Ancestry and Education

Indirect, direct, confounded and quasi-causal

Siqi Wang and a group of co-authors including Alex Young and Peter Visscher have published a preprint titled, “Direct effect of genetic ancestry on complex traits in a Mexican population“. For some time, people interested in the relationship between ancestry and complex human phenotypes, especially people interested in American Black-White IQ differences, have been using ancestry estimates as a way of addressing the problem. It is now possible to use GWAS data to come up with valid estimates of percentage of different ancestries in admixed individuals. (Of course all humans are “admixed” in the sense that no one has a pure ancestry. Here, admixture refers specifically to individuals with relatively recent mixed ancestry from distinct, identifiable populations.)

Admixture and Causation

Differences in admixture ratios, and their correlations with behavioral phenotypes, would seem to offer insight into the age-old question of whether any of the many observable differences in behavior among ancestral groups are related to genetic differences among the same groups. But there is a problem. Across unrelated people from different families, ancestral differences are related to a million other things. How do people get admixed in the first place? They are not intentionally and randomly bred. If an enslaved person has children with a master, the children of that union are going to inherit not only the master’s genes, but also the environmental advantages of having a privileged parent. Therefore ancestry effects are not simply genetic; they are confounded will all environmental effects related to ancestry. As such, the design is essentially useless for the purpose of distinguishing genetic from environmental causes of group differences.

These considerations parallel many recent advances in thinking about heritability and polygenic prediction. Polygenic scores based on unrelated individuals are associated with environmental effects associated with parents, extended families and cultures. So the field has recently turned to estimating SNP heritability and PGS within families rather than between them, using sibling pairs, parents and children, or both. Differences between siblings control for many between-family environmental effects, like population-stratification, parenting, and culture. As a result, heritabilities and PGS effect sizes for socially-mediated phenotypes like educational attainment are generally much smaller within families than between them.

In the same way, Wang et al show that you can use full siblings to estimate ancestry differences within families. The within-family variance of such differences is obviously much smaller than it is between families, so there will always be statistical power problems, but in principle it gives you a way to quantify ancestry differences while controlling for non-causal confounds at the family level.

“Quasi-causal”

The between and within family strategy has interesting analogies with what many of us were doing with twins in the oughts. In those days I was criticizing the use of twins to estimate heritability coefficients, mostly because I didn’t think the heritabilities offered any insight into causation once you estimated them. What was riding on the question of whether extraversion had a heritability of .3 or .6? Basically nothing, if you ask me. When I would give talks about those ideas, mainstream twin researchers would ask a reasonable question: OK, smart guy, if we aren’t supposed to compute heritability from twin studies, what should we do? Should we just stop doing twin studies?

At the time, I was working with my colleague Bob Emery about classic problems in developmental psychology. What were the effects of parental divorce on children? Was physical punishment detrimental to kids? The problem with questions like these is that there is usually no way to establish experimental control. You can’t randomly assign kids to divorced and non-divorced families. It was the inverse of the admixture problem: the presumed environmental effects of parental divorce were confounded by a zillion known and unknown genetic and environmental differences between divorced and intact families.

So we used the Australian Twin Register to find twin mothers who were discordant for marital status, and compared outcomes in their children. Just as in sibling-GWAS, comparing siblings controls for a raft of between-family confounds, both genetic and environmental. We worked hard on the methodology for these studies, and I am proud of the work we did, but the truth is we never reached any definitive causal conclusions, despite our clever twin controls. Why? Because the twin method, even as it improves on population-based methods that don’t use within-family controls, still leaves many things uncontrolled. We recognized that after you control for between family confounds, within family confounds remain. The obvious one in the divorce study is the fathers, who are obviously going to differ between divorced and non-divorced pairings. Twins are a quasi-experimental method, and like all such methods they help with but don’t really solve the problem of non-experimental confounding. I would not want to base marital policy on what we were able to establish. When we found an association that survived the between-within test, we called it “quasi-causal”. We intended this as a marker of epistemic, and especially causal, humility.

Sibling Admixture and Confounding

Back to the Wang et al study. They conducted their analysis for three phenotypes: height, Type-II diabetes, and educational attainment. For the first two, they found a strong signal for a within-family genetic ancestral effect; for education they found nothing. The authors played it straight, describing the within-family ancestry effect on educational attainment as “close to zero” and concluding that the group differences are apparently environmental and cultural. It is important to remember, however, that this is not the last time a study like this is going to be done, and that the studies are not always going to be about relatively low-impact questions like Type-II diabetes among indigenous Mexicans. The folks over on race-twitter, from Charles Murray to Cremieux, are already waxing rapturous about the report, which seems odd given its uniformly environmental conclusions about EA. But race-Twitter remains excited because they know there is more where that came from, and they are willing to wait.

That makes it important to emphasize something: within-family ancestral effects are still badly confounded, just like those identical twin mothers in our children of divorce study. An ongoing confusion in modern genomics is to draw a distinction between direct and indirect genetic effects operating within and between families, and then to label the direct effects as unambiguously causal. Here is Wang et al:

We used the natural experiment of randomization of genetic material during meiosis to estimate the causal effect of genetic ancestry differences on complex traits in an admixed sample from Mexico City. In principle this design can be applied whenever there is phenotypic and genome-wide genetic data from admixed families, enabling the disentanglement of causal effects of ancestry differences from correlations between genetic ancestry and phenotype due to other factors.

