How Genes Shape Environment
Interventions intended to change the environment assume that if the environment changes as planned, then, other things being equal, outcomes will be positively affected. Suppose that parents use inconsistent discipline, for example, which is known to be associated with negative outcomes for children. It is assumed that if the parents can be taught to use a reasonable disciplinary style consistently, those negative outcomes will be reduced. But that’s not the only way environmental effects work. They can also be mediated by the correlation (r) between the genotype (G) and the environment E, known in the trade as rGE. There are three types of rGE: passive, evocative, and active.
Passive rGE can be illustrated with the case of aggressive behavior in a child who was raised by a violent and abusive father. The child is likely to have gotten a double dose of bad luck: the environmental effects of the parental abuse, and a father with a genetic propensity for abusiveness, roughly half of which has also been passed down to the child.1 Evocative rGE (sometimes called reactive rGE) occurs when the child’s genetically influenced characteristics evoke a response from the environment (which includes other people). Consider the case of parents who use physical punishment on an aggressive child. It could be a case of evocative rGE, whereby parents are reacting to the child’s violence by using physical punishment, which then only makes matters worse. Active rGE occurs when children shape their environments—for example, by choosing peer groups—for genetically influenced reasons. An adolescent boy who has a genetic propensity for aggression is likely to be attracted to violent teenage gangs and become a member. The environmental influences of the gang then reinforce, or even amplify, the boy’s genetic propensity.
Now let’s make the example more complicated. In chapter 12, I referenced the Krapohl study of a large sample of British twins. The objective was to identify the causes of the differences among scores that the participants achieved on the General Certificates of Secondary Education. You may recall that IQ explained more of the variance in GCSE scores than any of the other individual measures, but the other variables collectively explained about as much as IQ.2 This would suggest that even if we don’t know how to raise IQ, outside interventions can do a lot to boost students on the other contributions to school achievement. But the authors found that the proportion of the covariance of these other factors with academic achievement was extremely high. “To the extent that children’s traits predict educational achievement, they do so largely for genetic reasons, for example, for personality (92%), behavior problems (81% for parent-rated, 89% for child-rated), intelligence (75%), self-efficacy (64%), and well-being (53%).”3 The authors went on to observe:
[T]hese results turn some fundamental assumptions about education upside down. For example, one of the reasons that the contribution of intelligence is sometimes considered controversial when discussing educational outcomes is that intelligence is viewed as genetic, whereas achievement is thought to be due to environmentally driven influences.… However, our results suggest the opposite: Genetic influence is greater for achievement than for intelligence, and other behavioral traits are related to educational achievement largely for genetic reasons.[4]
How does this affect the prospects for outside interventions? Correlations between the genotype and the environment lend themselves to two quite different interpretations.
One interpretation focuses on the ways in which genetic causes are mediated by parental behavior. Yes, bad parenting practices are partly driven by parental genes, but if it’s the parenting practices that are proximally causal, an outside intervention could still have an effect if it could change the parenting practices.
The other interpretation focuses on the ways in which what we once thought was an environmental cause turns out to be a partly genetic one. When an outside intervention sets out to change bad parenting practices, the already difficult task of changing a person’s behavior is fighting a genetic headwind. The behaviors that the intervention is trying to change aren’t occurring just because of ignorance about good parenting, but also because of genetic predispositions.
We have a natural experiment that lets us see how these competing forces work out in practice—adoption at birth. Adoption studies routinely show that the correlation between biological children’s IQ and the family’s socioeconomic status is around twice the correlation between adoptive children’s IQ and the adoptive family’s socioeconomic status.5 Can the benefits of competent parenting practices benefit the adopted child? Yes. But adoption is as good as it gets. In effect, adoption at birth to competent parents gives us a glimpse of what would happen if an outside intervention could magically be successful at changing a wide variety of parenting behaviors from bad to good. An outside intervention that makes modest improvements in a small proportion of parenting problems will have a far smaller effect.
“The First Premise Is Wrong for the Early Stages of Life”
Everything we know about human development says that humans are most malleable in the first years of life. The brain is still developing. Habits are not yet set in the child. Some parenting practices can be changed for the better and still have time to work their effects on the child. I fully share the view that if interventions are ever going to work, they’re going to work in infancy and early childhood. But it’s one thing to believe that; it’s another to confront the empirical findings about the difficulties and constraints that have attended a half century of attempts to intervene early in life.
Heritability and Socioeconomic Status
Recall the figure in chapter 12 showing the role of the shared environment in IQ from childhood to adulthood. It hit zero in adolescence, but the shared environment explained more than half of the variance for the preschool years. That finding implies that the second premise of my syllogism doesn’t necessarily apply to infants and young children.
