CPS 2018 Median earned income (000s)
White
In the labor force: $48
Employed full-time year-round: $50
Black
In the labor force: $35
Employed full-time year-round: $40
Latino
In the labor force: $30
Employed full-time year-round: $35
Asian
In the labor force: $54
Employed full-time year-round: $65
The ethnic differences are substantial, with fully employed blacks and Latinos earning only 80 percent and 70 percent of the white median respectively. Fully employed Asians made 30 percent more than the white median.
11. The next table shows the fitted values of earned income expressed in 2018 dollars for IQ set at 80, 100, and 120 for the two cohorts of the NLSY. Logged earned income was regressed on IQ and ethnicity. The sample was limited to persons ages 30–35 who reported being in the labor force (though not necessarily employed) throughout the previous calendar year and reported working at least one hour.
Earned income (2018 dollars): NLSY79, CY 1993
IQ: 80
White: $28,632
Black: $28,519
Latino: $26,973
IQ: 100
White: $40,514
Black: $40,355
Latino: $38,166
IQ: 120
White: $57,327
Black: $57,101
Latino: $54,005
Earned income (2018 dollars): NLSY97, CY 2014
IQ: 80
White: $29,883
Black: $25,144
Latino: $30,596
Asian: $46,975
IQ: 100
White: $40,584
Black: $34,147
Latino: $41,553
Asian: $63,796
IQ: 120
White: $55,117
Black: $46,375
Latino: $56,432
Asian: $86,641
For NLSY79, blacks and whites with comparable IQs had effectively identical fitted earned incomes, with Latinos only fractionally behind. For NLSY97, blacks had a fitted mean that was 84 percent of the white mean, Latinos and whites were effectively equal, and the fitted Asian mean was 57 percent higher than the white mean.
10: A Framework for Thinking About Heritability and Class
1. See Horowitz, Haynor, and Kickham (2018) for a survey of the ideological positions of 479 sociologists in U.S. colleges and universities.
2. For exceptions, see Moore and Shenk (2017) and Burt and Simons (2014).
3. Galton (1869). I am drawing from the account in Herrnstein (1973): chapter 4, which contains an excellent discussion of the history of scholarship about heritability of IQ as of the early 1970s.
4. Sewall Wright wrote the classic papers on measuring degrees of relatedness in 1921. For an accessible discussion of them, see Hill (1995).
5. For a nice illustration of correlations on phenotypic outcomes for different degrees of relatedness, see Cesarini and Visscher (2017): Fig. 1.
6. I adapted this example from Turkheimer, Pettersson, and Horn (2014): 519. “A heritability coefficient represents the proportion of phenotypic variability that is associated with variability in genotype. As such, it is an effect size, a variance ratio, an R2 coefficient; and like any variance ratio it is sensitive to characteristics of the population in ways that means are not. In particular, variance ratios depend crucially on the variability of both the predictor and the outcome. For example, the question, ‘How much of the variance in college performance is explained by differences in SAT scores?’ has no meaningful answer, other than, ‘It depends on the variability of SAT scores and other factors at the institutions where the study is conducted.’” Turkheimer, Pettersson, and Horn (2014): 519.
7. Tucker-Drob, Briley, and Harden (2013).
8. Technically, these influences are obvious potential components of the shared environment. They do not qualify as part of the shared environment unless they do in fact make persons raised together more similar. At an anecdotal level, consider the school that two siblings attend some years apart. It is “shared” in the sense that both siblings walked into the same building every day. But the siblings necessarily had different peer groups. Suppose also that the siblings shared none of the same teachers. Those elements of the school are not shared. There’s no a priori reason to assume that other elements of the school experience made “school” meaningfully shared between the two siblings born even a year or two apart. Now suppose that the solution to the ACE equation for large MZ and DZ twins’ samples indicates no role for the shared environment and a substantial role for the nonshared environment. It could well be that the effects of schooling were real, but they were part of the effect attributed to the nonshared environment.
