The use of heritability in policy development

The  heritability straw man has copped another bashing, this time in the Journal of Economic Perspectives. In it, Charles Manski picks up an old line of argument by Goldberger from 1979 and argues that heritability research is uninformative for the analysis of policy.

Manski starts by arguing that heritability estimates are based on the assumption that there is no gene-environment correlation. Manski writes:

The assumption that g and e are uncorrelated is at odds with the reasonable conjecture that persons who inherit relatively strong genetic endowments tend to grow up in families with more favorable environments for child development.

Any review of discussions of heritability, whether in the peer-reviewed literature or the blogosphere, will show that his claim is generally false. The proviso that the heritability estimate is only relevant to the existing environment is usually threaded through any discussion of heritability.

It is true that gene-environment covariance can affect estimates of heritability. Yet this does not mean that existing estimates have no value, nor that there are not methods that seek to account for the covariance. For example, the use of comparisons between misdiagnosed identical twins and actual identical twins allows for bounded estimates of heritability to be developed (pdf).

Manski’s broader claim, adopted directly from Goldberger, is that even if you knew the heritability of a trait, it tells you nothing about social policy. Manski uses Goldberger’s eyeglasses example as an illustration:

Consider Goldberger’s use of distribution of eyeglasses as the intervention. For simplicity, suppose that nearsightedness derives entirely from the presence of a particular allele of a specific gene. Suppose that this gene is observable, taking the value g = 0 if a person has the allele for nearsightedness and g = 1 if he has the one that yields normal sight.

Let the outcome of interest be effective quality of sight, where “effective” means sight when augmented by eyeglasses, should they be available. A person has effective normal sight either if he has the allele for normal sight or if eyeglasses are available. A person is effectively nearsighted if that person has the allele for nearsightedness and eyeglasses are unavailable.

Now suppose that the entire population lacks eyeglasses. Then the heritability of effective quality of sight is one. What does this imply about the usefulness of distributing eyeglasses as a treatment for nearsightedness? Nothing, of course. The policy question of interest concerns effective quality of sight in a conjectured environment where eyeglasses are available. However, the available data only reveal what happens when eyeglasses are unavailable.

Manski and Goldberger may be correct that the heritability estimate is uninformative as to the efficacy of distributing eyeglasses, but it is useful in assessing other policy responses to the problem and the trade-offs between them. Is it possible to prevent the eyesight loss in the first place? Is that policy cheaper and more effective than eyeglasses? If the heritability estimate was zero, you would look to the environmental causes and ask whether the eyesight problem is more appropriately dealt with by addressing the cause rather than by distribution of eyeglasses.

There is no shortage of other areas where heritability estimates might add value. Heritability estimates can inform whether it is an effective use of resources to make sure that everyone has a university degree or is over six-foot tall. Is everyone putty in the hands of the policy maker, or are there some constraints? On a personal level, Bryan Caplan’s use of heritability in Selfish Reasons to Have More Kids is a useful input to his parenting strategy.

ResearchBlogging.orgFor me, the most salient example of the usefulness of heritability research comes from examination of the heritability of IQ among children. Among high socioeconomic status families, the heritability tends to be high. Among low socioeconomic status families, it is significantly lower. This suggests that there is significant room to improve the outcomes of the children at the bottom of the socioeconomic ladder in the early years of their life (assuming those changes have effects that persist into adulthood). Increasing heritability of IQ might be evidence that environmental disadvantages are being ameliorated and opportunity equalised.

The latter part of Manski’s paper turns to the use of genes as covariates in statistical regressions. Regression identifies statistical association and not causation, which appears to be an important point in attracting Manski to this use. Noting the wealth of data being created and the possibility of observing changes in the effect of genes as the environment changes, Manski considers that these regression exercises may assist in examining how genes and environment interact.

I don’t disagree with Manski, but at present, genome association studies have plenty of issues. First, there is the missing heritability problem. To date, the magnitude of the identified effect of genes on most traits accounts for a miniscule proportion of the trait’s heritability. This points to the important role played by heritability research to provide direction to research on genes as covariates. It also indicates that until these genes are found, heritability estimates will be more informative for social policy.

A second issue is that with 30,000 odd genes and the ability to test so many of them for correlation with traits, many are found to have a statistically significant relationship through chance. As blogged about recently by Razib, this is shown when people seek to replicate earlier results – such as when it was found that most reported genetic associations with general intelligence are probably false positives (pdf).

