\doc\web\99\08\iqher.txt Chris Brand from The g Factor Newsletter, 4 viii '97 IQ Realism NEW BLOW FOR SOCIAL ENVIRONMENTALISTS The latest 'Nature' (31 vii '97) carries a new meta-analysis of kin correlations for intelligence by statisticians from Carnegie-Mellon University. According to Devlin, Daniels and Roeder, 212 studies covering some 50,000 pairings yield the following correlations for IQ between various kinds of relative. ---------------------------------------------------------- Monozygotic Twins reared together .85 Monozygotic Twins reared apart .74 Dizygotic Twins reared together .59 Siblings reared together .46 Mid-parent / Child together .50 Single parent / Child together .41 Single parent / Child apart .24 Adopting parent / Child together .20 ---------------------------------------------------------- In so far as these figures contain any surprises for a mainstream hereditarian, they are that the correlations between parents and children are on the low side -- especially when the parent and child are 'apart.' Thus the authors argue that the effect of whether or not relatives shared the same womb may have been underestimated previously; and they accordingly suggest Herrnstein and Murray's conclusion of a 60% heritability for IQ producing a stratified society to be "tenuous." As in 'The 'g' Factor', the authors make the important distinction between 'broad' and 'narrow' (breeding true, passed genetically from parent to child) heritability. They reckon narrow heritability may be around 35% -- whereas 'The 'g' Factor' gives around 45% (p. 117). However, there are several problems for the Devlin et al., mainly deriving from the fact that, perhaps as statisticians, they have more interest in analysing their data and 'model-fitting' than in actually presenting the data for proper inspection. (i) It is peculiar that Devlin et al. have no figure for DZA twins. In Pedersen et al. 1992, the DZA r for 88 65-year-olds was only .32 -- not much to show for DZs having the same womb in common. [Pedersen's DZT r was only .22 -- dramatically lower than Devlin et al.'s .59 for DZTs who are mainly children (and thus have arguably not yet had the opportunity to choose their own environments).] (ii) There can be very few studies of 'separated parents and children' -- the authors don't even give the number; and a father may be living apart from his family precisely because his own IQ is very different from that of the child -- e.g. if the child's intelligence has proved a big disappointment and source of family friction. (iii) The authors make no allowance for the point that MZ r's may all be underestimates because it is not uncommon for twins to *diverge* as they compete for maternal blood supply in the womb -- as is brought out in 'The 'g' Factor.' Devlin et al. are certainly right to criticize 'The Bell' for ignoring the difference between broad and narrow heritability and thus somewhat overestimating the degree to which caste-like stratification is likely to develop. But the major message of Devlin et al.'s study, properly construed, is that biological factors, whether genetic or prenatal, are those that account for at least three quarters of human variation in IQ. -- Helpfully, in a 'News and Views' comment for 'Nature', Matt McGue (pp. 417-8) nicely points out that Galton would have been well pleased with this new evidence for "a strong but malleable biological basis" for human intelligence. By contrast, social environmentalists will be gnashing their teeth: for Devlin et al.'s finding that living with your child raises the IQ correlation only from .24 to .41 shows that there is no significantly greater influence of the social environment on intelligence than is estimated in 'The 'g' Factor.' _____________________ ======================================================================== >IQ heritability, the portion of a population's IQ variability attributable to >the effects of genes1, has been investigated for nearly a century, yet it >remains controversial. Covariance between relatives may be due not only to >genes, but also to shared environments, and most previous models have assumed >different degrees of similarity induced by environments specific to twins, to >non-twin siblings (henceforth siblings), and to parents and offspring. We now >evaluate an alternative model that replaces these three environments by two >maternal womb environments, one for twins and another for siblings, along with >a common home environment. Meta-analysis of 212 previous studies shows that our >'maternal-effects' model fits the data better than the 'family-environments' >model. Maternal effects, often assumed to be negligible, account for 20% of >covariance between twins and 5% between siblings, and the effects of genes are >correspondingly reduced, with two measures of heritability