In this monograph, we have investigated the genetic and environmental origins of individual differences in performance in academic subjects (English, mathematics, and science) and general cognitive ability during the early school years. We began with the basic nature–nurture question about the relative influence of genes and environment (Chapter III). However, our main goal was to address three questions that go beyond this rudimentary question: (1) the etiological relationship between the normal (learning abilities) and the abnormal (learning disabilities), (2) genetic and environmental contributions to longitudinal stability and change from 7 to 10 years, and (3) genetic and environmental heterogeneity and homogeneity within and between learning abilities (English, mathematics, and science) as well as their relationship to general cognitive ability. These three themes were the topics of Chapters IV, V, and VI, respectively. In this final chapter, we begin by discussing our findings in relation to these three themes and then we return to more general issues related to nature and nurture that emerge from the results presented in Chapter III.
THE ABNORMAL IS NORMAL
The results presented in Chapter IV lead us to conclude that learning disabilities are the quantitative extreme of the same genetic and environmental influences that operate throughout the normal distribution of learning abilities. Stated more provocatively, these results suggest that there are no learning disabilities, just the low end of the normal distribution of learning abilities. Using reading disability as an example, when the genes responsible for the several replicated linkages are identified (McGrath, Smith, & Pennington, 2006; Fisher & Francks, 2006), we predict that these same genes will be associated with normal variation in reading ability, not just with reading disability. That is, even in pairs of siblings who are both good readers, we would expect that siblings with one or two copies of the “beneficial” allele will be better readers than their co-siblings who only have the other allele. Similarly, both shared and nonshared environmental factors associated with poor reading are also expected to be associated with variation throughout the normal distribution including good reading. This conclusion may appear counterintuitive, given the long tradition in psychology and education of viewing reading disability as a qualitatively distinct category. In fact, there is considerable convergence with an emerging cognitive view of variability in reading that emphasizes a continuum of variability. For example, poor readers have the same difficulties and make the same kinds of errors in reading as average readers, just more of them, and they last longer (cf., Catts & Kamhi, 2005).
Although we presented results using a 15% cutoff for reasons discussed in Chapter IV, we have conducted similar analyses using a 5% cutoff and found similar results. Of course, different results could emerge if different phenotypes were used for selection, such as selecting for a syndrome of multiple traits or selecting for specific learning disability in which children with poor performance are also required to have normal “g.” In a study of this latter type comparing specific versus nonspecific language impairment in 4-year-olds in TEDS, some differences appeared although power to detect such differences was modest (Hayiou-Thomas, Oliver, & Plomin, 2005). As indicated in Chapter VI, there is substantial genetic overlap between learning abilities and “g,” which suggests that specific language impairment with reduced variance in “g” could quite plausibly yield different results. One practical problem is that when selecting probands for a complex phenotype of this sort, it is difficult to know what quantitative trait to use for co-twins in DF extremes analyses.
Although we have focused on low learning ability because of its educational and societal importance, to what extent is high ability also the quantitative extreme of the same genetic and environmental factors responsible for normal variation in ability? High ability has been a nature–nurture battleground. For example, some theorists have argued that high performance is driven entirely by the time and effort spent developing relevant skills (e.g., Ericsson, Krampe, & Tesch-Romer, 1993; Howe, Davidson, & Sloboda, 1998). Others have argued for the primacy of innate brain-based differences (e.g., Geschwind & Galaburda, 1987). However, very little is actually known about the origins of high academic performance (Plomin & Thompson, 1993). The only genetic studies of this type focused on “g” rather than academic performance (e.g., Ronald, Spinath, & Plomin, 2002). In the first genetic study of high mathematics ability (Petrill, Kovas, Hart, Thompson, & Plomin, submitted), TEDS' web-based test data at 10 years yielded results similar to results reported in Chapter IV for the low end of the distribution and for the entire distribution of individual differences (Kovas, Haworth, Petrill, & Plomin, in press): substantial genetic influence, modest shared environmental influence, and moderate nonshared environmental influence.
As with low ability, it would be fallacious to pose the question of the source of high ability as a question of nature versus nurture. The substantial heritability of high ability does not mean that genes simply turn on and cause a child to perform at high levels. Although skills can be taught and high levels of performance can be attained regardless of genetic propensities, even at high levels of performance differences will remain and genetics is likely to play just as large a role at this high end of the distribution. Moreover, nature and nurture are not separate tracks in development. It is clear that high-performing children are more likely to engage in activities such as deliberate practice that enhance their abilities (Ericsson, Krampe, & Tesch-Romer, 1993). The substantial genetic influence at the high end of the distribution suggests that engaging in deliberate practice is in part a function of genes influencing ability indirectly, but powerfully, through motivation. Put more simply, genes code for appetites, not just aptitudes. Such gene–environment transactions are important for understanding why some children fail to benefit fully from enriched environments and why others reach high levels of performance despite environmental privation.