9. Plomin (2011): Table 1. It is adapted from Rowe and Plomin (1981).
10. Of these, different peer groups appear to be especially important, as documented by Harris (1998).
11. The 100 percent figure for MZ twins is accurate when rounded to within half a dozen decimal places. A study of SNPs in 66 adult twins found differences that amount to one SNP per 1.2 × 10–7 nucleotides—an exceedingly small fraction. Li, Montpetit, Rousseau et al. (2013). MZ twins can also fail to share 100 percent of their genes because of rare conditions.
12. DZ twins share only approximately 50 percent because the coin flips that decide the blocks of DNA donated by the mother and father don’t always produce exactly equal totals from each.
13. To see the logic, start by thinking about a trait that is 100 percent heritable and perfectly measured. The correlation of that trait among a sample of MZ twins, who share 100 percent of their genes, has to converge on +1.0. What is going to be the correlation among samples of DZ twins who share 50 percent of their genes? It will just as inevitably converge on +.5. That difference in shared genes in MZ and DZ twins gives the mathematical leverage for disentangling genes from the shared environment.
Let’s say that we want to understand the heritability of obsessive-compulsive disorder. Among a large sample of MZ twins raised together, we find the correlation of the diagnosis of obsessive-compulsive disorder to be +.53. If one MZ twin has obsessive-compulsive disorder, the other is much more likely than chance to exhibit the disorder as well. Next, we assemble a large sample of same-sex DZ twins raised together and discover that the correlation of their rate of obsessive-compulsive disorder is +.28. (These percentages are taken from Polderman, Benyamin, de Leeuw et al. (2015): Supplementary Table 21.) If obsessive-compulsive disorder were caused entirely by environmental influences, then the correlations for the samples of MZ and DZ twins would be the same. The larger correlation for the MZ twins must be caused by their additional shared genes.
This is a classic example of the distinction between specific people and large samples. For the sake of argument, assume that it is possible for parenting to produce obsessive-compulsive disorder. You observe that MZ twins both have obsessive-compulsive disorder. There is no way to tell whether it came about through shared genes or shared parents. But if you have 1,000 randomly chosen pairs of MZ twins and another 1,000 pairs of DZ twins, and you find that twice as many MZ twins as DZ twins have obsessive-compulsive disorder, then that difference (with an ascertainable margin of error) can be explained only by their additional shared genes. There’s no other possibility except that your samples weren’t really random after all.
The equations for implementing this logic that I give in the text are called the Falconer formulas after their originator, Douglas Falconer. What I’ve just given you is the most basic treatment of Falconer’s formulas. Falconer (1960). The results from the technical literature that I report in chapter 11 are usually based on fitted models that use more complicated statistics. This is necessary because, apart from any other reason, the basic ACE model must be elaborated whenever the correlation among MZ twins is more than twice the correlation among DZ twins—i.e., whenever solving for Falc
oner’s equations leads to a negative value for C. But these refinements generally tweak the results of the basic Falconer formulas by only a few percentage points.
14. Bouchard, Lykken, McGue et al. (1990). This is the study that you are likely to have heard about. In fact, the first scientific study of twins goes back to 1937. Newman, Freeman, and Holzinger (1937).
15. Segal (1999): 117–18.
16. The Minnesota study attracted charges that the heritability measures were radically too high (e.g., Taylor (1980), Farber (1981)). Bouchard and his colleagues responded with additional analyses that confirmed their main findings. See Bouchard (1982) and Bouchard (1983).
17. The three other assumptions are: (1) Either nonadditive genetic effects (D and I) or shared environmental effects (C) are zero; (2) there are no gene × environment (G×E) interactions and no gene-environment correlations (rGE); and (3) twins are representative of the non-twin population. For a discussion specifically of number 3, see Barnes and Boutwell (2013). For general discussion of these assumptions with additional references, see Verweij, Mosing, Zietsch et al. (2012) and Appendix A of Barnes, Wright, Boutwell et al. (2014).
18. Genetic assortative mating needs to be discriminated from cultural transmission, which also tends to inflate the estimate of shared environmental effects. For a discussion of the assumptions of the classical twin model and an empirical assessment of assortative mating for intelligence, see Vinkhuyzen, van der Sluis, Maes et al. (2012).