The proposition that heritability of IQ is lower for disadvantaged children than for children from ordinary backgrounds was first advanced by psychologist Sandra Scarr in 1971, who found provisional evidence for it.[6] Subsequent studies in 1980 and 1999 provided stronger evidence, but the role of the shared environment remained small even for disadvantaged children.[7]
In 2003, Eric Turkheimer published dramatic evidence that the opportunities for intervening are not only higher at young ages, but that they are especially high for disadvantaged children because of an interaction between heritability and SES.8 He and his coauthors used 390 twin pairs with excellent data on zygosity drawn from the National Collaborative Perinatal Project. The measure of IQ, administered at age seven, was the Wechsler Intelligence Scale for Children, the most highly regarded test of its kind. Their measure of the shared environment was an index of socioeconomic status, expressed as a scale going from 0 to 100. The index was based not just on parental education, but on a linear combination of information on parental education, occupational status, and income, using a well-established method.9
It was close to a gold-standard study, and it produced unequivocal evidence for a large gene × environment (hereafter G×E) interaction. In the best-fitting model, the expected value of heritability (h2) when the measure of SES was set to zero (i.e., the most impoverished, poorly educated families) was a tiny .02. When the measure of SES was set to 100, the expected heritability was an exceptionally high .90. The even more important finding was what happened to the role of the shared environment. When the measure of SES was set to 100, representing the wealthiest and most highly educated families, the expected value of the role of the shared environment (c2) was only .09. When the measure of SES was set to 0, the expected value of c2 was .62—almost three times greater than the role of the shared environment for the earlier studies. Furthermore, a large role for the shared environment remained throughout the bottom half of the SES distribution.10
SOME ADDITIONAL TECHNICAL TERMS
Set, as in “when the measure of SES was set to 100,” refers to the value assigned to an independent variabl
e, a definition easiest to understand by reading the following definition of expected value.
Expected value means a statistical best guess. All the results I discuss in this section are produced by mathematical equations applied to large datasets. The models themselves are often complicated, but the essence is simple. Suppose I have data for 1,000 adults on age, education, IQ, and income. I want to know the best guess of family income (the dependent variable) of someone 30 years of age with 12 years of school and an IQ of 120 (the independent variables). Multivariate statistical analysis can tell me the best guess for that dataset—the statistically expected value of the family income when age is set to 30, years of education is set to 12, and IQ is set to 120. The accuracy of the expected value depends on the magnitude of the correlations among the variables I’m working with. For more on correlation and multivariate statistics, see Appendix 1.
Interaction effect. In social science analyses, an interaction effect has occurred when two things acting in combination produce an effect in addition to their separate effects.
In the case of heritability and environment, here’s a simple example: Let’s assume that Alice and Becky have high and low genetic endowment for IQ respectively (A in the ACE model) and that reading during childhood raises IQ. Both children live in homes with many books (an element of C). Alice reads all of the books while Becky reads only a few of them. In such a case, it is possible (though not necessarily the case) that the effect that genes and books in the house have on IQ is not just A + C but A + C + (A × C). This is an extremely simplified description. The note adds details.[11]
Many subsequent studies have taken up the G×E interaction that Turkheimer’s 2003 study found. As I write, 11 of them are interpretable as attempts at replication.[12] Add in the original Turkheimer study, and we have 12 studies of interest. Five of them did not find a G×E interaction.13 The results for the seven that did are shown in the table below. Details on how the percentages were determined are given in the note.[14]
Study: Turkheimer et al. (2003)
Variance explained by the shared environment
Bottom SES percentile: 62%
Top SES percentile: 9%
Study: Harden et al. (2007)
Variance explained by the shared environment
Bottom SES percentile: 45%
Top SES percentile: 36%
Study: Tucker-Drob et al. (2011)
Variance explained by the shared environment
Bottom SES percentile: 78%
Top SES percentile: 38%
Study: Rhemtulla and Tucker-Drob (2012)
Variance explained by the shared environment
Bottom SES percentile: 77%
Top SES percentile: 50%
Study: Bates et al. (2013)
Variance explained by the shared environment
Bottom SES percentile: 12%
Top SES percentile: 10%
Study: Kirkpatrick et al. (2015)
Variance explained by the shared environment
Bottom SES percentile: 19%
Top SES percentile: 12%
Study: Tucker-Drob and Bates (2015)
Variance explained by the shared environment
Bottom SES percentile: 41%
Top SES percentile: 21%
Three of the studies (Harden, Bates, Kirkpatrick) found a statistically significant interaction effect, but the magnitude of the effect was small. Add in the five that failed to replicate the G×E effect, and eight out of the 12 did not find a substantively important interaction effect between the shared environment and SES. But four of them did. What’s going on?