Finally, genome based research is now feeding back into estimates of heritability. From a recent paper:

We conducted a genome-wide analysis of 3511 unrelated adults with data on 549 692 single nucleotide polymorphisms (SNPs) and detailed phenotypes on cognitive traits. We estimate that 40% of the variation in crystallized-type intelligence and 51% of the variation in fluid-type intelligence between individuals is accounted for by linkage disequilibrium between genotyped common SNP markers and unknown causal variants. These estimates provide lower bounds for the narrow-sense heritability of the traits.

Despite all the critiques about methodology, most new studies confirm that the old “methodologically poor” heritability estimates were in the right ballpark. The problem is not that the estimates are not useful, but rather that they are not used.

Manski, C. (2011). Genes, Eyeglasses, and Social Policy Journal of Economic Perspectives, 25 (4), 83-94 DOI: 10.1257/jep.25.4.83

13 comments

  1. Nice to see you are picking up new references!
    Although individual differences in cognitive ability are primarily due to genes, one cannot from this infer an intergenerational transmission. To inherit genes is not synonymous with inheriting traits, when the trait is dependent on a large number of interacting gene functions. (Certain combinations of alleles can be very favourable, while the very same gene variant that promotes cognitive ability in one individual can be detrimental in another genetic environment that can arise during meiosis)
    The intergenerational transmission of IQ, which is implied in your reasoning, has (still) not been sufficiently studied.
    Is this concept so extraordinary difficult to understand?

    1. My argument in the above does not rely on intergenerational transmission, only heritability. Intergenerational transmission does, however, increase the possible policy implications.

      As to the old chestnut of whether IQ is heritable and intergenerationally transmitted, I will simply link to our previous debate and repeat my challenge. Given the mountain of evidence on correlations between parent and offspring IQ – including from twin studies, family studies, adoption studies etc – can you provide a single reference that supports your contention that cognitive ability does not correlate to that of the father?

  2. Why is it of such crucial importance for policy if intelligence is transmitted between the generations as genetic elements or as social inheritance? (Because no-one is disputing that cognitive ability is transferred from parents to children).
    If parental intelligence is genetically transmitted to the offspring, it would make sense to have a policy where intelligent parents of high socioeconomic status could promote the education of their (genetically favourably predisposed) children – presuming that the whole society would benefit more from seeing the gifted reaching their full potential rather than the less gifted (actually a far-fetched assumption).
    On the other hand, if the combinations of common alleles (gene variants) that would predispose for exceptional cognitive abilities arose more randomly distributed in the population, if parental behaviour from the moment of conception until 2 or 5 or 10 years of age determined how the genetic potential could be played out, then optimal policy would be completely different. In the latter case, it would be wise policy to have extensive maternal care programmes, parental support programmes, and freely available, high quality day-care for small children, in order to catch the potential. In this case, no special attention would have to be given to children in privileged environments where they are already likely to find the stimulation to reach their potential.
    So, when you are talking of heritability of intelligence and the effect on policy, you cannot, as you keep doing, mix these two scenarios up.
    I can see a lot of wishful thinking amongst neo-liberal economist, and a lot of shoehorning in order to have reality fit the first scenario. But, there is no evidence for it. Rather the opposite.

    On top of this, epigenetic might be a complicating factor where people of low socioeconomic status might transfer unfavourable DNA modifications to their children. These modifications can be reversed with enhanced socioeconomic and/or psychosocial conditions (the specifics are really unclear here). If we want to promote an intelligent society, policy should therefore be to promote the socioeconomic status of affected groups. In many societies, the opposite is rather seen, when the already dispossessed take the hardest blow of the economic downturn. How will their stress response affect the cognitive ability of their children?

    1. As noted in my other comment, even without intergenerational transmission, heritability can inform this policy. If heritability were 1 across the full range on environments children were raised in, then all your maternal care and parental support programs would need to be shaped with that in mind. Let the rich parents spend their money but know that it is wasted. If heritability is less than one, there is plenty of room to debate what it means and the scope for environmental intervention.

      The funny thing is that you have used a heritability argument to suggest the options for maternal care, parental support etc. What differs is our estimates of heritability and intergenerational transmission. Even if we disagree on the interpretation, it still seems we find examination of heritability (and other genetic factors) useful – which was my overarching point.