19. A separate issue is the genome-wide genetic similarity of mates (e.g., see Domingue, Fletcher, Conley et al. (2014)). Here I am reporting evidence for phenotypic assortative mating on discrete traits that are known to be substantially heritable.
20. For citations, see the literature review in Barnes, Wright, Boutwell et al. (2014): 7.
21. Loehlin (1978): 72.
22. LoParo and Waldman (2014): 606.
23. E.g., Fosse, Joseph, and Richardson (2015); Hettema, Neale, and Kendler (1995); Kaprio, Koskenvuo, and Rose (1990).
24. For citations on each of these outcomes, see LoParo and Waldman (2014): 606–7. For a comprehensive list of studies of EEA as of 2014, see Barnes, Wright, Boutwell et al. (2014): Appendix D. The authors summarize the table as follows:
The studies included in Appendix D in the online supporting information tested for violations of the EEA across 1,233 environments and violations were detected in only 112 of them (9 percent). Of the 61 studies available, only 13 concluded that the EEA was invalid (21 percent), but of these only 6 performed any empirical analysis (10 percent), and none of these studies actually estimated the impact of the presence of unequal environments on heritability estimates. However, several studies examined directly the effect of violating the EEA on heritability estimates. Appendix D in the online supporting information includes 11 studies that estimated the impact of unequal environments on heritability estimates, with the average effect being an upward bias of about .012 (or about one percentage point) in the heritability estimate. What this necessarily means is that the widely cited heritability estimate of .50 for antisocial behaviors may be upwardly biased by .012 and the “true” A is actually closer to .488. However, we should note that these estimates do not take into account violations of other assumptions (e.g., assortative mating; the presence of evocative gene-environment correlation) that may downwardly bias heritability estimates. (Barnes, Wright, Boutwell et al. (2014): 11–12).
25. Felson (2014): 195.
26. These bodies of evidence will not prevent yet more assertions that twin studies are worthless. If you wish to get a sense of the scholarly rigor of the two sides of the debate, I recommend a matched pair of 2014 articles in the journal Criminology. For the case against twin studies, Callie Burt and Ronald Simons wrote “Pulling Back the Curtain on Heritability Studies” (Burt and Simons (2014)). For the defense, a team of seven scholars wrote “Demonstrating the Validity of Twin Research in Criminology” (Barnes, Wright, Boutwell et al. (2014)). The initial articles were followed by responses from both sides: Burt and Simons (2015) and Wright, Barnes, Boutwell et al. (2015). In my view, it is instructive as a case study. Put bluntly, I find the Burt and Simons articles to be transparently weak. But you don’t need to take my word for it. The issues regarding the Burt and Simons use of evidence are not subtle. If you have a basic grasp of statistical evidence, you can decide for yourself.
11: The Ubiquity of Heritability and the Small Role of the Shared Environment
1. Turkheimer (2000): 160.
2. The public’s intuitive judgment is remarkably accurate. Willoughby, Love, McGue et al. (2018) found that the public’s estimate of the heritability of 21 traits was correlated at +.77 with published heritabilities. With regard to cognitive repertoires (as opposed to physical resemblances), some parents hold out. In the words of psychologist Marvin Zuckerman, “All parents are environmentalists until they have their second child.” Quoted in Turkheimer (2019).
3. Harris (1998).
4. Pinker (2002).
5. E.g., Caplan (2011).
6. Loehlin and Nichols (1976): 92.
7. Rowe and Plomin (1981).
8. Scarr and Grajek (1982): 361.
9. For a review of Loehlin’s continuing exploration of the heritability of personality traits after the initial 1976 study, see Turkheimer, Pettersson, and Horn (2014).
10. I assert this, confident that it is true at the extreme, but I do not know of good studies proving it. Many badly abused children live seemingly normal adult lives. And for children who are obviously damaged, the problem is distinguishing between the effects of the abuse and a genetic confound when the child is the biological child of the abusers. That said, I continue to believe that my wording, “can damage children permanently,” is incontestable.