Quantitative Trait Loci (QTLs)
A model for understanding why the abnormal is normal is the QTL hypothesis, which suggests that a polygenetic continuum of genetic risk underlies a continuum of variation in behavior in the population and that common disorders lie at the extreme end of this normal variation (see Plomin et al., in press, for more detail). The QTL model refers to quantitative traits even in relation to disorders because if many genes affect a disorder, then it necessarily follows that there will be a quantitative distribution rather than a dichotomy. As with all of our conclusions based on the quantitative genetic research presented in this monograph, definitive proof that the abnormal is normal will come when genes identified for learning disabilities are found to be associated with the normal range of variation in learning abilities and vice versa.
The conclusion that the abnormal is normal is limited to common disorders. For all complex disorders—including medical disorders such as obesity and heart disease as well as behavioral problems such as mental retardation—there are rare, highly penetrant mutations that can create extreme versions of a disorder, which may show qualitative differences from normal variation. For example, contrary to the QTL hypothesis, an apparently unique genetic contribution to language impairment was hailed in the discovery of the FOXP2 mutation in the KE family (Lai, Fisher, Hurst, Vargha-Khadem, & Monaco, 2001). The FOXP2 mutation appears to be both necessary and sufficient for the 15 affected members of the KE family with an unusual type of speech–language impairment that includes deficits in oro-facial motor control. However, the FOXP2 mutation was not found in a single one of 270 children with low language ability in TEDS (Meaburn, Dale, Craig, & Plomin, 2002). More generally, hundreds of rare mutations with effects on “g” have been identified (Inlow & Restifo, 2004), but together these mutations appear to account for <1% of cases of mental retardation. We predict that many QTLs of small effect rather than one or two genes of large effect will account for most of the genetic variation in learning disabilities. We also predict that these QTLs will relate to variation in learning ability throughout the normal distribution.
Quantitative Trait Neural Processes (QTNs)
If learning disabilities involve many QTLs of small effect, then there are also likely to be many brain mechanisms that mediate the effects of these QTLs on learning disabilities. In other words, learning disabilities may be the extremes of the same brain and cognitive processes that are responsible for normal variation, as opposed to a “broken brain” with one malfunctioning part like a lesion that lights up in neuroimaging studies. We offer the term “qualitative trait neural processes” (QTNs) to highlight the possible parallels with QTLs (Kovas & Plomin, 2006). Both QTNs and QTLs support a shift of thinking about diagnosed abnormal individuals toward thinking about normal variation.
GENETIC STABILITY, ENVIRONMENTAL CHANGE
The results of longitudinal genetic analyses presented in Chapter V suggest that age-to-age stability is primarily mediated genetically whereas the environment, especially nonshared environment, contributes to change from age to age. Chapter V began by reporting remarkably similar quantitative ACE estimates at 7, 9, and 10 years of age, even for “g” for which the measures were as different as could be at 7 (telephone testing), 9 (mailed booklet), and 10 (web-based testing). It is striking that ACE estimates are so similar across this third of the children's lives despite major changes in their cognitive development and in the content of the measures. Nonetheless, ACE estimates could be similar from age to age even if different ACE factors operated at each age. Longitudinal analyses of the etiology of age-to-age change and continuity are key to understanding the development of individual differences in learning abilities and disabilities.
As discussed in Chapter V, longitudinal genetic analyses yield two types of genetic statistics: bivariate heritability and genetic correlation. Bivariate heritabilities, which indicate the proportion of the phenotypic correlation from age to age that is mediated genetically, are about .75 on average for NC teacher ratings across 7, 9, and 10 years. The reading tests from 7 to 10 years yield a bivariate heritability of .83. These bivariate heritabilities suggest that age-to-age stability of academic and cognitive abilities is largely mediated genetically.
Genetic correlations estimate the extent to which genetic influences at one age correlate with genetic influences at another age regardless of their heritability—that is, bivariate heritability could be low but genetic correlations could be high. Genetic correlations can be considered as the probability that a gene associated with a trait at one age is also associated with the trait at the other age. The genetic correlations from 7 to 10 years are .67 and .68 for NC teacher ratings for English and mathematics, respectively, .60 for reading tests, and .72 for “g.” These high genetic correlations across one-third of the children's lives indicate that genetic effects are largely stable, which is remarkable given the developmental changes during middle childhood. However, because the genetic correlations are <1.0, they also suggest some changes in genetic effects from age to age.
Molecular genetic studies that identify the genes responsible for the high heritability of learning abilities and disabilities will provide the definitive test of this conclusion derived from quantitative genetic analyses. These quantitative genetic analyses predict that the chances are about two-thirds that a gene found to be associated with learning abilities at 7 years would also be associated with learning abilities at 10 years.
