I think it depends on whether this mechanism exactly tracks genetic resemblance in monozygotic vs. dizygotic twins. If it doesn't, then twin studies continue to prove something about genes (and not this other thing), and molecular genetic studies obviously prove something about genes (and not this other thing), so the mismatch is still surprising.
If it does track genetic resemblance (ie much stronger in monozygotic than dizygotic twins) then I think it might work, although I'd have to think about it harder to be sure.
Epigenetic effects could be stronger in monozygotic twins because these twins split a few days after fertilization (and up to that point, they are the same group of cells).
It would be quite interesting to track the similarity of dichorionic vs. diamniotic vs. monoamniotic twins (which split at progressively later stages) on various traits and epigenetic marks. You mentioned this study: https://pubmed.ncbi.nlm.nih.gov/26410687/ which addresses some of this but l think there's more to be discovered here.
There is. Or rather, not exactly; it's still DNA, it just isn't human DNA. What about mitochondrial DNA- has anyone ever investigated whether this affects IQ?
Mitochondrial DNA is identical in both fraternal and identical twins, as it's inherited from the mother in a non-sexually-recombining way.
Wouldn't all twins (identical and fraternal) get the same mitochondrial DNA from their mother? It seems like this would cause us to underestimate heritability from these studies, not overestimate it.
Actually, this leads to a question: Do all children inherit the same mtDNA from their parents? Or is there some variation in this that might differ between identical and fraternal twins?
Transgenerational epigenetic inheritance doesn't seem to be important in humans. There are some epigenetic resets that occur early in embryonic development, plus the effects of epigenetic modifications acquired during life are much more important. The latter would show up as environment (some shared, some non-shared) in twin studies.
There *IS a non-DNA way to inherit traits. Many traits are inherited via the cytoplasm, e.g., IIUC, the structure of the ribosome. Also mitochondria are (essentially) only inherited via the cytoplasm, and that contributes to average energy level. There are certain to be lots of others.
I'm not going to pretend that I can understand everything discussed above - or rather - that I can digest it meaningfully in my first read through.
But I'm curious as to whether genome wide polygenetic studies are good at identifying gene-environment interactions.
Most people think of the environment piece of gene-environment interactions as something obviously related. For IQ, you might expect it to be the parenting strategy of the parents that will evoke the best response for the IQ genes. But that's not necessarily the case. And a gene that manifests higher IQ, might not look like a gene for higher IQ unless its unfolding in the right environment - which can be a wide range of things.
I know that twin studies to some degree are meant to understand this, and there are also twin studies of twins reared apart, which should more strongly take care of this. But its still the case that, even when twins are reared a part, the environmental component that elicits the effect of the gene is so common that twins reared apart don't functionally constitute twins reared in different environments.
"Dynastic effects" and "genetic nurture" are very different processes and do not belong in the same category. Dynastic effects basically refers to the kind of phenomenon you mention of Norman surnames correlating with high status many years later, while genetic nurture refers to direct genetic effects in parents affecting their influence on children. Gemini's summary is as good as any:
"Genetic nurture is about what your parents do for you environmentally because of their genes.
Dynastic effects are more about the broader social and economic standing and inherited environmental context of your family lineage over generations, which is correlated with your genetic ancestry."
The most recent work (e.g. Nature Hum Behav Nivard 2024) suggests that for educational attainment, genetic nurture is very weak, while dynastic effects are quite considerable.
I'd be mildly concerned that correcting for population structure in GWAS analysis may be "the sociologist's fallacy" at work again, though if twin studies vs. sib-regression and RDR is yielding contradictory estimates for kidney function then something fishy really is going on.
As someone who has not yet seriously looked into adoption but who is open to the idea, your child psychiatric hospital anecdote is terrifying. Isn't the whole point of adoption (for many people) to offer a better life to kids who otherwise might have had a terrible time due to awful parents/environment? But actually adopting those kids puts you at a much higher chance of raising antisocial parasites, even if you treat them exactly like your other kids? Such a depressing thought. If this is a grave oversimplification, I'm happy to be proven wrong.
I think you can mostly avoid this if you're careful about who/where you adopt from. Some children are in orphanages because their parents died in a war or something, and they're pretty normal. If it's because the parents were involved in crime or drugs or something, then yeah, I think there's a significant risk that the child will be pretty tough to raise.
How much of that is explained by FASD and the like? (Or the developmental effects of extreme neglect in early childhood?) I was about to say “not that that makes much difference to adoption pessimism”, but FASD is actually not so hard for a physician (or even a trained layperson: kindergarten teachers in some countries can tell) to diagnose, no?
Well, sure; I asked how much. (I am not a medical professional, unlike you; my sense that there is more "FASD and the like" than many assume comes from talking to medical professionals and other people who know more than I do. People in the US now may drink less than they used to, but they also take more drugs.)
I looked into this once as when researching the whole "alcohol-during-pregnancy" topic. My takeaway: the answer is "yes" for FAS (Fetal Alcohol Syndrome) and very much "no" for FASD (Fetal Alcohol Spectrum Disorder). The former is characterised by the typical FAS face (small eyes, smooth philtrum, thin upper lip), low birth weight, microcephaly and various pretty severe developmental disorders.
The latter spectrum's diagnostic criteria are ridiculously loose leading some studies to conclude that as many as 10% of children have FASD (https://pmc.ncbi.nlm.nih.gov/articles/PMC5839298/). Add to that everyone who is "on the spectrum" or has ADHS and nobody normal is left.
The Christians I am familiar with are eager to adopt the children of homeless drug addicts. The children of recidivist criminals, I’m not so sure.
Either way: they’ve imbibed the conventional wisdom about heredity and life outcomes and while it only strengthens their compassionate determination to provide better-than-the-best in terms of nutrition, screens (none), reading, homeschooling if school is harsh and so forth - it also means their mantra is that the goal is simply a happy life.
>Some children are in orphanages because their parents died in a war or something, and they're pretty normal.
This is extremely unlikely in a modern developed country. Even in the improbable event that a child is unfortunate enough to lose both parents in a war, custody will almost always be taken over by a surviving close relative. You can try to adopt from an orphanage in the third world, but even then, they tend to prioritize prospective adopters from the same nation over foreigners.
If a child is available for adoption by a stranger in the US in 2025, you can pretty safely assume that "the parents were involved in crime or drugs or something".
Anecdata: I went to a fancy Christian private school (because my parents bankrupted themselves to find a somewhat-tolerable schooling environment for their weird son, not because I grew up fancy meself), and two of my classmates—out of about 100 total—were adopted through a similar "help the worst-off" sort of idea on the part of the parents. It was /very evident/ that they were... different... from the rest; you'd guess they'd been adopted even if they hadn't freely admitted it. The girl, as far as I know, didn't do great but didn't get in trouble; her brother did, indeed, end up arrested for assault.
I suspect babies put up for adoption in the postwar golden age of adoption, such as myself, tended to be more promising on average because there were a lot of social custom reasons for not raising the baby yourself that aren't as stringent anymore.
This is an important point! The social context for why children get put up for adoption matters a lot in what you'd expect the heritable traits of the kids to be, and the fact that this changes over time in different societies probably makes all kinds of adoption studies done at different times hard to compare. US adoptees in 1950 and in 2000 are probably very different in a lot of ways because of the changes in why people gave up their kids for adoption.
A friend of mine said she'd deliberately chosen to adopt a kid with identifiable mental disorders (autism etc.). This is because, in the national adoption system (UK), healthy, happy kids from wonderful parents who tragically died together in a car crash, with no grandparents, aunties/uncles or family friends to absorb these kids, are vanishingly rare.
So, given that they're in the adoption system, all kids are problematic for *some* reason - whether their own or their parents. My friend judged that it would be less risky to choose a child whose problem is known and manageable.
Of course, waiting outside Ukrainian hospitals to whisk away the healthiest war orphans you can find is another option, but I suspect that it's a logistical nightmare.
Adopting from foreign countries seems to be the best way to improve a child’s environment. Twin studies are almost always comparing children in the same country, and cross country environments very significantly more.
There are almost no children left to adopt, unless you adopt severely disabled children, teenagers or do foster care (where the goal is reunification). There are like 30 times more prospective adoptive parents than there are children given up for adoption.
Read about the Baby Scoop Era and the investigations South Korea is now doing into their old adoption agencies. Even back then (~50's-90's), the demand for children was bigger than the supply. Most children "given up" were actually cases of hardcore pressuring young single mothers (by eg. only wanting to give them money or medical care if they gave up their baby) and even straight up kidnapping. Adoption was a business.
Not adopted but it still bothers me how uneducated the public is on the subject.
Maybe it's "woke," but as an educator with experience teaching "different" adopted children, and having adopted siblings myself, I find the idea of calling such children "antisocial parasites" both repulsive and dangerous.
"The children are always ours, every single one of them, all over the globe; and I am beginning to suspect that whoever is incapable of recognizing this may be incapable of morality."
I don't think it's woke, I think your reaction is normal and appropriate. I'd just like to mention that my antisocial parasite comment was strictly focused on this line from Scott's article: "Then they promptly proceeded to commit crime / get addicted to drugs / go to jail / abuse people". At face value, everything here except for "abuse people" is not that awful. But adoption, to me - and again I haven't given it much thought yet - is about choosing a child that you bring into your family. That choice will be made on your own terms, but I reckon a whole load of criteria that people routinely use when choosing a child to adopt are politically incorrect, or otherwise criteria they wouldn't feel super comfortable saying out loud. This can lead to certain subsets of kids being less adopted than others.
By using the word "antisocial parasite", I tried to capture the essence of the following feeling: If I am given a choice when it comes to adopting a kid, including the existential non-zero risk of the kid destroying my and my partner's lives, I will stay away from those that have (apparently) much higher risks of doing just that. It doesn't invalidate their existence and their right to a future and a loving family - but I won't be the one to take them on, and I suppose a lot of people wouldn't, if they knew about those risks upfront.
I think in an atmosphere where many young people with the means to have their own children, forgo doing so (or even marrying) because it feels like a risk - it would be a little strange to expect a wholesale anything-goes attitude toward adopting those ill-parented children society is producing in numbers.
As usual, it will be down to those despised folks, the fundamentalist Christians, always hopeful, this world not entirely real to them.
I'm always happy to see a "Much More Than You Wanted To Know"—and this one's even better than most, because "missing heritability" is something I've been quite puzzled about (& I trust our amigo Scott "Yvain" Alexander to make a good run at it if anyone can👌). Cheers.
On priors, I would suppose there's something we're missing with the newer methods—partly because that sort of thing happens all the time ("this thing actually worked this other way all along, oops!"), and partly because I agree that it's difficult to see exactly how major error could creep into the twin studies.
--------------------------
(Y'know, a little while ago I had an idea—possibly even a cromulent one!—for how to square some of the gap away; or, rather, for why it might not be as worrisome as it first appears... but I want to check me notes before I risk saying something totally wrong & dumb / something that turns out to be the first thing everyone thought of already. Now, where in my hundreds of untitled drafts /is/ it...)
"On priors, I would suppose there's something we're missing with the newer methods—partly because that sort of thing happens all the time ("this thing actually worked this other way all along, oops!"), and partly because I agree that it's difficult to see exactly how major error could creep in on the twin studies."
Yeah, these results should definitely make us less confident about the high heritability (though a large part of what is going on might be the focus on EA, like Scott said), but less confident in a case where we started out highly confident still can leave us confident. Though of course eventually the issue becomes harder and harder to ignore. I just don't think that the 'this definitely means something is wrong with adoption/ twin studies' can be said to have arrived until the system looking for genes is accounting for stuff other than common SNPs.
Perhaps something is wrong with the 100+ year history of twin and adoption studies. Alternatively, something might be not quite right with the 12 year history of GWAS studies, just as most of the candidate gene studies of 22 years ago flamed out.
"Cromulent" seems to be the 2025 World of the Year. I'm seeing it all the time in recent weeks. I had to look up what it means ("acceptable or adequate") a few weeks ago.
"Cromulent first appeared in the February 18, 1996 episode of The Simpsons called "Lisa the Iconoclast," in what could be considered a throw-away line given during the opening credits. The schoolchildren of Springfield are watching a film about the founding father of Springfield, Jebediah Springfield. The film ends with Jebediah intoning, “A noble spirit embiggens the smallest man.” One teacher at the back of the room leans over to another and says that she’d never heard the word embiggen before she moved to Springfield. “I don't know why,” the other teacher replies. “It's a perfectly cromulent word.”"
“Embiggen” was actually the word that entered my vocabulary from that scene, in near-earnest at this point.
Is that the one where the schoolkids have an essay or art contest or something and the top 67 out of 68 entries will be displayed in the administration building? I loved that more than I can say.
Even if we never understand DNA or the brain, we still had a golden age.
I’ve read a lot of behavioural genetics. When people cite the "missing heritability problem" I think they should add, "of course this also applies to height".
One reason for not citing it is it diminishes their "argument" (everyone knows that height is mostly genetic). I can’t think of another explanation, although they might be repeating someone else’s argument without being aware of the issue.
Stephen Hsu has a paper, "Determination of Nonlinear Genetic Architecture using Compressed Sensing" and another, "Accurate Genomic Prediction Of Human Height". Worth a read. My simplest summary, from reading these a while back - you need large sample sizes, and the relationship between accuracy of genomic prediction and sample size is not linear.
Everybody knows height is mostly genetic *except in environments in which nutritional deprivation is common*, in which case being short gets stigmatised in ways different from the way it gets stigmatised elsewhere (in some countries (most of them, in the recent past), it correlates strongly with the class you were born in, because of obvious, cause-and-effect non-genetic reasons).
I don't think this is drop-down obvious. We know that Japanese height was stunted in the early 20th century. But the Japanese weren't starving to death - they just didn't have enough protein or something. But if this can happen to seemingly-advanced populations meeting their calorie requirements, how do we know that we've finally got nutrition right or that remaining variance in nutrition isn't enough to make a big difference?
...I mean, I think we actually know because of twin studies, but if we didn't have them, or doubted them, I don't think we would know through common sense alone.
Barbara Tuchman's prequel to "The Guns of August," "The Proud Tower" about aristocratic Europe before the Great War, starts off with a chapter about how tall Lord Salisbury's Tory cabinet of 1895 was: about 5 or 6 inches taller than the average British man.
I don't think, however, that there are major class differences in height anymore, at least among whites and blacks. A large fraction of NBA stars came from the underclass (e.g., 6'8" 260 pound LeBron James was born into the bottom of society, which, by the way, is why I'm so impressed with his incredibly consistent career), although the latest generation tends to more bourgeois.
I was struck by what you said about black people and lactose intolerance. I hadn’t heard that. I don’t know if you meant black Americans or black Africans. But on the subject of nutrition, milk for children was the holy grail for a long time. For growth and good bones. I was still getting a cup of milk, not water, with dinner when I was 12 years old.
I don’t know what to think about the importance of milk in connection with lactose intolerance.
Is milk good so lots of people still drink it despite discomfort?
Was milk pointless?
Or just valuable for easy availability of calories to those who can drink it?
There have been a lot of identical twins in the NBA because height is so important to basketball success, but very few identical twins in major league baseball because baseball seems to be largely a subtle knack. Or something.
(I made this reply comment above): I thought of nonlinearity too. The twin studies compare the effect of different sets of genes. The DNA studies are predictions using different individual genes as variables. Are these linear predictions? The post said interaction terms don't explain anything. How about polynomials of the individual genes ? This is easy to do.
I was going to mention Height too, but it seems there *isnt* much missing DNA explanatory power there-- the post has a bar graph showing heritability of about 10 biological traits, of which height is one. On the other hand, most of those 10 DO have the same missing DNA explanatory power as IQ.
When you say "height is mainly genetic" you're making a lot of environmental assumptions. The average height of the Japanese increased dramatically after WWII.
I didn't see any mention of the gut microbiome! It's not genetic, but it is heritable, especially from the mother (there are special mechanisms for carrying gut microbiota up into the breastmilk). Seems like a good place to look.
I think this wouldn't explain "missing heritability" per se, because it means "twin study heritability that is missing in molecular methods". But since gut microbiome is inherited equally between MZ and DZ twins, it wouldn't show up as heritable in twin studies. So it can't explain why twin studies find more heritability than molecular methods.
Beat me to it! Only yesterday he posted an interesting article on how certain species of gut bacteria could release compounds which bring on or worsen symptoms of schizophrenia:
EEA criticisms are pretty weak, IMO. For one MZ and DZ twins really do have equally similar environments when it comes to all the things that environmentalists have been saying are the real cause of variation in cognitive traits: Parental income, wealth, and education, books in the home, neighborhood, schools, etc.
But also, if subtle excess variation in DZ twins' environments relative to MZ twins' really has such large effects, we'd expect to see sizable correlations among adopted siblings. It's hard to believe both that adoption studies don't work because range restriction makes adoptive households too similar to one another, and also that twin studies don't work between DZ twins' environments are too dissimilar.
This reminds me that I don't have a good sense of where Gusev and Turkheimer would say the missing variance is.
I think the strongest argument would be that it's non-shared environment - ie basically random, probably having to do with weird quirks of embryology - but I don't know if that's actually what they think. I agree that books and education are less plausible.
Yeah, Aporia did a pretty good takedown on the subject. The missing heritability in GWAS is a bit of a puzzle, alright, but the problems of missing environmentality are worse.
AFAIK Turkheimer also has stated repeatedly that high heritability would be disastrous for society, as it would vindicate the bad, old, obsolete theories of "bad apples", scientific racism, class-based societies etc.
This is something I've noticed repeatedly now. I'm working in genetics myself and have some contact with social scientists both professionally and privately (my wife, for one, studied psychology - thankfully she broadly is on my side), and not only is the definition of hereditarians often hilariously lopsided (especially for IQ, it's common to be called hereditarian as long as you consider genetics non-negligible), the anti-hereditarians almost always sooner or later start talking about how terrible it would be if IQ was genetic, how it's clearly part of a push to undermine education (by whom, for what purpose?), or how it's just a psychological defense mechanism by elitist, .... On the other hand, the hereditarians at most mention the possible positive applications such as improving IVF for couples struggling to conceive, but otherwise stick to scientific arguments.
Unlike most people who think about heredity, race, and IQ, I'm a sports fan. So I have a hard time worrying too much about Turkheimer's concern that if it turned out to be true that IQ tends to vary genetically by race, that would be the worst thing imaginable. Instead, it's pretty obvious from my 60 years of watching sports on TV that athletic abilities tend to vary by race, and yet ... we seem to be dealing with this revelation pretty well.
For example, in the NFL, only whites of the fleetest lineages seem to succeed at running back, such as Christian McCaffrey, whose father Ed was an All-Pro wide receiver and whose maternal grandfather, Dr. Dave Sime, won the silver medal in the 1960 Olympic 100 meter dash.
And yet, football fans (perhaps the broadest demographic in modern America) seem pretty cool with that.
Similarly, the NBA draft yesterday picked the first white American #1 draft pick, Cooper Flagg, since Kurt Benson in 1977. I wouldn't be surprised if the NBA would be 10% more popular if it weren't so racially disparate. But, still, it's awfully popular despite it's enormous racial tilt.
I know the portion of variance explained by education is tiny, but how tiny? What are the percentages explainable by education, income, parental education, etc.? THere must be some best estimate for each, even if it's statistically insignificant.
Last time I read Gusev, I came away with the impression that it should be interaction effects, as he argues that molecular genetic methods control for those. But I'm no expert on the topic, and maybe I'm misremembering.
Note that, although that the story you linked doesn't mention it, the study that found people with Norman names overrepresented at Oxford was by Greg Clark, who argues that very long-run persistence of social status is driven by genetics and highly assortative mating (0.8 on a latent measure of socioeconomic competence). So in this case it wouldn't be necessarily be confounding, because persistent differences between people with Norman and Saxon names might well be genetic.
The idea that truly exogenous economic endowments can persist over several generations has pretty weak empirical support, as I believe you mentioned in a post several years ago. Studies finding otherwise are likely confounded by genetics.
But there was also a St. Louis Fed study finding that the intergenerational correlation for residual wealth (i.e. wealth greater than or less than that predicted by earnings) is only about 0.2. Except perhaps in the most extreme cases, wealth has to be actively maintained and fortified by each successive generation, or even very wealthy families fall off the map within a few generations. Consider that less than a century after his death, there are no longer any Rockefellers on the Forbes 400 list.
Wouldn't your theory require that Normans started out smarter than Saxons? I'm not claiming they can't possibly be, just that I can't really figure out a reason to think this, and it would make more sense that the persistent Norman-Saxon difference is because Normans have always been the nobility (and therefore had more money) since the Norman Conquest.
What if social status gives you access to higher-quality mates? You're rich, so you marry other rich people, some of whom got rich through extraordinary ability. Through this mechanism, it seems like it should be possible to consolidate temporary social status into a permanent genetic advantage.
I always assumed this was part of the logic of noble birth. Old bloodlines with an accomplished past bring in new blood with an accomplished present, incentivied by the advantages of noble family, hopefully leading to stable excellence that adapts over time at a moderate rate.
Primary sources rarely talk about these things outright, but religion be damned, these people bred every animal they could for trait selection, and they knew what they were.
The Churchill lineage produced the great Duke of Marlborough, victor over Louis XIV's army at Blenheim, and his famous Duchess. Then they didn't do much for a couple of centuries, then many generations later they produced the brilliant but unsuccessful Randolph Churchill and the highly successful Winston Churchill.
Yes- this, exactly. The Normans who moved to England as a transplanted aristocracy were not the median Norman, and any genetic distinction could have been fortified in subsequent generations by strategic marriages and selection effects.