This is way too strong in my opinion. As an old twin researcher, I am always suspicious of claims of “natural experiments” that “disentangle” genetic effects from confounds. There are no natural experiments, only natural quasi-experiments. Within-family genetic associations are confounded in exactly the same way those twin correlations were confounded in our twin divorce studies. Studying associations within pairs controls for a good set of between-family confounds, like population stratification, socioeconomic status and parenting. That is a good thing, and sound scientific practice. But it doesn’t come close to controlling for all confounds. Back in our quasi-causal twin study days, we used to talke about “E confounds,” referring to non-causal confounds that work within pairs of identical twins. The uncontrolled husbands in our divorced mother design are an example.

In fact, the most famous confound of genetic effects, Christopher Jencks’ “red hair” effect, would stroll right through this “natural experiment”. For the uninitiated, the red hair effect refers to a hypothetical world in which redheads are discriminated against, and wind up with poor educational attainment. In that world, genes for red hair look like causal EA genes. Well, genes for hair color segregate within families, so in red-hair world ginger kids would fare worse than their siblings, and the genes would look direct and causal. And if there was an ancestral group with high incidence of red hair, we would be concluding that their deficits in EA are due to genetic differences.

The real lesson from human twin studies of behavior is that we don’t know squat about the mechanisms through which genetic effects operate. That’s fine, and it’s mostly inevitable in studies of human behavior, but as I said about quasi-causal twin studies, it is a reason to maintain a deep humility about our certainty regarding the causal meaning of our results. So when the day gets here that a little bit of the variance in some behavioral trait between group X and group Y is explained by a within-family ancestry model, p < .05, with the racists dancing in the end zone, I am going to remain skeptical. In fact, I would say I remain skeptical about the T2D findings in Wang et al. Not skeptical in the “I don’t believe it” sense, but skeptical in the “Wait and see” sense. Type-II diabetes has a lot of possibilities for within-family confounds.

What would it take to convince me? It would take a convincing causal account that rules out confounds in their various forms, especially gene by environment correlation and interaction. If we really know how a system works: here is a variant that regulates proteins that contribute to neural growth in mice, supported by human neuroscience, the effect observed around the world, in rich people and poor people, producing large and consistent effects. It’s hard to imagine because it never happens, not for the complex human phenotypes that behavioral scientists care about.

Mechanism is Unconfounded Causation

What is that kind of real causal knowledge called? It is called mechanism. Mechanism is a difficult concept about which there is a lot of good philosophy, but here is a quick and dirty definition: a mechanism is a causal process that is well enough understood that you don’t have to worry about uncontrolled confounds. Huntington’s disease has a mechanism from gene to brain to behavior. Because of that mechanism I don’t need to have serious worries about what the HD gene will do to people who live at high altitude or who get special education. Funny thing is, red hair is a mechanism. Let’s say the red hair hypothesis turned out to be true. As some people have insisted, that would mean that red hair genes were in in some distal sense “causal.” But it wouldn’t be the causal status of red hair genes that would undo them, it would be their mechanism, which would allow us to see that the association with EA is spurious from a genetic point of view.

Without mechanism, within-family ancestral effects will remain what I have always said this kind of behavioral genomics amounts to: social science. Studies will continue to churn out “results,” but they will not be decisive and they will not accumulate. Within the limits of humbly conducted social science, that’s fine, it’s the way things work and there is nothing we can do about it. But if we forget that these are quasi experiments, not natural ones, our exaggerated expectations for disentangling causation will continue to encourage the racists.

Two Epistemological Asymmetries

The prospect of demonstrating that there is or isn’t a genetic component to a group difference in behavior faces a couple of tricky asymmetries. The first is something we learned from the twin studies: the between-within family method works better as a means of denying hypotheses than it does for confirming them. This is just a corollary of the quasi-causal confounding problem. If your theory is X causes Y, but it turns out that X and Y are correlated between families but not within them, that puts a pretty big dent in your causal claim. On the other hand, if X and Y are correlated between and within families, there is still a lot of room for confounds to get in the way. That is why we didn’t make a big deal of the fact that kids of divorced MZ twin mothers drink more than kids of non-divorced co-twins. Interesting finding, but social science being social science, you wouldn’t want to take it to the bank. In the Wang et al case, they are, at least implicitly, looking for genetic effects. So the absence of an ancestral correlation within twin pairs for EA rings truer than the claim that genetic differences cause the differences in diabetes. I’m not saying they are wrong about diabetes, or that the environmental basis of the EA differences is completely settled, just that one would need a lot more evidence to make the strong positive causal claim.

The other asymmetry is something that comes up in the more toxic reaches of the race and IQ domain. Going back to Jensen and Rushton, and continuing now with Russell Warne, the race-hereditarians make the claim that the IQ gap is “>0% genetically caused.” This move is an attempt to put their opponents in the impossible position of defending the null hypothesis. Can you prove that vaccines have nothing to do with autism, or that the CIA had nothing to do with the Kennedy assassination? Of course not, and I can’t prove that the difference in EA between indigenous and colonial Mexicans is 0% genetic. It offers the race-hereditarians an enormous advantage. As I put it in Chapter 9 of my book, which is mostly about this problem:

That position justifies any lukewarm shred of circumstantial evidence that seems to tilt the scales just a little bit in their direction. Thus are conspiracy theories born. Anti-hereditarians, like everyone trying to defend against a conspiracy theory, are more often in a whack-a-mole crouch, fending off the inconclusive arguments of hereditarians, only occasionally mounting counterattacks with equally inconclusive empirical weapons.

I have a feeling that shred of lukewarm evidence is going to be coming our way soon.