Does Molecular Data Overturn IQ Twin Studies?

Replying to "No, intelligence is not like height"

Scott Alexander writes

Sasha Gusev has written an argument that twin studies are inaccurate and the heritability of IQ is much less than previously believed. It’s much better than the average obviously-dumb-and-motivated post trying to argue this, and contains lots of arguments I hadn’t seen before. See also subreddit comments on original, Gusev’s responses, comments on responses. My impression is that there’s enough circumstantial evidence (eg adoption studies) that this probably has to be wrong, but I don’t think any of the arguments against it land, and I don’t know enough statistical genetics to critique it myself. I’d be interested in seeing one of the more mathematically-inclined pro-heritability people give their impression.

I would say I know a lot of statistical genetics, so I thought I’d give my take.

Gusev has 5 main points:

1. IQ is much less heritable and more confounded than height

2. Genetic effects on IQ differ within families much more than for height

3. IQ estimates are much more biased by participation than height

4. The genetics of IQ is much more environmentally sensitive than height

5. Unlike height, no one knows what IQ is actually measuring

I’ve argued before against Gusev that classical twin studies are very solid, and if biased, they likely underestimate the broad-sense heritability of IQ. The narrow-sense heritability is usually estimated to be around 50%. The difference between these isn’t that important for the purposes of this article — we’ll just say that Scott is right that twin studies are like an enormous boulder that Gusev is tasked with moving.

Does Gusev succeed? Let’s take a look.

Is IQ is much less heritable and more confounded than height?

Source

Gusev cites the above study to show that the GWAS h^2 for IQ is confounded due to non-individual effects for IQ and not height. However, Gusev makes up his own chart to remove the error bars for this, to obscure the fact that the study found no evidence for this in IQ. The error bars for the population and within-sibship IQ h^2 overlapped. Thus, this study does not move the boulder of twin studies.

Note two things. First, the statistical significance for height’s within family effects being smaller than between family effects is GREATER than that of IQ in this study, yet Gusev represents it like this:

Second, the h^2 of IQ is highest of all traits in the study, 2nd to height.

So Gusev fails to demonstrate that IQ is “more confounded” than height. But does he show that IQ is less heritable?

There are reasons why the SNP h^2 for IQ is currently lower than the SNP h^2 for height. As background, neither are as high as the twin study estimate. This is because the detection power of current GWASs is limited. This means that if IQ is contributed by more genes with smaller effect sizes than height, then current polygenic scores will be missing more of the IQ genome, thus underestimating the IQ h^2 more than height.

Another problem is that in many biobanks, the IQ tests used may be inferior in terms of reliability to those used in twin studies. For example, the UK Biobank cognitive battery was unusually brief and had poor test-retest reliability.

Finally, height always had higher heritability in twin studies than IQ. The meta-analytic h^2 of IQ from the classical twin design is about .66 and for height it’s about .8. IQ score is therefore less heritable than height¹ to some extent, and it having the 2nd place SNP h^2 after height is what is expected from twin studies.

Consequently, Gusev has failed to show that intelligence is “much less” heritable than height or that GWAS shows that IQ is “much less” heritable than previously thought, relative to height. Meanwhile, molecular methods underestimate the heritability of height and IQ, relative to twin studies, to about the same degree.

Do genetic effects on IQ differ within families much more than for height?

Gusev writes:

Young et al. showed that some of the bias is due to population stratification, whereby the population effects are picking up correlations with where people tend to live and the environments therein⁵ and hypothesized that study participation bias may also be at play (see next section). In other words, what was initially thought to be caused by genetics turned out (to some extent) to be caused by zipcodes.

He cites this:

Source

What this shows is how much SNP effects on these traits correlate with each other when they are estimated using a within-family model or a whole population model. It shows the covariance between a SNP and IQ tends be a bit different depending on whether it’s among everyone or within a family. Meanwhile, there’s no difference for height.