My name Sailer is usually spelled Seiler (ropemaker), but an ancestor became mayor of a small town in Switzerland and decided his lineage deserved a classier surname (rather like more ambitious Smiths in Britain tended to change their names to Smythe or even Psmith).
Gregory Clark found that Smythes are more likely to graduate from Oxford or Cambridge than Smiths. Presumably they were all Smiths at one point, but the people who changed their names to Smythe were obvious social climbers.
Normans being by far the most effective in Europe and the Mediterranean in the medieval period seems some evidence there was something special about them. The Normans managed to kind of independently become the military elite in England, Ireland, Southern Italy and other places. Military skill is quite g loaded at least according to modern studies.
I don't know about lately, but when I last checked about 10 or 15 years ago, there were still four Hearsts on the Forbes 400. They trace their fortune to the Nevada silver rush of 1859, although they hit it rich again in mining strikes in the late 19th Century as well.
But ... most really rich guys on the Forbes 400 were semi-self made like Bill Gates, Jeff Bezos, or Elon Musk. They enjoyed well above average upbringings but then took huge advantage relative to other upper middle class individuals of their generation.
The US might also just be unusual in this regard. I once (peripherally) knew some members of the Fugger family. They are not crazy-rich but still wealthy enough that most of their energy seems to be dedicated to resolving within-family inheritance squabbles.
I wonder if there is a within-family version of the resource curse going on there. If you are extremely able and you come from a family with a big fortune, maybe there's little you can do that is as likely to pay off as well as to make sure to inherit the big fortune and manage it and the family's reputation well. Whereas if you don't come from vast wealth, maybe striking out on your own and starting a business is a better bet.
Thanks for such a wonderful write-up. One thing I’d love to see is more of a deep dive into the stats behind this. This seems like a very high dimensionality problem even without interactions, and I don’t see why interactions wouldn’t be as important as well. I could easily see k > n even with an n of three million, and it isn’t like k=n is a good situation. How the hell are they estimating this? LASSO? Something non-parametric? As an outsider, given the difficulty of the problem they’re facing, it seems like a no-brainier that regressions they’re running don’t add up.
For this "Alex Young thinks that once we get enough whole genomes sequenced (probably soon!) we might be able to use a technique called GREML-WGS to get more definitive answers about rare variants" is there any reason to think that even a sample size of 8 billion would be big enough to detect rare variant effects?
From an apriori point of view we should expect the relationship between genes and observed features to be incredibly complicated -- basically as complicated as the relationship between computer code and program operation -- and I'd argue the evidence we see is exactly what that model of highly complicated interactions predicts. Namely GWAS misses a bunch of twin study variation and so do simple non-linear models. Or to put the point differently the answer is the narrow/broad gap but broad hereditary is just really damn complicated.
Yes, people cite papers for the claim that non-linear effects don't seem to make the difference. But if you dig in the supposed evidence against non-linear effects (like you linked) are really only evidence against other very simple elaborations of the linear model. Showing that you don't predict much more of the variance by adding some simple non-linearity (eg dominance effects at a loci or quadratic terms etc) is exactly what you would expect if most effects are extremely complicated and neither model is even close to the complete causal story.
I mean, imagine that some OSS project like the Linux kernel gets bifrucated into a bunch of national variants which are developed independently with occasional merges between individual nation's versions (basically it acts like genes under recombination). Or better yet a Turing complete evolutionary programming experiment with complex behavior. Indeed this later one can be literally tested if people want.
We know in this case that code explains 100% of program variation and if you did the equivalent of a GWAS against it you would probably find some amount of correlations. I mean some perf improvements/behavior will be rarely discovered so they will be highly predicted by whether some specific code strings show up (they are all downstream of the initial mutation event) and other things like using certain approaches will have non-trivial correlation with certain keywords in the code.
But no linear model would actually even get close to capturing the true (indeed 100%) impact of code on those performance measurements. And it would look like there were no non-linear effects (in the sense of the papers claiming this in genetics) because adding some really simple addition to the linear model like "dominance loci" or whatever isn't going to do much better because you aren't anywhere close to the true causal model.
So the evidence we have is exactly what we should expect a priori -- genes have some really complicated (essentially Turing complete) relationship with observed behavior and simple models are going to guess at some of that but linear models will do about as well as simple non-linear ones until you break through some barrier.
I think certain twin/sibling studies have themselves been used to try and estimate the relative size of additive vs. non-additive genetic effects, and additive effects were maybe twice as large, though. Non-additive effects are also harder for natural selection to act on, which makes rapid evolution in a trait less likely.
Yes, but the problem is there are 2 types of studies that estimate non-additive effects.
1) Twin study type studies. Well that's exactly the high result which causes the gap.
2) Papers that test for non-additivity by elaborating the linear model in some way to allow for some simple non-linear effect. For instance, capturing the idea that maybe a gene can block the effect of other genes at the same loci.
But 2 is only evidence against the kind of non-linearity presumed in that model (more generally simple non-linearity). My point is that if the true relationship isn't something like mostly additive but maybe with some quadratic terms or multiplicative one but is actually more like a Turing complete interaction then you should expect that adding a bit of non-linearity doesn't improve the prediction much.
So as I said, we see exactly what we should expect consistent with the hypothesis of very complex genetic interaction s.
I'm probably not as immersed in the literature as you are, though I do work in the field of software so... I think I can partially intuit what you mean by non-linear program flow, although I think biological systems are both messier and generally include more built-in redundancy. However, I can certainly see the argument that complex interacting biological systems (such as brains, or even cellular metabolism) must have complex genetic correlates.
However, it's also the case that tiny regions of the genome can have counter-intuitively huge impacts on phenotype (e.g, humans vs. chimpanzees), so... I wouldn't assume in advance what percentage of gene-variants accounted for non-linear gene interactions, or what percentage of phenotypic variance those accounted for.
the hereditarians (the fact they call themselves this is funny, all geneticists are hereditarian) seem to have a weird idea that there's like a full genetic shuffle going on and evolution likes low-variance smooth fitness landscape. both are untrue afaict.
not a lot of shuffling is happening (1-3 crossovers per chromosome pair during human meiosis from memory) and with recombination hotspots the shuffle is somewhat predictable. so a lot of deep structure can be preserved and it's probably beneficial to have deep structure (you don't want an organism which rapidly loses genetic diversity to hit a local maximum and then all dies when the environment changes, similarly you don't want a function that's too full of 'holes' which the organism simply gets stuck in while maintaining diversity - you want an explorable and non-smooth fitness landscape).
note of course all these 'meta-parameters' like where the hotspots are, are themselves subject to selection which makes it harder to reason about.
i tend to agree with OP we should expect really complex interactions. the metaphor I'd use is that it's like we're analysing books by only doing word-counts. except in super high dimensional spaces it's often hard to intuit just how effective this is while still being completely wrong: for instance we can predict a lot from non-functional non-coding DNA simply because statistically it'll be colocated with genes that do matter. if you can make good predictions from junk DNA it's hard surprising you can do really well while wrongly assuming additivity.
i don't know much about the field but the impression i get is that the molecular geneticists largely take it for granted there's complex interactions because they've found really cool stuff in simple organisms. meanwhile you have more classical quantitative geneticists who have barely progressed since Fisher who are so far removed from the cutting edge that they barely interact with their much smarter brethren.
> But from an apriori point of view we should expect the relationship between genes and observed features to be incredibly complicated -- basically as complicated as the relationship between computer code and program operation -- so it's weird to think linear GWAS models should capture most of the variance.
If you only look at how genes get transcribed into proteins and what proteins actually do, and how long and convoluted the chain of cause and effect is to get anything we actually observe on a macro level, then your argument makes a lot of sense. It's ever crazier and weirder than human-readable computer programs.
But: remember that genes and chromosomes get reshuffled and stitched together and slightly mutated with every generation. That means for them to work, they need to be fairly robust.
If you want to think about a computational analogy, I wouldn't pick computer code ~ program execution, but perhaps like the weights in a neural network. Especially since methods 'dropout' (aka 'dilution') that randomly disable artificial neurons seem to work quite well, and not cripple them. See https://en.wikipedia.org/wiki/Dilution_(neural_networks)
(I guess it helps that 'dropout' in neural networks was inspired by an analogy with genes.)
Perhaps intelligence is the single thing that most requires intelligence to figure out? Hence, I'd hardly be surprised that we haven't figured out IQ completely yet.
I don't think the analogy with computer programs is on point.
The big difference is that the genetic code is recombinable. You can take two genomes, mash them together, add a few mutations, and it's still a working genome! You can't do that with computer code.(*) This is a pretty tough restriction, and it puts some severe restrictions on the complexity of interaction. Not that we would understand exactly how those restrictions look like. It doesn't just rule out "it's complicated" either, but it's enough that I find the comparison with computer code wrong.
(*)Actually you can, it's called Genetic Programming. The idea is that you evolve code via recombination and selective pressure by a fitness function, for example, to pass as many test cases s possible. It just doesn't work.(**) Since I work in a related field, I know a couple of people who work on Genetic Programming. The issue is precisely that it's so hard to recombine programs into another working program.
(**) My colleagues may beg to differ. But I would claim that it doesn't work unless you either work with absolutely tiny programs, or move so far away from the basic idea that you are no longer evolving code, but live in a totally different "genetic space".
Educational attainment is a meaningless measurement, at least from the internet survey versions I've seen with categories like "some college" or "bachelor's degree". This misses a lot of education that I think also correlates with intelligence. A Journeyman electrician requires a five-year apprenticeship that includes both on the job and classroom education, where would a high school dropout who joined the IBEW and became a Journeyman electrician be classified?
Both of my parents have a bachelor's degree, I chose instead to enlist in the US Navys Nuclear Power program which at the time was compared to the most difficult programs at top universities. Am I a strike against the heritability of educational attainment since I earned a nuclear Naval Enlisted Classification instead of a degree?
There's a difference between an imperfect proxy and a non-proxy.
Surveys of educational attainment don't classify *you* as an individual well or fairly. But they're still correlated with IQ, on the average. Getting 60% of the answers right on a true false test is not a perfect score. But it's not a 50% either. Surveys using educational attainment can get the equivalent of a 60% correct score and still be potentially useful.
They might be confounded, I suppose, if certain forms of intelligence predicted a person pursuing trade schools rather than college education. If you can be treated as noise, then a signal can still be extracted from the data. If you are a signal in your own right then the survey format might be a problem.
I guess using educational attainment may be...what's the phrase, directionally correct? I don't think this tells you anything more than you could learn from dog breeds, yes, behaviors can be inherited. A retriever will retrieve from not long after birth, and a shepherd breed will herd anything that moves which is always fun when a family with one invites a bunch of their child's friends to a birthday party and the dog will do its best to keep them all together.
I would think these studies aim for more useful information to quantify the degree of hereditariness or something and EA isn't going to get there. A degree may have correlated with intelligence for a period of time since that was the smooth path to a decent paying career, but when the people spending billions of dollars on AI datacenters say the bottlenecks are electrical generation and electricians to run all the cables, then over the next ten years intelligence may be more closely corelated with people who join the IBEW after high school than with those who earned a degree.
Back in 1986 I put an ad in the Chicago Tribune looking to hire a personal computer technician. The HR person asked about academic requirements. I thought about it for 5 seconds and said, "Bachelor's degree required."
But then I got a long letter from an applicant saying, I don't have a college degree because I enlisted in the Navy's submarine nuclear reactor program out of high school, but here are ten problems you are probably dealing with and how I'd solve them for you.
I hired him, and by the afternoon of the first day, it was clear he was smarter than me. While he solved my immediate problems, he also caused me a lot of unexpected problems because because within a few months my boss's boss, who had an MIT advanced degree, was calling me up, irate, to find out why the computer repairman I'd hired was having lunch with the Chairman of the Board to discuss corporate technology strategy.
Great story! I'm just ordinary smart, but yeah, we had more than our share of ridiculously intelligent people through that program, and a few of them also had interesting problems. We had one when I was an instructor at the NY prototype who never missed a single exam question and knew all the answers but had zero life skills. He couldn't drive, bought a plane ticket to Albany airport then started walking down the highway towards Saratoga Springs with his seabag on his back. The police picked him up and drove him to the site. We had to assign him roommates who could drive him to work because he hadn't made any friends who would do it voluntarily. He slowly got a little better, but wow, I had a much easier time assisting the ones who struggled academically.
Thanks for the write-up. I suspect that seeing this kind of confusion and disagreement is a good indicator that we are about to witness some nice chunk of progress from clearing this up, soon.
Especially if (one of the) limiting factors seems to be genetic sequencing and computing power: two things that are still getting cheaper and better quickly.
I know of two secret results I'm not supposed to talk about, by people claiming they've found very large amounts of "missing heritability". Not yet peer-reviewed or confirmed by anything except rumor. I expect one to be out within six months, and the other maybe eventually.
It's a funny coincidence, I came across this other post from Stephen Skolnick just yesterday arguing that the gut microbiome strongly predicts schizophrenia risk.
I will admit that Skolnick's essay sets off my 'too good to be true' alarms, but it still seems an idea worth exploring and he rigorously argues for the mechanism.
To what extent does the gut microbiome account for the outcomes we're seeing here, also? Gut microbiome is inherited, to some extent. It is also genetically influenced, and monozygotic twins have more similar gut microbiomes than dizygotic twins, with the divergence apparently increasing as life goes on.
Goodrich et al., 2014 (Cell)
SA: "Finally, she realizes she’s been scooped: evolution has been working on the same project, and has a 100,000 year head start. In the context of intense, recent selection for intelligence, we should expect evolution to have already found (and eliminated) the most straightforward, easy-to-find genes for low intelligence."
Doesn't this assume that increasing intelligence always increases reproductive fitness? In modern industrial societies there's a negative correlation between high intelligence and reproductive fitness, especially in women. I recognize that the same trend might not apply, pre-birth control in more traditional societies. But it at least seems worthwhile remembering that various types of mental facility and reproductive success may not always be positively related in all environments. This suggests at least the possibility of low hanging fruit if those gains involve tradeoffs with reproductive fitness.
Humans evolved to be more intelligent than other species because it was an advantage to us (perhaps because of our large social groups). But brains are metabolically expensive, and there are limits to how big our heads can be while still fitting through a birth canal (hence humans being relatively helpless at birth). So it's advantageous to have higher IQs, but there are tradeoffs and other traits being selected.
Did you ask the polygenic embryo selection folks about collider bias?
In the section comparing Kemper’s sib-regression estimate (14%) and Young’s Icelandic estimate (~40%), you note that the UK Biobank sample may be skewed toward healthier, higher-SES volunteers (so-called healthy volunteer bias, which commonly creates selection effects in medical research). But the implications of such selection effects extend far beyond variability in heritability estimates.
This kind of bias can flip signs when people think they're being clever by adjusting for confounds (collider stratification bias). This is especially a relevant risk in highly selected samples like the UK Biobank, where the same factors that influence participation (e.g., health, SES, education) may also affect the outcomes of interest (mental and physical health). Conditioning on variables that causally contribute to both participation and outcome can introduce more bias than it corrects for.
I wrote a bit about this here: https://wildetruth.substack.com/p/designer-baby-maybe. Munafò et al.'s schizophrenia example illustrates the mechanism more clearly than I can easily argue it for IQ: people at higher genetic risk for psychosis may be less likely to volunteer for studies ("they're out to get me, so I'm not giving them my blood"). This warps the apparent genetic architecture in large datasets. Doing embryo selection against schizophrenia risk based on such a sample, from which high-risk individuals disproportionately self-selected out, could backfire. And parents who paid $50K for it might not know for 20 years (or ever).
Hopefully the real-life practical implications are trivial: if you pay $50K for a prospective 3-point IQ gain, and collider stratification bias turns it into a 3-point loss, no big deal? You probably won't notice over dinner, as long as your kid slept well last night.
But I remain curious how the corporations selling embryo selection address the possibility that they're accidentally giving parents the exact opposite of what they're paying for. My guess is: they just don't tell anybody, and there's something in the fine print saying they could be wrong and aren't liable. Causality probably can't be proven in any individual case, and anyway you're paying for a best effort process, not a result.
This is entirely dependent on whether the “warping” by not having X group in a sample actually warps the sample. Because there’s a high chance it doesn’t warp it at all and the sample is still valid, even without X group present.
Jury is still out on the powers of nature versus nurture, basically. It's slightly complex.
What struck me is the questionable logic of EA as a measure. Scott got into this slightly with grade inflation and economic prospects, but I'd go further. Many of the twin studies took place between 1970 and 2000, or at least the kids were in school during that period, some of them earlier. This corresponds with a high point in educational prestige, and an opening of college to more people on a merit basis, while college was still rigorous. Today, getting a college degree is obviously less meaningful than it used to be, so the value presumed of its presence is almost irrelevant, and I'm thinking this is generational, a reflection of boomer-GenX social philosophy. Expecting the same effects thirty years on is going to be very muddled by social change.
EA is really, really sloppy from a standpoint of correlation with IQ.
One way or another, I still see nature and nurture about the same way. A person's biological design is full of complexities and redundancies, while their experiences will prod adaptation in various directions within the limits of that design. I know people don't want to believe that adaptation can be bracketed like that, but the evidence is everywhere, we maladapt to all sorts of things in the modern world, some of us more than others. People can push their limits, but it requires strong incentives and going too far results in trauma. Or failure. None of this is particularly compatible with Western ideals regarding individual agency and equality, but that's why it's controversial, right?
My sympathies to adopting parents who thought they were getting a blank slate.
Saw recently, don't recall the source, that the average IQ of college students has gone from almost 120 in the 1960's to just above 100 today due to pushing college enrollment on broader populations.
To the point about EA being full of landmines, wouldn't one of those be the observation that so many modern jobs expect a Bachelor's degree? That seems like a shared environmental component that would push swathes of the population to hit certain thresholds they might not otherwise.
A small technical note that I think is crucial. In the first figure (Tan et al 2024)
1. The x-axis is "Proportion of variance in population effects *uncorrelated* with direct genetic effects".
2. Being on the left side means that the variance in population (given by GWAS) can be correlated with direct genetic effects.
3. So for example, properties that are highly due to direct genetic effects are on the left (height, myopia, eczema)
4. Educational attainment is at ~0.5, this means that the direct effect variance is 0.5 of the population effect. The population effect is ~40% and thus the direct effect is 20%, exactly the value that GWAS converges to.
5. We can also verify this by looking on Tan et al 2024. In another figure they show that EA's genome wide correlation between direct genetic effects and population effects is ~0.75, if you take the square, you get ~0.5
How much effort has been put into algorithms decoding complex geneXgene interactions? Also how much of these things could be down to epigenetic expression, and how could this hypothesis be tested?
I looked into this once; the answer is that although Australia was founded by (a small number of) prisoners, almost all modern Australians are descended from non-prisoner immigrants who came later.
I conceptualise all this as: we have a new method sibregression that does not give the same results as twin studies; this is strong evidence that broad != narrow heritability. Which is very interesting and worthy of a lot of discussion about what it means.
To me it suggests that a significant amount of heritability is chaotic: the genes matter, but small differences have big effects, being "similar" is less informative than we might have expected.
And as you say:
"If nonadditive interactions are so important, why have existing studies had such a hard time detecting them?"
Is it possible that the *fact of being a twin* makes you weird? Twins are more likely to be premature and to die in infancy. Maybe twinhood is "bad for you" (perhaps via having a worse uterine environment) such that twins are more likely to have all kinds of health & behavior problems than singletons? Perhaps having a "bad" gene is more likely to result in a "bad" trait if you're a twin, which will push up the heritability of lots of traits in twin studies specifically?
That wouldn't affect adoption & pedigree studies, though, so if those also show high heritabilities for lots of traits this hypothesis doesn't help.
Twins used to have lower IQ than singletons, but have now converged (probably with better prenatal care). I can't prove that they don't have some kind of unique disposition which results in equal IQ on average but makes it have a different genetic structure, but seems like that would be surprising.
Could you respond to Lyman Stone's "Why Twin Studies Are Garbage"?
I think that Stone makes two basic points:
1. Heritability is not worth measuring or thinking about, because it doesn't mean what most people think it means. A trait can have perfect 1.00 heritability and proven genetic basis, and yet still be 100% environmental in the wild, due to gene-environment interactions.
2. Gene-environment interactions can be both very strong and very difficult to detect. We didn't notice for decades that peanut allergies were >90% environmental in practice, despite having >50% heritability and known genes for susceptibility. Just a tiny difference in the timing of diet accounts for all of that. Because of this, it's impossible to say that fraternal and identical twins have similar environments, which violates the first assumption of twin studies (as you have numbered them above).
Something can be worth measuring and thinking about even if "most people" don't know what it actually means. Most people don't understand quantum physics.
§4.3.3 Possibility 3: Non-additive genetics (a.k.a. “a nonlinear map from genomes to outcomes”) (a.k.a. “epistasis”)
…Importantly, I think the nature of non-additive genetics is widely misunderstood. If you read the wikipedia article on epistasis, or Zuk et al. 2012, or any other discussion I’ve seen, you’ll get the idea that non-additive genetic effects happen for reasons that are very “organic”—things like genes for two different mutations of the same protein complex, or genes for two enzymes involved in the same metabolic pathway.
But here is a very different intuitive model, which I think is more important in practice for humans:
• Genome maps mostly-linearly to “traits” (strengths of different innate drives, synaptic learning rates, bone structure, etc.)
• “Traits” map nonlinearly to certain personality, behavior, and mental health “outcomes” (divorce, depression, etc.)
