Warning: technical. ChatGPT is your friend. Or refer to Scott Alexander’s big summary for an introduction to some of the issues.
Sasha Gusev commented on my last post:
I think you cannot really talk about the magnitude/value of an effect size without first explaining what you intend to use it for. There are rare variants explaining a tiny amount of population variance but providing very important information for drug targets (which can then explain a large amount of population variance when intervened on). On the other hand, a smoking polygenic score might explain a decent amount of variance in lung cancer risk but will be useless in a model that already includes smoking itself. What is the intended use of the PGS? A lot of the existing applications in behavior genetics want to have a "clean" causal instrument, and the population PGS is clearly not that (and, I would argue, confounded in ways can be extremely misleading).
I agree with this, and specifically with the comment about the population PGS. Indeed, I recall telling people a few years back that family-based designs were gonna be essential in getting cleaner measures. That came from my background in economics, with its obsession with causal identification. Others more influential than me obviously thought the same way, and now we are getting a lot of family-based surveys, and work on using sibling data to get “purer” summary measures of the causal effects of genetic variants.
I do also think the behavioural genetics field could afford to be more self-conscious about the question of what our PGS are for. My impression, as an outsider, is that behaviour genetics is a kind of outgrowth of medical genetics, and they tend to carry over habits of thought uncritically from that, roughly like this:
You run a GWAS for breast cancer! Great, you can use it to target drugs on particular biological pathways starting from important causal variants. Also you can create a PGS and use it to predict individuals’ risk.
You run a GWAS for oppositional defiant disorder! Great, you can use it to target drugs for that, and make a PGS to predict children’s risk. Hmm, is this the right way to think about children being naughty — as a medical condition? Do we want to predict their “risk” in that way?
You have a GWAS for getting divorced/playing sports/being gay! Great, you can… wait, what?
In other words, behavioural genetics tends to start from quite individualistic assumptions. Genes → individual variation, and maybe in future we can (handwave handwave) use genetic discovery to “do something” about that variation. I’m often sceptical. So then, what is a PGS, specifically a behavioural PGS, for?
I think the starting point must always be a statement of the form: genetic variation has big effects on important outcomes.
Every word matters. The effects have to be big, or there’s nothing there worth thinking about. But big is measured on a substantive, issue-specific scale, not by R-squared, or standardized coefficients, or any other technical measure. A five or fifteen percentage point increase in a person’s chance of going to university is big. A much smaller increase in someone’s chance of going to jail would also be big.
The effects also have to be effects — truly causal. Or rather, if they aren’t effects but only correlations, that is interesting too and worth investigating, but you have to be clear which it is. Are genes captured by the PGS affecting the outcome, or just correlated with it? If both, then can you pinpoint just the causal effects using a sibling-based analysis?
Lastly, the outcomes have to be important, just as with any social science. Not that every piece of research has to be earth-shattering, but the point is the researcher should have a clear understanding of why the outcome matters. Educational attainment matters to individuals (it is an absolutely central life outcome) and also to whole societies (it is a basis for innovation which is key to economic growth). But EA itself, measured as years of education, is really a proxy for something deeper — actual learning and knowledge. You could automatically increase EA by mandating that everybody must remain in education until the age of 21. But that is not guaranteed to increase people’s knowledge! So you need to think what the outcome is that you actually care about.
If the statement above is true, for any topic, and you have a PGS which captures enough of the important genetic variation, then you have something worth investigating, simply because genetics are an important part of the story. That doesn’t mean that any particular design will work, but it does mean, by definition, that the genetics are worth understanding, and we won’t fully understand the topic until we understand the genetics.
I argued in my last post why the statement is true for educational attainment, for instance.
Specifically, there are multiple ways to use a PGS with substantive causal effects:
Obviously, you can use it in within-sibling regressions as an independent variable. Maybe the dependent variable is the phenotype the PGS was created to predict; maybe not. A PGS for one phenotype will also affect other, related phenotypes. Ideally you might want to do GWAS-style discovery with a whole bunch of dependent variables simultaneously, but using an off-the-shelf PGS is “cheap” in terms of data and lets you work in smaller samples.
Relatedly, you can use a PGS as an instrument. Does being fat make you happy? To avoid reverse causality, use a polygenic score for obesity as your instrument. Is this gonna work? Both Sasha and Scott Alexander expressed some scepticism, and I agree, irrespective of whether the instrument is strong or unconfounded by environmental variation. When phenotypes are polygenic, then genes are poly-phenotypic, i.e. they affect more than one thing, and that makes the exclusion restriction untenable, even in a within-sibling design. Maybe the statistical geneticists can work some magic to avoid this. Maybe? I would look very carefully at the assumptions built into any such model, though.
If you have a clean source of exogenous environmental variation, and a within-siblings design, you can look at gene-environment interactions (and do them right for once). That can also be also an indirect way to find out more about how a polygenic score “works”.
You can use it in a regression as a control, when your focal independent variable is something environmental. There is a dumb way to do this, which is to say “now we’ve controlled for genetics!” even though your PGS is a very very noisy measure. Aaaagh. Undoubtedly numerous social science studies will have been this stupid. A smarter way is to consider whether including the PGS changes the effect of the focal variable, and perhaps run an Oster 2019 style analysis of coefficient stability. If you feel you can confidently estimate the proportion of the “true genetic variance” captured by your PGS, that may help.
Flipping things around, you can use it as a dependent variable. This is what I like to do. For example, people’s birth order affects the polygenic score of their spouse, in both UK and Norway (not sure people have noticed, but we now have Norway data in there and it strongly confirms our UK story). Or, people with high polygenic scores for educational attainment (and many other things) left UK coalmining areas in the 20th century. Et cetera. If you have a score that affects something important, then it is interesting to understand how that genetic variation gets distributed and circulated, geographically, socially, whatever. This is also a way to move beyond the individualist bias of behaviour genetics and start thinking about large-scale social processes; or, you could think of it as doing population genetics with unusually recent time periods and fine-grained data. I think there is still lots to do here: in particular it would be interesting both to see how the truly causal genes circulate in society, and how the non-causal genes circulate (and how they got their non-causal association with the phenotype). So it will be interesting to look at polygenic scores for both “direct” and “indirect” effects.1
Those are some possible uses of a PGS. All of them more or less require that initial statement: the PGS must capture substantively big effects on something important. There are bad ways to do all of them, but good ways too (probably even for Mendelian instrumental variables, sometimes).
You’ll notice what isn’t in the list: embryo selection, and drug discovery. Personally those things don’t interest me, and I often find them quite creepy when applied to behavioural phenotypes. That doesn’t mean they can’t be done, and a “big” effect size is also an effect that parents may care about — again, irrespective of heritability statistics.
This is a lousy nomenclature, since direct effects aren’t necessarily direct and indirect effects aren’t effects. But it seems to have stuck :-(