It is unfortunate that while this post makes more of an effort to engage with modern data than the typical response, it also makes several egregious errors that ultimately undermine that effort.
1. "Gusev claims that “IQ is much less heritable” than height, based on molecular genetic studies with vastly different statistical power to detect genetic effects on the two traits"
This is incorrect. The estimation of molecular heritability does not depend on the power to detect genetic effects at all, it is a single genome-wide parameter aggregated over all variants in the study regardless of whether they are significant.
I explained this point in my first post on IQ with a reference (https://theinfinitesimal.substack.com/p/no-intelligence-is-not-like-height): "But prediction accuracy depends on sample size, could the findings drastically change with more samples in the future? In fact, through the magic of statistics, we actually know that this claim will always to be true. We know this because we have estimated a parameter called molecular heritability, which tells us the upper bound on what a genetic predictor could ever achieve".
And then I explained it again in my second post responding to comments (https://theinfinitesimal.substack.com/i/148251755/isnt-gwas-heritability-only-quantifying-the-mechanisms-we-currently-understand): "Isn’t GWAS heritability only quantifying the mechanisms we currently understand? This is a typical misconception: that GWAS only quantifies the heritability from individual significant associations or genes we understand. In fact, GWAS heritability is defined as the phenotypic variance explained by all genetic variation that has been measured, whether it is significant or not.".
I can explain it a third time, I guess. The population-scale molecular heritability of IQ is 10-20% and the population-scale molecular heritability of height is 40-50%. These estimates come with very low uncertainty, and the fact that 40% is larger than 20% is well established. The inability of this post to grapple with the basic operation of these methods nor their basic estimates should raise red flags about the other claims.
2. "He singles out one galaxy brain paper, Bingley et al 2023, as evidence that heritability can be made very low and shared environmental influence very high. This paper gets this result by assuming equal avuncular shared environments, in other words, by making the clearly wrong assumption that all differences between the families of cousins are environmental."
This is incorrect. Bingley et al. propose a quantitative method to directly test the equal environment assumption, something the author should be pleased to see. Their method assumes equal shared environments between twins and their nieces/nephews as between twin spouses and their nieces/nephews ("To avoid adding parameters, we assume that the degree of environmental sharing between twins and their niblings (Equation 15) is the same as that between twins’ spouses and niblings (Equation 16)"). This does not imply that "all differences between the families of cousins are environmental", it is a constraint *specifically* on the shared environments between these relationship classes. This constraint is clearly much less restrictive than the assumption from classic twin models that MZs and DZs have identical shared environments, which we know they do not. In addition to the Bingley paper, I describe a number of other studies that evaluated the EEA assumption with other extended family designs and found it to be violated (https://theinfinitesimal.substack.com/p/twin-heritability-models-can-tell). Again the author appears not to understand the model but it is confident that the model is wrong because it produces uncomfortable results.
3. "In the first, [Gusev] keeps making the argument that if a molecular method (RDR) finds lower heritability estimates then the quantitative estimate then the latter is wrong, which, as we saw, is a fallacy."
The author first presents RDR as "quantitative genetics with extra frills" but then later discards RDR because it is a molecular method. Well, which is it? What the fallacy is in reporting the RDR results is also never stated. It is odd to simply discount the strongest evidence against one's position as "a fallacy", typically this is the evidence one would addresses first and in most detail. I'll remind readers that the paper that introduced the RDR method is titled "Relatedness disequilibrium regression estimates heritability without environmental bias" (https://pubmed.ncbi.nlm.nih.gov/30104764/) and the entire motivation of the method was to estimate complete additive heritability (using Identical By Descent relatedness estimates) without the environmental assumptions of classic twin models. When applied to a large number of traits, the average RDR heritability was 32%, compared to an average twin heritability of 61% (and a SNP-based heritability of 26%); results that I also summarized in my post, which also goes into how RDR works (https://theinfinitesimal.substack.com/i/148251755/okay-but-why-do-you-think-twin-estimates-so-much-higher). Since RDR and twin estimates are biased to the same extent by assortative mating, this gap is explained by environmental bias in twin studies. So the quantitative genetics (with extra frills) models the author likes also tell us that twin estimates are 200% inflated on average, and that SNP/GWAS estimates are about 80% of the total heritability on average. These relative estimates are consistent with a variety of other studies and theory showing that common variants (i.e. those in GWAS) explain the majority of total heritability, so everything hangs together nicely.
---
TLDR: This post misunderstands how molecular genetic studies actually estimate heritability, then misunderstands how quantitative extended twin models have probed the equal environment assumption and found it violated, then finally does a sleight of hand by pretending that RDR is not a quantitative genetics method because the findings are inconvenient.
Finally, I do appreciate that this post finally acknowledges that behavioral GWAS were greatly inflated by population stratification and, for hereditarians, a disappointment. But I don't think we can simply let hereditarians off the hook, because they were the ones who bet hard on GWAS methods to begin with. Charles Murray predicted for over a decade that molecular studies were just about to prove him right (https://x.com/evopsychgoogle/status/1188441459860443137); Razib Khan argued that we just need to wait a little bit longer for GWAS to fully explain the genetics of intelligence (https://undark.org/2017/02/28/race-science-razib-khan-racism/); Steve Hsu predicted that GWAS would explain 60% of the total variance of intelligence by 2016 (https://x.com/SashaGusevPosts/status/1613724821480677378); Emil Kirkegaard has been running bogus polygenic score comparisons pretending stratification is a non-issue and getting the results completely backwards (https://theinfinitesimal.substack.com/p/how-population-stratification-led). Even setting aside motivated hereditarians, here's what internet celebrity Gwern said about the first GWAS of educational attainment: "Reading it was a revelation. The debate was over: behavioral genetics was right, and the critics were wrong. Kamin, Gould, Lewontin, Shalizi, the whole sorry pack—annihilated. IQ was indeed highly heritable, polygenic, and GWASes would only get better for it" (https://www.lesswrong.com/posts/sRchPdp6mCqY2ekJX/what-progress-gwern-s-10-year-retrospective). As we now know, this is not what actually happened.
And this ultimately gets at my issue with the modern hereditarian movement, my point is not to attack their *theories*, everyone is entitled to speculate about genetic processes and test out their speculation. What I am attacking is the pattern of overfitting and p-hacking in their *methods*. When GWAS provided a cudgel to beat up on Lewontin and Shalizi, then suddenly it was "a revelation" that will only get better and better. Now that GWAS falsifies these claims, it is deemed "something of an embarrassment". When the RDR method is first described, it is just "quantitative genetics with extra frills" (laudatory) but when it yields uncomfortable results, it is now "molecular genetics" (derogatory) and using it is "a fallacy"". Methods and findings are not used to understand how the world works, but to stitch together a narrative that supports the author's predetermined conclusion.
I take multiple issues with your comment on empirical grounds. I’ve followed some of your work since ~2016 give or take, but abandoned it later on as it seems you have switched to untenable positions without much evidence and I cannot really understand why (e.g., old acknowledgments that methods controlling for stratification fail to model it very well, are all linear etc but current support for PCs as a a proper stratification control). I believe this position that the lowest SNP_h2 estimates one can squeeze out are the most accurate ones to be extreme at best, and in-line with Bouchard’s (1982) description of what he coined “pseudoanalysis” at worst. I wouldn’t go far beyond this since I have not seen you post sources which contradict what you say about them or seen you lie (maybe some people will debate this), but I will go on to discuss what I think you get wrong often.
1. The fact that molecular h2 isn’t based on per-SNP significance doesn’t imply these are the “upper bound estimates we could ever achieve”. The estimates still only capture the variation tagged by common SNPs in LD. Painting this as if it has settled the debate is misleading, for reasons I will also get to below this point. These methods still have plenty of unidentified rare & structural variants as well as nonadditive genetic effects which are incredibly unlikely to be included in molecular designs anytime soon (goes against the whole point). But you already know all of this.
2. The Bingley method is arguably worse as it assumes equal environmental influence from twins and from the twins’ spouses on a niece/nephew (which bakes in indirect environmental sharing in a convoluted way). Whether the EEA holds is a matter relevant to the trait you are analyzing. You cannot generalize whether or not the EEA is practically true since a trait-relevant environment could be a trait-irrelevant one for a separate characteristic. For intelligence specifically, *g* is not affected by the commonly cited environmental variables alleged to violate the EEA (not adoption; not education, your post on the Ritchie paper notwithstanding; not prenatal effects and so on and so forth), so the EEA holds (and see Barnes et al., 2014; EEA violations do not invalidate the estimates empirically). Therefore high twin_h2 estimates for *g* (as in Panizzon et al., 2014) are valid wrt the EEA and within-family designs even moreso as they avoid other issues you allege.
3. You cannot conclude that since AM affects RDR and twin methods to the same extent, it must be the case that twin results are inflated by environmental variance. That’s a clear non-sequitur provided that there are other sources of variation (including modeling assumptions, like assuming away indirect effects under idealized conditions, as RDR does, thus underestimating h2 when they are violated) which explain the discrepancies. RDR estimates are also known to be less stable for traits with lower signal-to-noise (like intelligence). You have also been made aware of other methods that remove twin assumptions and estimate heritability to be high (Schwabe, Janss & van den Berg, 2017) to which you have not provided adequate criticisms that would explain discrepancies as high as >70%, such as assumptions regarding distant-relatives or C being zero (this would still put the heritability at >60% assuming both of these are addressed).
The hereditarian predictions you cite seem to have been wrong in a few ways, but the “fallacy” the OP describes was never citing molecular results. It was declaring them to be the final evidence and pretending they capture virtually all of the total heritability. So far I know of at least two papers which were able to attribute ~20% of the B-W gap in *g* to common associated SNPs. The argument is far from settled. But family-based estimates are still strong evidence and pedigree-based data has been shown to be able to bridge the “missing heritability” gap between methods (as in Hill et al., 2018) with h2 estimates reaching values >.5 even with bad assumptions (like that AM is totally environmental, as in the Hill paper). Where are the “biases” here which other methods capture? Where is the p-hacking? The “overfitting”, whatever that means? I hope it has been made clear that the “hereditarian” position (which, in my opinion, is an agnostic one) has way more ground than you are willing to admit.
I know I'm not Dr. Gusev, but I wanted to issue a response to this reply as it betrays several critical misunderstandings.
> The fact that molecular h2 isn’t based on per-SNP significance doesn’t imply these are the “upper bound estimates we could ever achieve”
To my knowledge, Dr. Gusev has never stated that SNP-heritability constitutes an upper bound on narrow-sense heritability. He specifically states that molecular heritability "tells us the upper bound on what a genetic predictor could ever achieve" where the term "genetic predictor" refers to *polygenic scores*. This context is made abundantly clear in the post that Dr. Gusev linked.
Why is Dr. Gusev talking about polygenic scores? Because East Hunter seems to be under the impression that, if genomic variants do not achieve statistical significance, they are not included in SNP-heritability. But this statement is incorrect. While non-significant variants would not be included in a polygenic score, they *would* be included when estimating molecular heritability. Thus, molecular heritability represents an upper bound on the predictive accuracy of polygenic scores.
> You have also been made aware of other methods that remove twin assumptions and estimate heritability to be high (Schwabe, Janss & van den Berg, 2017) to which you have not provided adequate criticisms
I'm happy to provide adequate criticism here.
Unlike RDR which leverages exogenous genetic variation (i.e. random inheritance of genetic variants), the paper you cite employs a pedigree-based method where, in the authors' own words, the "model is statistically equivalent to the ACE model for twins, except that the common environmental component C is assumed to be zero". In other words, this model relies on a statistically equivalent set of assumptions as twin studies *plus* the assumption that shared environment makes zero contribution to IQ. Thus, your statement that this study "removes twin assumptions" is unequivocally false.
Additionally, note that RDR makes no such assumption while still exploiting degrees of genetic relatedness across relative classes. In other words, RDR does exactly what the study you cite does *except* that it leverages an actually exogenous source of genetic variation as opposed to sources of genetic variation which are correlated with environmental variation.
> But family-based estimates are still strong evidence and pedigree-based data has been shown to be able to bridge the “missing heritability” gap between methods (as in Hill et al., 2018) with h2 estimates reaching values >.5 even with bad assumptions (like that AM is totally environmental, as in the Hill paper)
Hill et al is a pedigree-based method which both Dr. Gusev and Dr. Alexander Young (the lead author on the RDR paper) have responded to numerous times. Here is Dr. Gusev's response (which also includes Dr. Young's response):
The basic problem, once again, is that the paper does not leverage randomization to disentangle genetic and environmental processes, and as a result, is not a suitable alternative to RDR.
To make an analogy, your citation of pedigree-based methods is analogous to throwing away the results of an RCT because they don't agree with observational methods. I hope you can agree that no scientist worth their salt would do such a thing, all else being equal between the study designs.