As some examples: …
• I think the antisocial personality disorder (ASPD) diagnosis gets applied in practice to two rather different clusters of people, one basically with an anger disorder, the other with low arousal. So the map from the space of “traits” to the outcome of “ASPD” is a very nonlinear function, with two separate “bumps”, so to speak. The same idea applies to any outcome that can result from two or more rather different (and disjoint) root causes, which I suspect is quite common across mental health, personality, and behavior. People can wind up divorced because they were sleeping around, and people can wind up divorced because their clinical depression was dragging down their spouse. People can seek out company because they want to be widely loved, and people can seek out company because they want to be widely feared. Etc.
• I dunno, maybe “thrill-seeking personality” and “weak bones” interact multiplicatively towards the outcome of “serious sports injuries”. If so, that would be another nonlinear map from “traits” to certain “outcomes”.
All of these and many more would mathematically manifest as “gene × gene interactions” or “gene × gene × gene interactions”, or other types of non-additive genetic effects. For example, in the latter case, the interactions would look like (some gene variant related to thrill-seeking) × (some gene variant related to bone strength).
But that’s a very different mental image from things like multiple genes affecting the same protein complex, or the Zuk et al. 2012 “limiting pathway model”. In particular, given a gene × gene interaction, you can’t, even in principle, peer into a cell with a microscope, and tell whether the two genes are “interacting” or not. In that last example above, the thrill-seeking-related genes really don’t “interact” with the bone-strength-related genes—at least, not in the normal, intuitive sense of the word “interact”. Indeed, those two genes might never be expressed at the same time in the same cell….
As far as I can tell, if you call this toy example “gene × gene interaction” or “epistasis”, then a typical genetics person will agree that that’s technically true, but they’ll only say that with hesitation, and while giving you funny looks. It’s just not the kind of thing that people normally have in mind when they talk about “epistasis”, or “non-additive genetic effects”, or “gene × gene interactions”, etc. And that’s my point: many people in the field have a tendency to think about those topics in an overly narrow way.
…
§4.4.3 My rebuttal to some papers arguing against non-additive genetics being a big factor in human outcomes:
The first thing to keep in mind is: for the kind of non-additive genetic effects I’m talking about (§4.3.3 above), there would be a massive number of “gene × gene interactions”, each with infinitesimal effects on the outcome.
If that’s not obvious, I’ll go through the toy example from above. Imagine a multiplicative interaction between thrill-seeking personality and fragile bone structure, which leads to the outcome of sports injuries. Let’s assume that there are 1000 gene variants, each with a tiny additive effect on thrill-seeking personality; and separately, let’s assume that there’s a different set of 1000 gene variants, each with a tiny additive effect on fragile bones. Then when you multiply everything together, you’d get 1000×1000=1,000,000 different gene × gene interactions involved in the “sports injury” outcome, each contributing a truly microscopic amount to the probability of injury.
In that model, if you go looking in your dataset for specific gene × gene interactions, you certainly won’t find them. They’re tiny—miles below the noise floor. So absence of (that kind of) evidence is not meaningful evidence of absence.
The second thing to keep in mind is: As above, I agree that there’s not much non-additive genetic effects for traits like height and blood pressure, and much more for things like neuroticism and divorce. And many papers on non-additive genetics are looking at things like height and blood pressure. So unsurprisingly, they don’t find much non-additive genetics.…
[Then I discuss three example anti-epistasis papers including both of the ones linked in OP.]
I thought of nonlinearity too. The twin studies compare the effect of different sets of genes. The DNA studies are predictions using different individual genes as variables. Are these linear predictions? The post said interaction terms don't explain anything. How about polynomials of the individual genes ?
Just wanted to say this is a very good comment. It does seem to me that the more likely source of discrepancy is how we are defining a "trait," which is typically a more or less fuzzy concept.
>The biggest losers are the epidemiologists. They had started using polygenic predictors as a novel randomization method; suppose, for example, you wanted to study whether smoking causes Alzheimers. If you just checked how many smokers vs. nonsmokers got Alzheimers, your result would be vulnerable to bias; maybe poor people smoke more and get more Alzheimers. But (they hoped) you might be able to check whether people with the genes for smoking get more Alzheimers. Poverty can’t make you have more or fewer genes! This was a neat idea, but if the polygenic predictors are wrong about which genes cause smoking and what effect size they have, then the less careful among these results will need to be re-examined.
I've always been rather skeptical of mendelian randomization (MR) studies. In particular for MR to work, you need:
1. the genes to actually affect the phenotype (often true, but as this shows the effects can be weaker than previously thought)
2. no confounding by population stratification / assortative mating (hard to achieve)
3. no pleiotropic effects: e.g., maybe the genes for smoking overlap with the genes for Alzheimers (unlikely, but possible)
Usually when I see bad mendelian randomization studies (e.g. of ALDH2 variants for alcohol drinking -> cancer) they fail assumptions #2 and/or #3. But now it looks like assumption #1 is potentially worse than expected too!
Any time new types of testing and analysis are administered, and they give *wildly* different results than both multiple types of long used analysis and common sense, we should be skeptical of the analysis and results before we throw away anything of value that has gotten us through millennia of human history and centuries of the scientific era.
In a way this debate reminds me of the intelligent design debate, where one side (evolution believers) have mountains of evidence, but often it was more common sense and circumstantial, based on careful observation. This opened the door to claims about intelligent design that were simple at first, but were transformed by rebuttals until they were in many ways just evolution but “not evolution because evolution is wrong”. Anti-hereditarians are getting closer and closer to the intelligent design group in that they are slowly conceding ground in every debate so as to at least throw into doubt the tiniest bit of hereditarian, trading accuracy for vibes as it were.
“Maybe genetics does affect intelligence, but akshually it’s not your genetics, it’s your parents genetics that you also have which confounds even the claim I just made! Let’s talk around in circles and never admit it’s almost all genetic!”
Right, any speck of evidence on the X side is evidence that X is right, any speck of evidence on the Y side is just confounding and circumstantial. The (high) influence of heredity is not scientifically bulletproof, but the evidence *against* heredity is even weaker and more tenuous, thus not effectively invalidating the longstanding idea of “the apple doesn’t fall far from the tree” or “like father, like son” concepts that go back to antiquity.
There’s an implicit assumption that heritability is a fixed property of traits, rather than a model-dependent statistic that varies with environment, measurement, and population structure. If twin studies, GWAS, and sib-regression all yield different numbers, is that necessarily a “problem to be solved,” or might it reflect that traits like intelligence and EA aren’t stable enough constructs to expect convergence?
I'm seeing a fair few comments about genetic screening and 'careful adoption choices' that are giving me some serious liberal eugenics vibes, nasty. Not sure using anecdotes about criminal kids to defend IQ heritability estimates is that valid; I believe criminal behavior shows less heritability than IQ?
I wonder if there's a sampling bias here? Like only seeing kids whose biological parents got caught and ended up in the system. What about white-collar criminals, financial fraudsters, and corrupt executives who are smart/wealthy enough to avoid prison? If antisocial behavior is really heritable, wouldn't their kids show it too? But those kids aren't in adoption; they've probably been shipped of to elite prep schools to help continue their gene pool.
"In the degenerate case where the mother and father have exactly the same genes (“would you have sex with your clone?”) the fraternal twins will also share 100% of their genes."
Why is this so? If Mom and Dad both have genes (a, b) at one location, won't sometimes twin 1 get (a, a) while twin 2 will get (b, b)? Agree there's more commonality than normal because there's a possibility of 1 getting (a, b) while 2 gets (b, a), which isn't normally true.
Edit: I just saw that some of these points get covered in this post, thank you!
Someone with more human genomic knowledge please shoot me down.
Sequencing technology doesn't get discussed nearly enough in this area. Illumina short-read sequencing/SNP panels have been the major source of data for all of these studies, and they are absolutely delightful at finding SNPs but are crap at anything else. I think it will be appreciated that generally things that impact function aren't SNPs, they are broad changes, and so much human genomics seems to be hoping that the thing that is contributing to a change is able to spotted by being close to a SNP, instead of actually looking at the thing that is causing the change.
Genomes aren't lists of SNPs, they are mostly repeats and 2x150bp isn't going to get anywhere near close to capturing that variation, no matter how 'deep' you sequence. Long-read sequencing (PacBio & ONT, not Illumina's synthetic tech) is clearly better, and continues to demonstrate that there is massive variation that is easy to see when you have a bunch of 20kbp fragment, while almost impossible when you're just aligning little chunks of text to a 3gbp genome.
I work in non-model genomics and long-read sequencing is such a clear winner I keep getting surprised when Illumina gets contracts for these massive human studies. The Human Pangenome Consortium is going to be providing a dataset that is way more useful than anything that's come before. Anecdotally, I hear that for some non-European genomic data they know that ~10% of the data from an individual DOESN'T EVEN MAP to the human reference (but is human genomic data). This is all invisible to analysis, or even worse, just confounds things, as the 'true' causal SNP is somewhere in the data that doesn't get analysed, and so we're stuck looking at noise and trying to make sense of it.
I feel like this is such a blind-spot for human genomics, as it's always about the latest and greatest AI/ML method to try and get some information out, when it's the underlying data which just sucks and doesn't have a hope in actually being linked to function. There was even a point on an Open Thread a few weeks back (Open Thread 374 from tabulabio) asking for people to get in touch about Frontier Foundation Genomic models, with the focus being on fancy new ML architectures.
When I asked ChatGPT to write this comment for me ("Argue that sequencing technology could explain a lot of the Missing Heritability problem") it actually pushed back against me, trying to use the Wainschtein et al. 2022 paper as evidence that '...[this paper] used high-quality WGS (which includes better SVs than Illumina) and still found that adding rare and structural variants only modestly increased heritability estimates", which is NOT TRUE. Wainschtein uses the TOPMED dataset, which is from Illumina short reads. Yes, they do 'deep' sequencing, and yes it's analysed to the absolute hilt with the latest and greatest GatK pipeline and QC to the max. But that claim is false, it's just lists of SNPs, completely ignores huge chunks of the genome and just hopes that the thing contributing to a phenotype is is able to be fished out alongside a SNP.
Another anecdote - an older friend's wife died from a brain cancer. He was an old-school non-model genomicist and got all of the data from the oncologists and various tests and analysed things. All of it short-read, none of it turned anything up, either from his or the various doctors. Long-read sequencing run was eventually done after her death and indicated that it was a splice variant missed by short-reads. It was clear as day in the long-read data, since it didn't need to do fancy bioinformatic de bruijn graphs to figure out the splice isoform - it's just mapping the read against the genome and seeing it clear as day.
Thanks - can you explain more about whether things that are advertised as "whole genome sequencing" are the Illumina method you say is inadequate, or whether the WGS data is fine?
I'm probably not the first to think of this confounder, but it's been estimated that ~70% of human embryos turn out to be non-viable [https://pmc.ncbi.nlm.nih.gov/articles/PMC5443340/]. Do the sib-regression studies account for this? The 40-60% range of shared genes they measure is restricted to the siblings that survive to adulthood. But there may have been many 30% embryos that didn't survive the first month of gestation. Their educational achievement was 0.
In other words, there's a severe selection bias in any study that only looks at siblings who survived to adulthood. A genome that kills one of the siblings before they complete schooling or even get noticed on ultrasound is still an inherited trait.
I think this changes some extremely nitpicky form of heritability defined as "heritability across all embryos rather than just surviving ones", but I think as long as you define heritability as being across surviving humans, this definition is consistent and fine, and all of the existing studies remain comparable (and so it's still surprising when they don't match)
Would someone with strong quantitative chops be willing to explain whether GWAS (and related methods) have enough power to detect small effects—say, 0.5 or 1 percentage point—caused by combinations of 9, 11, or 53 SNPs? I’m particularly curious about scenarios where some of the SNPs are rare. In that case, the number of people with a given combination might be small, so the signal could be lost. If that sort of undetected combination effect happens multiple times—say, 30 or so—could that help explain the gap between GWAS results and what twin studies suggest? I’d love a mathematical explanation that someone with several semesters of college math/statistics could follow with effort.
Excellent post, I would just want to add that Sidorenko 2024 (https://pmc.ncbi.nlm.nih.gov/articles/PMC11835202/) that you cite for the discussion on rare variants also measure heritability using Sib-Regression although not for EA.
They show that correcting for assortative mating their estimates are coherent with twin studies :
"Therefore, if we account for assortative mating and assume that the resemblance between siblings is solely due to genetic effects (and not to common environmental), then our data are consistent with a heritability of 0.87 (s.e. 0.05) for height in the current population (Supplementary Note)."
they also show that shared environment estimates from twin studies are inflated by assortative mating :
"Our results for height and BMI agree with that conclusion in that we find no evidence of a residual sibling covariance (𝑐2) for BMI, while the significant 𝑐2 observed for height is largely explained by assortative mating."
And the WGS estimate is higher for height than the RDR estimates (55%), my take on this is that there is an assumption in RDR that posit that the environment of individuals are independent to each others and don't influence each others in correlation to the relatedness which since it's violated bias the estimate downward. (https://hereticalinsights.substack.com/i/146616013/disassortative-mating)
"Moreover, our estimates of 0.76 and 0.55 can also be compared to estimates from GWAS and WGS data. For height, the SNP-based estimate is about 0.55–0.60 (ref.41) and the WGS estimate ~0.70 (ref. 42; but with a large s.e. of ~0.10). These estimates imply that for height there is substantial genetic variation not captured by either SNP array and, to a lesser extent, sequence data, presumably ultra-rare variants (frequency <1/10,000 not included in ref.42)"
You are right, my bad, from the supplementary note 4 of Kemper 2021 (https://www.nature.com/articles/s41467-021-21283-4) they use the same equation to correct for AM and there is still a 0.1 gap in heritability for the two methods.
Sidorenko 2024 use a slightly different adjustment method to correct for the shared environment estimate becoming negative after correcting for assortative mating (another mystery to solve).
While writing this I just saw that the estimate twin heritability for height in Kemper 2021 is lower than the heritability for height in Sidorenko 2024 using sib-regression. Missing heritability solved I guess.
Gusev has a point when he says that when adjusted for assortative mating some of the high twin estimate for height or even sib regression get near or above 1 (when you read the supplementary note of Sidorenko 2024 the crude adjustement for a r=0.23 assortative matting in height push the estimate to 0.98 ! )
Turkheimer's quincunx metaphor is my personal mental model for the missing heritability phenomenon. A blog post is here: https://ericturkheimer.substack.com/p/a-speculative-explanation-of-the-e53 though you'll have to look at the few previous posts to fully understand it. It seemed like you dismissed this very briefly under non-additive effects and I don't have the technical background to grok those papers, but Turkheimer takes this hypothesis seriously so I figure it's serious.
Here's my summary:
Imagine a quincunx (aka a Galton board). You are a ball falling down the board. It bounces off a lot of pins, and it eventually falls into a bin at the bottom. That bin at the bottom is the outcome we're interested in (EA, IQ, etc). But the pins aren't random -- the pins are your genes, environmental variables, etc. So some pins are biased left or right, which influence the bin you end up in.
What Turkheimer shows is that in this model, if two people have the exact same quincunx (the exact same weights of pins), then they are very likely to have similar outcomes. Those are identical twins. If all your genes are the same, you are very likely to be similar. This reinforces the hereditarian position, and also makes a lot of intuitive sense. (One tricky part here is that it must be the case that mostly-the-same quincunxes (siblings, etc) cause similarity as well to a lesser degree. That's the biggest challenge to this mental model.)
But Turkheimer also shows that biases in pins are very very hard to detect in the non-identical-twin situation. Each pin only has an effect in the context of the entire system. For one person, a particular pin could have a large impact (maybe all the pins to the left bounce left, and all the pins to the right bounce right, so that pin matters a lot) but for another person that same pin has no influence at all (the pin to the left bounces right, the pin to the right bounces left, so it's like a funnel and that pin doesn't matter). So "in the wild," away from the twin context, it is very hard to come up with pins that matter in a GWAS type of way. But pins do matter! Genes do matter! They just happen to matter in this way where the whole package has a large influence on an individual, but individual pins are really hard to pick out in a practical way.
Turkheimer would emphasize that this is a simplified mental model. Our genes are not actually a quincunx -- but the interactions are much more complex, so if he can show that a quincunx has that sort of irreducible complexity, genetics must be even more complex.
I find this mental model really interesting and I actually wrote a blog about how I find it a helpful way to think about education and schools with some GIFs modeling the quincunx part if anyone is interested: https://fivetwelvethirteen.substack.com/p/the-quincunx
I think Turkheimer’s analogy here, given that he says that some of the quincunx pins are environmental, is dramatically failing to capture the fact that adult identical twins who were raised by different families in different cities are very obviously similar in all kinds of ways, indeed about as similar as if they were raised in the same family. (…With a few caveats & exceptions that I discuss in §2.2 here → https://www.lesswrong.com/posts/xXtDCeYLBR88QWebJ/heritability-five-battles )
That's a fair criticism of Turkheimer's position but I think you can rescue the broader idea by just saying that the quincunx is genetics only and the final bin is the sum of the genetic influence. To me, the main insight is that a complex system can have emergent properties that aren't easily correlated with the individual components of the system. And that's the core of the missing heritability problem: when we look at the whole system with twin studies we see high heritability, but when we try to find specific causal genes it becomes really messy.
I do think non-additive genetic effects are very important for adult personality, mental health and behavior (but not for other things like height and blood pressure). I talk about that a bunch in my post; see excerpts that I copied into a different comment → https://www.astralcodexten.com/p/missing-heritability-much-more-than/comment/129514879
Something that comes to my mind, having just read Pearl’s Book of Why, is whether collider bias or other common problems of regression in the hands of folks not very savvy about causal inference might explain the inconsistencies in the findings.
I mean, there is a lot of regression going on here, a lot of moderating and mediating variables, so it seems easy to make a mistake with the regression model.
Nice, I am a bioinformatics PHD working at least somewhat with GWAS and SNPs and from a Gell-Mann amnesia effect perspective this post was very reassuring.
My guess is also that the rare/ultra rare variants/combinations are responsible.
Lots of identical twins were raised to deny they were identical.
For example, in the 2004 Olympics Paul Hamm won the gold medal as the best all-around male gymnast while Morgan Hamm was 5th. But Paul and Morgan's parents doubted they were identical because their hair whorled in different directions.
Similarly, the nearly 7 foot tall Lopez twins of the NBA have done a great job over the decades of acting non-identical: they wear different hair styles and different clothes.
I think GxG is basically impossible to measure, so saying nobody found any GxG yet isn't really meaningful. It's harder to measure than rare variants; why believe in rare variants but doubt GxG?
Also, note that additiveness is sensitive to nonlinear functions of your measure. For example: if heritability of IQ is purely additive, then heritability of IQ^2 (the square of IQ) would not be. To posit no GxG at all is to say that the IQ scale is perfect and cannot be squared (and EA scale, and BMI, etc too).
A few final notes:
1. I think intuition from real life strongly suggests that everything is GxE (gene environment interaction); think, for example, about obesity. All the models assume GxE is zero in order to calculate the percent heritable.
2. You say assortative mating cannot bias downwards the heritability estimates from twin studies. That's only true if the twin studies don't already adjust for them! Many hereditarians cite 80% h^2 for IQ, but they get this by already adjusting for AM (and maybe even for attenuation). The can easily over-adjust! That would bias the estimates upwards!
3. Related to 2, I noticed that you quoted 60% for the typical twin-study heritability estimate, even though hereditarians (including you in the past) like to cite 80%. That's a substantial gap. Are you saying you now believe the 80% is wrong? Can you be more explicit about this disconnect?
I find the idea of reading modern scientific findings or social innovations into ancient myths and religions very compelling. As if the ancients through cultural evolution were able to find fundamental truths about reality without having to actually understand them. In light of that interest, here we go:
The ancient Greeks had the concept of generational sin. When someone committed a truly horrendous sin, like feeding your brother his own children as an act of revenge (https://en.wikipedia.org/wiki/Atreus), the gods may not punish the man directly, but leave punishment for a future generation.
Atreus (Progenitor of the house Atreides. Dune anyone?) was grandson of Tantalus, infamous for feeding Zeus his own son, and being punished by eternally having food and water just out of reach (origin of the word: tantalizing). Atreus echoed his grandfather's sin by feeding his brother his nephew's corpse and usurping his throne. Atreus' son, Agamemnon (famous from the Iliad) killed his own daughter, Clytaemnestra, in order to be allowed to attack Troy. In punishment for that, he was killed by his own wife upon returning victorious from Troy. His son, Orestes (grandson of Atreus) was given the "divine madness" (basically Schizophrenia) as punishment and it was only through the gods intervention that the cycle of inherited sin was synbolically broken. It was only ultimately ended when a new lineage, the Heracleidae (descendants of Herakles) invaded and taking over their lands, loosely representing an invasion of Northern Dorian peoples taking over the Achaean society that was there.
The lesson I take from this is, the Greeks understood the heritability of preponderancy for antisocial behavior. If your father or grandfather did something unspeakable, you were by extension unclean, and shunned by society because of that. If there is some actual truth to this, and children of men who are so antisocial as to practice cannibalism or other terrible sins, were poisoned by a Miasma that would carry over between generations, then perhaps there was a collective social benefit towards shunning the children and grandchildren of heinous criminals. The societies that practiced this, while fundamentally less just on an individual level, would have been able to better weed out antisocial behavior, and in the long run, outperform societies that didn't have this cultural meme.
The lesson at a societal level is that if your rulers have 6 generations of inherited sin and insanity, you will be invaded by your neighbors and made into slaves.