    2. “On the other hand, if the combinations of common alleles (gene variants) that would predispose for exceptional cognitive abilities arose more randomly distributed in the population, if parental behaviour from the moment of conception until 2 or 5 or 10 years of age determined how the genetic potential could be played out, then optimal policy would be completely different.”

      So, with a h^2 estimate, we have:
      Vg (genetic variance)/Vp (phenotypic variance)
      Where genetic variance equals:
      Va (additive) + Vd (dominance) + Vep (epistasis) + Vam (assortative mating)
      And phenotypic variance equals:
      Vg (genetic variance) + Ve (environmental variance) + Vge (non-additive environmental-genetic interaction) + Covge (covariance of genes and environment) + Verror (variance due to error).

      Could you restate your critique in the language of behavioral genetics? My point below was that Davies et al (2011) established the high Va/Vp (or narrow heritability
      of IQ). This establishes a high degree of “intergenerational transmission.” Other studies, using methods mentioned below, indicate that the 40% of the variance not due to additive genetics is due mostly to a combination of Vd + Vep + Vam + Ve (nonshared) + Verror. There’s not much room for “parental behavior” left, however you are conceptualizing that.

  3. “Nymnchen said:
    “Although individual differences in cognitive ability are primarily due to genes, one cannot from this infer an intergenerational transmission. To inherit genes is not synonymous with inheriting traits, when the trait is dependent on a large number of interacting gene functions.”

    Davies et al. (2011) already established the high narrow heritability of IQ. Narrow heritability (aka breeding value) is defined as the degree to which a trait is passed from parent to offspring. What’s your objection to the method used by Davies et al. (2011)? I would like to hear it.

    Jason,

    “It is true that gene-environment covariance can affect estimates of heritability. Yet this does not mean that existing estimates have no value, nor that there are not methods that seek to account for the covariance. For example, the use of comparisons between misdiagnosed identical twins and actual identical twins allows for bounded estimates of heritability to be developed (pdf).”

    For a detailed discussion of methods used to detect (passive, evocative, and active) g-e correlations, check out Polomin’s “The Nature of Nurture” in “Experience and Development: A Festschrift in Honor of Sandra Wood Scarr.” The major ones are: 1) compare the correlations between environmental factors and a child’s development in adoptive and non-adaoptive families, 2) compare biological parents’ traits with adoptive families’ environments, 3) use multivariate genetic analysis to analyze the covariance between environmental measures and traits. You’re just scratching the surface with “misdiagnosed identical twins.”

  4. “Manski writes:

    “The assumption that g and e are uncorrelated is at odds with the reasonable conjecture that persons who inherit relatively strong genetic endowments tend to grow up in families with more favorable environments for child development.”

    For example, the use of comparisons between misdiagnosed identical twins and actual identical twins allows for bounded estimates of heritability to be developed (pdf).”

    Jason,

    There are three types of GE correlations: passive, active, and evocative. Manski is pointing towards passive correlations, which are the easiest to test for. Your cite concerning misdiagnosed identical twins is more relevant to evocative or active GE. As for passive GE, one can look at the phenotypic correlations between virtual twins —
    –i.e. same-age, unrelated siblings reared together since infancy — or the correlation between their inverse, MZA, or the correlations between other adoptive/ non-adoptive kin/non kin pairs. In general, it’s difficult to take Manski seriously, given the evidence as it stands. (On semi-theoretical grounds, passive GE invokes shared environments as the cause of phenotypic similarities; this requires that shared environments are, in fact, correlated with the phenotypes; yet, by adulthood, it has routinely been found that this is not the case (hence, the “Gloomy prospects of behavioral genetics”). This, a priori, rules out passive GE as an explanation for variance.)

  5. Wouldn’t misidentified identical twins still look very much alike? If one was tall, the other would be tall, and plenty of research has shown that height correlates with how people treat you. They would have the same eye and hair color, similar skin tones, similar features- unfortunately, these all predict how people treat you, too. Don’t identical twins have different in utero experiences than fraternal twins, whether it is that they develop in the same amniotic sac, or just the fact that they developed from the same egg? What exactly does it mean to find “misclassified” identical twins- wouldn’t they likely have some idea they were identical? Might these shared characteristics lead them to identify with one another more than fraternal twins? Who are the parents who misclassify their twin children- isn’t that an implicit omitted variable that you can never get rid of here? If there are gene-environment interaction terms missing from the equation (and include a bunch of interactions with the unobservables I’ve listed here, and those I haven’t) wouldn’t you run into all the same problems with the “heritability” estimates?

