Wednesday 20 May 2015

Gone with the Wind


Every now and then a blockbuster paper comes along which, like the 1939 top grossing $1,640,602,400 movie “Gone with the Wind”, carries all before it. It may be that “Meta-analysis of the heritability of human traits based on fifty years of twin studies” by Tinca Polderman, Beben Benyamin, Christiaan de Leeuw, Patrick Sullivan, Arjen van Bochoven, Peter Visscher & Danielle Posthuma in Nature Genetics is such a paper. Although the title is not very snappy, it is a relief to read a paper on genetics where the author list is less than a thousand. This Magnificent Seven have ploughed through fifty years of twin studies to assemble 14,558,903 partly dependent twin pairs drawn from 2,748 publications. After all these monumental labours, perhaps their publication may yet get known as the “Gone with the Wind” paper, because it blows away much confusion, much prevarication and much obfuscation about twin studies.

Nature Genetics (2015) doi:10.1038/ng.3285

The authors find that estimates of heritability cluster strongly within functional domains, and across all traits the reported heritability is 49%. For a majority (69%) of traits, the observed twin correlations are consistent with a simple and parsimonious model where twin resemblance is solely due to additive genetic variation. The data are inconsistent with substantial influences from shared environment or non-additive genetic variation. (This finding has been under-reported).

All human traits contain a substantial heritable element. The blank slate is totally false. If you have colleagues who doubts the twin method or who have difficulty accepting the power of ancestry, shall I repeat for them Rhett Butler’s last words to Scarlett O'Hara right now, or is it better that I tell you a little more about the findings?

I expect you have an interest in the results on cognition, so rest easy, heritability is high, though not as strong as for skeletal, metabolic, ophthalmological, dermatological, respiratory, and neurological traits. Usually there is a big difference (top line of figures) between the high correlations for monozygotic and the lower correlations for dizygotic twins, showing a strong genetic effect. The exception is social values, in which the environment makes a bigger contribution than usual, though not quite as big as heredity.





Cognitive traits correlate 0.646 in identical twins, 0.371 in fraternal twins, with miniscule error terms of .01 in these enormous samples. An additive model seems appropriate for cognition.

They conclude: Our results provide compelling evidence that all human traits are heritable: not one trait had a weighted heritability estimate of zero. The relative influences of genes and environment are not randomly distributed across all traits but cluster in functional domains. In general, we showed that reported estimates of variance components from model-fitting can underestimate the true trait heritability, when compared with heritability based on twin correlations. Roughly two-thirds of traits show a pattern of monozygotic and dizygotic twin correlations that is consistent with a simple model whereby trait resemblance is solely due to additive genetic variation. This implies that, for the majority of complex traits, causal genetic variants can be detected using a simple additive genetic model.

Approximately one-third of traits did not follow the simple pattern of a twofold ratio of monozygotic to dizygotic correlations. For these traits, a simple additive genetic model does not sufficiently describe the population variance. An incorrect assumption about narrow-sense heritability (the proportion of total phenotypic variation due to additive genetic variation) can lead to a mismatch between the results from gene-finding studies and previous expectations. If the pattern of twin correlations is consistent with a substantial contribution from shared environmental factors, as we find for conduct disorders, religion and spirituality, and education, then gene-mapping studies may yield disappointing results. If the cause of departure from a simple additive genetic model is the existence of non-additive genetic variation, as is, for example, suggested by the average twin correlations for recurrent depressive disorder, hyperkinetic disorders and atopic dermatitis, then it may be tempting to fit non-additive models in gene-mapping studies (for example, GWAS or sequencing studies). However, the statistical power of such scans is extremely low owing to the many non-additive models that can be fitted (for example, within-locus dominance versus between-locus additive-by-additive effects) and the penalty incurred by multiple testing. Our current results signal traits for which an additive model cannot be assumed. For most of these traits, dizygotic twin correlations are higher than half the monozygotic twin correlations, suggesting that shared environmental effects are causing the deviation from a simple additive genetic model. Yet, data from twin pairs only do not provide sufficient information to resolve the actual causes of deviation from a simple additive genetic model. More detailed studies may identify the likely causes of such deviation and may as such uncover epidemiological or biological factors that drive family resemblance. To make stronger inferences about the causes underlying resemblance between relatives for traits that deviate from the additive genetic model, additional data are required, for example, from large population samples with extensive phenotypic and DNA sequence information, detailed measures of environmental exposures and larger pedigrees including non-twin relationships.