Nonetheless, this glass can also be seen as about one-third empty: The chances are about one-third that a gene associated at 7 years would not be associated at 10 years. What about molecular genetic studies with samples of a wide range, as is the case for most genetic studies? Longitudinal analyses typically yield a simplex pattern of correlations in which correlations are lower as age intervals increase. If age-to-age genetic correlations follow this simplex pattern, genetic stability will be less the longer the age interval. For a wide age interval—for example, from childhood to adulthood—age-to-age genetic correlations might be quite low. If this were the case, molecular genetic studies with a wide age range would only be able to detect the most age-general genes. Although it would be important to identify such age-general genes, given the evidence for age changes in genetic effects, many genetic effects across a wide age range would not be age-general and such studies would be unlikely to detect these age-specific genes. Given that genes largely contribute to stability both for ability and disability, longitudinally stable phenotypes—for example, children who have shown low performance for a particular learning ability throughout childhood—seem to be the best targets for molecular genetic studies. The most important benefit of identifying genes that put children at risk for developing learning disabilities is that the genes can be used as an early-warning system to predict problems before they occur. Genes associated with learning problems at 7 years can be used to predict early in life a child's risk for developing learning problems at 7 years. Although most of the genes associated with learning problems at 7 years will also be associated with learning problems at 10 years, the genetic prediction could be sharpened by focusing on those genes that are stably associated with learning problems at 7 and 10 years and beyond. In addition, genes that contribute to change from 7 to 10 years could be used to predict problems that are unlikely to develop until 10 years or transitory problems at 7 years that will be resolved by 10.
The value of early prediction is the opportunity it affords for prevention. Identifying children in early childhood who are genetically at risk for learning problems in middle childhood will encourage research that charts the developmental course of the learning problems and research that intervenes to change the course of development. This goal is achievable even in the case of skills such as reading that do not occur until later in development. Reading is a good example because there is a large and widely accepted body of evidence that phonology—and specifically the ability to reflect on the sound structure of spoken words—lies at the core of reading development and reading problems (Goswami & Bryant, 1990). These issues are discussed more fully later in this chapter.
Because about 75% of phenotypic stability from 7 to 10 years is mediated genetically, it necessarily follows that about 25% is mediated environmentally. Nearly all of this environmental stability is due to shared environment, as indicated by the bivariate shared environment estimates in Table 14. In terms of environmental correlations rather than bivariate environmental estimates, we found that shared environmental correlations from 7 to 10 years are almost as high as the genetic correlations: .71 for NC English, .52 for NC mathematics, and .45 for reading tests, but only .30 for “g.” However, nonshared environmental correlations are uniformly low: .26, .20, .11, and .03, respectively. In other words, nonshared environment largely contributes to change.
What are these nonshared environmental sources of change? Nonshared environment, which accounts for more variance than shared environment, is a major mystery for learning abilities and disabilities because the twins live in the same family, attend the same school, and are often even in the same classroom. Nonshared environment is discussed later in this chapter, but for now we simply mention another piece of this puzzle: Not only do nonshared environmental influences on learning abilities and disabilities make two children in the same family different from one another, they also make children at one age different from themselves at another age. The motivation for identifying significant nonshared environmental features should be at least as strong as the motivation for identifying DNA markers because nonshared environment appears to be the major source of change, and change is the essence of education.
GENERALIST GENES, SPECIALIST ENVIRONMENTS
Multivariate genetic analyses presented in Chapter VI lead to the conclusion that genes are generalists and nonshared environments are specialists. That is, genes largely contribute to similarity in performance within and between learning abilities, and between learning abilities and general cognitive ability, whereas nonshared environment contributes to differences in performance.
Within domains, genetic correlations were extraordinarily high: .87 on average for the three components of each domain of NC teacher ratings, .88 for the two subtests of the TOWRE, and .87 for the three components of the mathematics battery. This suggests that the components within each domain are nearly the same thing from a genetic perspective. Even more surprising were the high genetic correlations between domains. The average genetic correlation among NC teacher ratings of English, mathematics, and science at 7, 9, and 10 years was .79. The genetic correlation was .52 between the web-based tests of reading and mathematics at 10 years. Bivariate heritabilities were also substantial: .67 within domains and .64 between domains for NC ratings for the three ages, which indicates that about two-thirds of the phenotypic correlation between these domains is mediated genetically.
Our results are similar to those of other multivariate genetic studies on learning abilities and disabilities, which consistently yield high genetic correlations. For example, the first study in this area using standard measures of reading and mathematics reported a genetic correlation of .98 between reading and mathematics (Thompson, Detterman, & Plomin, 1991). In a recent review, genetic correlations varied from .67 to 1.0 for reading versus language (five studies), from .47 to .98 for reading versus mathematics (three studies), and from .59 to .98 for language versus mathematics (two studies) (Plomin & Kovas, 2005). The average genetic correlation between domains was about .70. We refer to these genetic effects as “generalist genes” in order to highlight the general effect of genes within and between learning abilities and disabilities (Plomin & Kovas, 2005).
We also found that some of these generalist genes that affect learning abilities are even more general in that they also affect other sorts of cognitive abilities included in the “g” factor. As reported in Chapter VI, the average genetic correlation between learning abilities and “g” is about .60. We argued there that academic performance is not just “g.” Although about a third of the genetic variance of English and mathematics is in common with “g,” about a third of the genetic variance is general to academic performance but not “g.”
Similar to issues discussed above in relation to genetic stability, the fact that genetic correlations are <1.0 means that there are also genes that contribute to predisposing children to perform better in one domain than another. Because genetic influence on learning abilities is substantial, such specialist genes contribute importantly to dissociations among learning abilities even though most genes are generalists.