11. Why do I ignore the advantages of choosing the best neighborhood and thereby increasing the chances of positive peer groups? Because defining what constitutes a “good” neighborhood is so intensely a matter of personal opinion and the childhood’s particular characteristics. For example, some parents yearn for a home in a prestigious zip code with the children enrolled in an exclusive private school. Other parents think that is a terrible environment for bringing up children and prefer a socioeconomically normal neighborhood and public schools filled with a wide range of children. I know of no data that could adjudicate these different views about the best places to raise children. For that matter, children with different traits will thrive in different environments.
12. Polderman, Benyamin, de Leeuw et al. (2015). The Center for Neurogenomics and Cognitive Research is part of Vrije Universiteit in Amsterdam.
13. Polderman, Benyamin, de Leeuw et al. (2015): 707.
14. The Polderman study’s “best” model was the standard ACE model when 2(rMZ − rDZ) > 0. When 2(rMZ − rDZ) < 0 (i.e., showing a negative value for the shared environment), the “best” model was an ADE model, with D standing for nonadditive genetic influences. This is a conservative approach to estimating h2. In the words of the study, “such per-study choices cause bias and can lead to a 10% downward bias in the reported estimates of h2 in comparison to those based on twin correlations, consistent with the observed discrepancy between our meta-analysis of variance component estimates calculated from twin correlations and the reported variance components.” Polderman, Benyamin, de Leeuw et al. (2015): 705.
15. Plomin (2011): 568.
16. Plomin (2011): 568.
17. Plomin (2011): 568.
18. Turkheimer, Pettersson, and Horn (2014).
19. Plomin (2011): 582.
20. Jencks, Smith, Acland et al. (1972): 66–67.
12: Abilities, Personality, and Success
1. Herrnstein (1973): 197–98.
2. The concept of g was famously derided by Stephen Jay Gould in his bestselling book The Mismeasure of Man. Gould (1981). The book was extremely successful in shaping public opinion about IQ and yet was irrelevant to the state of psychometrics at the time Gould was writi
ng. For technical reviews, see Blinkhorn (1982); Davis (1983); Humphreys (1983); and Carroll (1995). For a recent deep dive into the details of Gould’s errors, see Warne (2019). In 1998, Arthur Jensen published his magnum opus, The g Factor. Jensen (1998). In the 20 years since then, the biological reality of g has been established in new ways by advances in neuroscience. See Haier (2016).
3. Ritchie (2015).
4. See note 4 about test bias in the introduction to Part III.
5. A recent review of the effects of education on IQ is Ritchie and Tucker-Drob (2018).
6. On this issue, the Task Force on IQ that I cited extensively in the notes to Part III wrote:
Intelligence test scores are fairly stable during development. When Jones and Bayley (1941) tested a sample of children annually throughout childhood and adolescence, for example, scores obtained at age 18 were correlated r = +.77 with scores that had been obtained at age 6 and r = +.89 with scores from age 12. When scores were averaged across several successive tests to remove short-term fluctuations, the correlations were even higher. The mean for ages 17 and 18 was correlated r = +.86 with the mean for ages 5, 6, and 7, and r = +.96 with the mean for ages 11, 12, and 13. (Neisser, Boodoo, Bouchard et al. (1996): 81).
7. Jensen (1998): chapter 4.
8. Gottfredson (1997a).
9. I have not found a source with the actual phrase “Life is an IQ test,” but I (and many others who write about IQ) picked up the idea from Gordon (1997); Gottfredson (1997b); and Gottfredson (2003).
10. Gottfredson (2003).
11. Batty, Wennerstad, Smith et al. (2008).
12. The full discussion of this role of IQ is Gordon (1997).
13. High IQ also apparently reflects broader physiological well-being. In a 68-year follow-up of the Scottish Mental Survey of 1947, whose members were born in 1936, childhood intelligence was inversely associated with all major causes of death. Calvin, Batty, Der et al. (2017).
Human Diversity Page 56