Of course in the modern day we've moved past that, and if you think hard enough I think you can come up with a "reasonable" justification for literally any social practice, but I always find it interesting to think about.
1. I wonder when we're going to have a good computational model of personality to just directly track these issues on a gear level, like what tradeoffs are solved in what proportion, what inefficiencies are present in the Turing machine that corresponds to one's intelligence, etc. (Sounds easy, right?..)
2. Is it possible that any single study (i. e. the Iceland study) introduced a trivial calculation error, or a different calculating method that at least partially accounts for the discrepancy? Are the datasets in question open-source, did people reproduce their calculations and make sure it's not that?
I wonder how much in vitro fertilization might impact more recent twin studies. Also the fact that we have mitochondrial DNA as well. So more and more data to feed into the maw of the ai models. De novo mutations also seem to impact certain genetic areas as well. Maybe we have a gene that predisposes to de novo mutations.
Consider syndrome vs. disease. It seems to me that the likely cause is that most of the things being measured have multiple independent ways of causing the measured result, and the genes that facilitate one of them don't necessarily facilitate another. Dawkins calls this teams of genes that are "good traveling companions". E.g. both Tibetan and some Peruvians have genetic adaptation to high altitude, but they do it in different ways.
So under this assumption (which I think is fairly reasonable) if you measure one gene, you're measuring one part of a collection that *IF PROPERLY ASSEMBLED* would facilitate one trait. But someone else presenting that trait with a different underlying approach would not find that gene helpful.
"An alternative way of validating twin studies... is to check them against their close cousins, adoption studies..."
Using adoption studies to "validate" twin studies seems completely backwards to me. Adoption studies are so obviously flawed that I do not see how they could possibly act as a validation set for anything!
The population of children going into adoption are wildly unrepresentative of the general population of children, unless that population is "children who go into the adoption system". The population of parents adopting children are obviously unrepresentative of the general population of children, unless that population is "parents who are adopting children". Being an adopted child to a set of parents is not like being a biological child to the same set of parents. Losing or being given up by birth parents is generally kind of a big deal.
If anything, the fact that twin studies give the same answers as adoption studies should make us more skeptical of twin studies!
In any case, I applaud Scott for both a) being totally transparent about his completely bonkers epistemic commitment here, and b) giving a good-faith description of the current evidence against his aforementioned bonkers epistemic commitment.
I go to one of those fundamentalist churches where people adopt kids and raise them piously. I think everybody knows by now that there's a big danger the kids will turn out badly compared to their siblings, even though there's considerable success in training them to behave well as kids. These adopted kids are often foster care kids, and that's going to add a lot of noise to any study. They don't have normal environmental differences; they have huge ones-- fetal alcohol syndrome, autism, abandonment as babies, abuse, . . . So for finding the environmental share for the US on average, they're not ideal.
I am alarmed at this statement about why we care if gene studies show that heritability may be misunderstood.
“Not doctors. So far this research has only just barely begun to reach the clinic. But also, all doctors want to do is predict things (like heart attack risk)”
We absolutely care if doctors over-treat or apply dubious preventative treatments based on unreliable associations.
Do I want surgery removing my prostate and rendering me incontinent because of a false read on my prostate cancer risk?
Correlation versus causation always matters when a reading is used as the basis for an intervention.
The next paragraph criticizes the use of these scores in epidemiology. Actually, even if all you have is variable with a correlation, it can be useful as an instrument in multiple regression analysis.
I thought this little-read article (written by an astrophysicist) does an excellent job of explaining why we wouldn't expect GWAS studies in their current form to get anywhere close to "true" heritability: https://coel.substack.com/p/gwas-studies-underestimate-the-heritability. Notably, the author doesn't rest his argument on the "rare genes of large effect" factor.
To roughly summarize his article: We would expect the code for developing a human-level intelligence to be extremely complex. There are ~3 billion nucleotides in the human genome; these GWAS studies rest on the assumption that only a few hundred to a few thousand SNPs are entirely responsible for creating human intelligence. That seems wildly inaccurate. As the author argues, could you write code for developing a human-level intelligence with only a few thousand instructions? ChatGPT is based on hundreds of billions of neural-network weights (though admittedly this is the end-product, not the "recipe" for producing it like the human genome). If intelligence is based on tens of thousands (or even more) SNPs, then our current GWAS studies are simply statistically incapable of finding them.
On top of that, GWAS is only looking at a very simplistic model of gene-behavior causation. It assumes an "additive" model of heritability which simply sums the measured effect of each SNP, whereas in reality the recipe for intelligence is likely caused by subtle and complex interaction effects among genes as well. (The author also mentions that SNPs, which are all GWAS studies measure, are only one type of genetic variation. This seems like it could be important but I don't understand what it means.)
In contrast, twin studies take into account the complexity of developing intelligence by simply observing the output of the incredibly complex underlying genomic causes. It's therefore not surprising at all that they give much higher heritability scores, whereas GWAS studies, limited as they are, give nothing more than a lower bound for heritability.
How do they adjust EA for non-college education? In my job I interact with many elderly. I ask them what they did. I live in a high-tech area so frequently I hear 'engineer'. But when I ask them where they studied, I am more likely to hear 'I got out of the army and went to work for Sunstrand as a mechanic and they trained me' than to get the name of a university. Very smart folks, but formal education is high school.
Total non-expert speculation here. Has anyone looked into whatever the genetic equivalent of “sequence of returns risk” is? In the market it matters enormously if there is a crash immediately after you retire versus a decade after you retire and your investments have had extra time to grow.
I.e as the body is built, different genes will apply in sequence. It could be that early-in-the-sequence genes (about which we currently know little) will have an outsized effect. It may be that some genes only activate or have an impact conditional to other genes activating or having an impact. Etc. Or perhaps from another perspective the mistake is looking at the genome as if it is static and one dimensional when in fact it is self-referencing and applies across dimensions of time and space.
I'll be honest, I haven't read word for word your post, but I read well enough. I will continue to read as i have time. I think the answer doesn't exist yet but will exist soon considering ai. The true answer is having behavioral patterns linked to endocrinology. Many don't take psychology seriously, but it is what we have. I haven't seen any reliable studies regarding hormones and behavioral traits.
Regarding educational attainment, I think it's also important to consider diet. If people do not get enough protein and fats, they will not be at their peak intelligence. Addictions of any sort will lower iq.
I am of the mind that virtually everything is inherited. Even broad behaviors amongst the population. We can connect hormone levels to DNA and hormones to behavior.
Was surprised to find that Scott used o3 to help research this, and found it helpful, despite the continuing hallucination problem. Let's say it gives you certain hallucinations you know you need to check, like citations, and you check them and strike that, fine. What if it's negatively-hallucinating, insistent that something does NOT exist when it does? What if it has numerous unknown-unknowns that it doesn't realize are relevant to the research question?
It seems like you probably need at a *minimum* to have Scott's level of mastery and familiarity with the subject going into the project in order to have any chance at spotting the places that o3 screwed up, and knowing why, and knowing where to look to correct it. The entire field of research needs to have been reduced to known-unknowns for you the author before you can safely employ the assistance of AI tools.
Here he had extensive layers of human review/assistance by specialists in the field, and had previously researched the topic. But I just don't know why anyone would bother to use AI for research when the output is clearly unreliable absent such review and input, in ways that a researcher with incomplete knowledge won't be able to always detect.
Why do we expect genes to have independent linear effects? From gene to protein to trait to life outcome, there are many nonlinear effects and interactions that we could imagine — is the problem really our imaginations? Hunting for genes that linearly independently correlate with causally remote life outcomes seems like looking for your keys under the lamppost. But I’m not really qualified to dig into the studies that you say disprove interactiond.
One easy possible confounder for twin studies would be if some aspect of your early childhood environment mattered a lot for some trait (IQ, kidney function, BMI, whatever) and also a lot of that environment was a function of your appearance. That would give you inflated heritability results for identical vs fraternal twins.
It doesn't seem so likely to me that this is important, though--it seems like adoption studies make a pretty strong argument that there's not much impact of variation in early childhood environment within normal middle-class-and-up norms and most measurable outcomes. (Raise a kid in a cave, beat him regularly, and never show him a book, and you'll probably depress his IQ. But letting him watch 20% more TV or taking him to amusement parks instead of museums on vacation isn't likely to have any noticeable effect.). But I guess if really cute kids get extra attention at some critical point in their development relative to ugly kids, it might have some effect?
Well, I never thought I would see this day. Finally, the smart people are coming around to what I have been claiming for decades. Inheritability is an interaction effect, thus the main effects (variation due entirely to genes or learning) are not directly measurable. Also, the twin studies are deeply flawed.
Let's start with twin studies. They have problems with their independent variables (IV's), their dependent variables (DV's) and their analysis.
IV's like personality traits and IQ test scores are highly controversial. We don't know exactly what IQ scores measure--except that it isn't "intelligence" the way most people mean that term (something like "an individual's inherent capacity to solve real world intellectual problems"). No one has proven that such a capacity exists as a coherent cognitive process, let alone that IQ scores measure it. Every other personality inventory suffers from similar problems.
The DV is correlation in such traits between pairs of siblings. This is measured two different ways: by degree of genetic similarity, compared to degree of environmental similarity. One of these two measures (genes) is much more precise than the other. These days, they can not only tell you how many genes two siblings have in common, but often which ones.
But not environment. The original twin studies used household to determine similarity of childhood environment. This is far too crude. Many twins are adopted into different households, but by members of the same extended family (aunts, uncles, grandparents, etc.). This means that the twins will know each other, often live in the same neighborhood, go to the same school, play with each other, etc. Since these things are largely undocumented, the data is contaminated, and we can't know by how much.
Finally, there are strong theoretical reasons to believe that evolutionary effects should be interactions between genes and the environment. The environment, after all, is the nature that selects. Therefore one would expect that any species will evolve traits that engage features of the environment, which is the mechanism by which an organism can gain an advantage. IQ is pretty useless if it doesn't give one a decisive advantage *somewhere*, and if it's mostly useless, why would it be selected?
In terms of analysis, in the presence of a strong interaction effect, the main effects (the effect of each key variable by itself, in this case genes and the environment) lose their meaning. It's easy to illustrate: Imagine that intelligence is 100% inherited. Two children with identical genes grow up in two different locations: one is a wealthy suburb in a household with plenty of resources, the other in a war zone, rife with disease and natural disasters. Do you expect these two to do equally well on IQ tests, Educational Achievement, or any other measure of life performance? One is pretty clearly at risk of leaving fewer descendants than the other, but remember, their genes are identical. There's nothing here for natural selection to grab onto. The main effect of genes is meaningless.
What this implies is that inheritability scores are probably meaningless. Intelligence isn't 60% inherited and 40% learned. It's probably something more like 80% interaction between the two.
Another random thought that I suspect people in the relevant field have probably already looked into: (I'm a cryptographer, what the heck do I know about behavioral genetics?)
Suppose a lot of outcome differences between people are due to pathogens. For example, if IQ regularly gets depressed a few points if you get a particular viral infection at a particular critical point in your development. (I think there's some reason to believe this does happen, because the month in which you're born does seem to have a small correlation with life outcomes, and how that overlaps with flu season is a reasonable guess about the mechanism by which that works. Also, this is likely to be some of why breastfeeding correlates with good outcomes for kids, though a lot is probably tangled up with social class/education of the mother.)
If this is true, maybe the critical genes to look at for a lot of heritability are the ones that determine your immune response (which MHC receptors you have, what genes you start from when you shuffle stuff around randomly to get T-cell receptors and antibody binding regions, variants in innate immune response) and maybe any rare variants that confer some immunity to common circulating viruses.
There are like a gazillion viruses that circulate all the time, often giving us infections we never even notice, sometimes giving us annoying but not serious symptoms (colds, stomach bugs, cold sores, warts, etc.). And we know that some of these (rubella, polio, zika) can do nasty developmental things to fetuses, babies, or small children.
That would track with at least some of the Flynn effect as sanitation and vaccination made people in-general healthier.
- EA seems virtually worthless to pursue topics with the degree of precision desired here. It's not just noisy, it's that the noise is so systemically confounded (national, racial, class, time, wealth, etc.), so any cross-study comparison is likely to be accidentally catching systemic biases. Aren't those trying to do aggressive heritability work that are looking at EA doomed to failure?
- This field seems likely to be suffused with soft "biases" (perhaps there's a better word, less laden with perjorative connotation). Some researchers may despite the eugenic-parallel ideas associated with high IQ heritability. In contrast, commercial researchers pursuing genetic-therapy depend entirely upon that. Some researchers are experts solely in their technique, etc. and need the world to look like a nail to fit the eneds of their hammer. That's all fine in that clean research still drives growth in the overall scientific tree. My question is really whether that tree has visible (more fertile) away from a healthy branch that is pursuing the sunshine of the anti-heritability folks on one side, and another healthy branch on the opposite side pursuing the sunshine available from the "we can rewrite your DNA to control 100% of your superchild's IQ"? Not that I'm seeking some compromise median. I'm just wondering if (what has to be a pervasive) desire to see a specific outcome has limited the rest of the hypotheses/research that could be pursued.
- This work is heritability in humans. Humans offer a ton of advantages and disadvantages for such work. Do less complex and intelligent subjects allow an easier laboratory along some veins that could allow for better staging for extrapolation back to humans? I assume there's a ton of work there - just don't know if it's disproportionately useful or limited.
This problem should be getting as much attention as Hubble Tension, if not more.
Can we induce this measurement phenomenon in rats, or worms, something where we have control of the genome? A quick Google suggests identical twins are rare in nature, but we can possibly produce them artificially for some animals.
Nice overview! A few points I wanted to expand on.
1. I think the post conflates gene-gene and gene-environment interactions; the latter (specifically interactions between genes and the "shared" environment) also get counted by twin models as narrow sense heritability. While I agree there is very little evidence for gene-gene interactions (particularly dominance, as you cite [and, interestingly, twin/adoption studies actually forecast a huge amount of dominance -- another discrepancy we do not understand]) there is quote substantial evidence for gene-environment interactions including on educational attainment (see Cheesman et al: https://www.nature.com/articles/s41539-022-00145-8 ; Mostafavi et al: https://elifesciences.org/articles/48376), IQ, and BMI. In fact, Peter Visscher led a paper that came to the conclusion that twin estimates for the heritability of BMI are very likely to be overestimated by gene-environment interactions (https://pubmed.ncbi.nlm.nih.gov/28692066/). A large amount of GxE plus some amount of equal environment violation seems like a very plausible and parsimonious answer to the heritability gap.
2. You mention epidemiologists being the biggest losers of stratification in polygenic scores, but I think it is important to note a related group: the people who take polygenic scores trained in one population (with a ton of stratification) and directly apply them to other populations to make claims about innate abilities (see: (https://theinfinitesimal.substack.com/p/how-population-stratification-led). This is especially true for Edu/IQ GWAS, where every behavior geneticist has been screaming "do not do that!" since the very first study came out. People like Kirkegaard, Piffer, Lasker, etc. (and their boosters on social media like Steve Sailer and Cremieux) dedicated their careers to taking crappy GWAS data from and turning it into memes that show Africans on the bottom and Europeans on the top. These people also happen to be the court geneticists, so to speak, for SSC/ACX. I don't mean to come off as antagonistic and I'm sure some people will see this comment and immediately discount me as being an ideologue/Lysenkoist/etc so it does my broader position no favors, but this stuff has done and continues to do an enormous amount of damage to the field (including the now complete unwillingness of public companies like 23andme to collaborate on studies of sensitive traits).
3. I'm going to gently push back against the hereditarian/anti-hereditarian framing (which I understand is probably here as shorthand and scene setting). I am personally interested in accurate estimates that are free of assumptions. I believe twin study estimates are of low quality because the assumptions are untestable, not because they are high. I also think the public fixation on twin studies has created some real and damaging anti-genetics and anti-psychiatry backlash and wrong-headed Blank Slate views. People hear about twin studies, look up the literature and find that peanut allergy (or wearing sunglasses, or reading romance fiction) is estimated to be highly heritable and have minimal shared environment (https://lymanstone.substack.com/p/why-twin-studies-are-garbage), start thinking that the whole field is built on nonsense, and end up at quack theories about how schizophrenia is actually a non-genetic fungal condition or whatever. I've been very clear that there are direct genetic effects on essentially every trait out there, including behavioral traits and IQ. If someone were to run a large-scale RDR analysis of IQ tomorrow and got a heritability of 0.9 and it replicated and all that, I would say "okay, it looks like the heritability is 0.9 and we need to rethink our evolutionary models". If anything, large heritability estimates would make my actual day job much easier and more lucrative because I could confidently start writing a lot of grants about all the genome sequencing we should be doing.
4. Lastly, it's not clear to me where the conclusion that well-validated twin studies converge on "similar results" is coming from. To take one example: the leading lights of behavior genetics (Deary, McGue, Visscher, etc) ran a study looking at the relationship between intelligence and lifespan (https://pubmed.ncbi.nlm.nih.gov/26213105/). This is a nice study for us because they put together three large, modern, twin cohorts with IQ measurements, but the heritability of IQ was just a nuisance parameter for them, so they had no reason to scrutinize the findings or file-drawer them. If we look at their MZ/DZ correlations in Table S6 we find that the heritability of IQ was 0.36 in the US sample; 0.98 in the Swedish sample; 0.24 in the Danish sample; and ... 0.52 on average. In other words, all over the place (but averaging out to the nice "half nature half nurture" result you see in books); the authors themselves used an AE model in Table 2 and reported a range of 0.20 to 0.98. This is far greater than the variability we see with GWAS or Sib-Reg, so what are we to make of that?
The anecdotal evidence of “troubled child raised in a good Christian home” seems a strong contender for selection bias. You wouldn’t have encountered the hypothetical well-behaved children from awful parents raised in supportive households. If that population outnumbers those you worked with, the point becomes moot.
Extremely interesting essay overall; thanks for sharing!
Maybe there could be a self-reinforcing effect to silbing similarities? There's the sterotype of the twins that are always together and always do the same thing, maybe small differences especially in early childhood separates silbings and makes them more strangled so that in the future they affect each other less and less. So identical twins have some initial stronger bonds by being more alike and then that develops into a environmental stronger bond (ie they do most activities together and want to be equal), while fraternal twins are a little more distant between them because of initial differences and then don't have that bond that makes them also have a more alike environment.
Theory: IQ is very heritable but also extremely subtly disabled (potentially even environmentally, think "more males born after wars"), because IQ is terrible for reproduction (now as ever) and you need some people to keep passing it on.
Suppose there was some non-DNA way to inherit traits. Perhaps something epigenetic. Would this explain part of the gap?
I think it depends on whether this mechanism exactly tracks genetic resemblance in monozygotic vs. dizygotic twins. If it doesn't, then twin studies continue to prove something about genes (and not this other thing), and molecular genetic studies obviously prove something about genes (and not this other thing), so the mismatch is still surprising.
If it does track genetic resemblance (ie much stronger in monozygotic than dizygotic twins) then I think it might work, although I'd have to think about it harder to be sure.
Epigenetic effects could be stronger in monozygotic twins because these twins split a few days after fertilization (and up to that point, they are the same group of cells).
It would be quite interesting to track the similarity of dichorionic vs. diamniotic vs. monoamniotic twins (which split at progressively later stages) on various traits and epigenetic marks. You mentioned this study: https://pubmed.ncbi.nlm.nih.gov/26410687/ which addresses some of this but l think there's more to be discovered here.
See also: https://denovo.substack.com/p/a-tale-of-twin-sisters (where one twin likely stole primordial germ cells from the other).
There is. Or rather, not exactly; it's still DNA, it just isn't human DNA. What about mitochondrial DNA- has anyone ever investigated whether this affects IQ?
Mitochondrial DNA is identical in both fraternal and identical twins, as it's inherited from the mother in a non-sexually-recombining way.
Wouldn't all twins (identical and fraternal) get the same mitochondrial DNA from their mother? It seems like this would cause us to underestimate heritability from these studies, not overestimate it.
Actually, this leads to a question: Do all children inherit the same mtDNA from their parents? Or is there some variation in this that might differ between identical and fraternal twins?
Transgenerational epigenetic inheritance doesn't seem to be important in humans. There are some epigenetic resets that occur early in embryonic development, plus the effects of epigenetic modifications acquired during life are much more important. The latter would show up as environment (some shared, some non-shared) in twin studies.
As a reproductive developmental biologist, I agree. I've written about this if people want to know more: https://denovo.substack.com/p/epigenetic-inheritance-in-mammals https://denovo.substack.com/p/epigenetics-of-the-mammalian-germline
There *IS a non-DNA way to inherit traits. Many traits are inherited via the cytoplasm, e.g., IIUC, the structure of the ribosome. Also mitochondria are (essentially) only inherited via the cytoplasm, and that contributes to average energy level. There are certain to be lots of others.
Lares Familiares
I'm not going to pretend that I can understand everything discussed above - or rather - that I can digest it meaningfully in my first read through.
But I'm curious as to whether genome wide polygenetic studies are good at identifying gene-environment interactions.
Most people think of the environment piece of gene-environment interactions as something obviously related. For IQ, you might expect it to be the parenting strategy of the parents that will evoke the best response for the IQ genes. But that's not necessarily the case. And a gene that manifests higher IQ, might not look like a gene for higher IQ unless its unfolding in the right environment - which can be a wide range of things.
I know that twin studies to some degree are meant to understand this, and there are also twin studies of twins reared apart, which should more strongly take care of this. But its still the case that, even when twins are reared a part, the environmental component that elicits the effect of the gene is so common that twins reared apart don't functionally constitute twins reared in different environments.