As mentioned in the previous section, definitive proof of the importance of generalist genes will come from molecular genetic research. The prediction is clear: Most (but not all) genes found to be associated with a particular learning ability or disability (such as reading) will also be associated with other learning abilities and disabilities (such as mathematics). In addition, most (but not all) of these generalist genes for learning abilities (such as reading and mathematics) will also be associated with other cognitive abilities (such as memory and spatial).
When these generalist genes are identified, they will greatly accelerate research on general mechanisms at all levels of analysis from genes to brain to behavior. Implications of generalist genes for cognitive and brain sciences have recently been discussed (Kovas & Plomin, 2006).
Implications of generalist genes for translational research are also far-reaching. The most immediate implication is that, from a genetic perspective, learning disabilities are not distinct diagnostic entities.
Multivariate genetic research also has an interesting story to tell about environmental influences on learning abilities and disabilities. Shared environmental influences are also generalists: Shared environmental correlations are at least as high as genetic correlations. However, nonshared environmental correlations are on average half the magnitude of the genetic correlations, about .40 on average within and between learning abilities, although they vary considerably across domains. Nonshared environmental correlations between learning abilities and “g” are very low, about .10 on average.
We conclude that nonshared environmental influence is largely specific to each learning ability. This adds another piece to the puzzle of nonshared environment. As noted in the previous section, not only do nonshared environmental influences on learning abilities make two children in the same family different from one another, but they also make children at one age different from themselves at another age. Now we add the additional clue that these nonshared environmental influences also make children different across domains of learning. In other words, nonshared environments are specialists. It is difficult to imagine what such nonshared environmental influences might be; we return to this issue in the following section. As mentioned in the previous chapter, one implication of this conclusion is that educational influences might have their greatest impact on remediating discrepant performances among learning abilities such as differences in children's performance in reading and mathematics.
General limitations of the twin method and specific limitations of the present study were discussed in Chapter II. Three other limitations may be especially relevant to our finding of substantial heritability and modest shared environment. The first limitation involves the possibility of assortative mating. Assortative mating, which is the correlation between spouses, inflates the DZ twin correlation but does not affect the MZ twin correlation (Plomin et al., in press). Thus, assortative mating could have deflated our heritability estimates and inflated our shared environment estimates, even though the heritabilities are so high and the estimates of shared environment are so low. Assortative mating is substantial in the cognitive domain, about .40 for “g” (Jensen, 1978), although assortative mating for academic performance itself is not known. If assortative mating for parents' academic performance were also as high as .40, heritabilities adjusted for assortative mating could be as high as 80% and shared environment could be as low as 0% for learning abilities.
A second limitation that could also have inflated our estimates of shared environment is the possibility that twins share environmental experiences to a greater extent than nontwin siblings because twins are the same age and thus travel through life together. In early childhood, TEDS research indicated that for cognitive abilities, estimates of the role of shared environment were more than twice as large for twins as compared with nontwins siblings, suggesting that about half of twin study estimates of shared environment for cognitive abilities in early childhood are specific to twins (Koeppen-Schomerus, Spinath, & Plomin, 2003). We will be able to assess this possibility for learning abilities in middle childhood because younger siblings of the TEDS twins are also being assessed as they reach middle childhood.
Unlike the previous two limitations, which perversely suggest ways in which our high estimates of heritability could be even higher and our low estimates of shared environment could be even lower, the third limitation could explain why our heritability estimates are so high and our shared environmental estimates are so low. The U.K. National Curriculum provides similar curricula to all students, thus diminishing a potentially important source of environmental variation across schools, to the extent that the curriculum actually provides a potent source of environmental variation. In contrast, the educational system in the United States is one of the most decentralized national systems in the world. To the extent that these differences in educational policy affect children's academic performance, we would expect greater heritability and lower shared environment in the United Kingdom than in the United States. In other words, all other things being equal, greater equality in educational opportunity should lead to greater heritability. Differences in samples, ages, and measures among twin studies of learning abilities and disabilities make it difficult to compare U.K. and U.S. results. In particular, very large samples are needed to provide reasonable power to detect differences in heritability. The only large study other than TEDS is the U.S. study of bright twins in high school (Loehlin & Nichols, 1976). Results for that study are in the direction predicted by the “national curriculum” hypothesis: Heritability is lower than in TEDS' NC ratings (40% vs. 60%) and shared environment is higher (30% vs. 15%). One other U.S. study, with a smaller sample, also reported lower heritability (40%) and higher shared environment (40%) than TEDS (Thompson, Detterman, & Plomin, 1991). However, the national curriculum hypothesis is not supported by the results from studies in the Netherlands (Bartels, Rietveld, van Baal, & Boomsma, 2002b) where there is a national curriculum but one that is less prescriptive than the U.K. curriculum, and in Australia (Wainwright, Wright, Luciano, Geffen, & Martin, 2005) where there is no national curriculum other than for literacy (O'Donnell, 2004). The sample sizes in these other studies are not nearly large enough to provide adequate power to compare results across studies, however. We are currently coordinating our U.K. TEDS study with a U.S. study with the same measures at the same ages in order to be able to draw explicit comparisons between the United Kingdom and the United States (Petrill & Plomin, 2006). A comparison of ACE estimates for prereading skills and early literacy in United States, Australia, and Scandinavia generally found similar results across the countries but some evidence emerged in favor of the national curriculum hypothesis despite small sample sizes that limited the power to detect differences (Samuelsson et al., 2005).
