Nice job, Eric. It’s great that you can take this on. I just wish there was discussion other than pedantic statistical arguments, related to the fact that individual differences in behavioral traits are probably not significantly influenced by genetic variation. There are direct scientific, philosophical and psychological implications beyond just dealing with this statistical obfuscation that gets lost.
Steve, I disagree that these are "pedantic statistical arguments." You always want me to agree with you that genes are completely irrelevant to behavioral differences, and I just don't believe that. Frankly I find it hard to imagine that anyone does. Genetic and behavioral differences are related in subtle, complex, philosophically difficult ways. If that is pedantry I'll wear it. The interesting part for me is trying to work through the subtlety. Absolute answers may be rhetorically satisfying, but they are intellectually uninteresting, never mind wrong.
I’m referring more to the original argument you rebutted, which is both pedantic and desperate. But if you have a gene that contributes “significantly” (as I said above), to “educational attainment” in some meaningful and defined way, I’m all ears. I resent being pigeon-holed as an absolutist or romanticist, as the case may be. The field was founded on the idea that it was 60 to 70 percent heritability, when it now dropped to one percent (just saw someone pushing a .004 to .01 r2). So what exactly are you holding on to? It’s reasonably close to null. Thus it is not significant, and not just because embryo selection won’t work, but because there is nothing significant to find. If I would have said 30 years ago, “You’ll find no more than 2 percent heritability for behavioral traits,”I still would have been labeled an extremist. It’s an ad hominem, effectively, and the people still trying to hold up the field are making the same kind of correlative arguments that astrology proponents use in their studies. If I want you to admit anything, it’s that the traits are not “significantly” heritable, which you come close to saying.
Is there a more elementary resource for me to learn about this topic? I’m an economist, and as a (very) casual reader of behavior genetics discourse, the idea that fraction of variance explained is the main analytical tool for discerning the “cause” of genes on behavior has always been unintuitive to me. We tend to care much more about the statistical significance of parameter estimates over in the other social sciences.
Unless I'm missing something, this feels like an uncharacteristically bad post. I had expected better from you.
Let's take a different example: exam scores. If you are taking an exam, then your final score will depend partly on how good you are at the subject the exam is testing for, and partly on how much effort you put in at the exam.
The fraction of variance explained by effort depends partly on the effect size b of effort, but also on the residual variance v in ability, and on the variance e in effort. We could expect the R^2 to be something like b^2 e/(b^2 e + v). Clearly if v is big then the R^2 will be low.
In one of your simulations in the post, you seem to have set the parameters such that it's not uncommon to receive 35 years of education.
Unsurprisingly, I think you are missing something. The unstandardized b is an "effect size" in some sense, Grade points per hour studying or whatever, but you have to be careful. First, it is in units of Y per X, so it is not something that can be compared across situations. A Beta is in SDs of Y per SDs of X, which does introduce other problems but keeps things in a comparable scale. But the bigger problem, as I outlined in the post, is precision. The b is an ESTIMATE of the population parameter, and when R^2 it is a poor one; when R^2 is very low it is a meaningless one. Could I ask you to have a second look at the thread (ported from Twitter) I like to at the end of the post? That shows how R^2 is related to the CI on the prediction of EA that we make based on someone's PGS. If the PGS has a low R^2, the CI is very wide. Think about it: What happens when R^2=0? Then the CI for a prediction is the same as the CI around the mean of Y, because the predictor isn't doing anything to make it smaller. It matters of PGS are crummy predictors.
Oh, and it is hard to simulate data with enough error variance in EA, and also a reasonable range of EA. I'm working on that. But in the meantime, the unstandardized slope is still close to the original value of .75, even though the variables are almost completely unrelated.
I was looking into this the other day for a very different problem. Have you tried simulating the years of education as a mixture of several distributions? in the PGS vs. years of education plots, there is a lot of banding and a lot of the variance appears to be from the effect of matriculation into a new level of education as distinct from advancing through the years in that level. The raw and mean square standardized variance in years of education changes a good deal through the 20th century as educational policy changes to the required duration of education and incentivizing advanced credentialing.
I'm going to look at the residuals of some of the EA PGS regressions to see if they are chunky. I suspect they are and that doesn't bode well for the interpretation of the regression.
This is true but might reach more people if you defined PGS (polygenic scores) and EA3 (episodic ataxia type 3) at the outset. That way you reach not only the genetics crowd but also the stats and general research crowd. Your point about an imprecisely estimated slope is well-made.
