3 ms·

"Heritability" includes non-genetic factors, and combined with genetic factors accounts for ~50% of variation in intelligence. For genetic factors alone, the c

by pcrh 7y ago

"Heritability" includes non-genetic factors, and combined with genetic factors accounts for ~50% of variation in intelligence. For genetic factors alone, the consensus so far is that there is very weak correlation of genetic variation with intelligence, unless a not quite-yet-established polygenic score is constructed. To wit: >From the 1990s until 2017 no replicable associations were found. GPS from these early GWAS, which we refer to as ‘IQ1’, predicted only 1% of the variance of intelligence in independent samples. It became clear that the problem was power: the largest effect sizes of associations between individual single nucleotide polymorphisms (SNPs) and intelligence were extremely small, accounting for less than 0.05% of the variance of intelligence. However, things got "better": >in 2016, a second meta-analytic GWAS analysis with a sample size of 294,000 identified 74 significant loci30. This analysis produced a GPS, EA2, that predicted 3% of the variance in years of education on average in independent samples30. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5985927/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5985927/ It stretches the imagination how one would go from these numbers to claiming that "The best estimate of intelligence heritability in adulthood at the peak, using measurement-error-corrected estimates, is 70-80%.

gwern 7y ago
> "Heritability" includes non-genetic factors, and combined with genetic factors accounts for ~50% of variation in intelligence. No, it doesn't. Heritability is intended and defined to include genetic factors only, unless you are referring to indirect genetic effects which are mediated through the environment but are still genetic in origin. > For genetic factors alone, the consensus so far is that there is very weak correlation of genetic variation with intelligence, unless a not quite-yet-established polygenic score is constructed. To wit: You are again being extremely misleading and selective in your quote-mining, and you omit all of the qualifiers that Plomin & Stumm include, eg "Recent genome-wide association studies have successfully identified inherited genome sequence differences" - 'identified genome sequence differences' does not imply that they are the only genes involved. Merely that some have been identified, and more will be. As they then go on to point out in the parts of the review you ignore, future GWASes will surely make much more progress in identifying SNPs, and approach the SNP heritability ceiling of ~30% given enough data. > It stretches the imagination how one would go from these numbers to claiming that "The best estimate of intelligence heritability in adulthood at the peak, using measurement-error-corrected estimates, is 70-80%. Er, because I'm not talking about SNP heritability estimates. I am talking about estimates from twin studies, adoption studies, and MZA studies, which estimate total genetic contributions. The lower-bound is 50%, you don't get anything lower from any twin studies, and this is deflated by the fact that many twin studies are looking at young kids and the Wilson Effect has not yet kicked in; this is then further deflated by the fact that they use noisy IQ measurements, which inflate the error variance component and hide the shared-environment/heritability components. Correct for those, and you get up to 70-80%. All of this is covered in the Plomin textbook cited there, starting pg171, where they cover things like how MZA designs deliver 72% heritability estimates and they explicitly mention that measurement error biases downwards ('Corrected for unreliability of measurement, heritability estimates would be higher...'). They are being highly conservative by using a lower bound of 50%, and they are wrong to, because as I already mentioned family-GCTA/GREML-KIN indicates that semi-rare variants alone get you up to 50% heritability, so we can be sure that 50% is a considerable underestimate (which is consistent with their expanded textbook discussion noting that the real heritability is much higher).
- nbvcxwthrow 7y ago
  You have a very peculiar (read: incorrect) definition of "heritability". http://bactra.org/weblog/520.html http://bactra.org/weblog/520.html
  - gwern 7y ago
    No, I do not. Linking to Shalizi, who systematically misrepresents the issue and ignores all the evidence against his claims ( https://humanvarieties.org/2013/04/03/is-psychometric-g-a-myth/ https://humanvarieties.org/2013/04/03/is-psychometric-g-a-my... ) is not an impressive thing to do, nor does it defend in any way pcrh's persistent errors and misquotations and misrepresentations of SNP heritability & PGSes (which rhetorical strategies and arguments bear a remarkable resemblance to Shalizi's own attempts to FUD intelligence & heritability, which have failed so spectacularly he hasn't dared address the issue since Rietveld et al 2013 came out...).
    - nbvcxwthrow 7y ago
      The link does nothing to refute Shalizi's point. It basically amounts to saying: -Other models of intelligence as wrong and g is real because tests results correlate with each other, which is completely orthogonal to Shalizi's point. Because of the very nature of testing the results are going to correlate with each other; the most trivial example is, using Taleb's example, if dead people take the test. Dead people are dead and have an IQ of zero, and this is enough to increase the correlation (https://twitter.com/nntaleb/status/1114977580053147648 https://twitter.com/nntaleb/status/1114977580053147648). -Apparently confirmatory factor analysis means g is relevant because...confirmatory factor analysis means g is relevant? Nothing addresses the fact that g could be an artifact that arises in any set of measures that correlate with each other regardless of the origin. Adding to that, there is scant evidence of any link between g and the underlying neurobiology and/or actual causal genetic mechanisms, beyond very trivial cases such as obvious malformations etc. On the contrary, neurological findings are pointing to the opposite direction (https://www.ncbi.nlm.nih.gov/pubmed/23259956 https://www.ncbi.nlm.nih.gov/pubmed/23259956). -Even if we have no idea how g works it's still useful for predictions for stuff like educational attainment so it's relevant. Well duh, if you're good at taking tests you're going to do well in an environment where your attainment depends on your doing well in tests. Again, there is enormous circularity in reasoning. And as for the other outcomes, look at it this way: you could combine many measurements from your body (say, various analyzes from blood samples, but even trivial things like number of legs or amount of blood in your body could do), find they all correlate with each other (because when your body is breaking down it is likely damage to one organ that's spreading damage to other organs etc.), and from all that stuff you could draw a general "health quotient", or HQ. It would be highly heritable, there would be literal millions of SNPs that are associated with it. It would have immense predictive value for a variety of social outcomes. Does this mean that HQ reflects a unidimensional, physical reality in your body? Does this mean medical practitioners would have any use of it? And more crucially, does this mean that attempting to optimize for it (as opposed to, you know, optimize for actual real things that we can readily see) wouldn't be an utterly fruitless endeavor? One last thing, since people seem to love brandishing GWASes around like a magic wand: people in the community are becoming increasingly disillusioned with them these days. (The head of EMBO, Ewan Birney, recently spoke about this. Note that I'm talking about academic geneticists, not the small clique of psychologists/"independents" from institutes with shady funding whose names invariably pop up when dealing with the less savory elements of the debate.) Again, for a geneticist, you need to find an actual genetic mechanism to prove something is "genetic", otherwise you end up with ludicrous associations (https://twitter.com/SbotGwa/status/1151964472921219073 https://twitter.com/SbotGwa/status/1151964472921219073, or a GWAS for type of fat used in cooking). This has not deterred people from less rigorous fields such as psychology to attempt to GWAS for literally anything. A charitable interpretation is that people conducting such studies are unusually naive, a less charitable one would be to draw connections among the aforementioned small clique and correlate them with the views people in said clique have expressed, and reasonably infer that a motivated agenda might be at play here. Of course I'm not looking to convince you since you appear to have read a truly gigantic amount on the subject and probably have a list of well-prepared rebuttals for literally any argument so that anyone with less time for online posting would never have the last word, so this is more of a post for any passer-by who'd think that IQ has any relevance beyond some niches of psychology. It's certainly not a thing in mainstream popgen or any field of genomics and if someone presented an abstract at any conference dealing with the kind of stuff posted on humanvarieties they'd be laughed out of the room.