being less than 50%. >The shared maternal environment may explain the striking correlation between >the IQs of twins, especially those of adult twins that were reared apart. IQ >heritability increases during early childhood, but whether it stabilizes >thereafter remains unclear. A recent study of octogenarians2, for instance, >suggests that IQ heritability either remains constant through adolescence and >adulthood3, or continues to increase with age2. Although the latter hypothesis >has recently been endorsed4, it gathers only modest statistical support in our >analysis when compared to the maternal-effects hypothesis. Our analysis >suggests that it will be important to understand the basis for these maternal >effects if ways in which IQ might be increased are to be identified. > >Despite its conceptual simplicity, IQ heritability has engendered vitriolic >debates throughout this century, debates that are now recurring5 following the >publication of The Bell Curve by R. Herrnstein and C. Murray6. Part of this >controversy arises because IQ heritability is not well characterized. >Paradoxically, direct versus indirect analytic methods, both of which attempt >to estimate genetic effects on IQ unencumbered by environmental effects, yield >markedly different estimates7. Direct studies presumably eliminate >environmental covariance by evaluating relatives raised apart; indirect studies >presumably eliminate environmental covariance by contrasting results from >different study designs that have complementary environmental covariance terms. >Contrary to their identical objectives, the former typically produces notably >larger IQ heritability estimates. We postulate that this paradox is >attributable in large part to the presence of unacknowledged maternal effects >on IQ. > >In traditional quantitative genetic studies, the environment is divided into >maternal and external constituents1. We expand this dichotomy to take account >of the realities of IQ adoption studies, lumping maternal environment and any >shared external environment of adopted children into a single 'pre-separation >environment'. Because the duration of shared external environment is usually >small, pre-separation and maternal effects are roughly equivalent. Both are >distinct from maternal inheritance, which is sometimes called a maternal >effect8. > >Our IQ analyses focus on the magnitude of the additive and non-additive genetic >components1, estimated to explain 60-85% of the variation in IQ from adult >monozygotic twin studies, and the magnitude of maternal effects, usually >assumed to be negligible. To estimate these effects, we analysed 212 IQ studies >(or more precisely, correlations) based on 50,470 distinct pairings. The >analysis included 204 correlations from studies of zero and first-degree >relatives or their adoptive counterparts9. We supplemented this set with some >new twin studies published after 1981: a study of monozygotic twins reared >apart10, the Swedish adoption/twin study of aging of monozygotic twins reared >together and apart and dizygotic twins reared together11, and two studies of >monozygotic and dizygotic adult twins reared together12. > >Each IQ correlation and related sample size is classified by kind of study >(Fig. 1). We evaluate these data using a standard quantitative genetic model >for the components of variance (Table 1) and Bayesian meta-analysis13, a >standard technique for combining information across studies. Our model is built >on two levels of distributional assumptions: we assume a likelihood model for >the observed correlations among relatives in each type of study; and we specify >a prior distribution for the parameters of the model. We assume any >standardized component of variance (positive correlation) is a priori equally >likely to lie between zero and one. These prior distributions make the Bayesian >parameter estimates similar to maximum-likelihood estimates. > >All correlations were Fisher-transformed and, after transformation, observed >correlations from the same study design follow a normal distribution if they >differ only by measurement error. However, several studies are outliers under a >normal likelihood, suggesting that the studies are more heterogeneous with >respect to variability than is predicted by measurement error alone. >Heterogeneity is expected because some study methods differ markedly9. We >account for the heterogeneity by assuming that the transformed values follow a >t-distribution with two degrees of freedom. > >We fit four models to the data. The richest of these models, model IV, allows >for additive and non-additive genetic effects, twin and singleton maternal >effects, and external environmental effects for twins, siblings and >parent-offspring. The remaining models differ from model IV in that Models I >and II constrain