No, I’m sorry but this is a LARP and not a serious response. My post contains no misunderstandings outside of an extremely debatable semantic point regarding genetic predictions. You could have simply read what I typed and cited more carefully in order to avoid confusing yourself and others. But you chose not to, for some reason.
> refers to polygenic scores
Fine. I won’t even ask why he presents molecular results the way he does (obvious implications) or why he compares them to twin results right after this point and in other posts repeatedly. The statement is still just as wrong. There is no sensible way in which molecular results could be the “upper bound” of PGS since the former (1) literally informs the latter and (2) variance components could differ from within-group SNP_h2 to between-group_h2 across PGS group comparisons. This is only NOT a non-sequitur if you assume that molecular_h2 is not going to increase and better inform PGS, and it doesn’t even work because assuming that this is the case, PGS could still show higher heritabilities in population comparisons.
> I'm happy to provide adequate criticism
You know the funny part about this is that you stopped your quote of my statement right before I said “… that would explain discrepancies as high as >70%”. Just bringing this up because below it should be clearly demonstrated that none of the criticisms you allege explain anywhere close to that amount.
> In other words, this model relies on a statistically equivalent set of assumptions as twin studies *plus* the assumption that shared environment makes zero contribution
You say “in other words” and then proceed to completely misinterpret the paper. When the authors say that their *model* is statistically equivalent, it implies approximately equal fit, not similar assumptions (what in the world would it even mean for an assumption to be “statistically equivalent”?) which should have been clear from the title of the paper alone anyway. Why would the authors publish a paper attempting to validate twin studies using twin study methodology? They obviously apply different assumptions (no EEA, no MZ/DZ comparisons etc). You should know this provided that you acknowledged the method as a pedigree-based one, which literally requires a different kind of analysis to calculate heritability.
> note that RDR makes no such assumption
What assumption? You mean C being zero, which it basically is by adulthood? Again, I have to ask, did you read the paper? The authors are very clear about their modeling of C changing nothing about their heritability estimate:
> Note however, that the shared environmental component was rather small (8 and 16%, respectively) […] this would decrease heritability from 94% (that includes shared environmental variance) to around 77–85%, which would mean that the result is not at odds with the results found by earlier twin studies. Put differently, if we acknowledge that we can only estimate A + C together in our analysis (that is, in the empty model), and when we compare that to the A + C estimates based on the twin studies using the exact same birth cohorts, the results are rather similar: 94% here, 74 + 8 = 82% […] and 66 + 16 = 82%.
& tangentially:
> When we compare these heritability estimates to the estimate of 85% in this study, we can conclude that the high estimates resulting from the twin method are not simply an artifact of self-selection or because of any important difference between twins and singletons. Twin-based heritability estimates are not inflated, since an estimate based on a sample from the entire population (including twins and singletons) is even higher.
—
—
Continuing.
> RDR does exactly what the study you cite does *except* that it leverages an actually exogenous source of genetic variation as opposed to sources of genetic variation which are correlated with environmental variation
No, it does exactly what the study I cited does except that it assumes variance components with zero evidence. RDR is basically a method of “what if X variables affect heritability” materialized. What I’m citing is instead “which variables affect heritability”. The latter modeling is real, RDR’s modeling is just assuming h2 inflation. If your issue is rGE (you say “sources of genetic variation which are correlated with environmental variation”) then what I cited should not be a concern, since according to multiple different papers, rGE has almost no impact on heritability estimates (empirically; as stated in Conley, 2014 for example, or Willoughby et al., 2022 which found rGE was ~0.02 and an h2 of ~.62 for the ICAR-16 + >.53 for IQ in general; table S12, see here; https://imgur.com/a/yQu4UF1)!
> The basic problem, once again, is that the paper does not leverage randomization to disentangle genetic and environmental processes
Yes it does. If you had read the paper and the alleged criticisms you would probably know that ~0.027% of the sample comprised of related individuals according to table 1. The Hill paper is much closer to an RCT than RDR, which uses a dense family structure. And then you need to point to biases which *are there* (not just assumed to be there) which would ***massively*** inflate the estimate. To steelman your case a little, Gusev’s criticism of this is that assortative mating and indirect genetic effects might be biasing the results upwards, but again, the paper assumes all of the variance attributable to AM is environmental in origin, and almost none of the sample is related to any significant extent. The former modeling choice is obviously untrue, and does two things. For one, it underestimates heritability. For two, it gets around Gusev’s criticism. Young also provides no criticisms of the paper and instead simply claims that the results “could be” upwardly biased in the presence of indirect genetic effects. Really, that doesn’t tell me anything regarding what is the case empirically. I can only conclude from this that you either have bad judgment or you simply did not read/understand the responses, because none of these are valid criticisms or explain the other results (e.g. Schwabe paper, Panizzon paper &c).
Lastly, assuming certain variance components exist, like RDR does, is not the equivalent of an RCT. Saying “here I present you with heritability estimates which I will now model as being biased by effects I have not demonstrated are there” is, to me, the equivalent of making things up. Do you notice any of the methods I cited doing this? No. RDR is hypothetical. My methods are empirical. That’s the difference.
So no, to wrap this up, I have yet to be provided with criticisms that would invalidate the results I cited where h2 > 60%.
> Fine. I won’t even ask why he presents molecular results the way he does (obvious implications) or why he compares them to twin results right after this point and in other posts repeatedly.
I'll leave this for Dr. Gusev to respond to since it's a point about authorial intent.
> You say “in other words” and then proceed to completely misinterpret the paper. When the authors say that their *model* is statistically equivalent, it implies approximately equal fit, not similar assumptions ... Why would the authors publish a paper attempting to validate twin studies using twin study methodology? They obviously apply different assumptions (no EEA, no MZ/DZ comparisons etc).
First, the authors do actually assume the EEA because they set the shared environment component equal to zero across all modeled relative pairs. Thus, all modeled relative pairs share the trait-relevant environment to the same degree - zero.
Second, what I mean by "statistically equivalent" is that the set of assumptions which apply to twin studies are just extended to more degrees of relatedness (e.g. no gene-environment interactions, no gene-environment correlation, no assortative mating, EEA).
Third, you ask, "why would the authors publish a paper attempting to validate twin studies using twin study methodology?" and in a twist of irony, I think you've arrived at the answer yourself. When the authors acknowledge that their model assumes C=0, what do they do? Retreat back to twin studies to say that C is negligible!
The whole paper reeks of this kind of circularity in my opinion. They use a statistically analogous set of assumptions as a twin AE model just applied to a larger familial structure and then pretend that this allows one to validate the underlying assumptions of twin studies which are actually in contention.
> What assumption? You mean C being zero, which it basically is by adulthood? The authors are very clear about their modeling of C changing nothing about their heritability estimate:
Again, the authors traffic in circularity here. They claim to develop a methodology to "validate" twin studies and then, when pressed on the limitations of their own methodology, they just fall back to twin studies.
The reason they have to do this is stated in the manuscript:
"we have not taken into account whether sibling pairs and half-sibling pairs actually share the same home environment, nor whether cousins share the same household (could be true for some families), nor whether nonrelated children share the same household. We were therefore not able to estimate non-genetic variance that is shared due to being brought up in the same home environment"
My point is that RDR does not do this because it leverages the random inheritance of genetic variants (and thus, the relevant genetic influences are independent of environmental sharing).
> No, [RDR] does exactly what the study I cited does except that it assumes variance components with zero evidence. RDR is basically a method of “what if X variables affect heritability” materialized.
No, RDR merely *allows for* the existence of additional variance components. It does *not* assume that such components are non-zero. If the data empirically indicates that this is not the case, then the corresponding variance component will be set to zero.
The best way to see this is Supplementary Table 2. In a simulation where the trait in question is purely determined by additive genetic variation + noise (i.e. no genetic nurture). RDR correctly estimates 0 correlation between the environmental and parental genetic components of the phenotype.
> If you had read [Hill et al 2018] and the alleged criticisms you would probably know that ~0.027% of the sample comprised of related individuals according to table 1.
Yes, the sample consists predominantly of genetically-unrelated individuals but the percentage of heritability attributable to the sample of related individuals (h^2_kin / h^2) is relatively high (>50% in the case of general cognitive ability - see Figure 1). In other words, it is this narrow family-based slice of the sample that drives up the estimated heritability, and that's exactly what Dr. Gusev and Dr. Young take issue with.
While the authors state that they are able to replicate this result in the sample of fully unrelated individuals using GREML-MS, Evans et al points out that GREML-MS "requires that the same assumptions of GREML-SC hold within each [genomic relationship matrix]" and that GREML-SC assumes that "genetic similarity is uncorrelated with environmental similarity". In other words, GREML-MS relies on a stronger set of assumptions than RDR.
> The Hill paper is much closer to an RCT than RDR, which uses a dense family structure.
I disagree with this statement. I hate to beat the same drum again, but RDR leverages the *random* inheritance of IBD segments within relative pairs to estimate heritability. In other words, if I randomly inherit an IBD segment from my grandfather and my cousin doesn't, that is a difference in our genetic code that is, statistically speaking, independent of our environments and can lead to downstream differences in our phenotypes. Extrapolate that across N relative-pairs, and you get a heritability estimate that is based on exogenous genetic variation.
Conversely, Hill et al is not guaranteed to leverage purely random genetic variation. While both RDR and Hill et al involve analyses of pedigrees, only RDR leverages portions of the genome which are randomly inherited. Hill et al provides no such guarantee AFAIK. If I'm wrong about this, I'd love to see a proof explaining why.
> Lastly, assuming certain variance components exist, like RDR does, is not the equivalent of an RCT. Saying “here I present you with heritability estimates which I will now model as being biased by effects I have not demonstrated are there” is, to me, the equivalent of making things up. Do you notice any of the methods I cited doing this?
It's true that the methods you cite don't do this but neither does RDR. Just as the adoption study you cite *allows for* the existence of rGE, so too does RDR.
> the authors do actually assume the EEA because they set the shared environment component equal to zero
That’s not the same thing. The EEA is there so C and all other components are estimable. Otherwise, there would be no point for ACE models in twins. I’m not sure why you said this.
> Retreat back to twin studies to say that C is negligible
Unless you have a criticism of twin studies and multiple replications of the Wilson Effect and C being negligible across multiple family and within-family studies &c &c, this is not relevant. As was stated in the paper, this made little difference to their results. If you care to propose a C value you think would be plausible, do so. Even something as high as 30% still puts the h2 at >60-70% depending on a few other controls as well. The authors justified this pretty strongly. Please go and read the paper more carefully.
> The whole paper reeks of this kind of circularity
Evidence-based inference based on other independent modelings of C is not “circularity”. It is justified extrapolation based on cross-study convergence, which - again - made little difference to the results regardless of the magnitude (across plausible ranges).
> It does *not* assume that such components are non-zero
It assumes no indirect genetic effects despite its goal not having been to capture just the additive variance structure, which is also underestimated by RDR if it omits higher-order interactions. The most obvious assumption it hinges on (which is indeed the motivation behind the method) is that shared environmental effects are orthogonal, i.e. that these effects don’t vary with deviations in genetic relatedness. We know from many papers using reared-apart twins, heritability method comparisons and other designs (Lee, 2009; Visscher, Yang & Goddard, 2010; Beaver et al., 2014; Conley et al., 2014, see my original comment etc) that shared environment does not in fact track closely with true relatedness (and see Willoughby & Lee, 2010 as well as the replications, like by Kandler et al., 2016; Bates et al., 2018 &c, showing that parental PGS adds nothing after conditioning on child PGS). But even worse, if shared environment is correlated with IBD *even slightly* (which it is), RDR is even more biased downwards! It makes no sense to prefer this method over the one in the Schwabe paper.
> where the trait in question is purely determined by additive genetic variation + noise (i.e. no genetic nurture)
Only additive variance + noise doesn’t mean no genetic nurture. It ignores literally all broad and indirect genetic effects.
> the percentage of heritability attributable to the sample of related individuals (h^2_kin / h^2) is relatively high (>50% in the case of general cognitive ability - see Figure 1 […] GREML-MS relies on a stronger set of assumptions than RDR
What? This is not seen in Figure 1, I have no idea what you’re talking about. You’re claiming that over 95% of the estimated heritability in the Hill paper is significantly biased by less than 0.1% of the sample? This is an insane inquiry and requires evidence. Based on the papers I cited above, the idea that genetic similarity is uncorrelated with environmental similarity seems to be justified and not stronger than the assumptions RDR makes, so the GREML-Ms estimate in the Hill paper and the subsequent 2017 analysis is largely unbiased and probably not “biased” by the related individuals, but obviously driven by them due to the broader capturing of genetic effects.
