> This side of the veil, instead of looking for the “gene for intelligence”, we try to find “polygenic scores”. Given a person’s entire genome, what function best predicts their intelligence? The most recent such effort [citation in the original article] uses over a thousand genes and is able to predict 10% of variability in educational attainment. This isn’t much, but it’s a heck of a lot better than anyone was able to do under the old “dozen genes” model, and it’s getting better every year in the way healthy paradigms are supposed to.

I think that's a much more measured claim than you're giving credit for; "it works" is claiming to explain 10% of variance. Do you have specific complaints about the cited article? Or are you just arguing by analogy that since people were wrong in the past, they are probably wrong now?

What do you mean by "it works"? Polygenic scores?

I'm doubtful that polygenic scores work, in the sense that we all acknowledge that complex traits like intelligence are interactions between genetics and environment.

You can always decompose a function of two variables into three parts (with some hand-waving for notation):

f(a, b) = f(a) + f(b) + interaction term

where E(interaction term) = 0 in some statistical sense.

If the interaction term is zero, then great, your function is trivially decomposable. For our problem, take a = genetics and b = environment, that means you can precisely talk about someone's 'genetic score for intelligence' and someone's 'environmental score for intelligence', and never have to consider them together.

But I strongly suspect that the interaction between genes and environment is very very high; the latest effort to map genes to intelligence only accounts for 10% of variance not because the environment determines the other 90%, but probably the interaction term, which can be complicated and highly nonlinear.

Another thing: variance of your inputs can only be considered in a statistical sense so the relative importance of genes vs. environment won't be stable. If somehow the world became completely uniform (every single child in the world received the exact same education and upbringing), you'd expect genetic variation to account for everything just by definition.

You can't do that with things that interact, e.g.:

f(a,b) = b/a

f(a,b) = b/a = 0 + 0 + b/a = f(a) + f(b) + interaction term

In this case the system is contained entirely within the interaction term, and since you don't know the distributions of a and b there's not much motivation to go further. If you had the distributions of a and b you might be able to do something a little less trivial by skewing the interaction term to have an expected value of zero, potentially like:

a/b = a + b + (a/b - a - b) = f(a) + f(b) + (interaction term with E[term] = 0)

If by "interaction term" it is meant f(a,b), then I congratulate the author of that tautology.

Hidden in the parent's comment is the additional condition that f(a,b) have an expected value of zero. It's still pretty easy to prove, though.