By all standards of academic debate, this is the mother of “F*** Off” samples, which should lay low five decades of quibbling about the twin method. Author Beben Benyamin has given reassuring interviews saying that it is not a case of “genetics versus environment” but genetics with environments (which was always understood by researchers). The genetic component is increasingly understood, the environmental component remains vague, with ad hoc speculations about shared variance which are usually not validated.  Despite this massive paper being reported in The Guardian (though without drawing attention to the findings on cognitive ability) I fear that many journalists, commentators and researchers, in the attributed words of a British Trade Union leader turning down a management offer he considered too low, will treat it “with a complete and utter ignoral”.

Where do the heredity sceptics go now?

"Frankly, my dear, I don't give a damn"


  1. "The exception is social values, in which the environment makes a bigger contribution than usual, though not quite as big as heredity."

    Not so fast on that one. :) A big issue is that is this study makes no attempt to correct for the various factors that bias estimates from twin studies. One of them is assortative mating, which is pretty high for social values. This inflates the shared environment at the expense of heredity.

    This is in addition to other problems, like measurement error, which tends to attenuate heredity (appearing as unshared environment).

    This is a general problem with meta-analyses in general: you get out what you put in. They really can only correct for random sampling error, not much else.

    Now, all that said:

    "By all standards of academic debate, this is the mother of “F*** Off” samples, which should lay low five decades of quibbling about the twin method."


    1. @JayMan - you are absolutely correct in noting that the classical twin design cannot/does not take into account assortative mating and that if it is important it will inflate the estimate of shared environment at the cost of the heritability. We would need to include partners and we actually thought of including studies that had information on correlation between partners. However, such studies are still very rare and certainly not available for many traits. Overall, the estimate of c2 is low (17% when using the reported ACE estimates and 4% when using the less biased least squares estimates), so across all traits I do not expect that assortative mating has a large influence. Ofcourse, this can be different for specific traits, like the social ones.
      Same for m.e.: few studies incorporate this explicitly.
      D.P. [author on the NG study]

  2. AnonymousCoward20 May 2015 at 20:06

    "Roughly two-thirds of traits show a pattern of monozygotic and dizygotic twin correlations that is consistent with a simple model whereby trait resemblance is solely due to additive genetic variation. This implies that, for the majority of complex traits, causal genetic variants can be detected using a simple additive genetic model."

    Thomas, if you have access to the paywalled supplements and can gdrive them, or can see any interesting r values which make up the hundreds of traits constituting "higher-level cognitive functions", that would be awesome.

  3. James,

    Thank you for posting that article to google drive!

    Incidentally, you may be interested in reading a blog post I put together on "concentrated poverty" and the academic achievement gap and a brief addendum to it.


    1. Thanks. What an extremely detailed study you have done! Found the first few slides made strong points.

  4. Democracy is based on consensus. What does this mean ** It starts with the idea that the opinions of all the experts can produce a more holistic picture of what is only an analytical genius to do it. However, the consensus is manipulated from its root by means of ideology advertising. Therein lies much trouble. First, in fact, some kind of consensus should be imposed. For example, crime =homicide by futile motivation, is imposed as a consensus, rather than just something that everyone knows. If everyone knew then we would not have many people killing each other for futile motivation, and I'm not just talking about the low-functioning criminal.
    Therefore, the dualistic idea '' consensus is bad, non-consensus is good, '' is a lie, because depend on consensus.
    If one or more very smart people (insightful) have an insight into some ethical issue that can improve our lives in a qualitative way, then there is even a need for consensus to be applied. Of course, from the moment that this hypothetical ethical issue is fully analyzed, decanted.
    The mass media and reporters' opinion leaders' ', exist to manipulate the results, in bad faith or based on the most sincere good intentions of a typical silly, outgoing and self confident, the elite has managed to keep the cattle just through this deception tactics.
    The self declared '' opinion leaders '' are the propagandists of the new beliefs that have substantially meanings effects in the real world, like the idea of human races nonexistence. They need to be less intellectually perceptive and 'blindly' believers so that they can spend a credible idea, because of their profession (intellectual hierarchy) and his sincere belief.


