Hanging a cat on the Sabbath
Why I don't love regressions using "genetic ancestry"
This was inspired by writing a review of this paper, and reuses some material from my review. I don’t think that paper is very good, but my problems are with the whole approach, including much “classier” looking work published in better venues. This article also expands and repeats some arguments I made here.
With modern genomics you can try to find out things about DNA. The obvious thing is “what does it do?” What role does this genetic variant play in biology, in building a human, and thereby in affecting his disease risk, his behaviour or how he fits into his social environment? But another thing you can ask DNA is “where does it come from?” The reason this works is that much genetic mutation is just random drift, and that people tend to have children with others near them. So different parts of the world drifted in different ways. As a result, we can take a person’s DNA and guess where their ancestors came from; and we can even look within a single person and guess where different parts of their DNA came from: this chunk of chromosome 11, say, came from someone’s ancestors in Africa!
There is now a small industry of papers which take people’s genetic ancestry and use it to predict their outcomes, in what are called “admixture regressions”. I believe we should study genetic differences between ethnic groups, so you might expect me to be pleased by this. Let me try to explain my doubts.
In Scotland, a cat was once supposedly tried for catching a mouse on the Sabbath day:
The people all were horrified
And they were grieved sair;
And straightway led that wicked cat
Before the Minister.
The minister was horrified,
And unto it did say,
Oh, thou preversed pussy cat,
To break the Sabbath day.
The Sabbath’s been frae days of yore
An institution;
So they straightway led that wicked cat
To execution.
The right objection to this proceeding is not that the particular cat was innocent, but that cats are the wrong kind of thing to stand trial. This is also how I feel about admixture regressions. Irrespective of whether genetic ancestry is found innocent or guilty for causing intergroup differences, it is not the sort of thing we ought to be looking at. Admixture regressions answer the wrong question.
Put simply: ethnic groups are not ancestries. They correlate with ancestry, sure: if you’re black in the US, many of your forefathers probably came from Africa; if you’re white, many probably came from Europe. But not all of them. That’s what lets us do admixture regressions: we can separate out the effects of being in an ethnic group from the effects of having more or less genetic ancestry from that ethnic group! But it also means that if you use admixture regressions to learn about interethnic differences, then you are at risk of estimating the wrong thing.
Here’s a simple toy model.
Ethnic groups A and B start with a number of alleles which affect, say, their labour market earnings. Each allele can be identified by the genetic ancestry of its surrounding DNA; in both groups, each is equally likely to be earnings-increasing or not.
A person’s earnings are the sum of his earnings-increasing alleles, plus a bonus for someone in group A (due to discrimination in the labour market), plus a random shock. People marry by earnings: the highest earners pair off with each other. (Or there’s no discrimination in the labour market but there’s discrimination in the marriage market, so that A is seen as preferable: this is mathematically just the same.)
Ethnicity follows a one-drop rule: you only count as group A if both your parents are from group A.
After a few generations, researchers want to know whether genetics contributes to the difference in earnings between the groups, so they run some admixture regressions.
One research team realises that they can’t just compare earnings between groups —labour market discrimination may play a role. So instead they compare ancestry within group B. Some of group B have more group-A ancestry. They now find these people earn more!
The reason is that a group B person who marries a group A person has to have many earnings-increasing alleles on average, to compensate for the group A person’s advantage in the labour market. So children of these mixed marriages inherit more earnings-increasing alleles.
Is the regression right? Yes, after a few generations there will be genetic differences between the groups, even though there weren’t to start with. But the regression is underestimating the strength of the difference between groups! That is because group A people who marry group A people also have more earnings-increasing alleles than group A people who make mixed marriages. Here’s a graph of the simulation: blue dots are group B individuals, red dots are group A, and the x axis is the proportion of group A ancestry.
Another research team uses an even more powerful technique. They directly identify alleles and count them within group B’s individuals: how many alleles originally from each group are earnings-increasing?
They get a “surprising” result: among group B, B-ancestry alleles are more likely to be earnings-increasing. The reason is the same process of assortative mating. In mixed marriages, the group B partner typically had more earnings-increasing alleles: that’s how they equalled the earnings of the group A partner, who benefited from discrimination.
What does this result actually tell us? It’s not telling us about the evolutionary history of each ethnic group. (Remember, both groups started with exactly equal numbers of earnings-increasing alleles.) It’s also not informative about the differences between the groups, in fact it’s actively misleading if you map “ancestry within a person” to differences between ethnic groups. It tells you that one group’s “own” ancestry alleles are not to blame for their lower earnings. But “blaming” alleles, or ancestries, was a stupid idea to begin with, and so is declaring them innocent. What matters is which genes people from a given ethnic group actually have, not where those genes come from.
There may be interesting theoretical questions that can be answered by admixture regressions — for example, about the evolutionary history of a particular world region. (Not in my toy example, but maybe sometimes.) But the important practical question about ethnic groups is what causes the differences between them, whether those differences are in psychopathology, heart disease, or educational attainment. Admixture regressions don’t tell us that.
They also seem unlikely to help us find mechanisms. Individual genetic variants have a specific function, and even polygenic scores, which sum many variants, can be “enriched” for genes which are expressed in a particular biological mechanism. But the vast majority of “ancestry” differences are going to be pure drift. Even if some differences are due to natural selection, finding them will be like searching for a needle in a haystack. If our goal in explaining any given difference between ethnic groups is to do something about it, or just to find it how it works, then admixture regressions are unlikely to help us.1
In my toy example, admixture regressions will underestimate the effect of genetic differences between ethnic groups. It’s equally easy to imagine examples where they overestimate them. Suppose that instead of a unified marriage market, there is marriage market segregation. However, some elite members of group A still have children out-of-wedlock with group B. Then the A-ancestry alleles of group B will reflect those elites; the alleles may be more likely to be earnings-increasing than those of group A in general. The underlying problem is that an ethnic group’s genes, from any given ancestry, are not a random sample. (That is even true of their own ancestry: neither Europeans nor Africans who came to America were a random sample of their respective populations.)