Anyway, I'm just curious about that.
"Dynastic effects" and "genetic nurture" are very different processes and do not belong in the same category. Dynastic effects basically refers to the kind of phenomenon you mention of Norman surnames correlating with high status many years later, while genetic nurture refers to direct genetic effects in parents affecting their influence on children. Gemini's summary is as good as any:
"Genetic nurture is about what your parents do for you environmentally because of their genes.
Dynastic effects are more about the broader social and economic standing and inherited environmental context of your family lineage over generations, which is correlated with your genetic ancestry."
The most recent work (e.g. Nature Hum Behav Nivard 2024) suggests that for educational attainment, genetic nurture is very weak, while dynastic effects are quite considerable.
Thanks for catching that, I deleted the section on dynastic effects but forgot to delete it from the header.
Dynastic effects are also not necessarily non-genetic.
https://www.astralcodexten.com/p/missing-heritability-much-more-than/comment/129495557
I'd be mildly concerned that correcting for population structure in GWAS analysis may be "the sociologist's fallacy" at work again, though if twin studies vs. sib-regression and RDR is yielding contradictory estimates for kidney function then something fishy really is going on.
As someone who has not yet seriously looked into adoption but who is open to the idea, your child psychiatric hospital anecdote is terrifying. Isn't the whole point of adoption (for many people) to offer a better life to kids who otherwise might have had a terrible time due to awful parents/environment? But actually adopting those kids puts you at a much higher chance of raising antisocial parasites, even if you treat them exactly like your other kids? Such a depressing thought. If this is a grave oversimplification, I'm happy to be proven wrong.
I think you can mostly avoid this if you're careful about who/where you adopt from. Some children are in orphanages because their parents died in a war or something, and they're pretty normal. If it's because the parents were involved in crime or drugs or something, then yeah, I think there's a significant risk that the child will be pretty tough to raise.
How much of that is explained by FASD and the like? (Or the developmental effects of extreme neglect in early childhood?) I was about to say “not that that makes much difference to adoption pessimism”, but FASD is actually not so hard for a physician (or even a trained layperson: kindergarten teachers in some countries can tell) to diagnose, no?
I think some but not all.
Well, sure; I asked how much. (I am not a medical professional, unlike you; my sense that there is more "FASD and the like" than many assume comes from talking to medical professionals and other people who know more than I do. People in the US now may drink less than they used to, but they also take more drugs.)
I looked into this once as when researching the whole "alcohol-during-pregnancy" topic. My takeaway: the answer is "yes" for FAS (Fetal Alcohol Syndrome) and very much "no" for FASD (Fetal Alcohol Spectrum Disorder). The former is characterised by the typical FAS face (small eyes, smooth philtrum, thin upper lip), low birth weight, microcephaly and various pretty severe developmental disorders.
The latter spectrum's diagnostic criteria are ridiculously loose leading some studies to conclude that as many as 10% of children have FASD (https://pmc.ncbi.nlm.nih.gov/articles/PMC5839298/). Add to that everyone who is "on the spectrum" or has ADHS and nobody normal is left.
The Christians I am familiar with are eager to adopt the children of homeless drug addicts. The children of recidivist criminals, I’m not so sure.
Either way: they’ve imbibed the conventional wisdom about heredity and life outcomes and while it only strengthens their compassionate determination to provide better-than-the-best in terms of nutrition, screens (none), reading, homeschooling if school is harsh and so forth - it also means their mantra is that the goal is simply a happy life.
>Some children are in orphanages because their parents died in a war or something, and they're pretty normal.
This is extremely unlikely in a modern developed country. Even in the improbable event that a child is unfortunate enough to lose both parents in a war, custody will almost always be taken over by a surviving close relative. You can try to adopt from an orphanage in the third world, but even then, they tend to prioritize prospective adopters from the same nation over foreigners.
If a child is available for adoption by a stranger in the US in 2025, you can pretty safely assume that "the parents were involved in crime or drugs or something".
Anecdata: I went to a fancy Christian private school (because my parents bankrupted themselves to find a somewhat-tolerable schooling environment for their weird son, not because I grew up fancy meself), and two of my classmates—out of about 100 total—were adopted through a similar "help the worst-off" sort of idea on the part of the parents. It was /very evident/ that they were... different... from the rest; you'd guess they'd been adopted even if they hadn't freely admitted it. The girl, as far as I know, didn't do great but didn't get in trouble; her brother did, indeed, end up arrested for assault.
I suspect babies put up for adoption in the postwar golden age of adoption, such as myself, tended to be more promising on average because there were a lot of social custom reasons for not raising the baby yourself that aren't as stringent anymore.
This is an important point! The social context for why children get put up for adoption matters a lot in what you'd expect the heritable traits of the kids to be, and the fact that this changes over time in different societies probably makes all kinds of adoption studies done at different times hard to compare. US adoptees in 1950 and in 2000 are probably very different in a lot of ways because of the changes in why people gave up their kids for adoption.
A friend of mine said she'd deliberately chosen to adopt a kid with identifiable mental disorders (autism etc.). This is because, in the national adoption system (UK), healthy, happy kids from wonderful parents who tragically died together in a car crash, with no grandparents, aunties/uncles or family friends to absorb these kids, are vanishingly rare.
So, given that they're in the adoption system, all kids are problematic for *some* reason - whether their own or their parents. My friend judged that it would be less risky to choose a child whose problem is known and manageable.
Of course, waiting outside Ukrainian hospitals to whisk away the healthiest war orphans you can find is another option, but I suspect that it's a logistical nightmare.
Adopting from foreign countries seems to be the best way to improve a child’s environment. Twin studies are almost always comparing children in the same country, and cross country environments very significantly more.
There are almost no children left to adopt, unless you adopt severely disabled children, teenagers or do foster care (where the goal is reunification). There are like 30 times more prospective adoptive parents than there are children given up for adoption.
Read about the Baby Scoop Era and the investigations South Korea is now doing into their old adoption agencies. Even back then (~50's-90's), the demand for children was bigger than the supply. Most children "given up" were actually cases of hardcore pressuring young single mothers (by eg. only wanting to give them money or medical care if they gave up their baby) and even straight up kidnapping. Adoption was a business.
Not adopted but it still bothers me how uneducated the public is on the subject.
Maybe it's "woke," but as an educator with experience teaching "different" adopted children, and having adopted siblings myself, I find the idea of calling such children "antisocial parasites" both repulsive and dangerous.
"The children are always ours, every single one of them, all over the globe; and I am beginning to suspect that whoever is incapable of recognizing this may be incapable of morality."
- James Baldwin
I don't think it's woke, I think your reaction is normal and appropriate. I'd just like to mention that my antisocial parasite comment was strictly focused on this line from Scott's article: "Then they promptly proceeded to commit crime / get addicted to drugs / go to jail / abuse people". At face value, everything here except for "abuse people" is not that awful. But adoption, to me - and again I haven't given it much thought yet - is about choosing a child that you bring into your family. That choice will be made on your own terms, but I reckon a whole load of criteria that people routinely use when choosing a child to adopt are politically incorrect, or otherwise criteria they wouldn't feel super comfortable saying out loud. This can lead to certain subsets of kids being less adopted than others.
By using the word "antisocial parasite", I tried to capture the essence of the following feeling: If I am given a choice when it comes to adopting a kid, including the existential non-zero risk of the kid destroying my and my partner's lives, I will stay away from those that have (apparently) much higher risks of doing just that. It doesn't invalidate their existence and their right to a future and a loving family - but I won't be the one to take them on, and I suppose a lot of people wouldn't, if they knew about those risks upfront.
Thank you for your comment.
I think in an atmosphere where many young people with the means to have their own children, forgo doing so (or even marrying) because it feels like a risk - it would be a little strange to expect a wholesale anything-goes attitude toward adopting those ill-parented children society is producing in numbers.
As usual, it will be down to those despised folks, the fundamentalist Christians, always hopeful, this world not entirely real to them.
What a great post
Thanks a lot
I'm always happy to see a "Much More Than You Wanted To Know"—and this one's even better than most, because "missing heritability" is something I've been quite puzzled about (& I trust our amigo Scott "Yvain" Alexander to make a good run at it if anyone can👌). Cheers.
On priors, I would suppose there's something we're missing with the newer methods—partly because that sort of thing happens all the time ("this thing actually worked this other way all along, oops!"), and partly because I agree that it's difficult to see exactly how major error could creep into the twin studies.
--------------------------
(Y'know, a little while ago I had an idea—possibly even a cromulent one!—for how to square some of the gap away; or, rather, for why it might not be as worrisome as it first appears... but I want to check me notes before I risk saying something totally wrong & dumb / something that turns out to be the first thing everyone thought of already. Now, where in my hundreds of untitled drafts /is/ it...)
"On priors, I would suppose there's something we're missing with the newer methods—partly because that sort of thing happens all the time ("this thing actually worked this other way all along, oops!"), and partly because I agree that it's difficult to see exactly how major error could creep in on the twin studies."
Yeah, these results should definitely make us less confident about the high heritability (though a large part of what is going on might be the focus on EA, like Scott said), but less confident in a case where we started out highly confident still can leave us confident. Though of course eventually the issue becomes harder and harder to ignore. I just don't think that the 'this definitely means something is wrong with adoption/ twin studies' can be said to have arrived until the system looking for genes is accounting for stuff other than common SNPs.
Perhaps something is wrong with the 100+ year history of twin and adoption studies. Alternatively, something might be not quite right with the 12 year history of GWAS studies, just as most of the candidate gene studies of 22 years ago flamed out.
"Cromulent" seems to be the 2025 World of the Year. I'm seeing it all the time in recent weeks. I had to look up what it means ("acceptable or adequate") a few weeks ago.
You should watch the old Simpsons episode that coined it.
I don't think there's anything 2025 specific to it. But might just be weirdly common in whatever media you consume?
"Cromulent first appeared in the February 18, 1996 episode of The Simpsons called "Lisa the Iconoclast," in what could be considered a throw-away line given during the opening credits. The schoolchildren of Springfield are watching a film about the founding father of Springfield, Jebediah Springfield. The film ends with Jebediah intoning, “A noble spirit embiggens the smallest man.” One teacher at the back of the room leans over to another and says that she’d never heard the word embiggen before she moved to Springfield. “I don't know why,” the other teacher replies. “It's a perfectly cromulent word.”"
https://www.merriam-webster.com/wordplay/what-does-cromulent-mean#:~:text=Though%20'cromulent'%20originated%20as%20a,future%20entry%20in%20the%20dictionary.
See also https://tvtropes.org/pmwiki/pmwiki.php/Main/PerfectlyCromulentWord
“Embiggen” was actually the word that entered my vocabulary from that scene, in near-earnest at this point.
Is that the one where the schoolkids have an essay or art contest or something and the top 67 out of 68 entries will be displayed in the administration building? I loved that more than I can say.
Even if we never understand DNA or the brain, we still had a golden age.
I’ve read a lot of behavioural genetics. When people cite the "missing heritability problem" I think they should add, "of course this also applies to height".
One reason for not citing it is it diminishes their "argument" (everyone knows that height is mostly genetic). I can’t think of another explanation, although they might be repeating someone else’s argument without being aware of the issue.
Stephen Hsu has a paper, "Determination of Nonlinear Genetic Architecture using Compressed Sensing" and another, "Accurate Genomic Prediction Of Human Height". Worth a read. My simplest summary, from reading these a while back - you need large sample sizes, and the relationship between accuracy of genomic prediction and sample size is not linear.
Everybody knows height is mostly genetic *except in environments in which nutritional deprivation is common*, in which case being short gets stigmatised in ways different from the way it gets stigmatised elsewhere (in some countries (most of them, in the recent past), it correlates strongly with the class you were born in, because of obvious, cause-and-effect non-genetic reasons).
I don't think this is drop-down obvious. We know that Japanese height was stunted in the early 20th century. But the Japanese weren't starving to death - they just didn't have enough protein or something. But if this can happen to seemingly-advanced populations meeting their calorie requirements, how do we know that we've finally got nutrition right or that remaining variance in nutrition isn't enough to make a big difference?
...I mean, I think we actually know because of twin studies, but if we didn't have them, or doubted them, I don't think we would know through common sense alone.
Interestingly, there seems to be just as much (or perhaps even more?) sexual selection for height as for intelligence.
Barbara Tuchman's prequel to "The Guns of August," "The Proud Tower" about aristocratic Europe before the Great War, starts off with a chapter about how tall Lord Salisbury's Tory cabinet of 1895 was: about 5 or 6 inches taller than the average British man.
I don't think, however, that there are major class differences in height anymore, at least among whites and blacks. A large fraction of NBA stars came from the underclass (e.g., 6'8" 260 pound LeBron James was born into the bottom of society, which, by the way, is why I'm so impressed with his incredibly consistent career), although the latest generation tends to more bourgeois.
It's also fascinating how many NBA players have players who were also in the NBA.
I was struck by what you said about black people and lactose intolerance. I hadn’t heard that. I don’t know if you meant black Americans or black Africans. But on the subject of nutrition, milk for children was the holy grail for a long time. For growth and good bones. I was still getting a cup of milk, not water, with dinner when I was 12 years old.
I don’t know what to think about the importance of milk in connection with lactose intolerance.
Is milk good so lots of people still drink it despite discomfort?
Was milk pointless?
Or just valuable for easy availability of calories to those who can drink it?
The current state of this debate is that Gusev published a post called No, Intelligence Is Not Like Height ( https://theinfinitesimal.substack.com/p/no-intelligence-is-not-like-height ), but some other people say that the central finding in that post is based on a technical error that has been debunked by further research (see https://x.com/AlexTISYoung/status/1843288303325593923, search for "Our estimate of the SNP"). I haven't heard any response from Gusev and haven't dug into this further.
There have been a lot of identical twins in the NBA because height is so important to basketball success, but very few identical twins in major league baseball because baseball seems to be largely a subtle knack. Or something.
Nobody seems to know why there have been so few identical twins in major league baseball.
Except for Jose and Ozzie Canseco, which is likely due to Jose being "the Typhoid Mary of steroids" as a baseball agent told me over 30 years ago.
(I made this reply comment above): I thought of nonlinearity too. The twin studies compare the effect of different sets of genes. The DNA studies are predictions using different individual genes as variables. Are these linear predictions? The post said interaction terms don't explain anything. How about polynomials of the individual genes ? This is easy to do.
I was going to mention Height too, but it seems there *isnt* much missing DNA explanatory power there-- the post has a bar graph showing heritability of about 10 biological traits, of which height is one. On the other hand, most of those 10 DO have the same missing DNA explanatory power as IQ.
When you say "height is mainly genetic" you're making a lot of environmental assumptions. The average height of the Japanese increased dramatically after WWII.
I didn't see any mention of the gut microbiome! It's not genetic, but it is heritable, especially from the mother (there are special mechanisms for carrying gut microbiota up into the breastmilk). Seems like a good place to look.
(For anyone interested, researcher Stephen Skolnick has an excellent blog on this kind of thing: https://stephenskolnick.substack.com/)
I think this wouldn't explain "missing heritability" per se, because it means "twin study heritability that is missing in molecular methods". But since gut microbiome is inherited equally between MZ and DZ twins, it wouldn't show up as heritable in twin studies. So it can't explain why twin studies find more heritability than molecular methods.
Beat me to it! Only yesterday he posted an interesting article on how certain species of gut bacteria could release compounds which bring on or worsen symptoms of schizophrenia:
https://stephenskolnick.substack.com/p/schizophrenia
EEA criticisms are pretty weak, IMO. For one MZ and DZ twins really do have equally similar environments when it comes to all the things that environmentalists have been saying are the real cause of variation in cognitive traits: Parental income, wealth, and education, books in the home, neighborhood, schools, etc.
But also, if subtle excess variation in DZ twins' environments relative to MZ twins' really has such large effects, we'd expect to see sizable correlations among adopted siblings. It's hard to believe both that adoption studies don't work because range restriction makes adoptive households too similar to one another, and also that twin studies don't work between DZ twins' environments are too dissimilar.
This reminds me that I don't have a good sense of where Gusev and Turkheimer would say the missing variance is.
I think the strongest argument would be that it's non-shared environment - ie basically random, probably having to do with weird quirks of embryology - but I don't know if that's actually what they think. I agree that books and education are less plausible.
Gusev said on TwiX a week or two ago that it's obviously education. That really shattered the illusion that he might actually have a point.
Yeah, Aporia did a pretty good takedown on the subject. The missing heritability in GWAS is a bit of a puzzle, alright, but the problems of missing environmentality are worse.
https://www.aporiamagazine.com/p/sasha-gusev-is-wrong
AFAIK Turkheimer also has stated repeatedly that high heritability would be disastrous for society, as it would vindicate the bad, old, obsolete theories of "bad apples", scientific racism, class-based societies etc.
This is something I've noticed repeatedly now. I'm working in genetics myself and have some contact with social scientists both professionally and privately (my wife, for one, studied psychology - thankfully she broadly is on my side), and not only is the definition of hereditarians often hilariously lopsided (especially for IQ, it's common to be called hereditarian as long as you consider genetics non-negligible), the anti-hereditarians almost always sooner or later start talking about how terrible it would be if IQ was genetic, how it's clearly part of a push to undermine education (by whom, for what purpose?), or how it's just a psychological defense mechanism by elitist, .... On the other hand, the hereditarians at most mention the possible positive applications such as improving IVF for couples struggling to conceive, but otherwise stick to scientific arguments.
Unlike most people who think about heredity, race, and IQ, I'm a sports fan. So I have a hard time worrying too much about Turkheimer's concern that if it turned out to be true that IQ tends to vary genetically by race, that would be the worst thing imaginable. Instead, it's pretty obvious from my 60 years of watching sports on TV that athletic abilities tend to vary by race, and yet ... we seem to be dealing with this revelation pretty well.
For example, in the NFL, only whites of the fleetest lineages seem to succeed at running back, such as Christian McCaffrey, whose father Ed was an All-Pro wide receiver and whose maternal grandfather, Dr. Dave Sime, won the silver medal in the 1960 Olympic 100 meter dash.
And yet, football fans (perhaps the broadest demographic in modern America) seem pretty cool with that.
Similarly, the NBA draft yesterday picked the first white American #1 draft pick, Cooper Flagg, since Kurt Benson in 1977. I wouldn't be surprised if the NBA would be 10% more popular if it weren't so racially disparate. But, still, it's awfully popular despite it's enormous racial tilt.
I know the portion of variance explained by education is tiny, but how tiny? What are the percentages explainable by education, income, parental education, etc.? THere must be some best estimate for each, even if it's statistically insignificant.
Last time I read Gusev, I came away with the impression that it should be interaction effects, as he argues that molecular genetic methods control for those. But I'm no expert on the topic, and maybe I'm misremembering.
Note that, although that the story you linked doesn't mention it, the study that found people with Norman names overrepresented at Oxford was by Greg Clark, who argues that very long-run persistence of social status is driven by genetics and highly assortative mating (0.8 on a latent measure of socioeconomic competence). So in this case it wouldn't be necessarily be confounding, because persistent differences between people with Norman and Saxon names might well be genetic.
The idea that truly exogenous economic endowments can persist over several generations has pretty weak empirical support, as I believe you mentioned in a post several years ago. Studies finding otherwise are likely confounded by genetics.
But there was also a St. Louis Fed study finding that the intergenerational correlation for residual wealth (i.e. wealth greater than or less than that predicted by earnings) is only about 0.2. Except perhaps in the most extreme cases, wealth has to be actively maintained and fortified by each successive generation, or even very wealthy families fall off the map within a few generations. Consider that less than a century after his death, there are no longer any Rockefellers on the Forbes 400 list.
Wouldn't your theory require that Normans started out smarter than Saxons? I'm not claiming they can't possibly be, just that I can't really figure out a reason to think this, and it would make more sense that the persistent Norman-Saxon difference is because Normans have always been the nobility (and therefore had more money) since the Norman Conquest.
What if social status gives you access to higher-quality mates? You're rich, so you marry other rich people, some of whom got rich through extraordinary ability. Through this mechanism, it seems like it should be possible to consolidate temporary social status into a permanent genetic advantage.
Well, in a patriarchal society like Medieval England, it's more like you marry women whose fathers got rich/powerful through extraordinary ability.
Mr. Darcy, with his sexy Norman name, marries the smart Elizabeth Bennett in "Pride and Prejudice."
His great-great-great granddaughter, Marcy, did not marry as well.
https://marriedwithchildren.fandom.com/wiki/Marcy_D%27Arcy
I always assumed this was part of the logic of noble birth. Old bloodlines with an accomplished past bring in new blood with an accomplished present, incentivied by the advantages of noble family, hopefully leading to stable excellence that adapts over time at a moderate rate.
Primary sources rarely talk about these things outright, but religion be damned, these people bred every animal they could for trait selection, and they knew what they were.
The Churchill lineage produced the great Duke of Marlborough, victor over Louis XIV's army at Blenheim, and his famous Duchess. Then they didn't do much for a couple of centuries, then many generations later they produced the brilliant but unsuccessful Randolph Churchill and the highly successful Winston Churchill.
What should we learn from this?
Beats me.
"Wouldn't your theory require that Normans started out smarter than Saxons? "
Could be that smarter Saxons preferentially took on Norman surnames, through marriage or otherwise?
But also, maybe selective immigration after the Norman conquest? Norman elites move in, but Norman peasants remain in Normandy.
Yes- this, exactly. The Normans who moved to England as a transplanted aristocracy were not the median Norman, and any genetic distinction could have been fortified in subsequent generations by strategic marriages and selection effects.