Another, and conceptually more significant limitation of this Monograph is that it has not tackled an issue of great importance: the interplay between nature and nurture. Issues of gene–environment interaction and correlation interest us greatly (e.g., Asbury, Wachs, & Plomin, 2005; Plomin & Davis, 2006) and we believe that they will be important topics in relation to school environments and learning abilities and disabilities. However, our initial forays in this direction, mentioned above, have been disappointing. We found that school characteristics and children's perceptions of their school environment account for little variance in children's academic performance (Walker, Petrill, & Plomin, 2005; Walker & Plomin, 2006). We are currently attempting to improve our measures of children's perceptions of school environment by conducting interviews with 50 pairs of MZ twins for 10 consecutive school days in collaboration with David Almeida, who has developed the use of diary methods to assess daily stressors (Almeida, 2005).
In this final section, we return to more general issues about nature and nurture that emerge from the research reported in this monograph, beginning with three surprises.
(1) Substantial heritability and modest shared environment. The results surprised us by showing such substantial heritability and such modest shared environmental influence for learning abilities in the early school years. Heritabilities are about 65% for teacher assessments based on U.K. National Curriculum criteria and about 55% for test data. Heritabilities for learning abilities are considerably greater than for general cognitive ability (about 35%). The similarity of results across domains, across ages, and across methods of assessment indicates the robustness of these findings.
The modest contribution of shared environment was just as surprising because the twins grew up in the same family, attended the same school, and were often taught by the same teacher in the same classroom. Across domains and across age, the average estimate of shared environment is about 15% for the NC ratings and about 20% for the test data. In research reported elsewhere we have found that more than 80% of the shared environment for NC ratings can be accounted for by socioeconomic status (Walker, Petrill, & Plomin, 2005). The rest of the shared environment was accounted for by school characteristics as measured using U.K. government statistics on variables such as class size and student–teacher ratio, authorized and unauthorized absence, average NC achievement level, and percentage of students eligible for free school meals (Walker, Petrill, & Plomin, 2005).
Nonshared environment accounted for more variance than shared environment—20% for NC ratings and 25% for test data. Again, we note that nonshared environment includes error of measurement. However, nonshared environment is not solely error of measurement, as can be seen in the longitudinal and multivariate analyses in Chapters V and VI. The longitudinal analyses in Chapter V indicate that the average nonshared environmental correlation is .24 from 7 to 10 years for NC teacher ratings of English and mathematics (Table 14). This suggests that the chances are about one in four that a nonshared environmental factor associated with learning abilities at 7 will also be associated with learning abilities at 10. In other words, an environmental factor at 7 years that makes one member of an MZ pair better at mathematics than the co-twin also makes that same MZ co-twin better at mathematics at 10 years. In addition, the multivariate analyses in Chapter VI yield an average nonshared environmental correlation of .42 for NC teacher ratings of English and mathematics (Table 16). Although it is possible that such nonshared environmental factors could involve correlated error, they are at least systematic in their effect and independent of genetic and shared environmental influences and thus warrant further investigation.
What environmental factors could make siblings, even MZ twins, different from one another in learning abilities? As mentioned earlier, research reported in this monograph adds two more pieces to the puzzle of nonshared environment: Not only do nonshared environmental influences on learning abilities make two children in the same family different from one another, they also make children at one age different from themselves at another age and they make children different across domains of learning.
Nearly all research attempting to identify specific sources of nonshared environment has focused on family environments rather than school environments and on personality and behavior problems rather than learning abilities. Nonetheless, such research should be informative for future research that will attempt to identify nonshared school environments that affect learning abilities. A meta-analysis of 43 papers relating differential family experience of siblings to differential outcomes concluded that “measured nonshared environmental variables do not account for a substantial portion of nonshared variability” (Turkheimer & Waldron, 2000, p. 78). Although another review interpreted these results more optimistically (Plomin, Asbury, & Dunn, 2001), the search for nonshared environments in school might best begin outside the family.
For example, peers have been nominated as an important candidate for nonshared environment as siblings in a family make their own individual ways in the world outside their family (Harris, 1998), and initial research appears to confirm this prediction (Iervolino et al., 2002). However, peers would not seem likely to be able to explain why nonshared environmental factors change so much from year to year, nor would peers easily explain why nonshared environmental factors differ from one academic subject to another. We thought that children's perceptions of their school environment might be better able to address these new pieces to the puzzle because children's perceptions could differ across time and across subjects. We have conducted research that suggests that children's perceptions of their school environment are a potent source of nonshared environmental experience in school (Walker & Plomin, 2006). However, the problem is that these nonshared environmental experiences hardly relate to nonshared environmental variance in academic achievement.
