Missing Heritability Revisited

Following up on Sasha

Sasha Gusev had an excellent, informative post last week about all the new methods of estimating heritability and what they mean. Between families, within families, common variants, rare variants, whole genomes. I refer you to his post for all the technical parts. Here, I want to reflect for a second on what all this means for how we think about human behavioral differences.

When you get into a conversation about heritability, it is very easy to get sucked into a discussion of what the heritability of something is, as though the heritability of BMI or intelligence is some kind of evolved biological parameter that we are working to estimate as accurately as we can. But one of the most often-repeated and quickly-forgotten facts about heritability is that it is not a property of a trait, it is a property of a trait in a given population. That is repeated so often as to seem pedantic, but it is important. A heritability is a ratio of variances, and variances (especially human behavioral variances) depend on the time and place. We can ask, what is the variance of income in the US in 2025, but to ask in some general way what the variance of income is makes no sense. In a heritability coefficient we have to worry about this both in the numerator and the denominator, which just makes the indeterminacy worse.

To make his points about how different methods of estimating heritability change the values that result, Sasha makes the unusual decision to take the mean heritabilities across a wide range of phenotypes, broken down by method. I understand why he did this, and I learned a lot from all the sorting through the different ways to go about it, but it risks leaving the impression that there is some heritability floating out there that we will eventually nail down. There isn’t, not for individual phenotypes and certainly not for phenotypes in general.

The context for all of this is the missing heritability problem. Back in the twin study era, the heritabilities of most phenotypes were coming in shockingly high, at .6 if not higher. (Arthur Jensen corrected IQ scores for unreliability and arrived at .8) DNA-based molecular heritabilities are almost always much lower than that. Were the twin studies wrong? Are the molecular studies missing something important? Sasha’s post provides a clear path through all the modern work, showing that twin study estimates were almost certainly inflated, that rare variants probably don’t add much to the equation, and that the most up to date within family estimates of genetic transmission give the best guess as to how heritability (remember those caveats) are going to work out in the long run. Taking that funny mean across phenotypes it comes out around .3.

Like every non-zero estimate of heritability that has ever been suggested, this has led to celebration and gloating on hereditarian Twitter. Gusev admits that the heritability of IQ is .3! But first of all, that isn’t what he said— he said that the average phenotypic heritability was around .3. He states specifically that there still aren’t good estimates of within family heritability of IQ. Second, it is the hereditarians who are moving the goal posts here. Speaking for myself, I have never once suggested that the heritability of anything was going to turn out to be zero. Of course genes matter for how we differ in our intelligence. But it is a long way from Jensen’s .8 to .3, and I suspect that .3 is going to turn out to be on the high side anyway.

Another consideration is that what we are talking about here— differences in numerical coefficients— is only one leg, and probably the least important, of what Lucas Matthews and I have called the three legs of the missing heritability problem. The other two legs are the “prediction gap” and the “mechanism gap.” The prediction gap refers to the contrast between polygenic scores and what could be done by simply predicting phenotypes from the phenotypes of first degree relatives. Phenotypic parental IQ accounts for something like half the variance in offspring IQ. A properly controlled PGS accounts for 5% on a good day. (The embryo testing company Herasight has recently made big claims that they can do better, but I will wait until I see a peer-reviewed paper.) Sasha doesn’t get into polygenic scores in this post, but I think they are ultimately much more important than heritabilities, because they have real-world applications that heritability coefficients do not.

The third leg of the missing heritability problem is the mechanism gap. This refers to the fact that no matter what our estimate of heritability is, for human behavioral phenotypes we have essentially zero knowledge of how genetic differences produce behavioral differences. We can analyze variance, in Lewontin’s famous phrase, but we can’t analyze causes. All of the interesting within and between family work that Sasha catalogs is a step in the right direction, but the truth is there isn’t much progress toward actual mechanistic understanding. I personally doubt there ever will be (for complex human behavior), but that is another topic. The mechanism gap is crucial to my bet with Charles Murray, for example. Murray bet that by 2025 we would have IQ differences pretty much figured out, and that entails more than a reasonable estimate of a black-box heritability coefficient.

So where does all this leave us? For most human behavioral traits, genetic differences have a significant, non-ignorable effect on outcomes, but not the dominant, deterministic effects that have been anticipated by hereditarians since Galton. If you are conducting studies of how parental personalities are related to the personalities of their children, you can’t just correlate them and assume it is a matter of behavioral modeling. Try it in adoptive families and see what happens.

If that sounds disappointing coming from an anti-hereditarian like me, it shouldn’t be. I have always said it would be naive to expect the genetic effect on anything to be zero, and by now there is a deep pool of data showing that it never is. And the truth is, there aren’t many people writing these days who are willing to adopt a strictly zero-genetic-influence position. Jay Joseph is pretty much the only one I can think of.

On the other hand, our inability to define anything even resembling deterministic causal mechanisms underlying heritability coefficients places strict limitations on their application in the real world, or on our theoretical accounts of how human inequality (in the nonpejorative evolutionary sense) comes to be. If we ask, “if I had different genes, would I have a different IQ,” the answer is probably, yes. If we ask, “if I had these specific genes, how would my IQ be different?” (never mind why) the answer is almost entirely, we don’t know. We don’t know because it depends on a thousand other human contextual variables that we have no hope of controlling.