My name Sailer is usually spelled Seiler (ropemaker), but an ancestor became mayor of a small town in Switzerland and decided his lineage deserved a classier surname (rather like more ambitious Smiths in Britain tended to change their names to Smythe or even Psmith).
Gregory Clark found that Smythes are more likely to graduate from Oxford or Cambridge than Smiths. Presumably they were all Smiths at one point, but the people who changed their names to Smythe were obvious social climbers.
Yet, social climbing seems to work.
Normans being by far the most effective in Europe and the Mediterranean in the medieval period seems some evidence there was something special about them. The Normans managed to kind of independently become the military elite in England, Ireland, Southern Italy and other places. Military skill is quite g loaded at least according to modern studies.
The Normans got all the way to the New World in 1000.
Are you referring to Leif Erikson? He wasn't from Normandy.
I think that was a different batch of Vikings.
I don't know about lately, but when I last checked about 10 or 15 years ago, there were still four Hearsts on the Forbes 400. They trace their fortune to the Nevada silver rush of 1859, although they hit it rich again in mining strikes in the late 19th Century as well.
But ... most really rich guys on the Forbes 400 were semi-self made like Bill Gates, Jeff Bezos, or Elon Musk. They enjoyed well above average upbringings but then took huge advantage relative to other upper middle class individuals of their generation.
The US might also just be unusual in this regard. I once (peripherally) knew some members of the Fugger family. They are not crazy-rich but still wealthy enough that most of their energy seems to be dedicated to resolving within-family inheritance squabbles.
I wonder if there is a within-family version of the resource curse going on there. If you are extremely able and you come from a family with a big fortune, maybe there's little you can do that is as likely to pay off as well as to make sure to inherit the big fortune and manage it and the family's reputation well. Whereas if you don't come from vast wealth, maybe striking out on your own and starting a business is a better bet.
Thanks for such a wonderful write-up. One thing I’d love to see is more of a deep dive into the stats behind this. This seems like a very high dimensionality problem even without interactions, and I don’t see why interactions wouldn’t be as important as well. I could easily see k > n even with an n of three million, and it isn’t like k=n is a good situation. How the hell are they estimating this? LASSO? Something non-parametric? As an outsider, given the difficulty of the problem they’re facing, it seems like a no-brainier that regressions they’re running don’t add up.
Yeah, I am very bad at stats and not the right person to write this. Both Sasha Gusev and Alex Young are much better, and you might enjoy either of their work - https://x.com/SashaGusevPosts, https://theinfinitesimal.substack.com/, https://x.com/AlexTISYoung/, https://alextisyoung.github.io/
For this "Alex Young thinks that once we get enough whole genomes sequenced (probably soon!) we might be able to use a technique called GREML-WGS to get more definitive answers about rare variants" is there any reason to think that even a sample size of 8 billion would be big enough to detect rare variant effects?
I don't know for sure, but right now I think the best whole genome sample sizes are in the 6 digits, so there's a lot of room to go up!
From an apriori point of view we should expect the relationship between genes and observed features to be incredibly complicated -- basically as complicated as the relationship between computer code and program operation -- and I'd argue the evidence we see is exactly what that model of highly complicated interactions predicts. Namely GWAS misses a bunch of twin study variation and so do simple non-linear models. Or to put the point differently the answer is the narrow/broad gap but broad hereditary is just really damn complicated.
Yes, people cite papers for the claim that non-linear effects don't seem to make the difference. But if you dig in the supposed evidence against non-linear effects (like you linked) are really only evidence against other very simple elaborations of the linear model. Showing that you don't predict much more of the variance by adding some simple non-linearity (eg dominance effects at a loci or quadratic terms etc) is exactly what you would expect if most effects are extremely complicated and neither model is even close to the complete causal story.
I mean, imagine that some OSS project like the Linux kernel gets bifrucated into a bunch of national variants which are developed independently with occasional merges between individual nation's versions (basically it acts like genes under recombination). Or better yet a Turing complete evolutionary programming experiment with complex behavior. Indeed this later one can be literally tested if people want.
We know in this case that code explains 100% of program variation and if you did the equivalent of a GWAS against it you would probably find some amount of correlations. I mean some perf improvements/behavior will be rarely discovered so they will be highly predicted by whether some specific code strings show up (they are all downstream of the initial mutation event) and other things like using certain approaches will have non-trivial correlation with certain keywords in the code.
But no linear model would actually even get close to capturing the true (indeed 100%) impact of code on those performance measurements. And it would look like there were no non-linear effects (in the sense of the papers claiming this in genetics) because adding some really simple addition to the linear model like "dominance loci" or whatever isn't going to do much better because you aren't anywhere close to the true causal model.
So the evidence we have is exactly what we should expect a priori -- genes have some really complicated (essentially Turing complete) relationship with observed behavior and simple models are going to guess at some of that but linear models will do about as well as simple non-linear ones until you break through some barrier.
I think certain twin/sibling studies have themselves been used to try and estimate the relative size of additive vs. non-additive genetic effects, and additive effects were maybe twice as large, though. Non-additive effects are also harder for natural selection to act on, which makes rapid evolution in a trait less likely.
Yes, but the problem is there are 2 types of studies that estimate non-additive effects.
1) Twin study type studies. Well that's exactly the high result which causes the gap.
2) Papers that test for non-additivity by elaborating the linear model in some way to allow for some simple non-linear effect. For instance, capturing the idea that maybe a gene can block the effect of other genes at the same loci.
But 2 is only evidence against the kind of non-linearity presumed in that model (more generally simple non-linearity). My point is that if the true relationship isn't something like mostly additive but maybe with some quadratic terms or multiplicative one but is actually more like a Turing complete interaction then you should expect that adding a bit of non-linearity doesn't improve the prediction much.
So as I said, we see exactly what we should expect consistent with the hypothesis of very complex genetic interaction s.
I'm probably not as immersed in the literature as you are, though I do work in the field of software so... I think I can partially intuit what you mean by non-linear program flow, although I think biological systems are both messier and generally include more built-in redundancy. However, I can certainly see the argument that complex interacting biological systems (such as brains, or even cellular metabolism) must have complex genetic correlates.
However, it's also the case that tiny regions of the genome can have counter-intuitively huge impacts on phenotype (e.g, humans vs. chimpanzees), so... I wouldn't assume in advance what percentage of gene-variants accounted for non-linear gene interactions, or what percentage of phenotypic variance those accounted for.
the hereditarians (the fact they call themselves this is funny, all geneticists are hereditarian) seem to have a weird idea that there's like a full genetic shuffle going on and evolution likes low-variance smooth fitness landscape. both are untrue afaict.
not a lot of shuffling is happening (1-3 crossovers per chromosome pair during human meiosis from memory) and with recombination hotspots the shuffle is somewhat predictable. so a lot of deep structure can be preserved and it's probably beneficial to have deep structure (you don't want an organism which rapidly loses genetic diversity to hit a local maximum and then all dies when the environment changes, similarly you don't want a function that's too full of 'holes' which the organism simply gets stuck in while maintaining diversity - you want an explorable and non-smooth fitness landscape).
note of course all these 'meta-parameters' like where the hotspots are, are themselves subject to selection which makes it harder to reason about.
i tend to agree with OP we should expect really complex interactions. the metaphor I'd use is that it's like we're analysing books by only doing word-counts. except in super high dimensional spaces it's often hard to intuit just how effective this is while still being completely wrong: for instance we can predict a lot from non-functional non-coding DNA simply because statistically it'll be colocated with genes that do matter. if you can make good predictions from junk DNA it's hard surprising you can do really well while wrongly assuming additivity.
i don't know much about the field but the impression i get is that the molecular geneticists largely take it for granted there's complex interactions because they've found really cool stuff in simple organisms. meanwhile you have more classical quantitative geneticists who have barely progressed since Fisher who are so far removed from the cutting edge that they barely interact with their much smarter brethren.
> But from an apriori point of view we should expect the relationship between genes and observed features to be incredibly complicated -- basically as complicated as the relationship between computer code and program operation -- so it's weird to think linear GWAS models should capture most of the variance.
If you only look at how genes get transcribed into proteins and what proteins actually do, and how long and convoluted the chain of cause and effect is to get anything we actually observe on a macro level, then your argument makes a lot of sense. It's ever crazier and weirder than human-readable computer programs.
But: remember that genes and chromosomes get reshuffled and stitched together and slightly mutated with every generation. That means for them to work, they need to be fairly robust.
If you want to think about a computational analogy, I wouldn't pick computer code ~ program execution, but perhaps like the weights in a neural network. Especially since methods 'dropout' (aka 'dilution') that randomly disable artificial neurons seem to work quite well, and not cripple them. See https://en.wikipedia.org/wiki/Dilution_(neural_networks)
(I guess it helps that 'dropout' in neural networks was inspired by an analogy with genes.)
Thanks.
Perhaps intelligence is the single thing that most requires intelligence to figure out? Hence, I'd hardly be surprised that we haven't figured out IQ completely yet.
I don't think the analogy with computer programs is on point.
The big difference is that the genetic code is recombinable. You can take two genomes, mash them together, add a few mutations, and it's still a working genome! You can't do that with computer code.(*) This is a pretty tough restriction, and it puts some severe restrictions on the complexity of interaction. Not that we would understand exactly how those restrictions look like. It doesn't just rule out "it's complicated" either, but it's enough that I find the comparison with computer code wrong.
(*)Actually you can, it's called Genetic Programming. The idea is that you evolve code via recombination and selective pressure by a fitness function, for example, to pass as many test cases s possible. It just doesn't work.(**) Since I work in a related field, I know a couple of people who work on Genetic Programming. The issue is precisely that it's so hard to recombine programs into another working program.
(**) My colleagues may beg to differ. But I would claim that it doesn't work unless you either work with absolutely tiny programs, or move so far away from the basic idea that you are no longer evolving code, but live in a totally different "genetic space".
Educational attainment is a meaningless measurement, at least from the internet survey versions I've seen with categories like "some college" or "bachelor's degree". This misses a lot of education that I think also correlates with intelligence. A Journeyman electrician requires a five-year apprenticeship that includes both on the job and classroom education, where would a high school dropout who joined the IBEW and became a Journeyman electrician be classified?
Both of my parents have a bachelor's degree, I chose instead to enlist in the US Navys Nuclear Power program which at the time was compared to the most difficult programs at top universities. Am I a strike against the heritability of educational attainment since I earned a nuclear Naval Enlisted Classification instead of a degree?
There's a difference between an imperfect proxy and a non-proxy.
Surveys of educational attainment don't classify *you* as an individual well or fairly. But they're still correlated with IQ, on the average. Getting 60% of the answers right on a true false test is not a perfect score. But it's not a 50% either. Surveys using educational attainment can get the equivalent of a 60% correct score and still be potentially useful.
They might be confounded, I suppose, if certain forms of intelligence predicted a person pursuing trade schools rather than college education. If you can be treated as noise, then a signal can still be extracted from the data. If you are a signal in your own right then the survey format might be a problem.
I guess using educational attainment may be...what's the phrase, directionally correct? I don't think this tells you anything more than you could learn from dog breeds, yes, behaviors can be inherited. A retriever will retrieve from not long after birth, and a shepherd breed will herd anything that moves which is always fun when a family with one invites a bunch of their child's friends to a birthday party and the dog will do its best to keep them all together.
I would think these studies aim for more useful information to quantify the degree of hereditariness or something and EA isn't going to get there. A degree may have correlated with intelligence for a period of time since that was the smooth path to a decent paying career, but when the people spending billions of dollars on AI datacenters say the bottlenecks are electrical generation and electricians to run all the cables, then over the next ten years intelligence may be more closely corelated with people who join the IBEW after high school than with those who earned a degree.
Back in 1986 I put an ad in the Chicago Tribune looking to hire a personal computer technician. The HR person asked about academic requirements. I thought about it for 5 seconds and said, "Bachelor's degree required."
But then I got a long letter from an applicant saying, I don't have a college degree because I enlisted in the Navy's submarine nuclear reactor program out of high school, but here are ten problems you are probably dealing with and how I'd solve them for you.
I hired him, and by the afternoon of the first day, it was clear he was smarter than me. While he solved my immediate problems, he also caused me a lot of unexpected problems because because within a few months my boss's boss, who had an MIT advanced degree, was calling me up, irate, to find out why the computer repairman I'd hired was having lunch with the Chairman of the Board to discuss corporate technology strategy.
Great story! I'm just ordinary smart, but yeah, we had more than our share of ridiculously intelligent people through that program, and a few of them also had interesting problems. We had one when I was an instructor at the NY prototype who never missed a single exam question and knew all the answers but had zero life skills. He couldn't drive, bought a plane ticket to Albany airport then started walking down the highway towards Saratoga Springs with his seabag on his back. The police picked him up and drove him to the site. We had to assign him roommates who could drive him to work because he hadn't made any friends who would do it voluntarily. He slowly got a little better, but wow, I had a much easier time assisting the ones who struggled academically.
Thanks for the write-up. I suspect that seeing this kind of confusion and disagreement is a good indicator that we are about to witness some nice chunk of progress from clearing this up, soon.
Especially if (one of the) limiting factors seems to be genetic sequencing and computing power: two things that are still getting cheaper and better quickly.
I know of two secret results I'm not supposed to talk about, by people claiming they've found very large amounts of "missing heritability". Not yet peer-reviewed or confirmed by anything except rumor. I expect one to be out within six months, and the other maybe eventually.
Oh, neat! I had hoped for some progress, but hadn't expected anything quite so concrete and soon.
Interesting ...
It's a funny coincidence, I came across this other post from Stephen Skolnick just yesterday arguing that the gut microbiome strongly predicts schizophrenia risk.
https://stephenskolnick.substack.com/p/schizophrenia
I will admit that Skolnick's essay sets off my 'too good to be true' alarms, but it still seems an idea worth exploring and he rigorously argues for the mechanism.
To what extent does the gut microbiome account for the outcomes we're seeing here, also? Gut microbiome is inherited, to some extent. It is also genetically influenced, and monozygotic twins have more similar gut microbiomes than dizygotic twins, with the divergence apparently increasing as life goes on.
Goodrich et al., 2014 (Cell)
SA: "Finally, she realizes she’s been scooped: evolution has been working on the same project, and has a 100,000 year head start. In the context of intense, recent selection for intelligence, we should expect evolution to have already found (and eliminated) the most straightforward, easy-to-find genes for low intelligence."
Doesn't this assume that increasing intelligence always increases reproductive fitness? In modern industrial societies there's a negative correlation between high intelligence and reproductive fitness, especially in women. I recognize that the same trend might not apply, pre-birth control in more traditional societies. But it at least seems worthwhile remembering that various types of mental facility and reproductive success may not always be positively related in all environments. This suggests at least the possibility of low hanging fruit if those gains involve tradeoffs with reproductive fitness.
Apologies if I've missed your point, there.
Humans evolved to be more intelligent than other species because it was an advantage to us (perhaps because of our large social groups). But brains are metabolically expensive, and there are limits to how big our heads can be while still fitting through a birth canal (hence humans being relatively helpless at birth). So it's advantageous to have higher IQs, but there are tradeoffs and other traits being selected.
Did you ask the polygenic embryo selection folks about collider bias?
In the section comparing Kemper’s sib-regression estimate (14%) and Young’s Icelandic estimate (~40%), you note that the UK Biobank sample may be skewed toward healthier, higher-SES volunteers (so-called healthy volunteer bias, which commonly creates selection effects in medical research). But the implications of such selection effects extend far beyond variability in heritability estimates.
This kind of bias can flip signs when people think they're being clever by adjusting for confounds (collider stratification bias). This is especially a relevant risk in highly selected samples like the UK Biobank, where the same factors that influence participation (e.g., health, SES, education) may also affect the outcomes of interest (mental and physical health). Conditioning on variables that causally contribute to both participation and outcome can introduce more bias than it corrects for.
I wrote a bit about this here: https://wildetruth.substack.com/p/designer-baby-maybe. Munafò et al.'s schizophrenia example illustrates the mechanism more clearly than I can easily argue it for IQ: people at higher genetic risk for psychosis may be less likely to volunteer for studies ("they're out to get me, so I'm not giving them my blood"). This warps the apparent genetic architecture in large datasets. Doing embryo selection against schizophrenia risk based on such a sample, from which high-risk individuals disproportionately self-selected out, could backfire. And parents who paid $50K for it might not know for 20 years (or ever).
Hopefully the real-life practical implications are trivial: if you pay $50K for a prospective 3-point IQ gain, and collider stratification bias turns it into a 3-point loss, no big deal? You probably won't notice over dinner, as long as your kid slept well last night.
But I remain curious how the corporations selling embryo selection address the possibility that they're accidentally giving parents the exact opposite of what they're paying for. My guess is: they just don't tell anybody, and there's something in the fine print saying they could be wrong and aren't liable. Causality probably can't be proven in any individual case, and anyway you're paying for a best effort process, not a result.
This is entirely dependent on whether the “warping” by not having X group in a sample actually warps the sample. Because there’s a high chance it doesn’t warp it at all and the sample is still valid, even without X group present.
Jury is still out on the powers of nature versus nurture, basically. It's slightly complex.
What struck me is the questionable logic of EA as a measure. Scott got into this slightly with grade inflation and economic prospects, but I'd go further. Many of the twin studies took place between 1970 and 2000, or at least the kids were in school during that period, some of them earlier. This corresponds with a high point in educational prestige, and an opening of college to more people on a merit basis, while college was still rigorous. Today, getting a college degree is obviously less meaningful than it used to be, so the value presumed of its presence is almost irrelevant, and I'm thinking this is generational, a reflection of boomer-GenX social philosophy. Expecting the same effects thirty years on is going to be very muddled by social change.
EA is really, really sloppy from a standpoint of correlation with IQ.
One way or another, I still see nature and nurture about the same way. A person's biological design is full of complexities and redundancies, while their experiences will prod adaptation in various directions within the limits of that design. I know people don't want to believe that adaptation can be bracketed like that, but the evidence is everywhere, we maladapt to all sorts of things in the modern world, some of us more than others. People can push their limits, but it requires strong incentives and going too far results in trauma. Or failure. None of this is particularly compatible with Western ideals regarding individual agency and equality, but that's why it's controversial, right?
My sympathies to adopting parents who thought they were getting a blank slate.
Saw recently, don't recall the source, that the average IQ of college students has gone from almost 120 in the 1960's to just above 100 today due to pushing college enrollment on broader populations.
To the point about EA being full of landmines, wouldn't one of those be the observation that so many modern jobs expect a Bachelor's degree? That seems like a shared environmental component that would push swathes of the population to hit certain thresholds they might not otherwise.
A small technical note that I think is crucial. In the first figure (Tan et al 2024)
1. The x-axis is "Proportion of variance in population effects *uncorrelated* with direct genetic effects".
2. Being on the left side means that the variance in population (given by GWAS) can be correlated with direct genetic effects.
3. So for example, properties that are highly due to direct genetic effects are on the left (height, myopia, eczema)
4. Educational attainment is at ~0.5, this means that the direct effect variance is 0.5 of the population effect. The population effect is ~40% and thus the direct effect is 20%, exactly the value that GWAS converges to.
5. We can also verify this by looking on Tan et al 2024. In another figure they show that EA's genome wide correlation between direct genetic effects and population effects is ~0.75, if you take the square, you get ~0.5
How much effort has been put into algorithms decoding complex geneXgene interactions? Also how much of these things could be down to epigenetic expression, and how could this hypothesis be tested?
How do y'all explain Australia? Country literally founded by a prison population, now significantly safer than world average?
I looked into this once; the answer is that although Australia was founded by (a small number of) prisoners, almost all modern Australians are descended from non-prisoner immigrants who came later.
I conceptualise all this as: we have a new method sibregression that does not give the same results as twin studies; this is strong evidence that broad != narrow heritability. Which is very interesting and worthy of a lot of discussion about what it means.
To me it suggests that a significant amount of heritability is chaotic: the genes matter, but small differences have big effects, being "similar" is less informative than we might have expected.
And as you say:
"If nonadditive interactions are so important, why have existing studies had such a hard time detecting them?"
Is it possible that the *fact of being a twin* makes you weird? Twins are more likely to be premature and to die in infancy. Maybe twinhood is "bad for you" (perhaps via having a worse uterine environment) such that twins are more likely to have all kinds of health & behavior problems than singletons? Perhaps having a "bad" gene is more likely to result in a "bad" trait if you're a twin, which will push up the heritability of lots of traits in twin studies specifically?
That wouldn't affect adoption & pedigree studies, though, so if those also show high heritabilities for lots of traits this hypothesis doesn't help.
Twins used to have lower IQ than singletons, but have now converged (probably with better prenatal care). I can't prove that they don't have some kind of unique disposition which results in equal IQ on average but makes it have a different genetic structure, but seems like that would be surprising.
re: footnote 8, Graham Coop had a good critique in my opinion of the Reich paper when it was published: https://bsky.app/profile/gcbias.bsky.social/post/3l4oqx42hil2b
Could you respond to Lyman Stone's "Why Twin Studies Are Garbage"?
I think that Stone makes two basic points:
1. Heritability is not worth measuring or thinking about, because it doesn't mean what most people think it means. A trait can have perfect 1.00 heritability and proven genetic basis, and yet still be 100% environmental in the wild, due to gene-environment interactions.
2. Gene-environment interactions can be both very strong and very difficult to detect. We didn't notice for decades that peanut allergies were >90% environmental in practice, despite having >50% heritability and known genes for susceptibility. Just a tiny difference in the timing of diet accounts for all of that. Because of this, it's impossible to say that fraternal and identical twins have similar environments, which violates the first assumption of twin studies (as you have numbered them above).
