[aab0]

> Still doesn't make much sense to use a secondary trait like educational attainment (58% heritability at highest) over a primary one like IQ (>80% heritability).

Educational attainment is phenotyped much more often than is IQ, so it's both available in this dataset and available in the much larger datasets of the GWAS in question. Aside from being a phenotype of interest in its own right and a case of looking under the lamppost, EA also proxies for intelligence on both the genetic and phenotype level. (In some cohorts, like the UK Biobank, the cognitive performance test used has such low test-retest reliability that anyone who is looking for intelligence variants is going to want to use the education variable anyway!)

> Maybe I missed something, but how well does BMI measured at age 45-50 correlate with BMI at reproductive age?

I'm sure the correlation is <1. But as long as it's not an inverse correlation, it's fine. That is just measurement error which will bias down to 0 the estimate and lead to underestimate of selection.

> Wouldn't a far simpler approach be to measure allelic frequency at regions near GWAS QTLs in multiple generations?

How do you aggregate those QTLs? You can't examine them individually, that would be horrifically underpowered. If you wind up summing them, don't you get a polygenic score back and do what OP did?

[cSoze]

Considering that natural selection is generally considered to not operate on post-reproductive phenotype (prime case is antagonistic pleiotropy literature) I consider any conclusion based on these measures pretty suspect. Worse, specific heritability for these traits in this cohort, as far as I can tell, wasn't even estimated. Given that the supplemental says the mean in the literature is 40%, it could very well be sub 20%.

I suspect that selecting only known high effect QTLs would actually not impact your power all that much as it limits the number of tests performed. If you can't see selection at high effect alleles I'm not sure you'll be able to detect them in aggregate either, especially with very noisy phenotype data.

At the end of the day natural selection is a change in allele frequency driven by differences in phenotype. Sure, this maybe was an easy analysis to do given that the data was available, but I don't really see the value.

[aab0]

> Considering that natural selection is generally considered to not operate on post-reproductive phenotype (prime case is antagonistic pleiotropy literature) I consider any conclusion based on these measures pretty suspect.

Again, why? Is there any reason to think that late life BMI will inversely correlate with BMI such that the old fat people were actually the skinny young people? If it's just measurement error, then it's no worse than most variables which get used in health or sociology research.

> If you can't see selection at high effect alleles I'm not sure you'll be able to detect them in aggregate either, especially with very noisy phenotype data.

All of these are highly polygenic traits. A 'high effect allele' means explaining 1% or less of variance. Selection on such a trait is going to shift the frequency by a tiny amount. Grossly underpowered individually. You have to consider them jointly. Noisy phenotype makes that more, not less, true.

[cSoze]

The antagonistic pleiotropy theory actually advocates exactly that inverse correlation, more investment into young years for increased reproductive fitness at the expense of worse later years.

Further, I suspect that number of kids or even just a binary having kids or not would have a direct effect on BMI in old age. More kids > less time to excercise > compounded over many years. There's also a known correlation between educational attainment and number of children in the literature which is generally hypothesized to be causal.

I have concerns about the validity exactly because of the expected tiny effects. Polygenic traits with extremely high heritability (height being the prime example) top out at 10% of variance explained even when considering all alleles. That just highlights the need for better phenotypes and multi generational typing. I think it's Botstein who I've heard advocates for more time points over more replication when given the choice.

[aab0]

> The antagonistic pleiotropy theory actually advocates exactly that inverse correlation, more investment into young years for increased reproductive fitness at the expense of worse later years.

Which would be great if you had any evidence that BMI is largely influenced by such antagonistic pleiotropy. But since BMI is not a disease like Alzheimer's, it doesn't even have prima facie plausibility. This shouldn't be hard: does low BMI in youth predict high BMI at middle or old age? I've never seen any result ever hinting at this, and no one believes that.

> I have concerns about the validity exactly because of the expected tiny effects.

Yes, it does make it harder, but fortunately, that's what we have polygenic scores for. Do you have any power analysis suggesting that the sample size here is inadequate?

> Polygenic traits with extremely high heritability (height being the prime example) top out at 10% of variance explained even when considering all alleles.

Totally wrong, where did you get this idea? The height polygenic score is at >>10% of variance (see Wood or http://ije.oxfordjournals.org/content/45/2/417.full ) and BMI polygenic scores are already at 10% and will go up since all the twin and GCTA estimates indicate larger heritability than that.