Nice job, Eric. It’s great that you can take this on. I just wish there was discussion other than pedantic statistical arguments, related to the fact that individual differences in behavioral traits are probably not significantly influenced by genetic variation. There are direct scientific, philosophical and psychological implications beyond just dealing with this statistical obfuscation that gets lost.
Steve, I disagree that these are "pedantic statistical arguments." You always want me to agree with you that genes are completely irrelevant to behavioral differences, and I just don't believe that. Frankly I find it hard to imagine that anyone does. Genetic and behavioral differences are related in subtle, complex, philosophically difficult ways. If that is pedantry I'll wear it. The interesting part for me is trying to work through the subtlety. Absolute answers may be rhetorically satisfying, but they are intellectually uninteresting, never mind wrong.
I’m referring more to the original argument you rebutted, which is both pedantic and desperate. But if you have a gene that contributes “significantly” (as I said above), to “educational attainment” in some meaningful and defined way, I’m all ears. I resent being pigeon-holed as an absolutist or romanticist, as the case may be. The field was founded on the idea that it was 60 to 70 percent heritability, when it now dropped to one percent (just saw someone pushing a .004 to .01 r2). So what exactly are you holding on to? It’s reasonably close to null. Thus it is not significant, and not just because embryo selection won’t work, but because there is nothing significant to find. If I would have said 30 years ago, “You’ll find no more than 2 percent heritability for behavioral traits,”I still would have been labeled an extremist. It’s an ad hominem, effectively, and the people still trying to hold up the field are making the same kind of correlative arguments that astrology proponents use in their studies. If I want you to admit anything, it’s that the traits are not “significantly” heritable, which you come close to saying.
Is there a more elementary resource for me to learn about this topic? I’m an economist, and as a (very) casual reader of behavior genetics discourse, the idea that fraction of variance explained is the main analytical tool for discerning the “cause” of genes on behavior has always been unintuitive to me. We tend to care much more about the statistical significance of parameter estimates over in the other social sciences.
https://www.amazon.com/Understanding-Nature-Nurture-Debate-Life/dp/1108958168Amazon.com: Understanding the Nature‒Nurture Debate (Understanding Life): 9781108958165: Turkheimer, Eric: Books
Unless I'm missing something, this feels like an uncharacteristically bad post. I had expected better from you.
Let's take a different example: exam scores. If you are taking an exam, then your final score will depend partly on how good you are at the subject the exam is testing for, and partly on how much effort you put in at the exam.
The fraction of variance explained by effort depends partly on the effect size b of effort, but also on the residual variance v in ability, and on the variance e in effort. We could expect the R^2 to be something like b^2 e/(b^2 e + v). Clearly if v is big then the R^2 will be low.
In one of your simulations in the post, you seem to have set the parameters such that it's not uncommon to receive 35 years of education.
Unsurprisingly, I think you are missing something. The unstandardized b is an "effect size" in some sense, Grade points per hour studying or whatever, but you have to be careful. First, it is in units of Y per X, so it is not something that can be compared across situations. A Beta is in SDs of Y per SDs of X, which does introduce other problems but keeps things in a comparable scale. But the bigger problem, as I outlined in the post, is precision. The b is an ESTIMATE of the population parameter, and when R^2 it is a poor one; when R^2 is very low it is a meaningless one. Could I ask you to have a second look at the thread (ported from Twitter) I like to at the end of the post? That shows how R^2 is related to the CI on the prediction of EA that we make based on someone's PGS. If the PGS has a low R^2, the CI is very wide. Think about it: What happens when R^2=0? Then the CI for a prediction is the same as the CI around the mean of Y, because the predictor isn't doing anything to make it smaller. It matters of PGS are crummy predictors.
Oh, and it is hard to simulate data with enough error variance in EA, and also a reasonable range of EA. I'm working on that. But in the meantime, the unstandardized slope is still close to the original value of .75, even though the variables are almost completely unrelated.
I was looking into this the other day for a very different problem. Have you tried simulating the years of education as a mixture of several distributions? in the PGS vs. years of education plots, there is a lot of banding and a lot of the variance appears to be from the effect of matriculation into a new level of education as distinct from advancing through the years in that level. The raw and mean square standardized variance in years of education changes a good deal through the 20th century as educational policy changes to the required duration of education and incentivizing advanced credentialing.
I'm going to look at the residuals of some of the EA PGS regressions to see if they are chunky. I suspect they are and that doesn't bode well for the interpretation of the regression.
This is true but might reach more people if you defined PGS (polygenic scores) and EA3 (episodic ataxia type 3) at the outset. That way you reach not only the genetics crowd but also the stats and general research crowd. Your point about an imprecisely estimated slope is well-made.
Good point will fix tomorrow