  5. Garbage in = Garbage out.

    Applying fancy statistical methods to a simplistic cartoon model doesn't make it other than a simplistic cartoon model.

    The "null hypothesis" is that 2*r(MZ) – r(DZ) = 0 (variance in DZ twins is 2x that in MZ twins).

    “The proportion of single studies in which the pattern of twin correlations was consistent with the null hypothesis that 2rDZ = rMZ was 69%.”

    So, 31% of studies are not consistent with the null hypothesis? This “proves” the null hypothesis is correct?

    No, what it proves is that the simplistic models they are using (ACE, ANE and AE) are hopelessly simplistic (as has been known for decades).

    MZ twins share an environment where they synchronously grow from 1 cell to 10^12 cells. DZ twins share an environment, but the development of DZ twins is not synchronized, so the timing of environmental change on development is not in phase.

    We know that all aspects of development are non-linear, time varying, and coupled, and so will exhibit the butterfly effect. How many degrees of freedom are there? With ~10^4 genes and ~10^6 proteins, how many (non-linear) gene-gene and protein-gene and protein-protein interactions are there? (10^4)! That is a big number. Per cell. Assuming they are zero because there are too many to deal with gets us no closer to understanding the process of development.

    1. The purpose of twin studies, especially univariate ones, isn't to understand development. You seem to confuse individual development and individual differences. Twin studies are concerned with the latter. A failure to understand that twin studies deal with between-individual differences, not within-individual processes, is the source of so many nonsensical criticisms of the twin method. No matter how interactive the process of individual development is, that does not mean that between-individual differences cannot be explained by a primarily additive model.

  6. How do you square your views with evolution? Never mind, obviously the most parsimonious explanation is a conspiracy of butterflies!

    1. Evolution configured the human genome to generate a phenotype from two haploid gametes with a total of ~3x10^9 base pairs. How is that much information able to generate a brain of 10^11 cells with 10^15 connections? The vast majority cannot be specified genetically.

      The genome doesn't even specify the number of heads a person has.

      What basis is there for thinking the genome specifies the minutia details of how a brain is connected? Language is highly heritable (most people speak the same language their parents speak). We know language is not genetic because essentially any human infant can learn essentially any language as a first language with no accent.

    2. "Evolution configured the human genome to generate a phenotype from two haploid gametes with a total of ~3x10^9 base pairs. How is that much information able to generate a brain of 10^11 cells with 10^15 connections?"

      Umm, no. There are 4 possible base pairs so, technically, the (naive) upper bound estimate for the number of possible permutations is much much larger than that, i.e., 4^(3x10^9). In practice the number isn't quite that large, but still there are a phenomenally large number of permutations possible.

      "The vast majority cannot be specified genetically."

      You don't know that and there is no good reason to think that there needs to be a 1:1 correspondence between this number and the number of cells, connections, etc. In fact, there is much evidence that it doesn't work like this at all.

      "We know language is not genetic because essentially any human infant can learn essentially any language as a first language with no accent."

      And yet strangely we have failed to teach other species to speak English. I guess that's just stereotype threat (human privilege?) or something though, right? :-)

      "The genome doesn't even specify the number of heads a person has."

      I haven't known too many people born without heads, but perhaps you are an exception to the rule.

    3. Yes, you don't know anyone born without a head because being born without a head is fatal in a few hours. Being born without a head is not uncommon. Sometimes MZ twins are discordant for being born without a head. None of these twin studies considered twins with anencephaly because no individuals with anencephaly survived long enough to be in one of these studies.

      If MZ twins can be discordant for being born without a head, what basis is there for “having a head” being genetic?

      If the degree of difference of “having” vs “not-having” a head is not genetic, what basis is there for saying that much smaller differences in brain structure are genetic?

    4. If MZ twins can be discordant for being born without a head, what basis is there for “having a head” being genetic?

      Twin studies are concerned with the causes of variation in traits, or heritability and "environmentality." Heritability is not synonymous with "being genetic."