> RDR leverages the *random* inheritance of IBD segments within relative pairs to estimate heritability
Yes, but the point of this is that within-family variation is generally immune to AM, stratification etc. But again, so are within-sibling GREML designs, adoption studies, within-family GWAS and so on and so forth. It seems like special pleading to ignore all other evidence in favor of RDR if your concern is random segregation.
> Hill et al provides no such guarantee
It doesn’t need to but the related sample is really small to be biasing anything *too* significantly.
Again, to close off, the papers I cite finding high heritabilities don’t look like they are invalid. You have to propose stronger arguments and solutions to issues to discard the past 100 years of heritability research.
> That’s not the same thing. The EEA is there so C and all other components are estimable. Otherwise, there would be no point for ACE models in twins. I’m not sure why you said this
I think it's worth defining our terms here. The Equal Environments Assumption (EEA) stipulates that all modeled relative pairs share the trait-relevant environment to the same degree. For example, in the Classical Twin Design (CTD), the EEA entails that MZ and DZ twins share the trait-relevant environment to the same degree.
Now, the question we must ask ourselves is this: does assuming C = 0 (as Schwabe et al do) necessarily entail assuming that all modeled relative pairs share the trait-relevant environment to the same degree? The answer is yes. Just as the CTD assumes C_mz = C_dz, Schwabe et al assume that C_bio-sib = C_half-sib = ... = 0. In other words, the assumption that C = 0 necessarily entails the EEA.
Why do I bring this up? Because your initial claim was that Schwabe et al "removes twin assumptions" when it does no such thing. It merely extends the EEA (and other CTD assumptions) to a larger pedigree with the additional assumption that C is specifically equal to zero.
> [RDR] assumes no indirect genetic effects
Is this a typo? Because otherwise, this statement is unequivocally false. The v_e~g variance component estimated by RDR captures the variance driven by the environmental component of the phenotype which is correlated with paternal genotype (thus allowing for indirect genetic effects).
> despite its goal not having been to capture just the additive variance structure, which is also underestimated by RDR if it omits higher-order interactions
Could you clarify what you mean here? If you're talking about dominance or epistasis, then that is not relevant because Young et al demonstrates in simulations that RDR will recover the true narrow-sense heritability even in the presence of dominance and epistasis (see Supplementary Table 2).
> Only additive variance + noise doesn’t mean no genetic nurture. It ignores literally all broad and indirect genetic effects.
I realize I wasn't clear with my description of the simulation, so allow me to refer you to the caption underneath Supplementary Table 2:
"We simulated 500 replicates of each trait based on actual Icelandic genetic data for 10,000 individuals. Ten thousand SNPs with median frequency 23% were given additive effects for all the traits other than the rare SNPs trait, for which 2,200 SNPs with frequency between 0.1% and 1% (median 0.26%) were used. The true (narrow-sense) heritability of each trait was 40%. To this additive genetic component, only noise was added for the additive trait and the rare SNPs trait"
Young et al then describe various modifications they make to this simulated trait such as the introduction of genetic nurture:
"For the ‘genetic nurturing’ trait, the genotypes of the parents were also given effects to simulate ‘genetic nurturing’ effects"
How does RDR perform under such simulations? Well, as Supplementary Table 2 highlights, RDR does not estimate superfluous variance components as you implied. For the trait w/o genetic nurture (labeled "additive" in the table), v_e~g was correctly estimated at ~0.
> But even worse, if shared environment is correlated with IBD *even slightly* (which it is), RDR is even more biased downwards!
Ok, I'm genuinely not understanding what point you're making here or how you arrived at this conclusion. Earlier in your reply, you state:
"We know from many papers ... that shared environment does not in fact track closely with true relatedness"
But now, you are saying that shared environment *does* actually track relatedness ("if shared environment is correlated with IBD even slightly which it is") and somehow, this fact only biases RDR but not the other methods you cite even though RDR specifically only leverages relatedness that is *random* (i.e. independent of environmental influences) while the pedigree-based methods you cite make no such guarantee. At worst, your train of thought is completely contradictory. At best, the message is just getting garbled somewhere.
> What? This is not seen in Figure 1, I have no idea what you’re talking about.
Earlier, you accused me of not having read Hill et al. Now I must accuse you of the same. In the very first paragraph of the "Results" section, the authors state:
"For g, common SNPs (h2g) explained 23% (SE = 2%) of the phenotypic variation. Pedigree-associated genetic variants (h2kin) added an additional 31% (SE = 3%) to the genetic contributions to g, yielding a total contribution of genetic effects of 54% (SE = 3%) on g"
Note that 31/54 > 0.5. Hence, greater than 50% of the estimated heritability is driven by the pedigree-based sample. The authors re-emphasize this point in the "Discussion" section:
"The pedigree-associated genetic variants accounted for over half of the genetic effects in these phenotypes."
> [Hill et al] doesn’t need to but the related sample is really small to be biasing anything *too* significantly.
Your continued insistence on the small size of the pedigree-based sample is a complete non-sequitur. As the authors repeatedly point out, the pedigree-based sample makes a large contribution to the heritability of general cognitive ability in their sample.
Thanks for replying to the post! Braeann Danads said many things I wanted to stay. You raise very valid points but I disagree that these are "egregious errors" that undermine the point of my essay, which is that you can't use molecular genetic methods to falsify quantitative genetic findings of high heritability. 1) I was mainly thinking about within-sibship GWAS when I mean that different power can result in different heritability but I see now how this is not 100% clear. You are of course right that in an SNP heritability study (not a GWAS) low power should not bias effect sizes downward, although it should affect precision. PGS heritability should be affected by low power in the original GWAS though, even if you use all variants and not just the significant ones, because the effects of individual SNPs are not estimated precisely which affects the quality of the PGS. 2) You are probably correct that I got the Bingley et al model wrong (Unboxing Politics also left a comment about this). I'll read the paper again and update the post. 3) In the essay, I classified methods as quantitative or molecular based on the data they use (pedigree or SNPs). This makes RDR a molecular method, which I think I make clear. RDR is interesting but it still comes with the typical assumptions of molecular genetic studies.
Just to be clear, I also made the point that PGS accuracy is sample-size dependent and specifically advocated for focusing on heritability estimates for that exact reason ("But prediction accuracy depends on sample size, could the findings drastically change with more samples in the future? In fact, through the magic of statistics, we actually know that this claim will always to be true." ~ https://theinfinitesimal.substack.com/p/no-intelligence-is-not-like-height). I suspect part of the problem is outsourcing these critiques to the Bronski post, which makes many basic mathematical errors (some pointed out in Bronski's comments). Bronski read my explanation, proceeded to ignore it to make his arguments, then this worked through the broken telephone as a "for more, see:" reference in multiple response articles (yours and another in the quant racist journal Aporia). I don't know of this was intentional, but it has produced a tangled web of superficial "dunks" that never actually spell out what they're criticizing.
Regarding RDR (and your comments elsewhere), it is simply not the case that RDR is fancy Sib-Reg and/or fancy GCTA/GREML, though it shares some superficial aspects with those methods. RDR (like Sib-Reg) uses IBD segments to capture (nearly) complete additive heritability, but since it uses trios rather than siblings it is not biased by sibling indirect effects or interactions. RDR (like GCTA/GREML) uses variance components but because it is exploiting the random Mendelian transmission within families it is not biased by parental indirect effects or shared environment. Thus, RDR -- like the paper title says -- is able to probe the assumptions of twin studies without introducing new assumptions of its own. This is really fundamental to triangulating the estimates across these methods and the reason I have written so much about them (http://gusevlab.org/projects/hsq/#h.b480s3ce0s08).
I think it's wrong to get bogged down in the mathematical minutiae of statistical genetic models because we are missing the forest for the trees. As I wrote in the post: genetically identical people raised by different families end up very similar, while genetically related people raised by the same family don't. This is highly ecologically valid evidence based on minimal assumptions so we should weigh it strongly. RDR and other molecular methods are full of bells and whistles - you are not just testing if family members who end up slightly more or less genetically similar due to random meiotic forces are phenotypically similar, but you are trying to precisely quantify the relative importance of genetic forces based on the slope of this relationship! This is a neat idea but I'm not seeing how a result discordant to twin or adoption studies is evidence that the latter are wrong. I think you are wrong to respond with snark to Candide, he raises a point I agree with: the deflation of height heritability tells us that RDR misses something. If the twin literature is wrong about this and heritability is only 55%, then we are not only saying that shared environment accounts for about a third of family clustering of height - let's say that's possible - but also that MZs have unique family environmental experiences which make them more similar than genetics would dictate. Maybe height increases if the family eats well (height Flynn effect and all) but is it plausible that parents e.g. feed MZ twins so much more similarly to DZs that this ends up being half as important as genes?
If the goal of your post was to argue that we should only use evidence from adoption/raised apart studies and ignore any other inconvenient findings, that would have been a much shorter (and more honest) post! But that's not what you wrote. Your argument was that these models were fundamentally flawed and produced erroneous results. Now that they have been shown not to be flawed, at least in the ways you outlined, the reasonable thing to do is to correct your post and say "whoops, I guess I really misunderstood what was going on here, let me integrate this new information into how I think about the world and come up with an updated conclusion". Instead, the response appears to be "let's just ignore the findings from these models because they are inconvenient and focus on the findings that I do like". So we are back to exactly the criticism that I laid out at the beginning of my very first comment: modern hereditarianism is an overfitting exercise and now that it has been caught flat-footed making poor predictions, it is rapidly discarding the methods and findings that don't fit the narrative.
Nothing has been shown not to be flawed. My point was that assumption-free quantitative models (adoptees, MZAs, maybe also twins after so many assumption sanity checks) are more parsimonious than molecular genetic studies so they cannot be dismissed based on the latter. I stand by this point. I realized that our main contention is that you treat studies based on genome-based relatedness (RDR, sib-regression, GCTA) as exactly analogous to traditional pedigree studies. I don't, the fact that the former have lower estimates even for e.g. height tell me that this model misses some important aspect of relatedness that is important for phenotypic similarity. I'm not pretending to know what it is so this is just speculation: missing variants, non-linear effects, G*G, data errors. This is a much more parsimonious explanation of discrepant findings than that C>A despite all evidence to the contrary.
I think we're converging to the point. On the one hand we have Bingley et al. (making fewer assumptions than twin studies) and RDR (making no assumptions about environment at all). Both estimated in large, modern data sets and neither exhibiting the original flaws you had argued for (which, I'll just note again, really should be corrected in your article).
On the other hand we have (1) your hunch that height heritability of ~70% (after correcting for assortment) is too low; and (2) twin adoption studies that assume there is no selective placement, no GxE, no range restriction, no prenatal effects, and largely consist of bespoke analyses from the time of the replication crisis.
Well, okay, let's look at the adoption studies. First, the MISTRA estimated an MZA correlation of 0.62 and a DZA correlation of 0.50 for IQ (not sure why you omitted this from your post). In the classic ACE model, that produces a heritability of 24% and a selective placement effect of 38%! In an AD model it produces nonsensical estimates. Given the massive selective placement, even the 24% heritability estimate doesn't really tell us anything. Second, the SATSA estimated a DZA correlation of 0.32 which was paradoxically higher than the DZT correlation of 0.22; either implying that living together causes twins to be more different or, you guessed it, selective placement for the DZAs. When al relationship classes are fit, they also produce nonsensical estimates (Table A1): 0% additive genetic variance, 80% dominance genetic variance, and -38% non-shared environment -- that's right, a negative non-shared environment variance. Large selective placement effects (Ec) were also observed for multiple individual subtests (Table 3). In other words, the results are a complete mess. The model can't fit the observed correlations parsimoniously, so it introduces nonsensical negative variances to make it work, arbitrarily setting other paths to zero.
You are advocating that we put our faith in low quality observational data instead of genetic randomization and the observational data plainly violates your methodological assumptions to boot. Is it fair to consider these studies an embarrassment yet?
Call me stupid, but I just don't see which environmental effects could inflate heritability of height in Iceland from the 55% found in the original RDR paper to the standard ~80%. Iceland is not a poor country with inadequate nutrition and high wealth inequality - it is a rather rich country with superb nutrition and one of the most equal in Europe, one of those weird places where prime ministers ride to work on the common bus. Is it exercise equipment in the home, perhaps? (/s) The original RDR paper modeling (Table 1) gives 'maternal environment' and 'genetic nurturing' as examples of 'environmental effects' for which kinship methods significantly overestimate heritability. But if these effects are truly non-genetic, it would have been odd for them to cut off at or soon after birth, and if that was the case, then adoption studies would not have given similar heritability results to other methods.