the maternal effects to be zero, and models I and III >constrain the external environmental effects to be equal. > >The models are evaluated informally by comparing the average observed >correlations for each of the nine study designs to the models' predicted values >(Table 2). Model I predicts the correlation for the dizygotic twins raised >together very poorly. Except for monozygotic twins raised apart, the predicted >correlations for each study type are similar for models II-IV. The observed >correlations for monozygotic twins raised apart (Fig. 1) ranges from 0.62 to >0.78, with a mean of 0.74. The most extreme predicted correlation is that >obtained from model II (0.50), and the closest value is from model IV (0.74), >but model III also predicts the correlation well, with a value of 0.68. >Maternal effects seem to be essential for a good fit to the data. > >Models III and IV estimate a large maternal effect for twins and a smaller >effect for siblings (Table 3). Model III attributes 20% (95% Bayesian credible >interval: 15-24%) to the former and 5% (1-8%) to the latter. The difference >between these estimates is intuitively appealing. Twins share the womb >concurrently, whereas siblings share the womb serially. Hence, although a >mother may have similar physiological status and personal habits from one >pregnancy to another, the temporal separation between progeny apparently >diminishes the correlation of sibling IQ. > >Remarkably, the covariance attributed to environmental components is also >roughly equal for models II-IV (Table 3). The notable difference among the >models is how the covariance resulting from shared environments is partitioned. > >The models are compared formally using Bayes factors15,16 to obtain their >posterior probabilities. A priori each model is assumed equally likely. The >best model is model III, with a posterior probability of 0.95, followed by >model IV, with probability 0.05, and models I and II, with probabilities near >zero. Hence model II17, which is commonly used to describe the quantitative >genetics of IQ, is rejected when it is compared to model III, which involves a >common home environment and a maternal environment that varies depending on >whether the pair are twins or not. Model III is favoured over model II by a >posterior factor of almost 10,000 to 1, primarily on account of model II's fit >to the observed data from monozygotic twins reared apart: its predicted value >is about six standard deviations from the mean. Model IV, which combines >features of models II and III, seems to be overparameterized and its >environmental components make little sense (Table 3). > >Model III's parameter estimates (Table 3) define estimates of IQ heritability. >However, the concept of heritability has narrow-sense and broad-sense >interpretations. Narrow-sense heritability accounts for additive genetic >effects only; broad-sense heritability encompasses both additive and >non-additive genetic effects. This distinction is often lost in the 'IQ >debates', even though it is critical to the social implications of IQ >heritability18. Based on model III, broad- and narrow-sense heritability >estimates are 48% and 34%, respectively; both agree closely with estimates from >studies using standard quantitative genetic models19,20 and indirect estimation >methods7,21. > >An IQ enigma has been developed7 by contrasting the typically larger direct >heritability estimates with complementary, smaller indirect estimates. For >example, an indirect estimate of broad-sense heritability is obtained by >multiplying twice the difference between monozygotic and dizygotic twin >correlations (Table 2). For these data (Table 3), the broad-sense heritability >estimate of 0.52 is slightly greater than that obtained from model III, which >is expected because the contrast overestimates heritability by about 0.5D? (see >Table 1). The correlation for monozygotic twins reared apart yields a much >larger direct estimate of 0.74. The difference between estimates can be reduced >by subtracting the maternal effect from the direct estimate, yielding a value >of 0.54. > >Twice the observed correlation for siblings reared apart yields another direct, >biased estimate of broad-sense heritability; the bias is -1/2D?, estimated from >model III to be -7.5%. This direct estimate, 0.48, is much lower than the >direct estimate obtained from the monozygotic twins reared apart, but it is >similar to the indirect estimate and model III's estimate. The similarity to >model III's estimate can be attributed to the balance between the bias and the >5% sibling maternal effect, which the direct estimates ignores. Thus maternal >effects may resolve the IQ enigma by rectifying disparate direct and indirect >heritability estimates. > >It is not difficult to gather evidence for the