    5. "Being born without a head is not uncommon.....If the degree of difference of “having” vs “not-having” a head is not genetic, what basis is there for saying that much smaller differences in brain structure are genetic?"

      You are wrong and you have exceeded my kook quotient for the week, so I'll leave it at that. Good day.

  7. "Where do the heredity sceptics go now?"

    They'll probably continue trying to explain away, rather than explain.

    Plus, of course, they'll cry "Burn the heretics!".

  8. Learning capacity is like metaphorically speaking, elastics. Some people, when exposed a specific domain, can improve very well this capacity. For example, a mathematically gifted boy have larger elastics-learning than a average mathematically one. I thought specific part of brain can increase when is systematically exposed to certain knowledge. Of course, there are a qualitative and quantitative diversity of individual (and consequently collective ) capacity.
    Specific brain areas elasticity have genetic basis, of course.


  9. "They conclude: Our results provide compelling evidence that all human traits are heritable: not one trait had a weighted heritability estimate of zero."

    In a better world, they would be out of their jobs because there is no value in proving something that is already this ridiculously obvious.

    1. Why is it obvious? It has not traditionally been obvious at all, and the great achievement of behavior genetics has been to demonstrate the heritability of all human traits.

    2. I don't want to be that guy, but... "All human traits."

      OK, primary language spoken in the home? (Heritability = 0) Culturally specific food preferences? (Heritability is indistinguishable from 0) Side of the road you prefer to drive on? (Heritability = 0) Literacy? Well, if we look at the world population, over a reasonable period of time, rather than within a limited community then Heritability = 0 or close enough.

      Anyone tested hair-style yet? Across the world and across time? Just wondering.

      What I find astonishing is that anyone finds this study interesting -- it is a cool big data study, I suppose, but come on. Every damned twin study for the past 50 years has come up with heritability estimates of about .4-.6 for pretty much everything. Meta-analyses are interesting when a) there is dispute about a finding, with different studies pointing in different directions or b) when none of the individual studies reach statistical significance. In this case, neither holds. The result is entirely expected. (Cards on the table: I hate twin studies, but start every talk about them noting that for almost any trait that even might be heritable, they find heritability from .3-.7, with .4-.6 the most common... Everyone, even those of us that think the research is crap, agree that those are in fact the results of the published studies. So -- add them all up! I'm surprised that it is .5? Not so much, no.)

      If you think heritability is a garbage measure (except if you are trying to breed for a trait under a relatively constant environment), you'll think this study combines a bunch of garbage together into a big garbage pile. If you think heritability tells you something interesting, you shouldn't have had any doubts about the previous thousands of studies, all pointing in the same direction. If you thought it was kind of suspicious that the results clustered so much, but that for psychological traits we seem utterly unable to find even a tiny portion of the genes associated with the variance, you'll take the authors confident pronouncements that it is all obviously additive to be evidence that they are delusional. If you like the idea of most of our genome influencing every trait in tiny ways, you'll think their results fit in just fine.

    3. "What I find astonishing is that anyone finds this study interesting"

      The study isn't interesting. That the results are somehow controversial is very interesting.

    4. So somehow combining the results of over 2000 studies comprehending 14,000,000 subjects and a myriad of traits, both behavioral and physiological, just isn't interesting?

      I'll tell you why you don't find it "interesting" -- because you don't want to hear about, and deal with, the results. Even such a massive, pervasive set of findings assembled as they are in the paper into an overwhelming case for vast swaths of greatly heritable traits will not put you off from pushing an increasingly absurd emphasis on the importance of environment over all. Obviously, nothing will affect such a dogmatic view.

      And why is it in any way surprising that few of the genes involved in many behavioral traits have yet been identified? Given that this is also true for all kinds of other traits, from height to BMI to susceptibility to many diseases, why should we expect that behavioral traits will be different? In fact, of course, insofar as the effect sizes of the genes in most of these cases are extremely small and the number of genes involved is massive, identifying them would of course be nigh impossible until massive amounts of genomic data are involved. We will of course fairly soon have assembled data sets with a million individuals, and there's a good prospect that at that point we will be able to identify a substantial number of those genes. Who's willing to bet, I wonder, that they are going to find nothing? And when they do find something, what excuse is Mr. Anonymous going to use then?