Good question. People around here seem to have very well honed intuitions for what the *right* heritability should be for certain traits, I guess they are able to look around at all the twins they know and eyeball that 55% is just too low but 80% is just right for height -- I wish I had this ability. In terms of environmental influences, we already know that height can be significantly impacted by birth order including in wealthy European countries, so the family environment is influential and we also know there have been large secular changes in height in Europe as well (a kind of Flynn effect) so broader environmental influences can be involved too (I'll leave it to others to counter-argue that the increase in height is actually hollow because it is not on "h" the general height factor).
Setting environment aside, RDR and the classic twin design are both biased down by assortative mating; so while assortative mating will not explain the gap between RDR and twin estimates, it might explain why the RDR estimate feels low to you personally. It is possible to do a back-of-the-envelope correction for AM based on the observed mate-pair correlation (see: http://gusevlab.org/projects/hsq/#h.st55vexg74ep for details). For an RDR estimate of h2 = 55% and a mate correlation of 0.3 (taken from Horwitz et al. ; https://pubmed.ncbi.nlm.nih.gov/37653148/), the corrected h2 is 70%. Not quite the 80% you were aiming for but hopefully that helps you feel a bit more comfortable with RDR (or maybe the mate correlation was 0.4 in Iceland, the corrected estimate is 80% and everyone is happy)? On the other hand, if we take the twin estimate of 81% and a mate correlation of 0.3, the corrected h2 is ... well it is 139%. I know I said earlier that I can't eyeball heritabilities, but 139% still seems implausible. So that's something else for twin study enjoyers to think about.
With all respect, you are assuming things about a person you're seeing for the first time on the sole evidence of "place of observation". Isn't that stereotyping? I am not "aiming for" any specific value. I note a large discrepancy between standard results from older literature, which even someone as uninformed as myself has happened upon in my reading, and results from a relatively new method (RDR), a discrepancy that the original RDR paper thought worth remarking upon in its discussion section (it refers to estimates from Swedish twin studies as the closest available comparison group). I am asking questions because I do not have an intuition for what the "right" heritability is for a continuous trait like height, except that values too close to 0% and 100% feel wrong to me on general principles. Your statements about mate correlation sound interesting. I would like to understand how these are calculated. Could you perhaps suggest a good up-to-date graduate-level textbook? I have STEM training, no amount of mathematics biologists can come up with scares me.
> "we already know that height can be significantly impacted by birth order including in wealthy European countries, so the family environment is influential"
De-novo mutations also accumulate in parents with age, and a woman's physical health (which presumably has non-zero effects on pre-natal environment) also deteriorates with age. I wouldn't assume 'family environment' is the critical variable here.
Wanted to provide some constructive criticism on this piece given that I'm also quite interested in the nature vs nurture conversation (and am currently working on an article on this topic).
> He singles out one galaxy brain paper, Bingley et al 2023, as evidence that heritability can be made very low and shared environmental influence very high. This paper gets this result by assuming equal avuncular shared environments, in other words, by making the clearly wrong assumption that all differences between the families of cousins are environmental.
This interpretation of Bingley et al 2023 is flatly incorrect. In Equation 17, the authors set the covariance between cousins to be equal to 0.5^(1-z) * 0.25 * sigma_a^2 + sigma_cCgz where z=1 indicates monozygotic twins. The first term in the equation captures the covariance between cousins induced by genetic relatedness.
The assumption that the authors actually make (which Dr. Gusev has pointed out in a separate reply) is that consanguineous aunts/uncles (e.g. an uncle who I'm biologically related to) share the environment with their nieces/nephews to the same degree as their non-consanguineous partners (e.g. my uncle's wife who is not biologically related to me).
> In the first, he keeps making the argument that if a molecular method (RDR) finds lower heritability estimates then the quantitative estimate then the latter is wrong, which, as we saw, is a fallacy.
From reading this passage, it's not obvious to me what the fallacy is. The basic challenge that Gusev is posing to twin study proponents (as I understand it) is this: what explains the discrepancy between heritability estimates produced by RDR and the Classical Twin Design?
- The answer cannot be assortative mating (AM) as both methods are downwardly biased by AM.
I'm genuinely curious to understand your answer to this question as some of the answers that I have heard (such as "imputation quality could undermine RDR") seem specious.
Thank you for pointing this out. I may have indeed misunderstood the Bingley et al model. I'll read the paper again (they don't make too much of an effort to be understood) and correct the post.
My understanding is that RDR is just fancy sib regression (extended to more family members) which is just fancy GREML-GCTA. They all use the quantatitive genetic logic of measuring how phenotypic similarity scales up with genetic similarity after holding the shared environment constant at 0 (GREML-GCTA) or 1 (RDR, sib-regression). If causal genetic variants are not seen (rare or non-SNP) or their effects are nonlinear this might throw off estimates. My confidence on this is low though, not sure why molecular methods don't work better. The fact that there is substantial missing heritability even for well-studied, easy to measure, non-controversial traits like height tells me that these methods miss something. Very heretical thought but I did wonder about the reliability of converting the DNA sequence to digital data.
Turkheimer has made a career from being the one go-to hired gun expert witness who the media as lawyer always calls up to tell the jury that The Science is settled that genes don't matter and that anyone who says otherwise is a lying evil racist. His many false comments about the work of Charles Murray is more than enough to prove he's a snake.
Good effort from the author but these miss the point. Sure, low power makes GWAS less accurate. But Gusev et al focus on GCTA-style studies which use genomic data to estimate relatedness and see if phenotypical similarity scales up with that. Low power doesn't bias these downward, the same way a small twin study doesn't have a downward bias. In the second paper, the author seems to confuse rGE and G*E, so not a great post. I want to write more about those too.
This article assumes that twin studies are a valid measure of the effect of genes and environment without dealing with many well-known criticisms of twin studies. What about gene-environment correlations? Shared genes cause people to create similar environments, which amplify the effects of genetic similarity. You address how gene-environment correlations might be a problem for molecular genetics, but fail to consider how they might affect twin studies. Also, how does this author explain the Flynn Effect? I discuss these problems with twin studies here: https://open.substack.com/pub/eclecticinquiries/p/twin-studies-exaggerate-iq-heritability?r=4952v2&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false
But a big part of the post IS "dealing with many well-known criticisms of twin studies"! In one paper Plomin wrote that active & reactive rGE is just a form of gene expression: some genes code proteins which change you, others make you more likely to seek out trait-changing environments. I don't like this argument because it ignores how bullish such a reality would be for interventions. If smart people are smart because they are (genetically) predisposed to seeking out smartening environments then we can give these to anyone! The fact that this doesn't work - g, for example, doesn't respond to interventions even in the short term - also makes me skeptical about rGE. Just what ARE those people with the right genes doing?
The IQ Flynn effect is mostly just a measurement artifact: https://substack.com/@cremieux/p-161257638. But there is also a Flynn effect for height and other traits which make your point stronger. Heritability doesn't mean genetic determinism, it just means that in a population you can realistically study (e.g. people living in the same country at the same time) genetic effects explain most of the differences between people. Shared environments have a restriction of range (nobody living today in the US grew up like in North Korea or in the 1800s) which makes C effects weaker. I do believe that there are important environmental differences between countries and eras which affect traits, which is why I don't like people talking about African IQs of 60 as if they were an expression of genetic potential. Still, politicians, sociologists etc. have very strong assumptions about C effects visible to ordinary twin studies, such as growing up in poverty or being raised in a certain way, so it still is a big deal if these are low.
Whether you like the rGE argument or not, the point is that there are environmental effects that twin studies can't separate out, which makes them unreliable measures of heredity. If genes tend to build environments that amplify their effect, that amplification is going to show up as a genetic effect in twin studies, but it would seem to be environmental. I quote and link to a long article dealing with this problem in my essay. https://pmc.ncbi.nlm.nih.gov/articles/PMC5754247/
So Plomin apparently thinks that rGE is a genetic effect, whereas others think it's an environmental one. If we can't even agree what counts as environmental and what counts as genetic, what sense does this whole debate have? We would seem to be arguing about indefinite and meaningless concepts.
That article also proves you wrong about interventions. It gives multiple examples of adoption boosting children's IQ score by as much as 40 points.
If the Flynn Effect is a measurement artifact, that means that IQ is a poor measure of intelligence. Hereditarians seem to saving their argument from the Flynn Effect problem by making the whole issue of IQ irrelevant. If all that IQ tests are doing is measuring test-taking skills, then IQ is an uninformative number that isn't measuring anything important about people.
It boosts the _child’s_ IQ. IQ measurement before adulthood has to be adjusted by age. The adult has a potential which the child can reach more or less quickly depending on the environment. The upper limit doesn’t increase. Sadly, the environment _can_ stunt a child.
I don't agree that shared environment effects are more important than hard hereditarians think as Inquisitive Bird has written, but I will discuss why that is in the future.
For what reason is heritability research focused on humans, which show substantial variety even between siblings, not on animals, where variations are small? Are the findings inconvenient to both naturists and nuturists?
You write, "If within-family phenotypic similarity scales perfectly with genetic similarity (MZs are twice as similar as DZs), shared environment is zero."
I don't understand this assumption. Let's say MZ twins have phenotypes that are similar in exact proportion to how much more similar their genotypes are than DZ twins. To me this does not mean the contribution of shared environment is zero.
For example, take a gene for red hair. In the society these people are raised, having red hair is treated in a highly prejudicial way, and this tends to produce a range of reliable reactions in those who are persecuted/abused. It seems perfectly possible for MZ redheads to be more similar in a number of phenotypes than DZ twins due to environmental conditions triggered by a relatively small number of genes. The environment is the same in the way that twin studies define environments, but the way the people are treated is not the same.
This point about redheads can be generalized. Many genotypes may result in a phenotype that triggers divergent social reactions from other people living in the same environments, and thus impact phenotypic outcomes more broadly. I realize that the intention may be to be far stricter in defining a "shared environment" (to include the precise way everyone is treated), but then doesn't this contradict that first sentence above that I quote?
You are right! The scenario you described is an evocative gene-environment correlation (you genetically inherit a characteristic which triggers a causal environmental effect). Quantitative genetic studies assume that this is not happening, if it is and you are considering this as an environmental effect then they are wrong. (But something like this would throw off molecular genetic studies too.) The arguments against rGE are indirect. First, traits are hard to change even we are trying to do it - schools, psychotherapists, traumas etc. have short term effects at best, so it's a good question what others can be doing to you based on heritable visible traits that changes you so much. Second, poor look-alike similarity tells us that physical traits (like with the red hair example) don't do much to influence psychological characteristics, so it's another good question what exactly others are picking up about us which is not visible. Third, evocative rGE would be culture dependent - different cultures value different traits and some none at all - which would make heritability estimates vary a lot across places and eras.
Thanks for the clear and considerate response. And yes, it does seem that the best practical way to look for such "evocative" gene-environment correlations is to look for correlations that have a large swing in different cultures/settings.
The author mentioned, yet did not dispute, my charge that the Minnesota “reared-apart” twin researchers suppressed their DZ-apart control group IQ correlations to find above-zero IQ heritability. The author then cited a Swedish study that defined twins separated at 10 years of age as being “reared apart.” Twin studies are based false-assumptions (the EEA) and p-hacked results. Once the author recognizes this, they will understand why DNA-based studies of behavioral traits and psychiatric conditions have failed miserably.
East hunter, could you respond to this?It claims to refute race and intelligence research https://erikexamines.substack.com/p/iq-race-and-racism Tries to claim that Race realism it's false because the The black White Iq gap is narrowing
You would spend on your argument a little bit and show why he's wrong.Cause i'm not really an expert on this topic Bit of a layman. you address his main argument
Excellent detailed piece. For a lay audience, though, it's missing a So What? Doesn't seem to move the needle from the non expert take that intelligence, neuroticism and other traits with a neurological basis are part nature, part nurture.
Like athleticism. You can be born tall and be nothing but a couch potato.
It is unfortunate that while this post makes more of an effort to engage with modern data than the typical response, it also makes several egregious errors that ultimately undermine that effort.
1. "Gusev claims that “IQ is much less heritable” than height, based on molecular genetic studies with vastly different statistical power to detect genetic effects on the two traits"
This is incorrect. The estimation of molecular heritability does not depend on the power to detect genetic effects at all, it is a single genome-wide parameter aggregated over all variants in the study regardless of whether they are significant.
I explained this point in my first post on IQ with a reference (https://theinfinitesimal.substack.com/p/no-intelligence-is-not-like-height): "But prediction accuracy depends on sample size, could the findings drastically change with more samples in the future? In fact, through the magic of statistics, we actually know that this claim will always to be true. We know this because we have estimated a parameter called molecular heritability, which tells us the upper bound on what a genetic predictor could ever achieve".