A meta-analysis of all known studies of the G×E interaction (not just the ones that met the criteria for replications) conducted by psychologists Elliot Tucker-Drob and Timothy Bates in 2015 established one intriguing finding beyond doubt: The G×E effects in U.S. samples systematically differ from the effects in non-U.S. samples.15 The authors conducted extensive tests for the robustness of this finding, all of which it passed. Their main conclusion is worth quoting in full:
This meta-analysis of published and unpublished data provided clear answers to our three questions. First, studies from the United States supported a moderately sized Gene × SES interaction on intelligence and academic achievement. Second, in studies conducted outside the United States (in Western Europe and Australia), the best estimate for Gene × SES magnitude was very slightly negative and not significantly different from zero. Third, the difference in the estimated magnitude of the Gene × SES effect between the U.S. and the non-U.S. studies was itself significant.16
Why should the difference between the United States and the rest of the world be so marked? Tucker-Drob and Bates ran through the options: cross-national differences in the teaching of literacy and numeracy, educational quality, access to education and medical care, social mobility, and income support, each of which has been argued by sources they cite. But no one has done more than speculate about any of them. The note describes some differences in the samples that might also be relevant.[17]
The Empirical Record for Early Childhood Interventions
Few topics in social policy have received more intense empirical scrutiny than the effects of early childhood interventions. Unfortunately, few aspects of social policy have also been as intensely politicized. In 2013, a leading specialist in pre-K programs, economist Greg Duncan, and social policy scholar Katherine Magnuson published a comprehensive review of the evidence up to that time. The authors found these conclusions to be justified by the weight of the evidence:
Effect size at program exit. A meta-analysis of 84 evaluations of preschool programs for disadvantaged students found that “the simple average effects size for early childhood education on cognitive and achievement scores was .35 standard deviations at the end of the treatment periods, an amount equal to nearly half of race differences in the kindergarten achievement gap.”18 In other words, the average effect size was worth noticing on the exit test. But…
Trend in effect sizes over time. The same meta-analysis found that the exit effect sizes have been decreasing over time: “Programs beginning before 1980 produced significantly larger effect sizes (.33 standard deviations) than those that began later (.16 standard deviations).”19 The authors attributed this to improved conditions for children in the control group from the 1970s to the end of the century.
Fadeout. When participants in preschool programs are tracked after the end of the intervention, programs that achieved an impact at exit consistently show fadeout averaging about .03 standard deviations per year. “With end-of-treatment effect sizes averaging around .30 standard deviations, this implies that positive effects persist for roughly 10 years.”20
Head Start. From its beginning in 1965, Head Start generated many evaluations, often done by a single school system and poorly designed. As part of the Head Start reauthorization bill in 1998, Congress mandated a large and rigorously designed evaluation that would provide dispositive evidence. The final report of the evaluation was issued in 2010.21
After one academic year in the program, effect sizes in six language and literacy areas ranged from .09 to .31, but there was negligible impact on math skills or on children’s attention, antisocial, or mental health problems. The limited effects at exit disappeared within two years. “By the end of first grade, both achievement levels and behavioral ratings of treatment group children were essentially similar to achievement levels of control-group children.”22
Delayed effects. Duncan and Magnuson cite evidence from the Perry Preschool Project, the Abecedarian Project, and Head Start that some effects of the programs emerge only in adolescence or later. For example, a study of siblings found that children who attended Head Start were eight percentage points more likely to graduate from high school. “Taken together, these studies suggest that despite the decline in program impacts on achievement test scores as children progress through elementary school, there may be measurable and important effects of Head Start on childr
en’s life chances.”23
Duncan and Magnuson accurately stated the results of the various programs and were fair-minded in their interpretations of some of the inconsistencies and puzzles in the data. They did not, however, emphasize the reasons why even the modest successes warrant skepticism. In his review of the same programs, Grover Whitehurst, former head of the Department of Education’s Institute of Education Sciences, emphasized what Duncan and Magnuson did not:
Not one of the [pre-K] studies that has suggested long-term positive impacts of center-based early childhood programs has been based on a well-implemented and appropriately analyzed randomized trial, and nearly all have serious limitations in external validity. In contrast, the only two studies in the list with both high internal and external validity (Head Start Impact and Tennessee) find null or negative impacts, and all of the studies that point to very small, null, or negative effects have high external validity.[24]
In 2017, the Brookings Institution and the Duke Center for Child and Family Policy put together a task force of 10 of the leading scholars in the field to provide a consensus statement on the findings of the research to date. The specific consensus statements included two that led with the words “Convincing evidence.” They can serve as a summary for my account as well:
Convincing evidence shows that children attending a diverse array of state and school district pre-K programs are more ready for school at the end of their pre-K year than children who do not attend pre-K. Improvements in academic areas such as literacy and numeracy are most common; the smaller number of studies of social-emotional and self-regulatory development generally show more modest improvements in those areas.
Human Diversity Page 30