    5. What if I did a massive meta-analysis of every randomized clinical drug trial since 1950? It would be huge, and I could estimate the average chance that a drug was effective, and the average effect size for all drugs. Would you find that interesting? informative?

      Why is this study different in kind from that?

      (On another note: OK, we have *thousands* of traits, 2/3 of which, according to the study, are heritable via ordinary additive effects, but, given the failure to find any genes that explain more than a tiny fraction of variance, each of which, to be associated with ordinary additive effects, must be influenced by 1000s of genes. Do you see the problem? OK, sure, universal pleiotropy, but come on...)

    6. @Jonathan: I wouldn't be surprised if hairstyle did have a non-zero heritability. If aesthetic taste is influenced by genes, then what we make and do under the influence of our taste would probably reflect that.

      Obviously, twin studies are a second best option from a scientific point of view. An alien scientist not constrained by human ethical qualms or budgetary constraints would make a large number of human clones and raise them in precisely controlled environments to make a definite determination of what the limits are of environmental effects, what's genetic, and what's random.

    7. "Why is this study different in kind from that?"

      Because their counterparts in medicine generally recognize this data. The majority of health researchers don't routinely ignore pharmaceutical RCTs or pretend like drugs have no effects. If they were still talking about "bad humors" or the standard of care were still leeches, so to speak, the parallels would be much larger.

      "to be associated with ordinary additive effects, must be influenced by 1000s of genes. Do you see the problem? OK, sure, universal pleiotropy, but come on..."

      This problem isn't unique to behavioral genetics. The modest number of hits probably has little, to nothing, to do with pleiotropy and much to do with the fact that these are genes of small effect, i.e., we wouldn't expect them to find many hits given the sample sizes they've been working with. Also, many of these other behavioral phenotypes aren't as well measured and haven't been researched as much genetically (as compared to, say, intelligence). They can't very well produce hits on subjects they're barely even investigating.

    8. "I don't want to be that guy, but..."

      You *are* that guy.

      Gould on Morton, Redux: What can the debate reveal about the limits of data?

      "Lewis et al. (2011) attempted to restore the reputation of Samuel George Morton, a 19th century physician who reported on the skull sizes of different folk-races. Whereas Gould (1978) claimed that Morton’s conclusions were invalid because they reflected unconscious bias, Lewis et al. alleged that Morton’s findings were, in fact, supported, and Gould’s analysis biased. We take strong exception to Lewis et al.’s thesis that Morton was “right.” We maintain that Gould was right to reject Morton’s analysis as inappropriate and misleading, but wrong to believe that a more appropriate analysis was available."

    9. Jonathan, I have to say that your response just sounds like dogma talking.

      Look, I have no idea what you might ever find "interesting", unless, of course it is something you find congenial to believe. If a comprehensive study like this compiling data across so many traits and across so many individuals all tending to show the same thing, that it is additive effects of genes that seem to be doing the lion's share of predicting phenotypic traits leaves you unmoved, well, what does move you? To those of us in the real world, it's important to see how many traits are heritable, how much they are heritable, and in what sense they are heritable. God knows that there are vast benighted hordes of people, many of them with PhDs, many others prominent journalists and writers, who haven't even the slightest clue that this might be true. Perhaps the sheer mass of this evidence will penetrate their thick obdurate skulls or render them just the slightest bit less confident when they pronounce that everything is cultural and environmental.

      And you continue to harp on the fact that relatively few genes have yet been located to account for the heritability of various traits. Really, what is even your thinking here? That, say, contrary to what just about any competent person in medicine believes, there are no genetic disposition to any number of diseases because we haven't yet a good handle on what genes are involved? What is the idiocy here? And, again, if the genes involved in many traits are of exceedingly small effect but vast in number, then we would predict, as scientists, that they would not be located until the genes of perhaps a million of individuals are looked at. Your "argument" amounts to pointing out that indeed we are precisely in a situation we would expect to be in.

      Again, what will happen to you and your ilk when those million individual data points are assembled, and genes start pouring in? Will you all just collapse in a pained, confused heap, with your ideology in tatters?