What's your take?
Something can be worth measuring and thinking about even if "most people" don't know what it actually means. Most people don't understand quantum physics.
Environmental effects have been studied pretty exhaustively for things like IQ, if GxE was important we would know
> If nonadditive interactions are so important, why have existing studies had such a hard time detecting them?
Ooh, I have an extensive discussion of this in a recent post: https://www.lesswrong.com/posts/xXtDCeYLBR88QWebJ/heritability-five-battles Relevant excerpts follow:
§4.3.3 Possibility 3: Non-additive genetics (a.k.a. “a nonlinear map from genomes to outcomes”) (a.k.a. “epistasis”)
…Importantly, I think the nature of non-additive genetics is widely misunderstood. If you read the wikipedia article on epistasis, or Zuk et al. 2012, or any other discussion I’ve seen, you’ll get the idea that non-additive genetic effects happen for reasons that are very “organic”—things like genes for two different mutations of the same protein complex, or genes for two enzymes involved in the same metabolic pathway.
But here is a very different intuitive model, which I think is more important in practice for humans:
• Genome maps mostly-linearly to “traits” (strengths of different innate drives, synaptic learning rates, bone structure, etc.)
• “Traits” map nonlinearly to certain personality, behavior, and mental health “outcomes” (divorce, depression, etc.)
As some examples: …
• I think the antisocial personality disorder (ASPD) diagnosis gets applied in practice to two rather different clusters of people, one basically with an anger disorder, the other with low arousal. So the map from the space of “traits” to the outcome of “ASPD” is a very nonlinear function, with two separate “bumps”, so to speak. The same idea applies to any outcome that can result from two or more rather different (and disjoint) root causes, which I suspect is quite common across mental health, personality, and behavior. People can wind up divorced because they were sleeping around, and people can wind up divorced because their clinical depression was dragging down their spouse. People can seek out company because they want to be widely loved, and people can seek out company because they want to be widely feared. Etc.
• I dunno, maybe “thrill-seeking personality” and “weak bones” interact multiplicatively towards the outcome of “serious sports injuries”. If so, that would be another nonlinear map from “traits” to certain “outcomes”.
All of these and many more would mathematically manifest as “gene × gene interactions” or “gene × gene × gene interactions”, or other types of non-additive genetic effects. For example, in the latter case, the interactions would look like (some gene variant related to thrill-seeking) × (some gene variant related to bone strength).
But that’s a very different mental image from things like multiple genes affecting the same protein complex, or the Zuk et al. 2012 “limiting pathway model”. In particular, given a gene × gene interaction, you can’t, even in principle, peer into a cell with a microscope, and tell whether the two genes are “interacting” or not. In that last example above, the thrill-seeking-related genes really don’t “interact” with the bone-strength-related genes—at least, not in the normal, intuitive sense of the word “interact”. Indeed, those two genes might never be expressed at the same time in the same cell….
As far as I can tell, if you call this toy example “gene × gene interaction” or “epistasis”, then a typical genetics person will agree that that’s technically true, but they’ll only say that with hesitation, and while giving you funny looks. It’s just not the kind of thing that people normally have in mind when they talk about “epistasis”, or “non-additive genetic effects”, or “gene × gene interactions”, etc. And that’s my point: many people in the field have a tendency to think about those topics in an overly narrow way.
…
§4.4.3 My rebuttal to some papers arguing against non-additive genetics being a big factor in human outcomes:
The first thing to keep in mind is: for the kind of non-additive genetic effects I’m talking about (§4.3.3 above), there would be a massive number of “gene × gene interactions”, each with infinitesimal effects on the outcome.
If that’s not obvious, I’ll go through the toy example from above. Imagine a multiplicative interaction between thrill-seeking personality and fragile bone structure, which leads to the outcome of sports injuries. Let’s assume that there are 1000 gene variants, each with a tiny additive effect on thrill-seeking personality; and separately, let’s assume that there’s a different set of 1000 gene variants, each with a tiny additive effect on fragile bones. Then when you multiply everything together, you’d get 1000×1000=1,000,000 different gene × gene interactions involved in the “sports injury” outcome, each contributing a truly microscopic amount to the probability of injury.
In that model, if you go looking in your dataset for specific gene × gene interactions, you certainly won’t find them. They’re tiny—miles below the noise floor. So absence of (that kind of) evidence is not meaningful evidence of absence.
The second thing to keep in mind is: As above, I agree that there’s not much non-additive genetic effects for traits like height and blood pressure, and much more for things like neuroticism and divorce. And many papers on non-additive genetics are looking at things like height and blood pressure. So unsurprisingly, they don’t find much non-additive genetics.…
[Then I discuss three example anti-epistasis papers including both of the ones linked in OP.]
I thought of nonlinearity too. The twin studies compare the effect of different sets of genes. The DNA studies are predictions using different individual genes as variables. Are these linear predictions? The post said interaction terms don't explain anything. How about polynomials of the individual genes ?
Just wanted to say this is a very good comment. It does seem to me that the more likely source of discrepancy is how we are defining a "trait," which is typically a more or less fuzzy concept.
>The biggest losers are the epidemiologists. They had started using polygenic predictors as a novel randomization method; suppose, for example, you wanted to study whether smoking causes Alzheimers. If you just checked how many smokers vs. nonsmokers got Alzheimers, your result would be vulnerable to bias; maybe poor people smoke more and get more Alzheimers. But (they hoped) you might be able to check whether people with the genes for smoking get more Alzheimers. Poverty can’t make you have more or fewer genes! This was a neat idea, but if the polygenic predictors are wrong about which genes cause smoking and what effect size they have, then the less careful among these results will need to be re-examined.
I've always been rather skeptical of mendelian randomization (MR) studies. In particular for MR to work, you need:
1. the genes to actually affect the phenotype (often true, but as this shows the effects can be weaker than previously thought)
2. no confounding by population stratification / assortative mating (hard to achieve)
3. no pleiotropic effects: e.g., maybe the genes for smoking overlap with the genes for Alzheimers (unlikely, but possible)
Usually when I see bad mendelian randomization studies (e.g. of ALDH2 variants for alcohol drinking -> cancer) they fail assumptions #2 and/or #3. But now it looks like assumption #1 is potentially worse than expected too!
Any time new types of testing and analysis are administered, and they give *wildly* different results than both multiple types of long used analysis and common sense, we should be skeptical of the analysis and results before we throw away anything of value that has gotten us through millennia of human history and centuries of the scientific era.
In a way this debate reminds me of the intelligent design debate, where one side (evolution believers) have mountains of evidence, but often it was more common sense and circumstantial, based on careful observation. This opened the door to claims about intelligent design that were simple at first, but were transformed by rebuttals until they were in many ways just evolution but “not evolution because evolution is wrong”. Anti-hereditarians are getting closer and closer to the intelligent design group in that they are slowly conceding ground in every debate so as to at least throw into doubt the tiniest bit of hereditarian, trading accuracy for vibes as it were.
“Maybe genetics does affect intelligence, but akshually it’s not your genetics, it’s your parents genetics that you also have which confounds even the claim I just made! Let’s talk around in circles and never admit it’s almost all genetic!”
Kinda like how Creationists are very expert on the subject of petrified wood.
Right, any speck of evidence on the X side is evidence that X is right, any speck of evidence on the Y side is just confounding and circumstantial. The (high) influence of heredity is not scientifically bulletproof, but the evidence *against* heredity is even weaker and more tenuous, thus not effectively invalidating the longstanding idea of “the apple doesn’t fall far from the tree” or “like father, like son” concepts that go back to antiquity.
There’s an implicit assumption that heritability is a fixed property of traits, rather than a model-dependent statistic that varies with environment, measurement, and population structure. If twin studies, GWAS, and sib-regression all yield different numbers, is that necessarily a “problem to be solved,” or might it reflect that traits like intelligence and EA aren’t stable enough constructs to expect convergence?
I'm seeing a fair few comments about genetic screening and 'careful adoption choices' that are giving me some serious liberal eugenics vibes, nasty. Not sure using anecdotes about criminal kids to defend IQ heritability estimates is that valid; I believe criminal behavior shows less heritability than IQ?
I wonder if there's a sampling bias here? Like only seeing kids whose biological parents got caught and ended up in the system. What about white-collar criminals, financial fraudsters, and corrupt executives who are smart/wealthy enough to avoid prison? If antisocial behavior is really heritable, wouldn't their kids show it too? But those kids aren't in adoption; they've probably been shipped of to elite prep schools to help continue their gene pool.
If your thinking is based on "vibes", then you might not like blogs by decouplers.
whereas, I do seem to. Go figure? I seem to doom scroll on condescending passive aggressive flexing vocabulary intellectual gatekeeping vibes.
Iirc callous-unemotional traits (those responsible for most psychopathy and often criminal behaviour) shows high heritability
Interesting a trait that's useful in one environment gets pathologized in another. 'leadership material' i guess it is called.
This is my favourite ACX post in a hot minute
Thank you Scott!
"In the degenerate case where the mother and father have exactly the same genes (“would you have sex with your clone?”) the fraternal twins will also share 100% of their genes."
Why is this so? If Mom and Dad both have genes (a, b) at one location, won't sometimes twin 1 get (a, a) while twin 2 will get (b, b)? Agree there's more commonality than normal because there's a possibility of 1 getting (a, b) while 2 gets (b, a), which isn't normally true.
Thanks for the correction, fixed.
Edit: I just saw that some of these points get covered in this post, thank you!
Someone with more human genomic knowledge please shoot me down.
Sequencing technology doesn't get discussed nearly enough in this area. Illumina short-read sequencing/SNP panels have been the major source of data for all of these studies, and they are absolutely delightful at finding SNPs but are crap at anything else. I think it will be appreciated that generally things that impact function aren't SNPs, they are broad changes, and so much human genomics seems to be hoping that the thing that is contributing to a change is able to spotted by being close to a SNP, instead of actually looking at the thing that is causing the change.
Genomes aren't lists of SNPs, they are mostly repeats and 2x150bp isn't going to get anywhere near close to capturing that variation, no matter how 'deep' you sequence. Long-read sequencing (PacBio & ONT, not Illumina's synthetic tech) is clearly better, and continues to demonstrate that there is massive variation that is easy to see when you have a bunch of 20kbp fragment, while almost impossible when you're just aligning little chunks of text to a 3gbp genome.
I work in non-model genomics and long-read sequencing is such a clear winner I keep getting surprised when Illumina gets contracts for these massive human studies. The Human Pangenome Consortium is going to be providing a dataset that is way more useful than anything that's come before. Anecdotally, I hear that for some non-European genomic data they know that ~10% of the data from an individual DOESN'T EVEN MAP to the human reference (but is human genomic data). This is all invisible to analysis, or even worse, just confounds things, as the 'true' causal SNP is somewhere in the data that doesn't get analysed, and so we're stuck looking at noise and trying to make sense of it.
I feel like this is such a blind-spot for human genomics, as it's always about the latest and greatest AI/ML method to try and get some information out, when it's the underlying data which just sucks and doesn't have a hope in actually being linked to function. There was even a point on an Open Thread a few weeks back (Open Thread 374 from tabulabio) asking for people to get in touch about Frontier Foundation Genomic models, with the focus being on fancy new ML architectures.
When I asked ChatGPT to write this comment for me ("Argue that sequencing technology could explain a lot of the Missing Heritability problem") it actually pushed back against me, trying to use the Wainschtein et al. 2022 paper as evidence that '...[this paper] used high-quality WGS (which includes better SVs than Illumina) and still found that adding rare and structural variants only modestly increased heritability estimates", which is NOT TRUE. Wainschtein uses the TOPMED dataset, which is from Illumina short reads. Yes, they do 'deep' sequencing, and yes it's analysed to the absolute hilt with the latest and greatest GatK pipeline and QC to the max. But that claim is false, it's just lists of SNPs, completely ignores huge chunks of the genome and just hopes that the thing contributing to a phenotype is is able to be fished out alongside a SNP.
Another anecdote - an older friend's wife died from a brain cancer. He was an old-school non-model genomicist and got all of the data from the oncologists and various tests and analysed things. All of it short-read, none of it turned anything up, either from his or the various doctors. Long-read sequencing run was eventually done after her death and indicated that it was a splice variant missed by short-reads. It was clear as day in the long-read data, since it didn't need to do fancy bioinformatic de bruijn graphs to figure out the splice isoform - it's just mapping the read against the genome and seeing it clear as day.
Thanks - can you explain more about whether things that are advertised as "whole genome sequencing" are the Illumina method you say is inadequate, or whether the WGS data is fine?
I'm probably not the first to think of this confounder, but it's been estimated that ~70% of human embryos turn out to be non-viable [https://pmc.ncbi.nlm.nih.gov/articles/PMC5443340/]. Do the sib-regression studies account for this? The 40-60% range of shared genes they measure is restricted to the siblings that survive to adulthood. But there may have been many 30% embryos that didn't survive the first month of gestation. Their educational achievement was 0.
In other words, there's a severe selection bias in any study that only looks at siblings who survived to adulthood. A genome that kills one of the siblings before they complete schooling or even get noticed on ultrasound is still an inherited trait.
I think this changes some extremely nitpicky form of heritability defined as "heritability across all embryos rather than just surviving ones", but I think as long as you define heritability as being across surviving humans, this definition is consistent and fine, and all of the existing studies remain comparable (and so it's still surprising when they don't match)
Would someone with strong quantitative chops be willing to explain whether GWAS (and related methods) have enough power to detect small effects—say, 0.5 or 1 percentage point—caused by combinations of 9, 11, or 53 SNPs? I’m particularly curious about scenarios where some of the SNPs are rare. In that case, the number of people with a given combination might be small, so the signal could be lost. If that sort of undetected combination effect happens multiple times—say, 30 or so—could that help explain the gap between GWAS results and what twin studies suggest? I’d love a mathematical explanation that someone with several semesters of college math/statistics could follow with effort.
Excellent post, I would just want to add that Sidorenko 2024 (https://pmc.ncbi.nlm.nih.gov/articles/PMC11835202/) that you cite for the discussion on rare variants also measure heritability using Sib-Regression although not for EA.
They show that correcting for assortative mating their estimates are coherent with twin studies :
"Therefore, if we account for assortative mating and assume that the resemblance between siblings is solely due to genetic effects (and not to common environmental), then our data are consistent with a heritability of 0.87 (s.e. 0.05) for height in the current population (Supplementary Note)."
they also show that shared environment estimates from twin studies are inflated by assortative mating :
"Our results for height and BMI agree with that conclusion in that we find no evidence of a residual sibling covariance (𝑐2) for BMI, while the significant 𝑐2 observed for height is largely explained by assortative mating."
And the WGS estimate is higher for height than the RDR estimates (55%), my take on this is that there is an assumption in RDR that posit that the environment of individuals are independent to each others and don't influence each others in correlation to the relatedness which since it's violated bias the estimate downward. (https://hereticalinsights.substack.com/i/146616013/disassortative-mating)
"Moreover, our estimates of 0.76 and 0.55 can also be compared to estimates from GWAS and WGS data. For height, the SNP-based estimate is about 0.55–0.60 (ref.41) and the WGS estimate ~0.70 (ref. 42; but with a large s.e. of ~0.10). These estimates imply that for height there is substantial genetic variation not captured by either SNP array and, to a lesser extent, sequence data, presumably ultra-rare variants (frequency <1/10,000 not included in ref.42)"
Doesn't assortative mating hit twin studies and Sib-Regression about the same amount, so the uncorrected results should be comparable?
You are right, my bad, from the supplementary note 4 of Kemper 2021 (https://www.nature.com/articles/s41467-021-21283-4) they use the same equation to correct for AM and there is still a 0.1 gap in heritability for the two methods.
Sidorenko 2024 use a slightly different adjustment method to correct for the shared environment estimate becoming negative after correcting for assortative mating (another mystery to solve).
While writing this I just saw that the estimate twin heritability for height in Kemper 2021 is lower than the heritability for height in Sidorenko 2024 using sib-regression. Missing heritability solved I guess.
Gusev has a point when he says that when adjusted for assortative mating some of the high twin estimate for height or even sib regression get near or above 1 (when you read the supplementary note of Sidorenko 2024 the crude adjustement for a r=0.23 assortative matting in height push the estimate to 0.98 ! )
Turkheimer's quincunx metaphor is my personal mental model for the missing heritability phenomenon. A blog post is here: https://ericturkheimer.substack.com/p/a-speculative-explanation-of-the-e53 though you'll have to look at the few previous posts to fully understand it. It seemed like you dismissed this very briefly under non-additive effects and I don't have the technical background to grok those papers, but Turkheimer takes this hypothesis seriously so I figure it's serious.
Here's my summary:
Imagine a quincunx (aka a Galton board). You are a ball falling down the board. It bounces off a lot of pins, and it eventually falls into a bin at the bottom. That bin at the bottom is the outcome we're interested in (EA, IQ, etc). But the pins aren't random -- the pins are your genes, environmental variables, etc. So some pins are biased left or right, which influence the bin you end up in.
What Turkheimer shows is that in this model, if two people have the exact same quincunx (the exact same weights of pins), then they are very likely to have similar outcomes. Those are identical twins. If all your genes are the same, you are very likely to be similar. This reinforces the hereditarian position, and also makes a lot of intuitive sense. (One tricky part here is that it must be the case that mostly-the-same quincunxes (siblings, etc) cause similarity as well to a lesser degree. That's the biggest challenge to this mental model.)
But Turkheimer also shows that biases in pins are very very hard to detect in the non-identical-twin situation. Each pin only has an effect in the context of the entire system. For one person, a particular pin could have a large impact (maybe all the pins to the left bounce left, and all the pins to the right bounce right, so that pin matters a lot) but for another person that same pin has no influence at all (the pin to the left bounces right, the pin to the right bounces left, so it's like a funnel and that pin doesn't matter). So "in the wild," away from the twin context, it is very hard to come up with pins that matter in a GWAS type of way. But pins do matter! Genes do matter! They just happen to matter in this way where the whole package has a large influence on an individual, but individual pins are really hard to pick out in a practical way.
Turkheimer would emphasize that this is a simplified mental model. Our genes are not actually a quincunx -- but the interactions are much more complex, so if he can show that a quincunx has that sort of irreducible complexity, genetics must be even more complex.
I find this mental model really interesting and I actually wrote a blog about how I find it a helpful way to think about education and schools with some GIFs modeling the quincunx part if anyone is interested: https://fivetwelvethirteen.substack.com/p/the-quincunx
I think Turkheimer’s analogy here, given that he says that some of the quincunx pins are environmental, is dramatically failing to capture the fact that adult identical twins who were raised by different families in different cities are very obviously similar in all kinds of ways, indeed about as similar as if they were raised in the same family. (…With a few caveats & exceptions that I discuss in §2.2 here → https://www.lesswrong.com/posts/xXtDCeYLBR88QWebJ/heritability-five-battles )
That's a fair criticism of Turkheimer's position but I think you can rescue the broader idea by just saying that the quincunx is genetics only and the final bin is the sum of the genetic influence. To me, the main insight is that a complex system can have emergent properties that aren't easily correlated with the individual components of the system. And that's the core of the missing heritability problem: when we look at the whole system with twin studies we see high heritability, but when we try to find specific causal genes it becomes really messy.
I do think non-additive genetic effects are very important for adult personality, mental health and behavior (but not for other things like height and blood pressure). I talk about that a bunch in my post; see excerpts that I copied into a different comment → https://www.astralcodexten.com/p/missing-heritability-much-more-than/comment/129514879
Amazing post, thank you.
Something that comes to my mind, having just read Pearl’s Book of Why, is whether collider bias or other common problems of regression in the hands of folks not very savvy about causal inference might explain the inconsistencies in the findings.
I mean, there is a lot of regression going on here, a lot of moderating and mediating variables, so it seems easy to make a mistake with the regression model.
Nice, I am a bioinformatics PHD working at least somewhat with GWAS and SNPs and from a Gell-Mann amnesia effect perspective this post was very reassuring.
My guess is also that the rare/ultra rare variants/combinations are responsible.
Why is "principal components" in scare quotes?
Lots of identical twins were raised to deny they were identical.
For example, in the 2004 Olympics Paul Hamm won the gold medal as the best all-around male gymnast while Morgan Hamm was 5th. But Paul and Morgan's parents doubted they were identical because their hair whorled in different directions.
Similarly, the nearly 7 foot tall Lopez twins of the NBA have done a great job over the decades of acting non-identical: they wear different hair styles and different clothes.
Great post!
I think GxG is basically impossible to measure, so saying nobody found any GxG yet isn't really meaningful. It's harder to measure than rare variants; why believe in rare variants but doubt GxG?
Also, note that additiveness is sensitive to nonlinear functions of your measure. For example: if heritability of IQ is purely additive, then heritability of IQ^2 (the square of IQ) would not be. To posit no GxG at all is to say that the IQ scale is perfect and cannot be squared (and EA scale, and BMI, etc too).
A few final notes:
1. I think intuition from real life strongly suggests that everything is GxE (gene environment interaction); think, for example, about obesity. All the models assume GxE is zero in order to calculate the percent heritable.
2. You say assortative mating cannot bias downwards the heritability estimates from twin studies. That's only true if the twin studies don't already adjust for them! Many hereditarians cite 80% h^2 for IQ, but they get this by already adjusting for AM (and maybe even for attenuation). The can easily over-adjust! That would bias the estimates upwards!