by pcrh 7y ago

gwern 7y ago

> "Heritability" includes non-genetic factors, and combined with genetic factors accounts for ~50% of variation in intelligence. No, it doesn't. Heritability is intended and defined to include genetic factors only, unless you are referring to indirect genetic effects which are mediated through the environment but are still genetic in origin. > For genetic factors alone, the consensus so far is that there is very weak correlation of genetic variation with intelligence, unless a not quite-yet-established polygenic score is constructed. To wit: You are again being extremely misleading and selective in your quote-mining, and you omit all of the qualifiers that Plomin & Stumm include, eg "Recent genome-wide association studies have successfully identified inherited genome sequence differences" - 'identified genome sequence differences' does not imply that they are the only genes involved. Merely that some have been identified, and more will be. As they then go on to point out in the parts of the review you ignore, future GWASes will surely make much more progress in identifying SNPs, and approach the SNP heritability ceiling of ~30% given enough data. > It stretches the imagination how one would go from these numbers to claiming that "The best estimate of intelligence heritability in adulthood at the peak, using measurement-error-corrected estimates, is 70-80%. Er, because I'm not talking about SNP heritability estimates. I am talking about estimates from twin studies, adoption studies, and MZA studies, which estimate total genetic contributions. The lower-bound is 50%, you don't get anything lower from any twin studies, and this is deflated by the fact that many twin studies are looking at young kids and the Wilson Effect has not yet kicked in; this is then further deflated by the fact that they use noisy IQ measurements, which inflate the error variance component and hide the shared-environment/heritability components. Correct for those, and you get up to 70-80%. All of this is covered in the Plomin textbook cited there, starting pg171, where they cover things like how MZA designs deliver 72% heritability estimates and they explicitly mention that measurement error biases downwards ('Corrected for unreliability of measurement, heritability estimates would be higher...'). They are being highly conservative by using a lower bound of 50%, and they are wrong to, because as I already mentioned family-GCTA/GREML-KIN indicates that semi-rare variants alone get you up to 50% heritability, so we can be sure that 50% is a considerable underestimate (which is consistent with their expanded textbook discussion noting that the real heritability is much higher).

