Feb 13·edited Feb 13Liked by Emil O. W. Kirkegaard
I have three (somewhat) related questions about heritability:
1. Why do we use h^2(variance explained) and not "h"(correlation) with the breeder's equation? In other cases of regression towards the mean correlation is used right?
Is it because we have: parent phenotype (*h)-> parent genes (*1)-> child genes (*h)-> child phenotype. So regression towards the mean twice, h*1*h=h^2 in total?
2. Which mean is it that one actually regresses towards?
I've seen grandparent mean(would be unambiguous), family mean(what does this mean exactly?) and population mean(surely depends on how specific you are; White Americans or White West Virginians for example) mentioned.
3. I have seen a lower(0.5-0.6 vs the usual 0.8) heritability used with the breeder's equation with the explanation being that only additive/narrow-sense heritability matters for it.
Some family designs estimate broad heritability, sometimes written H², some estimate additive heritability, sometimes written h². Often they are mixed somewhat together. The classical twin design cannot distinguish them because the equations are unsolvable/underdetermined if you want to solve for additive and non-additive genetics at the same time. In theory, some of the larger family designs can estimate dominance and epistasis, but the precision is very low. Identical twins apart will yield H² directly. EA4 GWAS failed to find any evidence of dominance effects, which is IMO surprising, as it should have found something with a sample size of over 3 million people. Based on this, it seems dominance effects are very weak for EA, and thus also for intelligence. Based on this, one might be tempted to take estimates of heritability as reflecting basically only additive heritability. This would be good for us because it means training genetic models and using genomic selection (embryo selection) will be more efficient, and regression towards the mean weaker.
Thanks for this Emil. I've been trying to meander my way towards an understanding of individual heritability, but I have never seen it explained so clearly. Will definitely check out the book that you recommended.
Emil, thanks for the somewhat deep dive into individual heritability. This subject is fought with misunderstanding, and you have helped clear it up.
Nice post.
Ditto.
I have three (somewhat) related questions about heritability:
1. Why do we use h^2(variance explained) and not "h"(correlation) with the breeder's equation? In other cases of regression towards the mean correlation is used right?
Is it because we have: parent phenotype (*h)-> parent genes (*1)-> child genes (*h)-> child phenotype. So regression towards the mean twice, h*1*h=h^2 in total?
2. Which mean is it that one actually regresses towards?
I've seen grandparent mean(would be unambiguous), family mean(what does this mean exactly?) and population mean(surely depends on how specific you are; White Americans or White West Virginians for example) mentioned.
3. I have seen a lower(0.5-0.6 vs the usual 0.8) heritability used with the breeder's equation with the explanation being that only additive/narrow-sense heritability matters for it.
Make sense but then apparently additive heritability('A') is what is estimated in twin studies and that there's no non-additive genetic effects for intelligence(for example table 1 here: https://www.stevestewartwilliams.com/p/12-things-everyone-should-know-about).
So is additive heritability ~0.8 or lower at like 0.6?
Some family designs estimate broad heritability, sometimes written H², some estimate additive heritability, sometimes written h². Often they are mixed somewhat together. The classical twin design cannot distinguish them because the equations are unsolvable/underdetermined if you want to solve for additive and non-additive genetics at the same time. In theory, some of the larger family designs can estimate dominance and epistasis, but the precision is very low. Identical twins apart will yield H² directly. EA4 GWAS failed to find any evidence of dominance effects, which is IMO surprising, as it should have found something with a sample size of over 3 million people. Based on this, it seems dominance effects are very weak for EA, and thus also for intelligence. Based on this, one might be tempted to take estimates of heritability as reflecting basically only additive heritability. This would be good for us because it means training genetic models and using genomic selection (embryo selection) will be more efficient, and regression towards the mean weaker.
Basically natural aristocracy
Thanks for this Emil. I've been trying to meander my way towards an understanding of individual heritability, but I have never seen it explained so clearly. Will definitely check out the book that you recommended.