And then I explained it again in my second post responding to comments (https://theinfinitesimal.substack.com/i/148251755/isnt-gwas-heritability-only-quantifying-the-mechanisms-we-currently-understand): "Isn’t GWAS heritability only quantifying the mechanisms we currently understand? This is a typical misconception: that GWAS only quantifies the heritability from individual significant associations or genes we understand. In fact, GWAS heritability is defined as the phenotypic variance explained by all genetic variation that has been measured, whether it is significant or not.".
I can explain it a third time, I guess. The population-scale molecular heritability of IQ is 10-20% and the population-scale molecular heritability of height is 40-50%. These estimates come with very low uncertainty, and the fact that 40% is larger than 20% is well established. The inability of this post to grapple with the basic operation of these methods nor their basic estimates should raise red flags about the other claims.
2. "He singles out one galaxy brain paper, Bingley et al 2023, as evidence that heritability can be made very low and shared environmental influence very high. This paper gets this result by assuming equal avuncular shared environments, in other words, by making the clearly wrong assumption that all differences between the families of cousins are environmental."
This is incorrect. Bingley et al. propose a quantitative method to directly test the equal environment assumption, something the author should be pleased to see. Their method assumes equal shared environments between twins and their nieces/nephews as between twin spouses and their nieces/nephews ("To avoid adding parameters, we assume that the degree of environmental sharing between twins and their niblings (Equation 15) is the same as that between twins’ spouses and niblings (Equation 16)"). This does not imply that "all differences between the families of cousins are environmental", it is a constraint *specifically* on the shared environments between these relationship classes. This constraint is clearly much less restrictive than the assumption from classic twin models that MZs and DZs have identical shared environments, which we know they do not. In addition to the Bingley paper, I describe a number of other studies that evaluated the EEA assumption with other extended family designs and found it to be violated (https://theinfinitesimal.substack.com/p/twin-heritability-models-can-tell). Again the author appears not to understand the model but it is confident that the model is wrong because it produces uncomfortable results.
3. "In the first, [Gusev] keeps making the argument that if a molecular method (RDR) finds lower heritability estimates then the quantitative estimate then the latter is wrong, which, as we saw, is a fallacy."
The author first presents RDR as "quantitative genetics with extra frills" but then later discards RDR because it is a molecular method. Well, which is it? What the fallacy is in reporting the RDR results is also never stated. It is odd to simply discount the strongest evidence against one's position as "a fallacy", typically this is the evidence one would addresses first and in most detail. I'll remind readers that the paper that introduced the RDR method is titled "Relatedness disequilibrium regression estimates heritability without environmental bias" (https://pubmed.ncbi.nlm.nih.gov/30104764/) and the entire motivation of the method was to estimate complete additive heritability (using Identical By Descent relatedness estimates) without the environmental assumptions of classic twin models. When applied to a large number of traits, the average RDR heritability was 32%, compared to an average twin heritability of 61% (and a SNP-based heritability of 26%); results that I also summarized in my post, which also goes into how RDR works (https://theinfinitesimal.substack.com/i/148251755/okay-but-why-do-you-think-twin-estimates-so-much-higher). Since RDR and twin estimates are biased to the same extent by assortative mating, this gap is explained by environmental bias in twin studies. So the quantitative genetics (with extra frills) models the author likes also tell us that twin estimates are 200% inflated on average, and that SNP/GWAS estimates are about 80% of the total heritability on average. These relative estimates are consistent with a variety of other studies and theory showing that common variants (i.e. those in GWAS) explain the majority of total heritability, so everything hangs together nicely.
---
TLDR: This post misunderstands how molecular genetic studies actually estimate heritability, then misunderstands how quantitative extended twin models have probed the equal environment assumption and found it violated, then finally does a sleight of hand by pretending that RDR is not a quantitative genetics method because the findings are inconvenient.
Finally, I do appreciate that this post finally acknowledges that behavioral GWAS were greatly inflated by population stratification and, for hereditarians, a disappointment. But I don't think we can simply let hereditarians off the hook, because they were the ones who bet hard on GWAS methods to begin with. Charles Murray predicted for over a decade that molecular studies were just about to prove him right (https://x.com/evopsychgoogle/status/1188441459860443137); Razib Khan argued that we just need to wait a little bit longer for GWAS to fully explain the genetics of intelligence (https://undark.org/2017/02/28/race-science-razib-khan-racism/); Steve Hsu predicted that GWAS would explain 60% of the total variance of intelligence by 2016 (https://x.com/SashaGusevPosts/status/1613724821480677378); Emil Kirkegaard has been running bogus polygenic score comparisons pretending stratification is a non-issue and getting the results completely backwards (https://theinfinitesimal.substack.com/p/how-population-stratification-led). Even setting aside motivated hereditarians, here's what internet celebrity Gwern said about the first GWAS of educational attainment: "Reading it was a revelation. The debate was over: behavioral genetics was right, and the critics were wrong. Kamin, Gould, Lewontin, Shalizi, the whole sorry pack—annihilated. IQ was indeed highly heritable, polygenic, and GWASes would only get better for it" (https://www.lesswrong.com/posts/sRchPdp6mCqY2ekJX/what-progress-gwern-s-10-year-retrospective). As we now know, this is not what actually happened.
And this ultimately gets at my issue with the modern hereditarian movement, my point is not to attack their *theories*, everyone is entitled to speculate about genetic processes and test out their speculation. What I am attacking is the pattern of overfitting and p-hacking in their *methods*. When GWAS provided a cudgel to beat up on Lewontin and Shalizi, then suddenly it was "a revelation" that will only get better and better. Now that GWAS falsifies these claims, it is deemed "something of an embarrassment". When the RDR method is first described, it is just "quantitative genetics with extra frills" (laudatory) but when it yields uncomfortable results, it is now "molecular genetics" (derogatory) and using it is "a fallacy"". Methods and findings are not used to understand how the world works, but to stitch together a narrative that supports the author's predetermined conclusion.
I take multiple issues with your comment on empirical grounds. I’ve followed some of your work since ~2016 give or take, but abandoned it later on as it seems you have switched to untenable positions without much evidence and I cannot really understand why (e.g., old acknowledgments that methods controlling for stratification fail to model it very well, are all linear etc but current support for PCs as a a proper stratification control). I believe this position that the lowest SNP_h2 estimates one can squeeze out are the most accurate ones to be extreme at best, and in-line with Bouchard’s (1982) description of what he coined “pseudoanalysis” at worst. I wouldn’t go far beyond this since I have not seen you post sources which contradict what you say about them or seen you lie (maybe some people will debate this), but I will go on to discuss what I think you get wrong often.
1. The fact that molecular h2 isn’t based on per-SNP significance doesn’t imply these are the “upper bound estimates we could ever achieve”. The estimates still only capture the variation tagged by common SNPs in LD. Painting this as if it has settled the debate is misleading, for reasons I will also get to below this point. These methods still have plenty of unidentified rare & structural variants as well as nonadditive genetic effects which are incredibly unlikely to be included in molecular designs anytime soon (goes against the whole point). But you already know all of this.
2. The Bingley method is arguably worse as it assumes equal environmental influence from twins and from the twins’ spouses on a niece/nephew (which bakes in indirect environmental sharing in a convoluted way). Whether the EEA holds is a matter relevant to the trait you are analyzing. You cannot generalize whether or not the EEA is practically true since a trait-relevant environment could be a trait-irrelevant one for a separate characteristic. For intelligence specifically, *g* is not affected by the commonly cited environmental variables alleged to violate the EEA (not adoption; not education, your post on the Ritchie paper notwithstanding; not prenatal effects and so on and so forth), so the EEA holds (and see Barnes et al., 2014; EEA violations do not invalidate the estimates empirically). Therefore high twin_h2 estimates for *g* (as in Panizzon et al., 2014) are valid wrt the EEA and within-family designs even moreso as they avoid other issues you allege.
3. You cannot conclude that since AM affects RDR and twin methods to the same extent, it must be the case that twin results are inflated by environmental variance. That’s a clear non-sequitur provided that there are other sources of variation (including modeling assumptions, like assuming away indirect effects under idealized conditions, as RDR does, thus underestimating h2 when they are violated) which explain the discrepancies. RDR estimates are also known to be less stable for traits with lower signal-to-noise (like intelligence). You have also been made aware of other methods that remove twin assumptions and estimate heritability to be high (Schwabe, Janss & van den Berg, 2017) to which you have not provided adequate criticisms that would explain discrepancies as high as >70%, such as assumptions regarding distant-relatives or C being zero (this would still put the heritability at >60% assuming both of these are addressed).
The hereditarian predictions you cite seem to have been wrong in a few ways, but the “fallacy” the OP describes was never citing molecular results. It was declaring them to be the final evidence and pretending they capture virtually all of the total heritability. So far I know of at least two papers which were able to attribute ~20% of the B-W gap in *g* to common associated SNPs. The argument is far from settled. But family-based estimates are still strong evidence and pedigree-based data has been shown to be able to bridge the “missing heritability” gap between methods (as in Hill et al., 2018) with h2 estimates reaching values >.5 even with bad assumptions (like that AM is totally environmental, as in the Hill paper). Where are the “biases” here which other methods capture? Where is the p-hacking? The “overfitting”, whatever that means? I hope it has been made clear that the “hereditarian” position (which, in my opinion, is an agnostic one) has way more ground than you are willing to admit.
I know I'm not Dr. Gusev, but I wanted to issue a response to this reply as it betrays several critical misunderstandings.
> The fact that molecular h2 isn’t based on per-SNP significance doesn’t imply these are the “upper bound estimates we could ever achieve”
To my knowledge, Dr. Gusev has never stated that SNP-heritability constitutes an upper bound on narrow-sense heritability. He specifically states that molecular heritability "tells us the upper bound on what a genetic predictor could ever achieve" where the term "genetic predictor" refers to *polygenic scores*. This context is made abundantly clear in the post that Dr. Gusev linked.
https://theinfinitesimal.substack.com/i/148059447/iq-is-much-less-heritable-and-more-confounded-than-height
Why is Dr. Gusev talking about polygenic scores? Because East Hunter seems to be under the impression that, if genomic variants do not achieve statistical significance, they are not included in SNP-heritability. But this statement is incorrect. While non-significant variants would not be included in a polygenic score, they *would* be included when estimating molecular heritability. Thus, molecular heritability represents an upper bound on the predictive accuracy of polygenic scores.
> You have also been made aware of other methods that remove twin assumptions and estimate heritability to be high (Schwabe, Janss & van den Berg, 2017) to which you have not provided adequate criticisms
I'm happy to provide adequate criticism here.
Unlike RDR which leverages exogenous genetic variation (i.e. random inheritance of genetic variants), the paper you cite employs a pedigree-based method where, in the authors' own words, the "model is statistically equivalent to the ACE model for twins, except that the common environmental component C is assumed to be zero". In other words, this model relies on a statistically equivalent set of assumptions as twin studies *plus* the assumption that shared environment makes zero contribution to IQ. Thus, your statement that this study "removes twin assumptions" is unequivocally false.
Additionally, note that RDR makes no such assumption while still exploiting degrees of genetic relatedness across relative classes. In other words, RDR does exactly what the study you cite does *except* that it leverages an actually exogenous source of genetic variation as opposed to sources of genetic variation which are correlated with environmental variation.
> But family-based estimates are still strong evidence and pedigree-based data has been shown to be able to bridge the “missing heritability” gap between methods (as in Hill et al., 2018) with h2 estimates reaching values >.5 even with bad assumptions (like that AM is totally environmental, as in the Hill paper)
Hill et al is a pedigree-based method which both Dr. Gusev and Dr. Alexander Young (the lead author on the RDR paper) have responded to numerous times. Here is Dr. Gusev's response (which also includes Dr. Young's response):
https://theinfinitesimal.substack.com/i/148251755/what-about-kinship-studies-why-do-we-need-to-control-for-relatedness
The basic problem, once again, is that the paper does not leverage randomization to disentangle genetic and environmental processes, and as a result, is not a suitable alternative to RDR.
To make an analogy, your citation of pedigree-based methods is analogous to throwing away the results of an RCT because they don't agree with observational methods. I hope you can agree that no scientist worth their salt would do such a thing, all else being equal between the study designs.
No, I’m sorry but this is a LARP and not a serious response. My post contains no misunderstandings outside of an extremely debatable semantic point regarding genetic predictions. You could have simply read what I typed and cited more carefully in order to avoid confusing yourself and others. But you chose not to, for some reason.
> refers to polygenic scores
Fine. I won’t even ask why he presents molecular results the way he does (obvious implications) or why he compares them to twin results right after this point and in other posts repeatedly. The statement is still just as wrong. There is no sensible way in which molecular results could be the “upper bound” of PGS since the former (1) literally informs the latter and (2) variance components could differ from within-group SNP_h2 to between-group_h2 across PGS group comparisons. This is only NOT a non-sequitur if you assume that molecular_h2 is not going to increase and better inform PGS, and it doesn’t even work because assuming that this is the case, PGS could still show higher heritabilities in population comparisons.
