[DANK_YACHT]

> The point is not whether including polygenic scores is better than nothing. The point is whether it's good enough to justify the claims they are making. It's not.

What is the justification for this assertion? If polygenic scores are simply "noisy," then, as the GP mentioned, they may be good enough when used in aggregate. There can be a lot of signal in noisy data. Ask any ML practitioner.

[dash2]

Polygenic scores are themselves an aggregation. And they contain a ton of noise. From twin studies, R2 of the "true" polygenic score on education is about 40%. The R2 of our best actual polygenic score is 4%. IIRC these guys aren't even using the education polygenic score. They're using the PGS for IQ. That's even noisier (because it was created using smaller samples, and you need biiig samples for this to work).

[DANK_YACHT]

> Polygenic scores are themselves an aggregation. And they contain a ton of noise.

Again, noisy data can still be useful. For instance, generate a perfect single-variable normal distribution, sample along the x-axis and perturb each point randomly in the y direction either up or down. Depending on the range of random values used to move the sampled points, you can still see the underlying distribution even though the data is noisy.

Two possible arguments you might make:

1) The data is noisy and the paper's authors haven't collected enough data to account for the amount of noise. Usually people will do something like null-hypothesis significance testing to measure this.

2) The noise isn't uniformly random and has some underlying bias that is affecting the results.

[dash2]

Noisy data can be useful. What it can't do is be useful enough to say "we have controlled for X". There will be a large unmeasured component of genetic variation - probably about 90% of it - that controlling for the PGS doesn't capture. Given that, throwing a PGS into the regression is not worth the candle.