Nicholas Kristof has an interesting Op-Ed piece this week in the Times. Reporting on University of Michigan Professor Richard Nisbett’s new book, Intelligence and How to Get It, Kristof argues for the general malleability of intelligence. He writes,
If intelligence were deeply encoded in our genes, that would lead to the depressing conclusion that neither schooling nor antipoverty programs can accomplish much. Yet while this view of I.Q. as overwhelmingly inherited has been widely held, the evidence is growing that it is, at a practical level, profoundly wrong.
I think that this is an important point that is worth pursuing. There is indeed a widely held view that intelligence is “genetically determined” (whatever that means — how you define it matters), perhaps most infamously articulated in Charles Murray and Richard Herrnstein’s book, The Bell Curve. This idea comes from numerous studies of the correlation of relatives’ scores in standardized intelligence tests, the most common design for which is the twin study. The basic idea is that you compare the concordance in test scores of monozygotic (i.e., genetically identical) twins with dizygotic twins who share only 50% of their genes. The assumption is that both monozygotic and dizygotic twins will share the same rearing environment. Therefore, differences that appear in the observed concordance should be attributable to genes.
Twin studies show that IQ, like many other features of human behavior, is moderately “heritable.” Now, a key to understanding this field and the debate that it has spawned is understanding what is meant by heritability. Geneticists posit two different conceptions of heritability. The first is the common parlance sense: heritability simply means that a trait is genetically determined and can therefore be inherited from one’s parents. This is known as “broad-sense heritability.” In contrast, “narrow sense heritability” has a fairly precise technical meaning. Narrow sense heritability is the fraction of total phenotypic variance attributable to additive genetic variance. Based simply on this definition, laden with unfamiliar terms, you can see why most people think in terms of broad-sense heritability.
So let’s parse out the definition of narrow-sense heritability. First, “total phenotypic variance” simply means the total observed variance in the trait in question (e.g., IQ) for some well-defined population (e.g., the sample of individuals in the study). This variance arises from a variety of sources, some genetic, some environmental, some both. It is very important to note that variance is central to both definitions of heritability. A trait can be completely genetically determined (whatever that means) but have no variance in a population. Think head-number among human beings. This trait is so deeply developmentally canalized that there is no variance (everybody has one) and, thus, zero heritability.
As sexual beings, when we reproduce, our alleles (variants of genes) reshuffle whenever we generate our gametes, or reproductive cells (i.e., eggs and sperm), during the process of meiosis. One of the principles that Mendel is known for is the principle of independent assortment. This is the idea that when our alleles get reshuffled during meiosis, they appear in any given gamete independently of what the other alleles that show up in that gamete are. It turns out that independent assortment is not in any way universal. Some alleles assort independently while others are linked to other alleles, typically because they are near each other on a chromosome (but sometimes for more interesting reasons). The additive genetic variance in the definition of heritability refers to the variance attributable to just the alleles that assort independently. These are the so-called additive effects. Additivity arises from the independence of the different allelic effects. We care so much about the additive effects because these are what let us make predictive models. When an animal breeder wants to know the response to selection of some quantitative trait (e.g., body size, milk fat percentage, age at maturity), she uses an equation that multiplies the narrow-sense heritability and the selective advantage of the trait in question. Now, our scientific interest in the heritability of intelligence ostensibly arises from the desire to create predictive and explanatory models like this breeder’s equation. In the absence of explanatory or predictive power, I don’t see much scientific value.
Genes can express their effects in ways other than through their additive effects. For example, there is that familiar concept from Mendelian genetics, dominance. Dominance is a type of allele-allele interaction, just limited to the special case of occurring within a single locus. A more general case is allelic interactions is epistasis. An epistatic gene is one that affects the expression of one or more other genes. The epistatic gene is a regulator which can either increase or decrease (possibly turn off) the effect of other genes. These interactions are harder to predict and typically go into the error term in the breeder’s equation.
The real gotcha in heritability analysis though is the existence of genotype-environment (GxE) interactions. These are generally not measured and can be quite large. Lewontin, in his classic (1974) paper, first suggested that GxE interactions (in addition to other types of difficult-to-measure interactions like those arising from epistasis) might actually be large. Much of the thought that followed has supported this idea (see, e.g., Pigliucci 2001). In twin study designs, GxE interactions are non-identifiable, meaning that we don’t have enough information to simultaneously estimate the interaction, genetic, and environmental effects, so they are generally assumed to be zero. I think that it is fair to say that the consensus among population geneticists is that heritability analyses, as done though twin studies, for example, are misleading at best because of this fundamental flaw.
In my mind, the fundamental problem with twin studies of the heritability of intelligence is that they can’t begin to measure GxE interactions and therefore their estimates of heritability are hopelessly suspect.
Where is heritability of intelligence likely to be large and not quite as fraught with the problems of unmeasurable and potentially large GxE interactions? One possibility is in homogenous, affluent communities, not entirely unlike Palo Alto. Kristof notes in his Op-Ed piece that “Intelligence does seem to be highly inherited in middle-class households.” In such communities, external (“environmental”) sources of variation are relatively small. Most kids have stable homes with (two) college educated parents who place high value of achievement in school, go to safe, well-funded schools with motivated and highly trained teachers, eat nutritious food and live fairly enriched lives. When the total variance is low, whatever variance is explained by additive genetic effects is likely to be a higher fraction of the total variance. Hence, high heritability. This is a quite general point: the more environmentally homogenous a population is, the higher we should expect heritability to be.
It is very, very important, however, to note that this is generally not the case. When we move out of relatively homogenous and affluent communities, the sources of environmental variance increase and compound. The fact that a trait with such high measured heritability can be modified so extensively as discussed in Nisbett’s book suggests that intelligence is a trait with an enormous environmental effect and, I’m betting, a huge GxE interaction effect. It seems to me that the Flynn effect, the observation that IQ increases with time, provides further suggestive evidence for a massive environmental interaction. While the genomic evidence for recent strong selection on humans is mounting (in contrast to the bizarre idea that somehow selection came to a screeching halt with the advent of the Holocene), I doubt that there have been significant selective changes in the genes for intelligence (whatever that means) in the past century. Now, the environment certainly has changed in the last 100 years. This is what makes me thing big GxE interactions.
So, in a phrase, sure, genes help determine intelligence. But the action of these genes is so fundamentally tied up in environmental interactions that it seems that the explanatory power of simple genetic models for intelligence and other complex social traits such as political and economic behavior or social network measures is very low indeed. Moreover, the predictive power of these models in changing environments is low. Without explanatory or predictive potential, we are left with something that isn’t really science. I applaud efforts to more deeply understand how productive environments, good schools, and healthy decisions can maximize human potential. Heritability studies of IQ (and I worry about these other fashionable areas) seem to provide an excuse for the inexcusable failure to deal with the fundamental social inequalities that continue to mar our country — and the larger world.
Lewontin, R. 1974. The analysis of variance and the analysis of causes. American Journal of Human Genetics 26: 400-411.
Pigliucci, M. 2001. Phenotypic Plasticity: Beyond Nature and Nurture. Baltimore, Johns Hopkins University Press.