> I'm happy to provide adequate criticism
You know the funny part about this is that you stopped your quote of my statement right before I said “… that would explain discrepancies as high as >70%”. Just bringing this up because below it should be clearly demonstrated that none of the criticisms you allege explain anywhere close to that amount.
> In other words, this model relies on a statistically equivalent set of assumptions as twin studies *plus* the assumption that shared environment makes zero contribution
You say “in other words” and then proceed to completely misinterpret the paper. When the authors say that their *model* is statistically equivalent, it implies approximately equal fit, not similar assumptions (what in the world would it even mean for an assumption to be “statistically equivalent”?) which should have been clear from the title of the paper alone anyway. Why would the authors publish a paper attempting to validate twin studies using twin study methodology? They obviously apply different assumptions (no EEA, no MZ/DZ comparisons etc). You should know this provided that you acknowledged the method as a pedigree-based one, which literally requires a different kind of analysis to calculate heritability.
> note that RDR makes no such assumption
What assumption? You mean C being zero, which it basically is by adulthood? Again, I have to ask, did you read the paper? The authors are very clear about their modeling of C changing nothing about their heritability estimate:
> Note however, that the shared environmental component was rather small (8 and 16%, respectively) […] this would decrease heritability from 94% (that includes shared environmental variance) to around 77–85%, which would mean that the result is not at odds with the results found by earlier twin studies. Put differently, if we acknowledge that we can only estimate A + C together in our analysis (that is, in the empty model), and when we compare that to the A + C estimates based on the twin studies using the exact same birth cohorts, the results are rather similar: 94% here, 74 + 8 = 82% […] and 66 + 16 = 82%.
& tangentially:
> When we compare these heritability estimates to the estimate of 85% in this study, we can conclude that the high estimates resulting from the twin method are not simply an artifact of self-selection or because of any important difference between twins and singletons. Twin-based heritability estimates are not inflated, since an estimate based on a sample from the entire population (including twins and singletons) is even higher.
—
—
Continuing.
> RDR does exactly what the study you cite does *except* that it leverages an actually exogenous source of genetic variation as opposed to sources of genetic variation which are correlated with environmental variation
No, it does exactly what the study I cited does except that it assumes variance components with zero evidence. RDR is basically a method of “what if X variables affect heritability” materialized. What I’m citing is instead “which variables affect heritability”. The latter modeling is real, RDR’s modeling is just assuming h2 inflation. If your issue is rGE (you say “sources of genetic variation which are correlated with environmental variation”) then what I cited should not be a concern, since according to multiple different papers, rGE has almost no impact on heritability estimates (empirically; as stated in Conley, 2014 for example, or Willoughby et al., 2022 which found rGE was ~0.02 and an h2 of ~.62 for the ICAR-16 + >.53 for IQ in general; table S12, see here; https://imgur.com/a/yQu4UF1)!
> The basic problem, once again, is that the paper does not leverage randomization to disentangle genetic and environmental processes
Yes it does. If you had read the paper and the alleged criticisms you would probably know that ~0.027% of the sample comprised of related individuals according to table 1. The Hill paper is much closer to an RCT than RDR, which uses a dense family structure. And then you need to point to biases which *are there* (not just assumed to be there) which would ***massively*** inflate the estimate. To steelman your case a little, Gusev’s criticism of this is that assortative mating and indirect genetic effects might be biasing the results upwards, but again, the paper assumes all of the variance attributable to AM is environmental in origin, and almost none of the sample is related to any significant extent. The former modeling choice is obviously untrue, and does two things. For one, it underestimates heritability. For two, it gets around Gusev’s criticism. Young also provides no criticisms of the paper and instead simply claims that the results “could be” upwardly biased in the presence of indirect genetic effects. Really, that doesn’t tell me anything regarding what is the case empirically. I can only conclude from this that you either have bad judgment or you simply did not read/understand the responses, because none of these are valid criticisms or explain the other results (e.g. Schwabe paper, Panizzon paper &c).
Lastly, assuming certain variance components exist, like RDR does, is not the equivalent of an RCT. Saying “here I present you with heritability estimates which I will now model as being biased by effects I have not demonstrated are there” is, to me, the equivalent of making things up. Do you notice any of the methods I cited doing this? No. RDR is hypothetical. My methods are empirical. That’s the difference.
So no, to wrap this up, I have yet to be provided with criticisms that would invalidate the results I cited where h2 > 60%.
> Fine. I won’t even ask why he presents molecular results the way he does (obvious implications) or why he compares them to twin results right after this point and in other posts repeatedly.
I'll leave this for Dr. Gusev to respond to since it's a point about authorial intent.
> You say “in other words” and then proceed to completely misinterpret the paper. When the authors say that their *model* is statistically equivalent, it implies approximately equal fit, not similar assumptions ... Why would the authors publish a paper attempting to validate twin studies using twin study methodology? They obviously apply different assumptions (no EEA, no MZ/DZ comparisons etc).
First, the authors do actually assume the EEA because they set the shared environment component equal to zero across all modeled relative pairs. Thus, all modeled relative pairs share the trait-relevant environment to the same degree - zero.
Second, what I mean by "statistically equivalent" is that the set of assumptions which apply to twin studies are just extended to more degrees of relatedness (e.g. no gene-environment interactions, no gene-environment correlation, no assortative mating, EEA).
Third, you ask, "why would the authors publish a paper attempting to validate twin studies using twin study methodology?" and in a twist of irony, I think you've arrived at the answer yourself. When the authors acknowledge that their model assumes C=0, what do they do? Retreat back to twin studies to say that C is negligible!
The whole paper reeks of this kind of circularity in my opinion. They use a statistically analogous set of assumptions as a twin AE model just applied to a larger familial structure and then pretend that this allows one to validate the underlying assumptions of twin studies which are actually in contention.
> What assumption? You mean C being zero, which it basically is by adulthood? The authors are very clear about their modeling of C changing nothing about their heritability estimate:
Again, the authors traffic in circularity here. They claim to develop a methodology to "validate" twin studies and then, when pressed on the limitations of their own methodology, they just fall back to twin studies.
The reason they have to do this is stated in the manuscript:
"we have not taken into account whether sibling pairs and half-sibling pairs actually share the same home environment, nor whether cousins share the same household (could be true for some families), nor whether nonrelated children share the same household. We were therefore not able to estimate non-genetic variance that is shared due to being brought up in the same home environment"
My point is that RDR does not do this because it leverages the random inheritance of genetic variants (and thus, the relevant genetic influences are independent of environmental sharing).
> No, [RDR] does exactly what the study I cited does except that it assumes variance components with zero evidence. RDR is basically a method of “what if X variables affect heritability” materialized.
No, RDR merely *allows for* the existence of additional variance components. It does *not* assume that such components are non-zero. If the data empirically indicates that this is not the case, then the corresponding variance component will be set to zero.
The best way to see this is Supplementary Table 2. In a simulation where the trait in question is purely determined by additive genetic variation + noise (i.e. no genetic nurture). RDR correctly estimates 0 correlation between the environmental and parental genetic components of the phenotype.
> If you had read [Hill et al 2018] and the alleged criticisms you would probably know that ~0.027% of the sample comprised of related individuals according to table 1.
Yes, the sample consists predominantly of genetically-unrelated individuals but the percentage of heritability attributable to the sample of related individuals (h^2_kin / h^2) is relatively high (>50% in the case of general cognitive ability - see Figure 1). In other words, it is this narrow family-based slice of the sample that drives up the estimated heritability, and that's exactly what Dr. Gusev and Dr. Young take issue with.
While the authors state that they are able to replicate this result in the sample of fully unrelated individuals using GREML-MS, Evans et al points out that GREML-MS "requires that the same assumptions of GREML-SC hold within each [genomic relationship matrix]" and that GREML-SC assumes that "genetic similarity is uncorrelated with environmental similarity". In other words, GREML-MS relies on a stronger set of assumptions than RDR.
https://pmc.ncbi.nlm.nih.gov/articles/PMC5934350/
> The Hill paper is much closer to an RCT than RDR, which uses a dense family structure.
I disagree with this statement. I hate to beat the same drum again, but RDR leverages the *random* inheritance of IBD segments within relative pairs to estimate heritability. In other words, if I randomly inherit an IBD segment from my grandfather and my cousin doesn't, that is a difference in our genetic code that is, statistically speaking, independent of our environments and can lead to downstream differences in our phenotypes. Extrapolate that across N relative-pairs, and you get a heritability estimate that is based on exogenous genetic variation.
Conversely, Hill et al is not guaranteed to leverage purely random genetic variation. While both RDR and Hill et al involve analyses of pedigrees, only RDR leverages portions of the genome which are randomly inherited. Hill et al provides no such guarantee AFAIK. If I'm wrong about this, I'd love to see a proof explaining why.
> Lastly, assuming certain variance components exist, like RDR does, is not the equivalent of an RCT. Saying “here I present you with heritability estimates which I will now model as being biased by effects I have not demonstrated are there” is, to me, the equivalent of making things up. Do you notice any of the methods I cited doing this?
It's true that the methods you cite don't do this but neither does RDR. Just as the adoption study you cite *allows for* the existence of rGE, so too does RDR.
> the authors do actually assume the EEA because they set the shared environment component equal to zero
That’s not the same thing. The EEA is there so C and all other components are estimable. Otherwise, there would be no point for ACE models in twins. I’m not sure why you said this.
> Retreat back to twin studies to say that C is negligible
Unless you have a criticism of twin studies and multiple replications of the Wilson Effect and C being negligible across multiple family and within-family studies &c &c, this is not relevant. As was stated in the paper, this made little difference to their results. If you care to propose a C value you think would be plausible, do so. Even something as high as 30% still puts the h2 at >60-70% depending on a few other controls as well. The authors justified this pretty strongly. Please go and read the paper more carefully.
> The whole paper reeks of this kind of circularity
Evidence-based inference based on other independent modelings of C is not “circularity”. It is justified extrapolation based on cross-study convergence, which - again - made little difference to the results regardless of the magnitude (across plausible ranges).
> It does *not* assume that such components are non-zero
It assumes no indirect genetic effects despite its goal not having been to capture just the additive variance structure, which is also underestimated by RDR if it omits higher-order interactions. The most obvious assumption it hinges on (which is indeed the motivation behind the method) is that shared environmental effects are orthogonal, i.e. that these effects don’t vary with deviations in genetic relatedness. We know from many papers using reared-apart twins, heritability method comparisons and other designs (Lee, 2009; Visscher, Yang & Goddard, 2010; Beaver et al., 2014; Conley et al., 2014, see my original comment etc) that shared environment does not in fact track closely with true relatedness (and see Willoughby & Lee, 2010 as well as the replications, like by Kandler et al., 2016; Bates et al., 2018 &c, showing that parental PGS adds nothing after conditioning on child PGS). But even worse, if shared environment is correlated with IBD *even slightly* (which it is), RDR is even more biased downwards! It makes no sense to prefer this method over the one in the Schwabe paper.
> where the trait in question is purely determined by additive genetic variation + noise (i.e. no genetic nurture)
Only additive variance + noise doesn’t mean no genetic nurture. It ignores literally all broad and indirect genetic effects.
> the percentage of heritability attributable to the sample of related individuals (h^2_kin / h^2) is relatively high (>50% in the case of general cognitive ability - see Figure 1 […] GREML-MS relies on a stronger set of assumptions than RDR
What? This is not seen in Figure 1, I have no idea what you’re talking about. You’re claiming that over 95% of the estimated heritability in the Hill paper is significantly biased by less than 0.1% of the sample? This is an insane inquiry and requires evidence. Based on the papers I cited above, the idea that genetic similarity is uncorrelated with environmental similarity seems to be justified and not stronger than the assumptions RDR makes, so the GREML-Ms estimate in the Hill paper and the subsequent 2017 analysis is largely unbiased and probably not “biased” by the related individuals, but obviously driven by them due to the broader capturing of genetic effects.
> RDR leverages the *random* inheritance of IBD segments within relative pairs to estimate heritability
Yes, but the point of this is that within-family variation is generally immune to AM, stratification etc. But again, so are within-sibling GREML designs, adoption studies, within-family GWAS and so on and so forth. It seems like special pleading to ignore all other evidence in favor of RDR if your concern is random segregation.
> Hill et al provides no such guarantee
It doesn’t need to but the related sample is really small to be biasing anything *too* significantly.
Again, to close off, the papers I cite finding high heritabilities don’t look like they are invalid. You have to propose stronger arguments and solutions to issues to discard the past 100 years of heritability research.