We also need to consider the possibility that chance contributes to nonshared environment in terms of random noise, idiosyncratic experiences, or the subtle interplay of a concatenation of events (Plomin, Asbury, & Dunn, 2001). Chance is the most obvious candidate for explaining the age-specific and subject-specific nature of nonshared environment in learning abilities. Nonetheless, our view is that chance is the null hypothesis and that systematic sources of nonshared environment need to be thoroughly examined before we dismiss it as chance. Moreover, chance might only be a label for our current ignorance about the environmental processes by which children—even pairs of MZ twins—in the same family and same classroom come to be so different. Using differences within pairs of MZ twins is a particularly powerful strategy for identifying nonshared environmental effects independent of genetics (Asbury, Dunn, Pike, & Plomin, 2003; Asbury, Dunn, & Plomin, 2006).
(2) Similar results for boys and girls. The results are similar for boys and girls as well as for same-sex and opposite-sex twins, suggesting that neither quantitative nor qualitative sex differences play an important role in the origins of individual differences in learning abilities.
(3) Similar results for same teacher and different teachers. Heritability estimates for NC teacher ratings were similar when the same teacher assessed both members of a twin pair and when different teachers assessed them, which provides strong support for the validity of the heritability estimates.
Three unsolved puzzles also emerged from our analyses:
(1) Why are teacher ratings of academic performance at 10 years more heritable than test scores? For teacher ratings of reading at 10 years, heritability was 52%; for the web-based PIAT test of reading comprehension, heritability was 39%. For mathematics, heritability was 64% for teacher ratings and 49% for the web-based mathematics composite test score. Neither of these differences was significant; moreover, the heritability difference was not seen for reading at 7 years where heritabilities were 68% for NC ratings and 70% for TOWRE. Nonetheless, the heritability differences at 10 years warrant further consideration because they are substantial and consistent, especially if we find similar differences in the future when the twins are assessed again at 12 years.
Although teacher ratings and test scores at 10 years correlate about .50 phenotypically and about .60 genotypically, this leaves plenty of room for differences in the two types of measures. We explored this difference by comparing patterns of correlations with other variables in TEDS for teacher ratings versus test scores. For example, the heritability differences might be due to the possibility that teachers' year-long evaluation of children yield deeper insights into children's capabilities, including their appetites as well as their attitudes. However, in analyses of children's self-perceptions of reading and mathematics ability and their liking of these subjects (Spinath, Spinath, Harlaar, & Plomin, 2006), correlations with teacher ratings were not greater than correlations with test scores. We also considered the possibility that web-based tests show less heritability because they entail more artifactual shared environmental influence due to differences in the testing situation in the home such as more or less chaotic homes or more or less experience and comfort with computers. In support of this hypothesis, shared environment is slightly greater for reading and mathematics test scores (25% and 19%) than for teacher ratings (20% and 12%). However, in multiple regression analyses of home measures such as chaos, parental discipline and socioeconomic status, we again found similar correlations for teacher ratings and test scores.
Because reliability creates a ceiling for heritability estimates, a third possibility is that teacher ratings might be more reliable than test scores. This hypothesis is supported by the finding that nonshared environment, which includes measurement error, is slightly but significantly lower for teacher ratings than for test scores: 28% and 24% for NC teacher ratings of reading and mathematics and 36% and 32% for tests of reading and mathematics (Table 11). However, as discussed in Chapter II, our web-based test scores show high internal consistency, and test–retest reliability of the PIAT across 7 months was .66 in a study of 55 TEDS children. Moreover, a study of 30 TEDS children yielded a correlation of .92 between our web-based mathematics test and a standard version of the test administered in person 2 months later, which suggests that the web-based test is both highly reliable and valid (Haworth et al., 2007). A direct test of this hypothesis of differential reliability and stability will be possible in TEDS when the 10-year results for teacher ratings and test scores can be compared with similar measures that will be included in the 12-year assessment.
(2) Why is the TOWRE measure of word recognition at 7 years significantly more heritable than PIAT reading comprehension at 10 years? Heritability estimates are 70% for the TOWRE at 7 years and 39% for the PIAT at 10 years, a significant difference in heritability. We had expected the reverse pattern of results—that is, the TOWRE would be less heritable than the PIAT—based on our naïve assumption that early skill at reading words (TOWRE) is more a matter of exposure and training and that later reading comprehension (PIAT) involves “g” to a greater extent. The simplest explanation for the finding is that our assumptions were wrong and early word recognition is in fact much more heritable than later reading comprehension. At the level of genetic correlations as well, our assumptions about differences between the tests were wrong: the TOWRE and the PIAT are highly correlated genetically (.60), even though the tests assessed such apparently different cognitive processes and were administered 3 years apart.