3. Related to 2, I noticed that you quoted 60% for the typical twin-study heritability estimate, even though hereditarians (including you in the past) like to cite 80%. That's a substantial gap. Are you saying you now believe the 80% is wrong? Can you be more explicit about this disconnect?
I find the idea of reading modern scientific findings or social innovations into ancient myths and religions very compelling. As if the ancients through cultural evolution were able to find fundamental truths about reality without having to actually understand them. In light of that interest, here we go:
The ancient Greeks had the concept of generational sin. When someone committed a truly horrendous sin, like feeding your brother his own children as an act of revenge (https://en.wikipedia.org/wiki/Atreus), the gods may not punish the man directly, but leave punishment for a future generation.
Atreus (Progenitor of the house Atreides. Dune anyone?) was grandson of Tantalus, infamous for feeding Zeus his own son, and being punished by eternally having food and water just out of reach (origin of the word: tantalizing). Atreus echoed his grandfather's sin by feeding his brother his nephew's corpse and usurping his throne. Atreus' son, Agamemnon (famous from the Iliad) killed his own daughter, Clytaemnestra, in order to be allowed to attack Troy. In punishment for that, he was killed by his own wife upon returning victorious from Troy. His son, Orestes (grandson of Atreus) was given the "divine madness" (basically Schizophrenia) as punishment and it was only through the gods intervention that the cycle of inherited sin was synbolically broken. It was only ultimately ended when a new lineage, the Heracleidae (descendants of Herakles) invaded and taking over their lands, loosely representing an invasion of Northern Dorian peoples taking over the Achaean society that was there.
The lesson I take from this is, the Greeks understood the heritability of preponderancy for antisocial behavior. If your father or grandfather did something unspeakable, you were by extension unclean, and shunned by society because of that. If there is some actual truth to this, and children of men who are so antisocial as to practice cannibalism or other terrible sins, were poisoned by a Miasma that would carry over between generations, then perhaps there was a collective social benefit towards shunning the children and grandchildren of heinous criminals. The societies that practiced this, while fundamentally less just on an individual level, would have been able to better weed out antisocial behavior, and in the long run, outperform societies that didn't have this cultural meme.
The lesson at a societal level is that if your rulers have 6 generations of inherited sin and insanity, you will be invaded by your neighbors and made into slaves.
Of course in the modern day we've moved past that, and if you think hard enough I think you can come up with a "reasonable" justification for literally any social practice, but I always find it interesting to think about.
Educational Attainment is obviously both part Nature and part Nurture (e.g., did your ancestors pony up enough money for you to finish college?)
1. I wonder when we're going to have a good computational model of personality to just directly track these issues on a gear level, like what tradeoffs are solved in what proportion, what inefficiencies are present in the Turing machine that corresponds to one's intelligence, etc. (Sounds easy, right?..)
2. Is it possible that any single study (i. e. the Iceland study) introduced a trivial calculation error, or a different calculating method that at least partially accounts for the discrepancy? Are the datasets in question open-source, did people reproduce their calculations and make sure it's not that?
I wonder how much in vitro fertilization might impact more recent twin studies. Also the fact that we have mitochondrial DNA as well. So more and more data to feed into the maw of the ai models. De novo mutations also seem to impact certain genetic areas as well. Maybe we have a gene that predisposes to de novo mutations.
This is folly isn’t it? Still looking for scientific bases for subjective social constructs after all these years…
Consider syndrome vs. disease. It seems to me that the likely cause is that most of the things being measured have multiple independent ways of causing the measured result, and the genes that facilitate one of them don't necessarily facilitate another. Dawkins calls this teams of genes that are "good traveling companions". E.g. both Tibetan and some Peruvians have genetic adaptation to high altitude, but they do it in different ways.
So under this assumption (which I think is fairly reasonable) if you measure one gene, you're measuring one part of a collection that *IF PROPERLY ASSEMBLED* would facilitate one trait. But someone else presenting that trait with a different underlying approach would not find that gene helpful.
"An alternative way of validating twin studies... is to check them against their close cousins, adoption studies..."
Using adoption studies to "validate" twin studies seems completely backwards to me. Adoption studies are so obviously flawed that I do not see how they could possibly act as a validation set for anything!
The population of children going into adoption are wildly unrepresentative of the general population of children, unless that population is "children who go into the adoption system". The population of parents adopting children are obviously unrepresentative of the general population of children, unless that population is "parents who are adopting children". Being an adopted child to a set of parents is not like being a biological child to the same set of parents. Losing or being given up by birth parents is generally kind of a big deal.
If anything, the fact that twin studies give the same answers as adoption studies should make us more skeptical of twin studies!
In any case, I applaud Scott for both a) being totally transparent about his completely bonkers epistemic commitment here, and b) giving a good-faith description of the current evidence against his aforementioned bonkers epistemic commitment.
I go to one of those fundamentalist churches where people adopt kids and raise them piously. I think everybody knows by now that there's a big danger the kids will turn out badly compared to their siblings, even though there's considerable success in training them to behave well as kids. These adopted kids are often foster care kids, and that's going to add a lot of noise to any study. They don't have normal environmental differences; they have huge ones-- fetal alcohol syndrome, autism, abandonment as babies, abuse, . . . So for finding the environmental share for the US on average, they're not ideal.
Throughout this entire article, the “computational intractability” buzzer was ringing in my head, yet you only got there at the very end.
Isn’t it obvious that the interaction of genes quickly reaches a combinatorial explosion?
Is there any psychometric data set that comes even close to covering even the tiniest fraction of theoretical gene combinations?
Don’t we have good ideas for addressing the theoretical limits of these gene studies?
I am alarmed at this statement about why we care if gene studies show that heritability may be misunderstood.
“Not doctors. So far this research has only just barely begun to reach the clinic. But also, all doctors want to do is predict things (like heart attack risk)”
We absolutely care if doctors over-treat or apply dubious preventative treatments based on unreliable associations.
Do I want surgery removing my prostate and rendering me incontinent because of a false read on my prostate cancer risk?
Correlation versus causation always matters when a reading is used as the basis for an intervention.
The next paragraph criticizes the use of these scores in epidemiology. Actually, even if all you have is variable with a correlation, it can be useful as an instrument in multiple regression analysis.
I thought this little-read article (written by an astrophysicist) does an excellent job of explaining why we wouldn't expect GWAS studies in their current form to get anywhere close to "true" heritability: https://coel.substack.com/p/gwas-studies-underestimate-the-heritability. Notably, the author doesn't rest his argument on the "rare genes of large effect" factor.
To roughly summarize his article: We would expect the code for developing a human-level intelligence to be extremely complex. There are ~3 billion nucleotides in the human genome; these GWAS studies rest on the assumption that only a few hundred to a few thousand SNPs are entirely responsible for creating human intelligence. That seems wildly inaccurate. As the author argues, could you write code for developing a human-level intelligence with only a few thousand instructions? ChatGPT is based on hundreds of billions of neural-network weights (though admittedly this is the end-product, not the "recipe" for producing it like the human genome). If intelligence is based on tens of thousands (or even more) SNPs, then our current GWAS studies are simply statistically incapable of finding them.
On top of that, GWAS is only looking at a very simplistic model of gene-behavior causation. It assumes an "additive" model of heritability which simply sums the measured effect of each SNP, whereas in reality the recipe for intelligence is likely caused by subtle and complex interaction effects among genes as well. (The author also mentions that SNPs, which are all GWAS studies measure, are only one type of genetic variation. This seems like it could be important but I don't understand what it means.)
In contrast, twin studies take into account the complexity of developing intelligence by simply observing the output of the incredibly complex underlying genomic causes. It's therefore not surprising at all that they give much higher heritability scores, whereas GWAS studies, limited as they are, give nothing more than a lower bound for heritability.
How do they adjust EA for non-college education? In my job I interact with many elderly. I ask them what they did. I live in a high-tech area so frequently I hear 'engineer'. But when I ask them where they studied, I am more likely to hear 'I got out of the army and went to work for Sunstrand as a mechanic and they trained me' than to get the name of a university. Very smart folks, but formal education is high school.
Total non-expert speculation here. Has anyone looked into whatever the genetic equivalent of “sequence of returns risk” is? In the market it matters enormously if there is a crash immediately after you retire versus a decade after you retire and your investments have had extra time to grow.
I.e as the body is built, different genes will apply in sequence. It could be that early-in-the-sequence genes (about which we currently know little) will have an outsized effect. It may be that some genes only activate or have an impact conditional to other genes activating or having an impact. Etc. Or perhaps from another perspective the mistake is looking at the genome as if it is static and one dimensional when in fact it is self-referencing and applies across dimensions of time and space.
I'll be honest, I haven't read word for word your post, but I read well enough. I will continue to read as i have time. I think the answer doesn't exist yet but will exist soon considering ai. The true answer is having behavioral patterns linked to endocrinology. Many don't take psychology seriously, but it is what we have. I haven't seen any reliable studies regarding hormones and behavioral traits.
Regarding educational attainment, I think it's also important to consider diet. If people do not get enough protein and fats, they will not be at their peak intelligence. Addictions of any sort will lower iq.
I am of the mind that virtually everything is inherited. Even broad behaviors amongst the population. We can connect hormone levels to DNA and hormones to behavior.
Great post.
Was surprised to find that Scott used o3 to help research this, and found it helpful, despite the continuing hallucination problem. Let's say it gives you certain hallucinations you know you need to check, like citations, and you check them and strike that, fine. What if it's negatively-hallucinating, insistent that something does NOT exist when it does? What if it has numerous unknown-unknowns that it doesn't realize are relevant to the research question?
It seems like you probably need at a *minimum* to have Scott's level of mastery and familiarity with the subject going into the project in order to have any chance at spotting the places that o3 screwed up, and knowing why, and knowing where to look to correct it. The entire field of research needs to have been reduced to known-unknowns for you the author before you can safely employ the assistance of AI tools.
Here he had extensive layers of human review/assistance by specialists in the field, and had previously researched the topic. But I just don't know why anyone would bother to use AI for research when the output is clearly unreliable absent such review and input, in ways that a researcher with incomplete knowledge won't be able to always detect.
Twin study consistency in EA might be because EA used to be a better measure in the past and twin studies are disproportionately from the past?
Why do we expect genes to have independent linear effects? From gene to protein to trait to life outcome, there are many nonlinear effects and interactions that we could imagine — is the problem really our imaginations? Hunting for genes that linearly independently correlate with causally remote life outcomes seems like looking for your keys under the lamppost. But I’m not really qualified to dig into the studies that you say disprove interactiond.
One easy possible confounder for twin studies would be if some aspect of your early childhood environment mattered a lot for some trait (IQ, kidney function, BMI, whatever) and also a lot of that environment was a function of your appearance. That would give you inflated heritability results for identical vs fraternal twins.
It doesn't seem so likely to me that this is important, though--it seems like adoption studies make a pretty strong argument that there's not much impact of variation in early childhood environment within normal middle-class-and-up norms and most measurable outcomes. (Raise a kid in a cave, beat him regularly, and never show him a book, and you'll probably depress his IQ. But letting him watch 20% more TV or taking him to amusement parks instead of museums on vacation isn't likely to have any noticeable effect.). But I guess if really cute kids get extra attention at some critical point in their development relative to ugly kids, it might have some effect?
Well, I never thought I would see this day. Finally, the smart people are coming around to what I have been claiming for decades. Inheritability is an interaction effect, thus the main effects (variation due entirely to genes or learning) are not directly measurable. Also, the twin studies are deeply flawed.
Let's start with twin studies. They have problems with their independent variables (IV's), their dependent variables (DV's) and their analysis.
IV's like personality traits and IQ test scores are highly controversial. We don't know exactly what IQ scores measure--except that it isn't "intelligence" the way most people mean that term (something like "an individual's inherent capacity to solve real world intellectual problems"). No one has proven that such a capacity exists as a coherent cognitive process, let alone that IQ scores measure it. Every other personality inventory suffers from similar problems.
The DV is correlation in such traits between pairs of siblings. This is measured two different ways: by degree of genetic similarity, compared to degree of environmental similarity. One of these two measures (genes) is much more precise than the other. These days, they can not only tell you how many genes two siblings have in common, but often which ones.
But not environment. The original twin studies used household to determine similarity of childhood environment. This is far too crude. Many twins are adopted into different households, but by members of the same extended family (aunts, uncles, grandparents, etc.). This means that the twins will know each other, often live in the same neighborhood, go to the same school, play with each other, etc. Since these things are largely undocumented, the data is contaminated, and we can't know by how much.
Finally, there are strong theoretical reasons to believe that evolutionary effects should be interactions between genes and the environment. The environment, after all, is the nature that selects. Therefore one would expect that any species will evolve traits that engage features of the environment, which is the mechanism by which an organism can gain an advantage. IQ is pretty useless if it doesn't give one a decisive advantage *somewhere*, and if it's mostly useless, why would it be selected?
In terms of analysis, in the presence of a strong interaction effect, the main effects (the effect of each key variable by itself, in this case genes and the environment) lose their meaning. It's easy to illustrate: Imagine that intelligence is 100% inherited. Two children with identical genes grow up in two different locations: one is a wealthy suburb in a household with plenty of resources, the other in a war zone, rife with disease and natural disasters. Do you expect these two to do equally well on IQ tests, Educational Achievement, or any other measure of life performance? One is pretty clearly at risk of leaving fewer descendants than the other, but remember, their genes are identical. There's nothing here for natural selection to grab onto. The main effect of genes is meaningless.
What this implies is that inheritability scores are probably meaningless. Intelligence isn't 60% inherited and 40% learned. It's probably something more like 80% interaction between the two.
Good lord, have a gander at the Sib Regression error bars in part 4! I believe the 95% error bars for menarche contain ~95% of all possible values.
Another random thought that I suspect people in the relevant field have probably already looked into: (I'm a cryptographer, what the heck do I know about behavioral genetics?)
Suppose a lot of outcome differences between people are due to pathogens. For example, if IQ regularly gets depressed a few points if you get a particular viral infection at a particular critical point in your development. (I think there's some reason to believe this does happen, because the month in which you're born does seem to have a small correlation with life outcomes, and how that overlaps with flu season is a reasonable guess about the mechanism by which that works. Also, this is likely to be some of why breastfeeding correlates with good outcomes for kids, though a lot is probably tangled up with social class/education of the mother.)
If this is true, maybe the critical genes to look at for a lot of heritability are the ones that determine your immune response (which MHC receptors you have, what genes you start from when you shuffle stuff around randomly to get T-cell receptors and antibody binding regions, variants in innate immune response) and maybe any rare variants that confer some immunity to common circulating viruses.
There are like a gazillion viruses that circulate all the time, often giving us infections we never even notice, sometimes giving us annoying but not serious symptoms (colds, stomach bugs, cold sores, warts, etc.). And we know that some of these (rubella, polio, zika) can do nasty developmental things to fetuses, babies, or small children.
That would track with at least some of the Flynn effect as sanitation and vaccination made people in-general healthier.
Wow for a rich post. Thx Scott.
Three somewhat orthogonal questions:
- EA seems virtually worthless to pursue topics with the degree of precision desired here. It's not just noisy, it's that the noise is so systemically confounded (national, racial, class, time, wealth, etc.), so any cross-study comparison is likely to be accidentally catching systemic biases. Aren't those trying to do aggressive heritability work that are looking at EA doomed to failure?
- This field seems likely to be suffused with soft "biases" (perhaps there's a better word, less laden with perjorative connotation). Some researchers may despite the eugenic-parallel ideas associated with high IQ heritability. In contrast, commercial researchers pursuing genetic-therapy depend entirely upon that. Some researchers are experts solely in their technique, etc. and need the world to look like a nail to fit the eneds of their hammer. That's all fine in that clean research still drives growth in the overall scientific tree. My question is really whether that tree has visible (more fertile) away from a healthy branch that is pursuing the sunshine of the anti-heritability folks on one side, and another healthy branch on the opposite side pursuing the sunshine available from the "we can rewrite your DNA to control 100% of your superchild's IQ"? Not that I'm seeking some compromise median. I'm just wondering if (what has to be a pervasive) desire to see a specific outcome has limited the rest of the hypotheses/research that could be pursued.
- This work is heritability in humans. Humans offer a ton of advantages and disadvantages for such work. Do less complex and intelligent subjects allow an easier laboratory along some veins that could allow for better staging for extrapolation back to humans? I assume there's a ton of work there - just don't know if it's disproportionately useful or limited.
This problem should be getting as much attention as Hubble Tension, if not more.
Can we induce this measurement phenomenon in rats, or worms, something where we have control of the genome? A quick Google suggests identical twins are rare in nature, but we can possibly produce them artificially for some animals.
Nice overview! A few points I wanted to expand on.
1. I think the post conflates gene-gene and gene-environment interactions; the latter (specifically interactions between genes and the "shared" environment) also get counted by twin models as narrow sense heritability. While I agree there is very little evidence for gene-gene interactions (particularly dominance, as you cite [and, interestingly, twin/adoption studies actually forecast a huge amount of dominance -- another discrepancy we do not understand]) there is quote substantial evidence for gene-environment interactions including on educational attainment (see Cheesman et al: https://www.nature.com/articles/s41539-022-00145-8 ; Mostafavi et al: https://elifesciences.org/articles/48376), IQ, and BMI. In fact, Peter Visscher led a paper that came to the conclusion that twin estimates for the heritability of BMI are very likely to be overestimated by gene-environment interactions (https://pubmed.ncbi.nlm.nih.gov/28692066/). A large amount of GxE plus some amount of equal environment violation seems like a very plausible and parsimonious answer to the heritability gap.
2. You mention epidemiologists being the biggest losers of stratification in polygenic scores, but I think it is important to note a related group: the people who take polygenic scores trained in one population (with a ton of stratification) and directly apply them to other populations to make claims about innate abilities (see: (https://theinfinitesimal.substack.com/p/how-population-stratification-led). This is especially true for Edu/IQ GWAS, where every behavior geneticist has been screaming "do not do that!" since the very first study came out. People like Kirkegaard, Piffer, Lasker, etc. (and their boosters on social media like Steve Sailer and Cremieux) dedicated their careers to taking crappy GWAS data from and turning it into memes that show Africans on the bottom and Europeans on the top. These people also happen to be the court geneticists, so to speak, for SSC/ACX. I don't mean to come off as antagonistic and I'm sure some people will see this comment and immediately discount me as being an ideologue/Lysenkoist/etc so it does my broader position no favors, but this stuff has done and continues to do an enormous amount of damage to the field (including the now complete unwillingness of public companies like 23andme to collaborate on studies of sensitive traits).
3. I'm going to gently push back against the hereditarian/anti-hereditarian framing (which I understand is probably here as shorthand and scene setting). I am personally interested in accurate estimates that are free of assumptions. I believe twin study estimates are of low quality because the assumptions are untestable, not because they are high. I also think the public fixation on twin studies has created some real and damaging anti-genetics and anti-psychiatry backlash and wrong-headed Blank Slate views. People hear about twin studies, look up the literature and find that peanut allergy (or wearing sunglasses, or reading romance fiction) is estimated to be highly heritable and have minimal shared environment (https://lymanstone.substack.com/p/why-twin-studies-are-garbage), start thinking that the whole field is built on nonsense, and end up at quack theories about how schizophrenia is actually a non-genetic fungal condition or whatever. I've been very clear that there are direct genetic effects on essentially every trait out there, including behavioral traits and IQ. If someone were to run a large-scale RDR analysis of IQ tomorrow and got a heritability of 0.9 and it replicated and all that, I would say "okay, it looks like the heritability is 0.9 and we need to rethink our evolutionary models". If anything, large heritability estimates would make my actual day job much easier and more lucrative because I could confidently start writing a lot of grants about all the genome sequencing we should be doing.
4. Lastly, it's not clear to me where the conclusion that well-validated twin studies converge on "similar results" is coming from. To take one example: the leading lights of behavior genetics (Deary, McGue, Visscher, etc) ran a study looking at the relationship between intelligence and lifespan (https://pubmed.ncbi.nlm.nih.gov/26213105/). This is a nice study for us because they put together three large, modern, twin cohorts with IQ measurements, but the heritability of IQ was just a nuisance parameter for them, so they had no reason to scrutinize the findings or file-drawer them. If we look at their MZ/DZ correlations in Table S6 we find that the heritability of IQ was 0.36 in the US sample; 0.98 in the Swedish sample; 0.24 in the Danish sample; and ... 0.52 on average. In other words, all over the place (but averaging out to the nice "half nature half nurture" result you see in books); the authors themselves used an AE model in Table 2 and reported a range of 0.20 to 0.98. This is far greater than the variability we see with GWAS or Sib-Reg, so what are we to make of that?
The anecdotal evidence of “troubled child raised in a good Christian home” seems a strong contender for selection bias. You wouldn’t have encountered the hypothetical well-behaved children from awful parents raised in supportive households. If that population outnumbers those you worked with, the point becomes moot.
Extremely interesting essay overall; thanks for sharing!
Maybe there could be a self-reinforcing effect to silbing similarities? There's the sterotype of the twins that are always together and always do the same thing, maybe small differences especially in early childhood separates silbings and makes them more strangled so that in the future they affect each other less and less. So identical twins have some initial stronger bonds by being more alike and then that develops into a environmental stronger bond (ie they do most activities together and want to be equal), while fraternal twins are a little more distant between them because of initial differences and then don't have that bond that makes them also have a more alike environment.
Theory: IQ is very heritable but also extremely subtly disabled (potentially even environmentally, think "more males born after wars"), because IQ is terrible for reproduction (now as ever) and you need some people to keep passing it on.