> That’s not the same thing. The EEA is there so C and all other components are estimable. Otherwise, there would be no point for ACE models in twins. I’m not sure why you said this
I think it's worth defining our terms here. The Equal Environments Assumption (EEA) stipulates that all modeled relative pairs share the trait-relevant environment to the same degree. For example, in the Classical Twin Design (CTD), the EEA entails that MZ and DZ twins share the trait-relevant environment to the same degree.
Now, the question we must ask ourselves is this: does assuming C = 0 (as Schwabe et al do) necessarily entail assuming that all modeled relative pairs share the trait-relevant environment to the same degree? The answer is yes. Just as the CTD assumes C_mz = C_dz, Schwabe et al assume that C_bio-sib = C_half-sib = ... = 0. In other words, the assumption that C = 0 necessarily entails the EEA.
Why do I bring this up? Because your initial claim was that Schwabe et al "removes twin assumptions" when it does no such thing. It merely extends the EEA (and other CTD assumptions) to a larger pedigree with the additional assumption that C is specifically equal to zero.
> [RDR] assumes no indirect genetic effects
Is this a typo? Because otherwise, this statement is unequivocally false. The v_e~g variance component estimated by RDR captures the variance driven by the environmental component of the phenotype which is correlated with paternal genotype (thus allowing for indirect genetic effects).
> despite its goal not having been to capture just the additive variance structure, which is also underestimated by RDR if it omits higher-order interactions
Could you clarify what you mean here? If you're talking about dominance or epistasis, then that is not relevant because Young et al demonstrates in simulations that RDR will recover the true narrow-sense heritability even in the presence of dominance and epistasis (see Supplementary Table 2).
> Only additive variance + noise doesn’t mean no genetic nurture. It ignores literally all broad and indirect genetic effects.
I realize I wasn't clear with my description of the simulation, so allow me to refer you to the caption underneath Supplementary Table 2:
"We simulated 500 replicates of each trait based on actual Icelandic genetic data for 10,000 individuals. Ten thousand SNPs with median frequency 23% were given additive effects for all the traits other than the rare SNPs trait, for which 2,200 SNPs with frequency between 0.1% and 1% (median 0.26%) were used. The true (narrow-sense) heritability of each trait was 40%. To this additive genetic component, only noise was added for the additive trait and the rare SNPs trait"
Young et al then describe various modifications they make to this simulated trait such as the introduction of genetic nurture:
"For the ‘genetic nurturing’ trait, the genotypes of the parents were also given effects to simulate ‘genetic nurturing’ effects"
How does RDR perform under such simulations? Well, as Supplementary Table 2 highlights, RDR does not estimate superfluous variance components as you implied. For the trait w/o genetic nurture (labeled "additive" in the table), v_e~g was correctly estimated at ~0.
> But even worse, if shared environment is correlated with IBD *even slightly* (which it is), RDR is even more biased downwards!
Ok, I'm genuinely not understanding what point you're making here or how you arrived at this conclusion. Earlier in your reply, you state:
"We know from many papers ... that shared environment does not in fact track closely with true relatedness"
But now, you are saying that shared environment *does* actually track relatedness ("if shared environment is correlated with IBD even slightly which it is") and somehow, this fact only biases RDR but not the other methods you cite even though RDR specifically only leverages relatedness that is *random* (i.e. independent of environmental influences) while the pedigree-based methods you cite make no such guarantee. At worst, your train of thought is completely contradictory. At best, the message is just getting garbled somewhere.
> What? This is not seen in Figure 1, I have no idea what you’re talking about.
Earlier, you accused me of not having read Hill et al. Now I must accuse you of the same. In the very first paragraph of the "Results" section, the authors state:
"For g, common SNPs (h2g) explained 23% (SE = 2%) of the phenotypic variation. Pedigree-associated genetic variants (h2kin) added an additional 31% (SE = 3%) to the genetic contributions to g, yielding a total contribution of genetic effects of 54% (SE = 3%) on g"
Note that 31/54 > 0.5. Hence, greater than 50% of the estimated heritability is driven by the pedigree-based sample. The authors re-emphasize this point in the "Discussion" section:
"The pedigree-associated genetic variants accounted for over half of the genetic effects in these phenotypes."
> [Hill et al] doesn’t need to but the related sample is really small to be biasing anything *too* significantly.
Your continued insistence on the small size of the pedigree-based sample is a complete non-sequitur. As the authors repeatedly point out, the pedigree-based sample makes a large contribution to the heritability of general cognitive ability in their sample.
Thanks for replying to the post! Braeann Danads said many things I wanted to stay. You raise very valid points but I disagree that these are "egregious errors" that undermine the point of my essay, which is that you can't use molecular genetic methods to falsify quantitative genetic findings of high heritability. 1) I was mainly thinking about within-sibship GWAS when I mean that different power can result in different heritability but I see now how this is not 100% clear. You are of course right that in an SNP heritability study (not a GWAS) low power should not bias effect sizes downward, although it should affect precision. PGS heritability should be affected by low power in the original GWAS though, even if you use all variants and not just the significant ones, because the effects of individual SNPs are not estimated precisely which affects the quality of the PGS. 2) You are probably correct that I got the Bingley et al model wrong (Unboxing Politics also left a comment about this). I'll read the paper again and update the post. 3) In the essay, I classified methods as quantitative or molecular based on the data they use (pedigree or SNPs). This makes RDR a molecular method, which I think I make clear. RDR is interesting but it still comes with the typical assumptions of molecular genetic studies.
Just to be clear, I also made the point that PGS accuracy is sample-size dependent and specifically advocated for focusing on heritability estimates for that exact reason ("But prediction accuracy depends on sample size, could the findings drastically change with more samples in the future? In fact, through the magic of statistics, we actually know that this claim will always to be true." ~ https://theinfinitesimal.substack.com/p/no-intelligence-is-not-like-height). I suspect part of the problem is outsourcing these critiques to the Bronski post, which makes many basic mathematical errors (some pointed out in Bronski's comments). Bronski read my explanation, proceeded to ignore it to make his arguments, then this worked through the broken telephone as a "for more, see:" reference in multiple response articles (yours and another in the quant racist journal Aporia). I don't know of this was intentional, but it has produced a tangled web of superficial "dunks" that never actually spell out what they're criticizing.
Regarding RDR (and your comments elsewhere), it is simply not the case that RDR is fancy Sib-Reg and/or fancy GCTA/GREML, though it shares some superficial aspects with those methods. RDR (like Sib-Reg) uses IBD segments to capture (nearly) complete additive heritability, but since it uses trios rather than siblings it is not biased by sibling indirect effects or interactions. RDR (like GCTA/GREML) uses variance components but because it is exploiting the random Mendelian transmission within families it is not biased by parental indirect effects or shared environment. Thus, RDR -- like the paper title says -- is able to probe the assumptions of twin studies without introducing new assumptions of its own. This is really fundamental to triangulating the estimates across these methods and the reason I have written so much about them (http://gusevlab.org/projects/hsq/#h.b480s3ce0s08).
I think it's wrong to get bogged down in the mathematical minutiae of statistical genetic models because we are missing the forest for the trees. As I wrote in the post: genetically identical people raised by different families end up very similar, while genetically related people raised by the same family don't. This is highly ecologically valid evidence based on minimal assumptions so we should weigh it strongly. RDR and other molecular methods are full of bells and whistles - you are not just testing if family members who end up slightly more or less genetically similar due to random meiotic forces are phenotypically similar, but you are trying to precisely quantify the relative importance of genetic forces based on the slope of this relationship! This is a neat idea but I'm not seeing how a result discordant to twin or adoption studies is evidence that the latter are wrong. I think you are wrong to respond with snark to Candide, he raises a point I agree with: the deflation of height heritability tells us that RDR misses something. If the twin literature is wrong about this and heritability is only 55%, then we are not only saying that shared environment accounts for about a third of family clustering of height - let's say that's possible - but also that MZs have unique family environmental experiences which make them more similar than genetics would dictate. Maybe height increases if the family eats well (height Flynn effect and all) but is it plausible that parents e.g. feed MZ twins so much more similarly to DZs that this ends up being half as important as genes?
If the goal of your post was to argue that we should only use evidence from adoption/raised apart studies and ignore any other inconvenient findings, that would have been a much shorter (and more honest) post! But that's not what you wrote. Your argument was that these models were fundamentally flawed and produced erroneous results. Now that they have been shown not to be flawed, at least in the ways you outlined, the reasonable thing to do is to correct your post and say "whoops, I guess I really misunderstood what was going on here, let me integrate this new information into how I think about the world and come up with an updated conclusion". Instead, the response appears to be "let's just ignore the findings from these models because they are inconvenient and focus on the findings that I do like". So we are back to exactly the criticism that I laid out at the beginning of my very first comment: modern hereditarianism is an overfitting exercise and now that it has been caught flat-footed making poor predictions, it is rapidly discarding the methods and findings that don't fit the narrative.
Nothing has been shown not to be flawed. My point was that assumption-free quantitative models (adoptees, MZAs, maybe also twins after so many assumption sanity checks) are more parsimonious than molecular genetic studies so they cannot be dismissed based on the latter. I stand by this point. I realized that our main contention is that you treat studies based on genome-based relatedness (RDR, sib-regression, GCTA) as exactly analogous to traditional pedigree studies. I don't, the fact that the former have lower estimates even for e.g. height tell me that this model misses some important aspect of relatedness that is important for phenotypic similarity. I'm not pretending to know what it is so this is just speculation: missing variants, non-linear effects, G*G, data errors. This is a much more parsimonious explanation of discrepant findings than that C>A despite all evidence to the contrary.
I think we're converging to the point. On the one hand we have Bingley et al. (making fewer assumptions than twin studies) and RDR (making no assumptions about environment at all). Both estimated in large, modern data sets and neither exhibiting the original flaws you had argued for (which, I'll just note again, really should be corrected in your article).
On the other hand we have (1) your hunch that height heritability of ~70% (after correcting for assortment) is too low; and (2) twin adoption studies that assume there is no selective placement, no GxE, no range restriction, no prenatal effects, and largely consist of bespoke analyses from the time of the replication crisis.
Well, okay, let's look at the adoption studies. First, the MISTRA estimated an MZA correlation of 0.62 and a DZA correlation of 0.50 for IQ (not sure why you omitted this from your post). In the classic ACE model, that produces a heritability of 24% and a selective placement effect of 38%! In an AD model it produces nonsensical estimates. Given the massive selective placement, even the 24% heritability estimate doesn't really tell us anything. Second, the SATSA estimated a DZA correlation of 0.32 which was paradoxically higher than the DZT correlation of 0.22; either implying that living together causes twins to be more different or, you guessed it, selective placement for the DZAs. When al relationship classes are fit, they also produce nonsensical estimates (Table A1): 0% additive genetic variance, 80% dominance genetic variance, and -38% non-shared environment -- that's right, a negative non-shared environment variance. Large selective placement effects (Ec) were also observed for multiple individual subtests (Table 3). In other words, the results are a complete mess. The model can't fit the observed correlations parsimoniously, so it introduces nonsensical negative variances to make it work, arbitrarily setting other paths to zero.
You are advocating that we put our faith in low quality observational data instead of genetic randomization and the observational data plainly violates your methodological assumptions to boot. Is it fair to consider these studies an embarrassment yet?
Call me stupid, but I just don't see which environmental effects could inflate heritability of height in Iceland from the 55% found in the original RDR paper to the standard ~80%. Iceland is not a poor country with inadequate nutrition and high wealth inequality - it is a rather rich country with superb nutrition and one of the most equal in Europe, one of those weird places where prime ministers ride to work on the common bus. Is it exercise equipment in the home, perhaps? (/s) The original RDR paper modeling (Table 1) gives 'maternal environment' and 'genetic nurturing' as examples of 'environmental effects' for which kinship methods significantly overestimate heritability. But if these effects are truly non-genetic, it would have been odd for them to cut off at or soon after birth, and if that was the case, then adoption studies would not have given similar heritability results to other methods.
Good question. People around here seem to have very well honed intuitions for what the *right* heritability should be for certain traits, I guess they are able to look around at all the twins they know and eyeball that 55% is just too low but 80% is just right for height -- I wish I had this ability. In terms of environmental influences, we already know that height can be significantly impacted by birth order including in wealthy European countries, so the family environment is influential and we also know there have been large secular changes in height in Europe as well (a kind of Flynn effect) so broader environmental influences can be involved too (I'll leave it to others to counter-argue that the increase in height is actually hollow because it is not on "h" the general height factor).