So far, I’ve been assuming the best possible scenario for these regressions, which is that there are no ordinary confounds and they can accurately estimate true effects (of whatever they are estimating). But of course real genetic data has plenty of confounds, and this is especially true for admixture data. By definition, your genetic ancestry tells you what proportion of ancestors you had from a particular area. But then it must correlate with anything that is passed down from those ancestors, genetically or not. We worry about “chopsticks gene” effects for polygenic scores, but at least for polygenic scores, we try to estimate the effect of genetic variants on some target phenotype. Genetic ancestry is more like a deliberate search for chopstick genes! We’re actively trying to find genes that correlate with geography, and hence with everything else that goes with it. The authors of the paper I reviewed are commendably cautious about interpreting their results:
We situate the findings within a non-specific vertical transmission framework that accommodates genetic, cultural, and other intergenerational mechanisms without assuming their relative contributions.
That’s nice, but vertical transmission can cover a wide variety of mechanisms:
Ethnic groups have different genetic ancestry, and their different genes cause them to have different outcomes.
Ethnic groups have different cultures, which are transmitted parent-child. Cultural ancestry correlates with genetic ancestry, and those cultures cause them to have different outcomes.
Some ethnic groups were discriminated against in the past, had lower income, and accumulated less capital. Capital is inherited parent-child, and causes different outcomes. Genetic ancestry tracks the number of low-capital ancestors.
Some ethnic groups were discriminated against in the past, and lived in poorer areas with worse schools. People tend to live near where their parents lived, and those areas still have worse schools, causing differences in EA. Genetic ancestry tracks the number of ancestors living in poor areas.
Admixture regressions are, in the death-knell phrase loved by article referees, “hard to interpret”.
Examples
I don’t want to remain at generalities, so here are some specific papers that use this approach. They were chosen for convenience, not because I think they were especially bad or especially important.
Wang et al. “Direct effect of genetic ancestry on complex traits in a Mexican population”
This paper examines a genetically mixed sample in Mexico City. It starts boldly with:
Human populations differ in phenotype means for complex traits…, but we know very little about the relative contribution of genetic and environmental factors, and their interactions, to these differences.
That sounds as if they are going to use admixture regressions to find out, and indeed they run within-family regressions to capture the “direct effects of ancestry”. (That is, they compare siblings where one sibling randomly got more DNA from one particular ancestry.) This avoids the problem of confounds from transmitted culture, discrimination, etc. The paper is genuinely estimating a causal effect of genetic ancestry (which presumably averages many different small effects from many parts of the genome). But what do we learn about what causes human populations to differ? The authors are too smart not to know about the problems I list above, and the conclusion rows back:
It is not known if our findings translate to phenotypic differences between non-admixed IAM [Indigenous American] and EUR [European] populations. Our inference is strictly on MCPS participants from Mexico City, and our estimates are weighted towards offspring of parents with more ancestry differences between their paternal and maternal genomes….
So then, what have we learned?
Parcha et al. “Genetic European Ancestry and Incident Diabetes in Black Individuals: Insights From the SPRINT Trial”
To show I’m not just picking on behaviour genetics, here’s a paper looking at ancestry and diabetes. They look at black people in the US and divide them up by proportions of European ancestry. Those with more European ancestry had less diabetes risk. They conclude “The higher risk of incident diabetes in Black individuals may have genetic determinants in addition to adverse social factors.” Well, maybe! It seems less likely that mixed-ethnicity parents were non-randomly selected in terms of diabetes risk, compared to in terms of education or income, say. But it’s still not that unlikely, given that diabetes is associated with socio-economic status. All we can really get is a “maybe genetics matters to this difference”. How much does it matter? We can’t tell, because a study solely within black people isn’t going to tell us about black-white differences. Plus, unlike Wang et al., this study has the usual problems that genetic ancestry is confounded with other things — since diet preferences, access to food options, etc. may all be passed down within families without being genetic.
There is another way
Look, we know how to estimate the causal effects of any different genetic measure. You run a within-family regression. So if you want to estimate how much of the difference between ethnic groups on any variable is due to a given genetic measure:
Run a within-family regression within ethnic group A. Estimate the causal effect, leveraging the lottery of meiosis. Multiply it by the differences in the measure between group A and group B. Express the result as a fraction of the overall between-group difference in the variable.
Do the same thing in ethnic group B.
These two techniques answer the two questions “what if group A had the same genes as group B on this measure?” and “what if group B had the same genes as group A?” These aren’t guaranteed to give the same answer (because the two groups’ environments, or other genes, may differ). And they aren’t perfect approaches to the question. There are ways that within-family regressions can potentially be misleading; and for now, no single measure is likely to capture every genetic difference that matters. But they are simpler and more direct than faffing around with admixture. Why not do that?
I want to be careful here: I’m not making the broader claim that understanding ancestry differences can never help us learn about genetics. For example, it’s important to learn about why polygenic scores are not “portable” across different populations, and here’s a paper using local genetic ancestry to do that.
A little above my head but we'll worth reading. Thank you.