    Perhaps I misunderstand the claims being made (I’m certainly not a geneticist), but it sounds a lot like causality, when all I see is a correlation with many confounders.

    1. Wouldn’t misidentified identical twins still look very much alike? If one was tall, the other would be tall, and plenty of research has shown that height correlates with how people treat you. They would have the same eye and hair color, similar skin tones, similar features- unfortunately, these all predict how people treat you, too. …

      What exactly does it mean to find “misclassified” identical twins- wouldn’t they likely have some idea they were identical? Might these shared characteristics lead them to identify with one another more than fraternal twins?

      The similarity between misidentified identical twins is one of the features the researchers are after. They want fraternal twins that are treated like identical twins, meaning that they have equally similar environments. If the misidentified identical twins identify with each other as identical twins do, that’s also good for the test.

      Don’t identical twins have different in utero experiences than fraternal twins, whether it is that they develop in the same amniotic sac, or just the fact that they developed from the same egg?

      Mono-amniotic twins are a small percentage of identical twins, so that is a minor confound at best. I’m not aware of any research that points to the in-utero environment being more similar for identical twins than for fraternal twins, and that in turn affecting the traits that we’re examining to any noticeable degree.

      Who are the parents who misclassify their twin children- isn’t that an implicit omitted variable that you can never get rid of here? If there are gene-environment interaction terms missing from the equation (and include a bunch of interactions with the unobservables I’ve listed here, and those I haven’t) wouldn’t you run into all the same problems with the “heritability” estimates?

      Heritability of behavioural traits has been examined from many directions – twin studies, adoption studies, family studies, GREML, etc and they all turn up the same result. There’s a reason why the first law of behavioural genetics is that all human behavioural traits are heritable.

      Lastly, there are also confounds that may lead us to be underestimating heritability.

  6. You say: “They want fraternal twins that are treated like identical twins, meaning that they have equally similar environments” but N=16 on that arm of the sample, so I was only talking about the other direction. I’ll skip the 10 minute brainstorm on the problems with a hypothetical sample of fraternal twins who think they are identical. As far as I can tell, the questions I raised are unanswerable.

    One reason these studies might all turn up the same result is that they all are looking at the same correlation. I’m not sure they do give the same result, either.

    You cite an article from 2000, which contains the memorable passage regarding the Human Genome Project: “What should we expect from this endeavor? Behavior geneticists anticipate vindication: At long last, statistical variance components will be rooted in the actual causal consequences of actual genes. Critics of behavior genetics expect the opposite, pointing to the repeated failures to replicate associations between genes and behavior as evidence of the shaky theoretical underpinnings of which they have so long complained.” As far as I know, none of that came true. In place of “actual genes” was substituted this GREML stuff, which is the sort of high-R-squared-means-good-model methods economics threw out in the 1970s. At least the referees at PNAS made them write it up as a correlation that “possibly corroborates” these other theories. You might consider citing it accordingly.

    1. One reason these studies might all turn up the same result is that they all are looking at the same correlation.

      You mean the correlation between behaviour and genotype?

      “What should we expect from this endeavor? Behavior geneticists anticipate vindication: At long last, statistical variance components will be rooted in the actual causal consequences of actual genes. Critics of behavior genetics expect the opposite, pointing to the repeated failures to replicate associations between genes and behavior as evidence of the shaky theoretical underpinnings of which they have so long complained.” As far as I know, none of that came true.

      Genetics has moved beyond 2000-era debates. The only ones still expecting the “gene for intelligence” etc. are the media.

      In place of “actual genes” was substituted this GREML stuff, which is the sort of high-R-squared-means-good-model methods economics threw out in the 1970s. At least the referees at PNAS made them write it up as a correlation that “possibly corroborates” these other theories. You might consider citing it accordingly.

      You mean the R-squared that is still given in almost every economic paper with econometric analysis? The reason GREML resembles R-squared is that heritability resembles R-squared – the proportion of phenotypic variation that is due to genetic variation. That’s the measure of interest. As for “partially corroborates”, GREML provides a lower bound, so of course it only partially corroborates.

      I’m not sure they do give the same result, either.

      Please point me to the studies that suggest behavioural traits are not affected by a person’s genotype. I’d be pleased to read them.

Comments are closed.