The same two methodological issues discussed above could contribute to the difference in heritability between the TOWRE and PIAT. That is, differences in the testing situation in the home could contribute to the difference in heritability by increasing shared environmental variance for the web-based tests at the expense of heritability. Support for this hypothesis comes from the greater shared environment estimate for the web-based PIAT (.25) as compared with the telephone-administered TOWRE (.15). The other methodological hypothesis is that the web-based PIAT might be less reliable than the telephone-administered TOWRE. Support for this hypothesis comes from the greater nonshared environment estimate for the PIAT (.36) than for the TOWRE (.15).
A more interesting possibility is that, despite the genetic correlation of .60 between TOWRE at 7 years and PIAT at 10 years, different cognitive processes contribute to the two measures. The early stages of word recognition are known to be closely related to levels of phonological awareness, the knowledge that the sound of a word is made up of smaller pieces of sound (Snowling & Hayiou-Thomas, 2006). Learning to decode is therefore less related to “g” than later stages of reading that emphasize reading comprehension (Vanderwood, McGrew, Flanagan, & Keith, 2001). Several studies have shown that the heritability of phonological awareness at the beginning of school is approximately .6 (e.g., Hohnen & Stevenson, 1999; Petrill, Deater-Deckard, Thompson, DeThorne, & Schatschneider, 2006). This is much higher than the heritability of “g” (.36, .36, .41 at 7, 9, 10 years in our sample). And the genetic correlation between the TOWRE at 7 and “g” (.47) is lower than the genetic correlation between the PIAT at 10 and “g” (.63). In other words, word recognition might depend, more than comprehension, on a highly heritable specific foundational skill of phonological awareness.
In summary, we are again left with a puzzle. And, again, the 12-year TEDS assessment should help to solve the puzzle because the TOWRE and the PIAT, as well as other reading measures (although not phonological awareness), are included concurrently in the ongoing 12-year assessment.
(3) Why is science at 10 years less heritable and more influenced by shared environment than English or mathematics? In contrast to the heritabilities of 60% for NC English and 64% for NC mathematics at 10 years, the heritability of NC science is 48%. Shared environment estimates for English and mathematics are 20% and 12%, whereas for science the estimate is 27% (Table 11). These differences might not be reliable because they are not significant and they did not occur at 9 years. Nonetheless, throughout our longitudinal and multivariate genetic analyses (Chapters V and VI), science at 10 years often yields results that differ from results for English and mathematics. For example, our multivariate genetic analyses with “g” and NC ratings indicate that science performance at 10 years has more to do with “g” genetically than do English and mathematics.
The possibility that science performance is etiologically different from other academic subjects warrants further exploration because hardly any genetic research has addressed science performance as compared with the many studies of reading and an increasing number of studies of mathematics. Although science is a very broad and diffusely defined domain, we are especially interested in science as a way of understanding the world rather than as a body of facts. From little problems of daily life to problems in the field of cosmology, the scientific method—that is, posing logical, testable, and falsifiable hypotheses; testing the hypotheses as convincingly as possible; and interpreting the results reasonably—is the best way to solve problems, at least those problems that are amenable to empirical solutions. Getting children to see that the scientific method works in a practical way may help to counter the trend toward declining interest in science during the school years (Murphy & Beggs, 2003) and the general societal trend toward mysticism (Koch & Smith, 2006).
In a recent survey, more than 90% of parents and teachers perceive genetics to be at least as important as the environment for learning abilities and disabilities (Plomin & Walker, 2003). However, if there are any parents, teachers or policymakers who do not yet realize the important contribution that genetics makes to learning abilities and disabilities, these findings are important at this most rudimentary level of nature and nurture.
Even though teachers appear to recognize substantial genetic influence on academic achievement, there is a wide gap between education and genetics. The field of education scarcely acknowledges genetics despite the evidence for its importance, which is unfortunate because schools are the primary societal mechanism for fostering cognitive development (Rutter & Maughan, 2002). For example, a 2003 review of major educational psychology textbooks revealed that no text included more than three pages devoted to the topic of genetics (Plomin & Walker, 2003). Also, very few papers on genetics have been published in educational psychology journals.
Some of the reluctance to embrace genetics may be specific to the history and epistemology of education and educational psychology (Wooldridge, 1994). However, much of the reluctance is likely to involve general misconceptions about what it means to say that genetics is important (Rutter & Plomin, 1997). A key misconception is environmental nihilism, that is, if a disorder is heritable there is nothing that can be done about it environmentally (Sternberg & Grigorenko, 1999). The myth of environmental nihilism feeds into a related myth that finding genetic influence will serve to justify social inequality. We do not accept this view. Knowledge alone does not account for educational, societal, or political decisions—values are just as important in the decision-making process. We are aware that the relationship between knowledge and values is a complicated area of philosophy; here we are merely making the simple point that decisions, both good and bad, can be made with or without knowledge. For example, finding specific genes associated with reading disability obviously does not mean that we ought to put all available resources into educating those children with the most favorable genes and forget the rest of the children. Depending on our values, genetics could be used to argue for the opposite policy: We need to devote more resources to helping disadvantaged children. Whether or not our view of policymaking is naïve, surely it cannot be good for the science of education to pretend that genetic differences do not exist. And it may not be too Pollyannaish to hope that better policy decisions can be made with knowledge than without.