Setting environment aside, RDR and the classic twin design are both biased down by assortative mating; so while assortative mating will not explain the gap between RDR and twin estimates, it might explain why the RDR estimate feels low to you personally. It is possible to do a back-of-the-envelope correction for AM based on the observed mate-pair correlation (see: http://gusevlab.org/projects/hsq/#h.st55vexg74ep for details). For an RDR estimate of h2 = 55% and a mate correlation of 0.3 (taken from Horwitz et al. ; https://pubmed.ncbi.nlm.nih.gov/37653148/), the corrected h2 is 70%. Not quite the 80% you were aiming for but hopefully that helps you feel a bit more comfortable with RDR (or maybe the mate correlation was 0.4 in Iceland, the corrected estimate is 80% and everyone is happy)? On the other hand, if we take the twin estimate of 81% and a mate correlation of 0.3, the corrected h2 is ... well it is 139%. I know I said earlier that I can't eyeball heritabilities, but 139% still seems implausible. So that's something else for twin study enjoyers to think about.
With all respect, you are assuming things about a person you're seeing for the first time on the sole evidence of "place of observation". Isn't that stereotyping? I am not "aiming for" any specific value. I note a large discrepancy between standard results from older literature, which even someone as uninformed as myself has happened upon in my reading, and results from a relatively new method (RDR), a discrepancy that the original RDR paper thought worth remarking upon in its discussion section (it refers to estimates from Swedish twin studies as the closest available comparison group). I am asking questions because I do not have an intuition for what the "right" heritability is for a continuous trait like height, except that values too close to 0% and 100% feel wrong to me on general principles. Your statements about mate correlation sound interesting. I would like to understand how these are calculated. Could you perhaps suggest a good up-to-date graduate-level textbook? I have STEM training, no amount of mathematics biologists can come up with scares me.
I linked to a description of how to calculate these corrections, which provides the derivations as well as the original sources:
http://gusevlab.org/projects/hsq/#h.st55vexg74ep
specifically the Border and Kemper papers cited. That primer also includes discussions of most contemporary heritability methods.
I'm not aware of a textbook that has integrated findings this recent, but Lynch & Walsh (https://global.oup.com/academic/product/genetics-and-analysis-of-quantitative-traits-9780878934812?cc=us&lang=en&) is a generally good quantitative genetics text, although quite dense, that would give you a very solid foundation for the field up through 1998.
> "we already know that height can be significantly impacted by birth order including in wealthy European countries, so the family environment is influential"
De-novo mutations also accumulate in parents with age, and a woman's physical health (which presumably has non-zero effects on pre-natal environment) also deteriorates with age. I wouldn't assume 'family environment' is the critical variable here.
Nice post. But I'm a bit bummed because I have been writing a similar piece for some time...
I guess I have to make it better now.
Well Gusev has some comments on this post: so incorporating those into your post would do wonders
Wanted to provide some constructive criticism on this piece given that I'm also quite interested in the nature vs nurture conversation (and am currently working on an article on this topic).
> He singles out one galaxy brain paper, Bingley et al 2023, as evidence that heritability can be made very low and shared environmental influence very high. This paper gets this result by assuming equal avuncular shared environments, in other words, by making the clearly wrong assumption that all differences between the families of cousins are environmental.
This interpretation of Bingley et al 2023 is flatly incorrect. In Equation 17, the authors set the covariance between cousins to be equal to 0.5^(1-z) * 0.25 * sigma_a^2 + sigma_cCgz where z=1 indicates monozygotic twins. The first term in the equation captures the covariance between cousins induced by genetic relatedness.
The assumption that the authors actually make (which Dr. Gusev has pointed out in a separate reply) is that consanguineous aunts/uncles (e.g. an uncle who I'm biologically related to) share the environment with their nieces/nephews to the same degree as their non-consanguineous partners (e.g. my uncle's wife who is not biologically related to me).
> In the first, he keeps making the argument that if a molecular method (RDR) finds lower heritability estimates then the quantitative estimate then the latter is wrong, which, as we saw, is a fallacy.
From reading this passage, it's not obvious to me what the fallacy is. The basic challenge that Gusev is posing to twin study proponents (as I understand it) is this: what explains the discrepancy between heritability estimates produced by RDR and the Classical Twin Design?
- The answer cannot be assortative mating (AM) as both methods are downwardly biased by AM.
- The answer is unlikely (though not strictly impossible) to be rare variants as Young et al applied RDR to IBD segments which should capture the majority of rare variant effects (https://geneticvariance.wordpress.com/2017/11/15/rdr-and-rare-variants/).
I'm genuinely curious to understand your answer to this question as some of the answers that I have heard (such as "imputation quality could undermine RDR") seem specious.
Thank you for pointing this out. I may have indeed misunderstood the Bingley et al model. I'll read the paper again (they don't make too much of an effort to be understood) and correct the post.
My understanding is that RDR is just fancy sib regression (extended to more family members) which is just fancy GREML-GCTA. They all use the quantatitive genetic logic of measuring how phenotypic similarity scales up with genetic similarity after holding the shared environment constant at 0 (GREML-GCTA) or 1 (RDR, sib-regression). If causal genetic variants are not seen (rare or non-SNP) or their effects are nonlinear this might throw off estimates. My confidence on this is low though, not sure why molecular methods don't work better. The fact that there is substantial missing heritability even for well-studied, easy to measure, non-controversial traits like height tells me that these methods miss something. Very heretical thought but I did wonder about the reliability of converting the DNA sequence to digital data.
Turkheimer has made a career from being the one go-to hired gun expert witness who the media as lawyer always calls up to tell the jury that The Science is settled that genes don't matter and that anyone who says otherwise is a lying evil racist. His many false comments about the work of Charles Murray is more than enough to prove he's a snake.
Outstanding post
Here's another defense of a similar thesis that I thought was good:
https://coel.substack.com/p/gwas-studies-underestimate-the-heritability
https://coel.substack.com/p/estimates-of-the-heritability-of
Good effort from the author but these miss the point. Sure, low power makes GWAS less accurate. But Gusev et al focus on GCTA-style studies which use genomic data to estimate relatedness and see if phenotypical similarity scales up with that. Low power doesn't bias these downward, the same way a small twin study doesn't have a downward bias. In the second paper, the author seems to confuse rGE and G*E, so not a great post. I want to write more about those too.
Here are two more articles explaining Dr. Gusev's errors:
https://hereticalinsights.substack.com/p/the-time-has-come
https://menghu.substack.com/p/sometimes-biased-but-not-systematically
The second one is very good!
This was incredibly well written! You explained these studies and how they work in a way that makes them easy to understand even to layman.
This article assumes that twin studies are a valid measure of the effect of genes and environment without dealing with many well-known criticisms of twin studies. What about gene-environment correlations? Shared genes cause people to create similar environments, which amplify the effects of genetic similarity. You address how gene-environment correlations might be a problem for molecular genetics, but fail to consider how they might affect twin studies. Also, how does this author explain the Flynn Effect? I discuss these problems with twin studies here: https://open.substack.com/pub/eclecticinquiries/p/twin-studies-exaggerate-iq-heritability?r=4952v2&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false
But a big part of the post IS "dealing with many well-known criticisms of twin studies"! In one paper Plomin wrote that active & reactive rGE is just a form of gene expression: some genes code proteins which change you, others make you more likely to seek out trait-changing environments. I don't like this argument because it ignores how bullish such a reality would be for interventions. If smart people are smart because they are (genetically) predisposed to seeking out smartening environments then we can give these to anyone! The fact that this doesn't work - g, for example, doesn't respond to interventions even in the short term - also makes me skeptical about rGE. Just what ARE those people with the right genes doing?
The IQ Flynn effect is mostly just a measurement artifact: https://substack.com/@cremieux/p-161257638. But there is also a Flynn effect for height and other traits which make your point stronger. Heritability doesn't mean genetic determinism, it just means that in a population you can realistically study (e.g. people living in the same country at the same time) genetic effects explain most of the differences between people. Shared environments have a restriction of range (nobody living today in the US grew up like in North Korea or in the 1800s) which makes C effects weaker. I do believe that there are important environmental differences between countries and eras which affect traits, which is why I don't like people talking about African IQs of 60 as if they were an expression of genetic potential. Still, politicians, sociologists etc. have very strong assumptions about C effects visible to ordinary twin studies, such as growing up in poverty or being raised in a certain way, so it still is a big deal if these are low.
Whether you like the rGE argument or not, the point is that there are environmental effects that twin studies can't separate out, which makes them unreliable measures of heredity. If genes tend to build environments that amplify their effect, that amplification is going to show up as a genetic effect in twin studies, but it would seem to be environmental. I quote and link to a long article dealing with this problem in my essay. https://pmc.ncbi.nlm.nih.gov/articles/PMC5754247/
So Plomin apparently thinks that rGE is a genetic effect, whereas others think it's an environmental one. If we can't even agree what counts as environmental and what counts as genetic, what sense does this whole debate have? We would seem to be arguing about indefinite and meaningless concepts.
That article also proves you wrong about interventions. It gives multiple examples of adoption boosting children's IQ score by as much as 40 points.
If the Flynn Effect is a measurement artifact, that means that IQ is a poor measure of intelligence. Hereditarians seem to saving their argument from the Flynn Effect problem by making the whole issue of IQ irrelevant. If all that IQ tests are doing is measuring test-taking skills, then IQ is an uninformative number that isn't measuring anything important about people.
It boosts the _child’s_ IQ. IQ measurement before adulthood has to be adjusted by age. The adult has a potential which the child can reach more or less quickly depending on the environment. The upper limit doesn’t increase. Sadly, the environment _can_ stunt a child.
Fantastic work, nicely done!
I don't agree that shared environment effects are more important than hard hereditarians think as Inquisitive Bird has written, but I will discuss why that is in the future.
For what reason is heritability research focused on humans, which show substantial variety even between siblings, not on animals, where variations are small? Are the findings inconvenient to both naturists and nuturists?
You write, "If within-family phenotypic similarity scales perfectly with genetic similarity (MZs are twice as similar as DZs), shared environment is zero."
I don't understand this assumption. Let's say MZ twins have phenotypes that are similar in exact proportion to how much more similar their genotypes are than DZ twins. To me this does not mean the contribution of shared environment is zero.
For example, take a gene for red hair. In the society these people are raised, having red hair is treated in a highly prejudicial way, and this tends to produce a range of reliable reactions in those who are persecuted/abused. It seems perfectly possible for MZ redheads to be more similar in a number of phenotypes than DZ twins due to environmental conditions triggered by a relatively small number of genes. The environment is the same in the way that twin studies define environments, but the way the people are treated is not the same.
This point about redheads can be generalized. Many genotypes may result in a phenotype that triggers divergent social reactions from other people living in the same environments, and thus impact phenotypic outcomes more broadly. I realize that the intention may be to be far stricter in defining a "shared environment" (to include the precise way everyone is treated), but then doesn't this contradict that first sentence above that I quote?
You are right! The scenario you described is an evocative gene-environment correlation (you genetically inherit a characteristic which triggers a causal environmental effect). Quantitative genetic studies assume that this is not happening, if it is and you are considering this as an environmental effect then they are wrong. (But something like this would throw off molecular genetic studies too.) The arguments against rGE are indirect. First, traits are hard to change even we are trying to do it - schools, psychotherapists, traumas etc. have short term effects at best, so it's a good question what others can be doing to you based on heritable visible traits that changes you so much. Second, poor look-alike similarity tells us that physical traits (like with the red hair example) don't do much to influence psychological characteristics, so it's another good question what exactly others are picking up about us which is not visible. Third, evocative rGE would be culture dependent - different cultures value different traits and some none at all - which would make heritability estimates vary a lot across places and eras.
Thanks for the clear and considerate response. And yes, it does seem that the best practical way to look for such "evocative" gene-environment correlations is to look for correlations that have a large swing in different cultures/settings.
The author mentioned, yet did not dispute, my charge that the Minnesota “reared-apart” twin researchers suppressed their DZ-apart control group IQ correlations to find above-zero IQ heritability. The author then cited a Swedish study that defined twins separated at 10 years of age as being “reared apart.” Twin studies are based false-assumptions (the EEA) and p-hacked results. Once the author recognizes this, they will understand why DNA-based studies of behavioral traits and psychiatric conditions have failed miserably.
East hunter, could you respond to this?It claims to refute race and intelligence research https://erikexamines.substack.com/p/iq-race-and-racism Tries to claim that Race realism it's false because the The black White Iq gap is narrowing
Almost everything in this post is wrong and the arguments have been addressed going back to the time of Jensen.
You would spend on your argument a little bit and show why he's wrong.Cause i'm not really an expert on this topic Bit of a layman. you address his main argument
If proband genes are real, family-environment-effects are genetic too, aren't them?
Excellent detailed piece. For a lay audience, though, it's missing a So What? Doesn't seem to move the needle from the non expert take that intelligence, neuroticism and other traits with a neurological basis are part nature, part nurture.
Like athleticism. You can be born tall and be nothing but a couch potato.
What's the point of this debate, exactly?
East Hunter Hey, I have a question.Could I ask you to take a look at this paper or this substack post?It's not very good and get your thoughts