importance of maternal >environment. There is substantial brain growth in utero, and the brain has 70% >of its final mass within a year of birth. Additionally, various studies >indicate that IQ can be affected by prenatal environment: IQ is positively >correlated with birth weight even after controlling for gestational age22,23; >twins, who usually weigh less than singletons, average 4-7 points lower on IQ >tests24; some dietary supplements can raise IQs25,26; alcohol, drug and >cigarette consumption may lower IQs27,29; and lead exchange from mother to >fetus may reduce IQ30. > >In contrast, the large direct estimates derived from studies of adult twins >raised apart have been explained by the conjecture that, as twins get older, IQ >becomes more heritable. Evidence has been provided to support this 'age >hypothesis'4,12, although the results are inconclusive and relationship >unclear. > >To compare the age- and maternal-effects hypotheses, we related the IQ >correlations to subject age whenever possible. We classified the studies by age >into three categories, namely youths (<14 years), adolescents (14-18 years) and >adults (>18 years), based on median or mean subject age for each study; we >assigned 71% of the twin and sibling studies. Because they are >cross-generational, assigning parent-offspring studies was problematic: in one >treatment we included all such studies as adolescents, and in another we >excluded all such studies except for parent/child apart. We then extended model >II. This age-effect model allowed heritability to increase with age by two >adjustments: a multiplicative increase in the genetic components of variance, >namely b-1/2 for youths, b0 for adolescents, and b2 for adults; and a >multiplicative decrease in the shared environment components of variance, >namely c1/2 for youths, c0 for adolescents, and c-2 for adults. For b and c > >1, these multiplicative functions have forms consistent with those hypothesized >in ref. 12. Our prior distribution for b and c favoured the age hypothesis >slightly by putting uniform mass between (1/2,2). > >The parameter estimates for b (1.13,[1.05-1.19]) and c (1.21,[1.03-1.38]), >based on the analysis of all types of study, are consistent with the age >hypothesis. Because Bayes factors (BF) impose a penalty for added complexity we >also considered a one-parameter age-effect model with b = c. This model is >favoured over the two-parameter model (BF = 2.5), and is also favoured to a >small degree over model II (BF = 11, b = 1.10, [1.04-1.16]). Similar results >are obtained when parent-offspring studies are excluded. However, no version of >model II is competitive with model III by Bayes factor. Moreover, age-effects >models fail to fit the data better than a simpler model that invokes maternal >effects. > >Our statistical analyses cannot, of course, be considered definitive. Age may >affect IQ heritability, even though the age model does not gain much support >from our analyses. Moreover, our analyses do not preclude other, unmodelled >factors, such as cultural inheritance and interaction between genes and the >environment, from having important effects on IQ. Despite these caveats though, >several important conclusions emerge from our results. > >Adoption designs are a popular means of estimating IQ heritability. Associated >analyses, however, usually assume negligible maternal effects. By contrast, our >results show that 20% of twin and 5% of sibling covariance may be attributable >to maternal effects. These results have two implications: a new model may be >required regarding the influence of genes and environment on cognitive >function; and interventions aimed at improving the prenatal environment could >lead to a significant increase in the population's IQ. Moreover, some of >Herrnstein and Murray's conclusions6 regarding human evolution such as the >development of cognitive castes and IQ dysgenics, arise from their belief that >IQ heritability is at least 60%, and is probably closer to the 80% values >obtained from adoption studies. Our results suggest far smaller heritabilities: >broad-sense heritability, which measures the total effect of genes on IQ, is >perhaps 48%; narrow-sense heritability, the relevant quantity for evolutionary >arguments because it measures the additive effects of genes, is about 34%. >Herrnstein and Murray's evolutionary conclusions are tenuous in light of these >heritabilities18. > >1. Falconer, D. S. Introduction to Quantitative Genetics (Longman, New York, >1981). >2. McClearn, G. E. et al. Substantial genetic influence on cognitive abilities >in twins 80 or more years old. Science 276, 1560-1563 (1997). >3. Gottesman, I. I. Twins: en route to QTLs for cognition. Science 276, >2522-1523 (1997). >4. Task Force for Scientific Affairs. Intelligence: Knowns and Unknowns >(American Psychological Association, Washington DC, 1995). >5. Devlin, B., Fienberg, S., Resnick, D. & Roeder, K. (eds) Intelligence, Genes >and Success: Scientists Respond to The Bell Curve (Springer, New York, 1997). >6. Herrnstein, R. J. & Murray, C. The Bell Curve: Inteligence and Class >Structure in American Life (Free Press, New York, 1994). >7. Plomin, R. & Loehlin, J. C. Direct and indirect IQ heritability studies: a >puzzle. Behav. Genet. 