Even further back in the shadows is a general uneasiness about genetics in terms of our view of the essence of humanity. Are we not all created equal? The authors of the U.S. Declaration of Independence did not mean that we are all created identical—we obviously differ in height, for example. They clearly meant that in a democracy we should all be treated equally before the law and, more optimistically, that we should all have equal opportunities including educational opportunity (Husén, 1978). Indeed, if we were all identical there would be no need for a legal equality because its purpose is to ensure equality of treatment despite our differences (Pinker, 2002).
It would be a pity if the nature–nurture battles fought two decades ago in other areas of the behavioral sciences had to be refought in the field of education. By coming late to genetics, educational psychology can from the start embrace a more balanced position that acknowledges the importance of nature as well as nurture and uses genetic research to ask questions that go beyond heritability, such as the developmental and multivariate questions that are the focus of this monograph. Finding genetic influence will not denigrate the role of education; it will suggest new ways of thinking about effective education, such as recognizing that children create their own experience within the educational process in part on the basis of their genetic propensities.
Just as important as finding substantial genetic influence on learning abilities and disabilities is the finding that shared environment accounts for <20% of the variance. In contrast, research and discussion of the environmental origins of academic performance has focused almost entirely on family background and the school and classroom viewed as shared environmental effects. It is time to change that assumption. In addition, the finding that nonshared environment accounts for at least as much environmental variation as does shared environment opens up a new area of research that considers how children in the same school—even clones (MZ twins) in the same classroom—experience different environments.
The most important implications of these findings will come to the fore when specific genes are identified that contribute to the high heritability of learning abilities and disabilities. Although progress has been slow, recent developments in molecular genetics are promising (Plomin, 2005). For reading disability, for example, four candidate genes are currently the target of intense research (McGrath, Smith, & Pennington, 2006; Fisher & Francks, 2006). Reports are also beginning to appear of genes associated with normal variation in cognitive abilities (Plomin, Kennedy, & Craig, 2006). Although few educational and psychological researchers are likely to become involved in the quest to find genes associated with learning abilities and disabilities, when the genes are found they will be widely used in research as DNA risk indicators in much the same way that demographic risk indicators are currently used (Plomin & Walker, 2003). It should be emphasized that, like demographic risk indicators, genetic prediction will be probabilistic because there will be many genes of small effect size, as suggested by the QTL model described earlier.
Acceptance of genetic influence will come more readily because identifying specific genes provides evidence for genetic influence that is much more direct than the evidence provided by quantitative genetic research such as twin studies. Moreover, DNA has a unique causal status in that correlations between DNA differences and behavioral differences can only be explained causally in one direction: DNA differences cause behavioral differences. This causal status of DNA is unique in the sense that correlations involving other biological variables such as brain variables are just correlations that can be explained in either causal direction—behavioral differences can cause brain differences. However, variation in DNA sequence, which is the basis of heredity, is not changed by behavior, biology, or the environment.
Although finding specific genes associated with learning disabilities is unlikely to have much direct effect on teachers in the classroom confronted with a particular child with a learning disability, such findings will have far-reaching ramifications in terms of diagnosis, treatment, and prevention. Finding genes responsible for the high heritability of learning disabilities will lead to new diagnostic classifications that are based on etiology rather than symptomatology. As discussed earlier, two crucial examples emerged from the research described in this monograph: Learning disabilities are the quantitative extreme of the same genetic and environmental factors responsible for normal variation in learning abilities and the same set of genes influence most learning disabilities and abilities.
In terms of treatment, genes will be used clinically or educationally to the extent that response to treatment depends on genetic risk. This goal is part of a “personalized medicine” movement toward individually tailored treatments rather than treatments that are “one size fits all” (Abrahams, Ginsburg, & Silver, 2005). As noted earlier, the most important benefit of identifying genes that put children at risk for developing learning disabilities is that the causal nature of genes means that they can serve as an early-warning system. This should facilitate research on interventions that prevent problems, rather than waiting until problems are so severe that they can no longer be ignored. The goal of early intervention fits with a general trend toward preventative medicine. Because vulnerability to learning disabilities involves many genes of small effect, genetic engineering is unimaginable for learning disabilities; interventions will rely on environmental engineering, primarily educational interventions.
When genes associated with learning abilities and disabilities are found, the next step in research is to understand how these genes have their effect, called functional genomics. Functional genomics is usually considered in terms of the bottom-up agenda of molecular biology, which begins with the analysis of molecules in cells. However, the behavioral level of analysis is also useful for understanding how genes have their effect in relation to the development of the whole child, for example, in understanding interactions and correlations between genes and environment as they affect development and in leading to new diagnoses, treatments and interventions. The phrase behavioral genomics has been proposed to emphasize the importance of such top-down levels of analysis for understanding how genes have their effect on behavior (Plomin & Crabbe, 2000). Bottom-up and top-down levels of analysis of developmental pathways between genes and behavior will eventually meet in the brain. The grandest implication for science is that DNA will serve as an integrating force across diverse life sciences relevant to understanding learning abilities and disabilities.