about how sparse traits like skin color won't actually correlate with African admixture. This makes zero sense. And you'll notice the model is never actually derived, just argued on faith.

When we look in papers that actually derive a model (medrxiv.org/content/10.110…), we can see that the derivation does not depend on sparsity at all. The analysis estimates the difference in direct genetic values regardless of the trait architecture.

                            8
                        
Or we can use our intuition: for a sparse trait, siblings will either have no trait differences or very large differences; for a non-sparse trait, siblings will often have small trait differences -- it's a wash. Sparsity impacts the variance and not the estimate.

Or we can just simulate it (gist.github.com/sashagusev/b92…): generate admixed families with a heritable trait that has allelic differences between populations, and confirm that the sibling analysis accurately recovers the true difference. It does! Regardless of sparsity.

                            10
                        
So the analysis is clearly bogus from theory, intuition, or simulation. Why invent a ridiculous explanation for it? Because the fake result is *useful*. If ancestry doesn't impact skin darkness, then one can claim that 'colorism' is not a relevant mediating factor.

                            11
                        
One other point you might notice in the simulation is it takes tens of thousands of siblings to get an accurate estimate. Within-family ancestry variance is low, and so the number of families needed for statistical power is very high. This is also a well known result.

The recent Wang et al. analysis used ~17,000 siblings with more ancestry variance and still had a standard error of ~0.25. Earlier, the blogger gwern ran power calculations and came to a similar conclusion (gwern.net/note/statistic…):

The UK Biobank analysis purports to have a much tighter standard error (~0.13) so how many sibling pairs did they use? Hundreds of thousands? Nope ... 175! That's right. They looked at 175 families and concluded that racism is real once and for all.

At this point it should be clear that we're dealing with complete bullshit. For the interested, a longer discussion (), with multiple derivations coming to the same conclusion. The accounts behind the analysis don't even bother to respond.

Receipts Only@BuzzLightworth

If I'm wrong on any of this, point out where. Calling it "crankery" without modeling/diagnosing selection (i.e. whether UKB participation S depends jointly on ancestry A and cognition Y) is hand-waving. In UKB, S depends on education/health/identity; conditioning on S=1 opens A→S←Y. Sibling FE removes shared family background, not that collider - and requiring both sibs (S1=S2=1) is a second selection step that strengthens it. Am I missing something?

If this is truly full-sibling within-family, the within-pair ancestry SD is ~1–3 percentage points. With that variance and a slope ~0.15 IQ points per pp (≈0.01 SD/pp), you typically need ~20k+ sib pairs to reach p≈.01 even before correcting for ancestry measurement error, which should attenuate the slope further. If strong p's survive with far fewer pairs, the live explanations are half-sib leakage (reintroducing confounds) or selection doing the work.

So to be perfectly clear, my critique is: (i) UKB selection opens A→S←Y even within true full-sib pairs, and (ii) strong p's with small N are inconsistent with full-sib ΔA variance unless half-sibs slipped in or selection is driving results.

Easy way to settle this would be to:
1. disclose the kinship filter (including an IBD2>0 full-sib criterion) and the ΔA histogram;
2. report mean ΔA, σ within(ancestry), residual σ, and an explicit selection strategy (IPW or bounds) with diagnostics;
3. show expected within-family shrinkage for PGS vs admixture in the same sample/spec;
4. if you used a latent g, demonstrate (partial) measurement invariance / DIF checks across groups.

To preempt a misread: there's no basis to invoke any "social channel" within true full-sibs; that's a between-household pathway and would matter only if half-sibs/step-households were mixed in. If I'm wrong, show the DAG/selection model where A→S←Y is irrelevant and the IBD2-based QC that rules out half-sibs, and I'll drop it.

I'm belaboring the point is because it is endemic of the broader quantitative racism project: find data that fits the narrative, fake it if necessary, and when someone finally points out the errors, the time for argument is past and we move on to a new spectacular claim.

First, it was individual candidate genes that had massive influences on behavior and cognition. When these failed to replicate spectacularly, the excuse was that the *theory* still made sense (ironically, this claim was recently deleted!).

Sasha Gusev@SashaGusevPosts

Winegard spends most of his time advocating for white identity politics based on discredited science (e.g. 5-HTTLPR) and the rest of his time advocating for scientists to be less political and more evidence-based. It's a great gig!

Then it was the claim that heritability estimates themselves could be manipulated to prove the genetic differences between groups. When this was mathematically disproven (pubmed.ncbi.nlm.nih.gov/38470926/), suddenly people like Murray lack the rigor to evaluate it.

Then it was the claim (borrowed from Jensen) that IQ tests cannot be driven by culture because more g-loaded subsets are more "culture free". When this too was shown empirically to be the exact opposite of the truth ... crickets.

Sasha Gusev@SashaGusevPosts

Murray: Race differences are higher on culture-free and g loaded tests like the Ravens

"Every part of this is incorrect"

Murray: Happy to take a look, what's the cite?

*months later, on bluesky*

Murray: Race differences are higher on abstract, culture-free tests

Then it was the claim that genetic scores for IQ are lower in African samples. When a proper analysis revealed the results to be completely *flipped*, this did not lead to a sudden interest in African genius -- the earlier claim was simply forgotten.

Sasha Gusev@SashaGusevPosts

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.

Then it was the claim that admixture correlates with IQ. When this too was pointed out to be environmentally confounded, lo and behold: a set of 175 sibling pairs was immediately identified and manipulated to make a bold new result. And here we are.

Sasha Gusev@SashaGusevPosts

Admixture analysis, where % genetic ancestry is correlated with some outcome, is a favorite of online racists because it can make environmental influences appear to be genetic. This was always known conceptually, but a new paper demonstrates it empirically.🧵

In each instance, the goal was not to get at some underlying truth, but to add more bullshit to the conversation faster than careful people can refute it. Create the appearance that the prophecy is always coming true but ~they~ are hiding it from you.

evo psych googling@evopsychgoogle

charles murray, prophesying the arrival of evidence vindicating the bell curve

1994: in three or four years we'll see whos right ;)
1995: five years from now
2014: probably less than a decade
2019: within a few years...
2049: (moments before death) any minute now

                            22
                        
Coming back to Murray and Jensen. Murray, frankly, should know better. But at this point he's basically a victim of elder abuse: high-tech scams that he himself does not even pretend to understand and merely echoes because they fit the narrative he spent his life cultivating.

                            23
                        
Jensen, for all his ugliness, at least attempted to do research. He went out into the field and collected data and tried to report the results. But this new generation of Jensen-ites is different: truth is irrelevant, they can simply put on the performance of doing research ...

                            24
                        
... grabbing data, forcing it to fit, constructing elaborate epicycles to explain inconvenient results, then abandoning them as convenient.

And now that multiple rounds of these pseudo-analyses have resulted in failure, they've finally graduated to ... just making shit up. /x

Murray and most of race twitter has apparently been fooled by this...

Actions

What You Can Do

Actions

What You Can Do

More in Science & Research

More from @SashaGusevPosts