      How long do you all expect to show your faces in public when you've been arguing such junk, and you are finally found out?

    10. Let me continue the point regarding the fact that we haven't yet located the loci (or at most a small number of them) related to many diseases with an hereditary component.

      What would Jonathan have doctors tell patients with a family history of these diseases, but who have no gene already established to be connected to that disease? That they should take no zero extra precautions because there just isn't any reason to believe that they have a heightened risk? That they shouldn't undergo more frequent colonoscopies or mammograms, on the authority of Jonathan?

      Really, what kind of insanity is being argued here?

    11. Thanks for your lucid point about the "no genes found yet" argument.

  10. "It has not traditionally been obvious at all": good grief, have you never talked to farmers, or people who keep pets?

    1. Good grief, have you never talked to social scientists or historians? There is a good prima facie case for non-heritability of human traits, namely the enormous fluctuations in trait means and distributions over time and place when holding genetic ancestry constant.

    2. That case is only good to people who don't understand basic genetics.

    3. I'm curious, Anonymous, what do you mean when you mention enormous fluctuations in trait means? Have these fluctuations been observed over timespans measured in years, or centuries? Remember, allele frequencies are not static over time. If they were, we would still be living in trees with the chimps. (Also, chimps would still be drifting around in the sea with the other descendants of the last universal ancestor, and I don't know what there would be to talk about.)

    4. Height, weight, and many related anthropometric traits have changed a lot over the last, say, 100 years. IQ scores have increased by up to a couple of standard deviations. Political and social attitudes have changed enormously.

    5. Interesting; the all-knowing Internet tells me that human height has increased around 1.5 inches over the last 100 years, but this is well less than a single standard deviation. IQ scores have increased by as much as 2 standard deviations, and that's quite a lot, but since four standard deviations are needed even to span 97% of the entire Gaussian distribution, this really can't qualify as enormous. You lastly say that political and social attitudes have changed enormously, but I'm not sure how this would be measured. Yes, majority opinion has changed in way that might seem enormous, but even a shift of, say, 10% on any given issue can completely replace one majority with another.

      But more to the point, none of these environmental shifts eliminated variation. It would be necessary for short to converge in height with tall, for poor to converge with rich, or for right to converge with left for zero heritability to have been plausible. Instead, as education, medical care, and social services become standardized over the last century, variation in most traits we're discussing has remained or even increased. So I'll reiterate what DK wrote above: If historians and social scientists understood basic genetics, they would never have made any case for zero heritability, and this meta analysis would be widely seen as redundant with common understanding.

      The real issue has never been whether most traits are heritable; the real issue has been whether most people can grasp concepts they aren't really interested in or don't particularly care for. The existence and reception of this meta analysis suggests the answer may very well be no. :(

    6. Public opinions Always change enormously and is not a direct natural event but a crowd natural event. Public opinions Express cultures and culture can be compared metaphorically with bird seasonal migrations. People not only express contextually righteous opinions but also use it in their everyday life. Culture is a way of Life to transcende or to be purified. Non-wise human culture is extremist, a way to purify every dissonant point of views, especially the extreme dualistic 'western' Culture.


    7. @other Anonymous: Secular change of a trait over a small number of generations is not evidence that a trait is not heritable. It is merely evidence that the heritability of the trait is below 1.0. The variance in the trait could be 90% due to genes, or 10%. Common sense should tell you that it is impossible that the heritability of a trait such as height might be zero. The question cannot be whether it is zero or not, but only whether it is low or high, and that's where twin studies come in. They help us get a good estimate of the _degree_ of heritability of a trait.

    8. Yep.

      And of course the value that results is dependent on the population under study. In a sample of 100 identical twins grown from the same fertilized egg - a Cylon experiment carried out aboard the resurrection ship, maybe, or business as usual in Aldous Huxley's Brave New World - we would expect that, for all traits, a~0, c~0, and e~1.

      The irony is that, in order to arrive at values that dovetail with assumptions common to the social sciences, we need to propose a dystopian science-fiction setting for everyone to be living in. That people speak as though this were axiomatically desirable is really strange.

  11. "non-heritability of human traits": do you hold humans to be exempt from the patterns in other creatures? If so, why?