nbvcxwthrow 7y ago
You have a very peculiar (read: incorrect) definition of "heritability". http://bactra.org/weblog/520.html http://bactra.org/weblog/520.html
- gwern 7y ago
  No, I do not. Linking to Shalizi, who systematically misrepresents the issue and ignores all the evidence against his claims ( https://humanvarieties.org/2013/04/03/is-psychometric-g-a-myth/ https://humanvarieties.org/2013/04/03/is-psychometric-g-a-my... ) is not an impressive thing to do, nor does it defend in any way pcrh's persistent errors and misquotations and misrepresentations of SNP heritability & PGSes (which rhetorical strategies and arguments bear a remarkable resemblance to Shalizi's own attempts to FUD intelligence & heritability, which have failed so spectacularly he hasn't dared address the issue since Rietveld et al 2013 came out...).
  - nbvcxwthrow 7y ago
    The link does nothing to refute Shalizi's point. It basically amounts to saying: -Other models of intelligence as wrong and g is real because tests results correlate with each other, which is completely orthogonal to Shalizi's point. Because of the very nature of testing the results are going to correlate with each other; the most trivial example is, using Taleb's example, if dead people take the test. Dead people are dead and have an IQ of zero, and this is enough to increase the correlation (https://twitter.com/nntaleb/status/1114977580053147648 https://twitter.com/nntaleb/status/1114977580053147648). -Apparently confirmatory factor analysis means g is relevant because...confirmatory factor analysis means g is relevant? Nothing addresses the fact that g could be an artifact that arises in any set of measures that correlate with each other regardless of the origin. Adding to that, there is scant evidence of any link between g and the underlying neurobiology and/or actual causal genetic mechanisms, beyond very trivial cases such as obvious malformations etc. On the contrary, neurological findings are pointing to the opposite direction (https://www.ncbi.nlm.nih.gov/pubmed/23259956 https://www.ncbi.nlm.nih.gov/pubmed/23259956). -Even if we have no idea how g works it's still useful for predictions for stuff like educational attainment so it's relevant. Well duh, if you're good at taking tests you're going to do well in an environment where your attainment depends on your doing well in tests. Again, there is enormous circularity in reasoning. And as for the other outcomes, look at it this way: you could combine many measurements from your body (say, various analyzes from blood samples, but even trivial things like number of legs or amount of blood in your body could do), find they all correlate with each other (because when your body is breaking down it is likely damage to one organ that's spreading damage to other organs etc.), and from all that stuff you could draw a general "health quotient", or HQ. It would be highly heritable, there would be literal millions of SNPs that are associated with it. It would have immense predictive value for a variety of social outcomes. Does this mean that HQ reflects a unidimensional, physical reality in your body? Does this mean medical practitioners would have any use of it? And more crucially, does this mean that attempting to optimize for it (as opposed to, you know, optimize for actual real things that we can readily see) wouldn't be an utterly fruitless endeavor? One last thing, since people seem to love brandishing GWASes around like a magic wand: people in the community are becoming increasingly disillusioned with them these days. (The head of EMBO, Ewan Birney, recently spoke about this. Note that I'm talking about academic geneticists, not the small clique of psychologists/"independents" from institutes with shady funding whose names invariably pop up when dealing with the less savory elements of the debate.) Again, for a geneticist, you need to find an actual genetic mechanism to prove something is "genetic", otherwise you end up with ludicrous associations (https://twitter.com/SbotGwa/status/1151964472921219073 https://twitter.com/SbotGwa/status/1151964472921219073, or a GWAS for type of fat used in cooking). This has not deterred people from less rigorous fields such as psychology to attempt to GWAS for literally anything. A charitable interpretation is that people conducting such studies are unusually naive, a less charitable one would be to draw connections among the aforementioned small clique and correlate them with the views people in said clique have expressed, and reasonably infer that a motivated agenda might be at play here. Of course I'm not looking to convince you since you appear to have read a truly gigantic amount on the subject and probably have a list of well-prepared rebuttals for literally any argument so that anyone with less time for online posting would never have the last word, so this is more of a post for any passer-by who'd think that IQ has any relevance beyond some niches of psychology. It's certainly not a thing in mainstream popgen or any field of genomics and if someone presented an abstract at any conference dealing with the kind of stuff posted on humanvarieties they'd be laughed out of the room.