I guess Gusev is somewhat right here — IQ is different than height in one important way. It’s a fundamental basis for class. This means smarter people have more resources and higher mate value. Therefore, they get mates with better genes on other traits, like health. Therefore, a SNP that causally reduces heart disease risk comes to predict IQ in the population but not within families. This a possible mechanism of population stratification.

This is a cool theory² and the paper Gusev cites is okay circumstantial evidence for it. It’s not evidence that IQ is not heritable and is actually just caused by nice neighborhoods — it’s still probably genetics, Gusev cited the SNP h^2 of the within family SNP effects in the first section and it didn’t significantly differ from the population SNP h^2.

Classical twin studies aren’t touched by this any more than they are by the first section.

Are IQ estimates are much more biased by participation than height?

Source

Participants aren’t representative here by either height or IQ. They are a little less representative, but not by much — about .2 SDs (.6 - .4) at 1 SD of participation, the equivalent of .6 inches or 3 IQ points.

Gusev says this is a source of bias but forgot to point out that typically this kind of bias is a downward pressure on effect sizes, the downwardness increasing with the correlation between the trait and participation.

What happens to your correlation as you range restrict X. Low X variance, lower correlation with Y.

There’s no reason to think this bias is inflating IQ heritability estimates, and so if anything this could explain part of the gap between h^2 SNP and twin study h^2 estimates for IQ and height.

Are the genetics of IQ is much more environmentally sensitive than height?

He could provide evidence for samples being higher IQ than the general population inflated h^2 estimates but next he does the opposite. He says IQ is more heritable among lower SES (lower IQ) people. So that would mean his earlier h^2 SNP estimates are underestimated, since they come from higher IQ samples. Which fails to overturn twin studies.

I’ll let you in on a secret though: interaction effects are usually p-hacked, so depending on how you look at it this is either more or less of an error than if it wasn’t p-hacked.

Source

Hypothesis: IQ h^2 SNP for this sample is .27. Do any groups differ from this significantly? No.

There is a significant effect for education, but education is not IQ so we will not discuss that here.

Unlike height, no one knows what IQ is actually measuring

This can’t possibly move our priors on twin studies, because people have been arguing about this since the twin study days. I direct readers to lookup old articles and blogposts on the predictive validity of IQ and so on. It’s too boring and hashed out to redo here. The gist is that, yes, IQ is measured in terms of deviations from the mean, while height has an absolute scale. IQ is also harder to measure than height — it’s more expensive, and takes longer to do right. These problems all contribute noise to the measurement, which reduces the heritability. So if anything, we should expect people who believe intelligence is more heritable than IQ estimates of heritability to be pointing out the flaws in measuring IQ, not people who think IQ heritability is overestimated relative to intelligence.

Conclusion

Gusev fails to move my priors on twin studies. While IQ differs from height in some obvious ways, none of these relate to IQ’s heritability — molecular methods seem to chronically underestimate the heritability of all traits, including height, compared to twin study estimates.

Gusev concludes:

Nearly a decade of genetic studies of cognitive outcomes were done under the assumption that the effects being identified were direct when that was largely not the case. Prior to that, over a century of race science argued that intelligence is just like any other biological trait, with individual differences explained by simple genetic causes that are easily quantifiable and culturally immutable. It turns out molecular genetics has (for lack of a better word) thoroughly debunked that view. But when this cold fact is pointed out the response is to simply deny it ever happened.

Molecular genetics hasn’t thoroughly debunked the idea that IQ is somewhat less heritable than height, and about as “culturally mutable” as height. No where in Gusev’s post does he demonstrate the “cultural mutability” of IQ or height. 2 out of his 4 main citations are not statistically significant and the other 2 are evidence of population stratification and range restriction in on IQ in GWAS. The latter decreases h^2 SNP estimates and the former is not evidence of “cultural mutability.”

1. This does not mean intelligence is less heritable than height. The test-retest reliability of the best IQ tests is the same as the IQ correlation between identical twins reared apart. Most of the lower heritability of IQ compared to height comes from measurement error being dumped into e^2.

2. Classes being miniature ethnicities is an idea someone like Gusev would probably hate.