19, 331-342 (1989). >8. Rose, R. J. et al. Data from kinships of monozygotic twins indicate maternal >effects on intelligence. Nature 283, 375-377 (1980). >9. Bouchard, T. J. Jr & McGue, M. Familial studies of intelligence: a review. >Science 250, 223-228 (1990). >10. Bouchard, T. J. Jr, Lykken, D. T., McGue, M., Segal, N. L. & Tellegen, A. >Sources of human psychological differences: the Minnesota study of twins reared >apart. Science 250, 223-228 (1990). >11. Pedersen, N. L., Plomin, R., Nesselroade, J. R. & McClearn, G. E. A >quantitative genetic analysis of cognitive abilities during the second half of >the life span. Psychol. Sci. 3, 346-353 (1992). letters to nature >12. McGue, M., Bouchard, T. J. Jr, Iacono, W. G. & Lykken, D. T. in Nature, >Nurture and Psychology (eds Plomin, R. & McClearn, G. E.) 59-76 (American >Psychological Association, Washington DC, 1993). >13. National Research Council. Combining Information. Statistical Issues and >Opportunities for Research (National Academy Press, Washington DC, 1992). >14. Smith, A. F. M. & Roberts, G. O. Bayesian computation via the Gibbs sampler >and related Markov chain Monte Carlo methods. J. R. Stat. Soc. B 55, 3-24 >(1993). >15. Kass, R. E. & Raftery, A. E. Bayes Factors. J. Am. Stat. Assoc. 90, 773-795 >(1995). >16. Verdinelli, I. &Wasserman, L. Computing Bayes factors by using a >generalization of the Savage-Dickey density ratio. J. Am. Stat. Assoc. 90, >614-618 (1995). >17. Plomin, R., DeFries, J. C. & McClearn, G. E. Behavioral Genetics: A Primer >(Freeman, New York, 1990). >18. Daniels, M., Devlin, B. & Roeder, K. in Intelligence, Genes and Success: >Scientists respond to The Bell Curve (eds Devlin, B., Fienberg, S., Resnick, D. >& Roeder, K.) (Springer, New York, 1997). >19. Chipuer, H. M., Rovine, M. J. & Plomin, R. LISREL modeling: genetic and >environmental influences on IQ revisited. Intelligence 14, 11-29 (1990). >20. Rao, D. C., Morton, N. E., Lalouel, J. M. & Lew, R. Path analysis under >generalized assortative mating. II. American IQ. Genet. Res. 39, 187-198 >(1982). >21. McCartney, K., Harris, M. J. & Bernieri, F. Growing up and growing apart: a >developmental meta-analysis of twin studies. Psychol. Bull. 107, 226-237 >(1990). >22. Lynn, R. & Hattori, K. The heritability of intelligence in Japan. Behav. >Genet. 20, 545-546 (1990). >23. Churchill, J. A. The relationship between intelligence and birth weight in >twins. Neurology 15, 341-347 (1965). >24. T. Husen, T. Intra-pair similarities in the school achievement of twins. >Scand. J. Psychol. 4, 108-144 >(1963). >25. Harrell, R. F., Woodyard, E. & Gates, A. I. The Effects of Mothers' Diets >on the Intelligence of Offspring (Teacher's College, New York, 1955). >26. Rush, D., Stein, Z., Susser, M. & Brody, N. in Diet in Pregnancy: A >Randomized Controlled Trial of Nutritional Supplements (eds Rush, D., Stein, Z. >& Susser, M.) 88-131 (Liss, New York, 1980). >27. Reinisch, J., Sanders, S. A., Mortensen, E. L. & Rubin, D. B. In utero >exposure to phenobarbital and intelligence deficits in adult men. J. Am. Med. >Assoc. 274, 1518-1525. >28. Streissguth, A. P., Barr, H. M., Sampson, P. D., Darby, B. L. & Martin, D. >C. IQ at age 4 in relation to maternal alcohol and smoking during pregnancy. >Dev. Psychol. 25, 3-11 (1989). >29. Olds, D. L., Henderson, C. R. Jr & Tatelbaum, R. Intellectual impairment in >children of women who smoke cigarettes during pregnancy. Pediatrics 93, 221-227 >(1994). >30. Baghurst, P. A. et al. Environmental exposure to lead and children's >intelligence at the age of seven >years. N. Engl. J. Med. 327, 1279-1284 (1992). > >Acknowledgements. We thank T. J. Bouchard Jr for compilations of IQ data, and >J. C. Loehlin, M. McGue, S. Fienberg and R. Kass for comments on a previous >draft of the manuscript. This work was supported by grants from the National >Science Foundation and the National Institutes of Health. > >Correspondence and requests for materials should be addressed to B.D. e-mail: >devlinbj@msx.upmc.edu -- Chris Brand Edinburgh. Author of 'The 'g' Factor' [Wiley DePublisher, 1996]. For overview, CV, psychology articles and access to McDougall NewsLetters [for neoliberalism], see http://www.crispian.demon.co.uk. -- "Where the race is right, the place is right". Ralph Waldo EMERSON (1803-1882). The Complete Writings of Ralph Waldo Emerson. New York : Win, Wise & Co., 1929. --