The genetics of specific cognitive abilities. Francesca Procopioa et al. Intelligence, Volume 95, November–December 2022, 101689. https://doi.org/10.1016/j.intell.2022.101689
Highlights
• Specific cognitive abilities (SCA) are 56% heritable, similar to g.
• Some SCA are significantly more heritable than others, 39% to 64%.
• Independent of g (‘SCA.g’), SCA remain substantially heritable (∼50%).
• SCA do not show the dramatic developmental increase in heritability seen for g.
• Genomic research on SCA.g is needed to create profiles of strengths and weaknesses.
Abstract: Most research on individual differences in performance on tests of cognitive ability focuses on general cognitive ability (g), the highest level in the three-level Cattell-Horn-Carroll (CHC) hierarchical model of intelligence. About 50% of the variance of g is due to inherited DNA differences (heritability) which increases across development. Much less is known about the genetics of the middle level of the CHC model, which includes 16 broad factors such as fluid reasoning, processing speed, and quantitative knowledge. We provide a meta-analytic review of 747,567 monozygotic-dizygotic twin comparisons from 77 publications for these middle-level factors, which we refer to as specific cognitive abilities (SCA), even though these factors are not independent of g. Twin comparisons were available for 11 of the 16 CHC domains. The average heritability across all SCA is 56%, similar to that of g. However, there is substantial differential heritability across SCA and SCA do not show the developmental increase in heritability seen for g. We also investigated SCA independent of g (SCA.g). A surprising finding is that SCA.g remain substantially heritable (53% on average), even though 25% of the variance of SCA that covaries with g has been removed. Our review highlights the need for more research on SCA and especially on SCA.g. Despite limitations of SCA research, our review frames expectations for genomic research that will use polygenic scores to predict SCA and SCA.g. Genome-wide association studies of SCA.g are needed to create polygenic scores that can predict SCA profiles of cognitive abilities and disabilities independent of g.
Keywords: Specific cognitive abilityIntelligencemeta-analysisTwin studyHeritability
4. Discussion
Although g is one of the most powerful constructs in the behavioural sciences (Jensen, 1998), there is much to learn about the genetics of cognitive abilities beyond g. Our meta-analysis of 747,567 twin comparisons yielded four surprising findings. One of the most interesting findings about g is that its heritability is similar to that of SCA. The heritability of g is about 50% (Knopik et al., 2017) and the average heritability of SCA from our meta-analysis is 56%.
We focused on three additional questions: Are some SCA more heritable than others (differential heritability)? Does the heritability of SCA increase during development as it does for g? What is the heritability of SCA independent of g?
4.1. Differential heritability
We conclude that some SCA are more heritable than others. The estimates ranged from 39% for auditory processing (Gt) to 64% for quantitative knowledge and processing speed (Gs). Our expectation that domains conceptually closer to g would have higher heritability than ones more conceptually distinct from g led us to be surprised which SCA were most heritable.
For example, we hypothesised that acquired knowledge would be less heritable than fluid reasoning. This is because acquired knowledge is a function of experience, while fluid reasoning involves the ability to solve novel problems. To the contrary, our results indicate that acquired knowledge is the most heritable grouping of CHC domains, with an average heritability of 58%. In contrast, fluid reasoning has a comparatively low heritability estimate of 40%.
We were also surprised to find significantly large differences in heritability between SCA within the same functional grouping. For example, processing speed (Gs), one of the most heritable CHC domains, is within the functional grouping of general speed. Processing speed is defined as ‘the ability to automatically and fluently perform relatively easy or over-learned elementary cognitive tasks, especially when high mental efficiency (i.e., attention and focused concentration) is required’ (McGrew, 2009, p. 6). In contrast, reaction and decision speed (Gt), another CHC domain within the functional grouping of general speed for which twin comparisons were available, yielded one of the lowest heritabilities of 42%. It is defined as ‘the ability to make elementary decisions and/or responses (simple reaction time) or one of several elementary decisions and/or responses (complex reaction time) at the onset of simple stimuli’ (McGrew, 2009, p. 6). Why is reaction and decision speed (Gt) so much less heritable than processing speed (Gs) (42% vs 64%)? One possibility is that processing speed picks up ‘extra’ genetic influence because it involves more cognitive processing than reaction time, as suggested by their definitions. Moreover, Schneider and McGrew (2018) propose a hierarchy of speeded abilities (Kaufman, 2018, p. 108) in which Gs (which they call broad cognitive speed) has a higher degree of cognitive complexity than Gt (broad decision speed). But this would not explain why processing speed is more heritable than fluid reasoning (40%), which seems to involve the highest level of cognitive processing such as problem solving and inductive and deductive reasoning.
One direction for future research is to understand why some SCA are more heritable than others. A first step in this direction is to assess the extent to which differential reliability underlies differential heritability because reliability, especially test-retest reliability rather than internal consistency, creates a ceiling for heritability. For example, the least heritable SCA is short-term memory (Gsm), for which concerns about test-retest reliability have been raised (Waters & Caplan, 2003).
If differential reliability is not a major factor in accounting for differential heritability, a substantive direction for research on SCA is to conduct multivariate genetic analyses investigating the covariance among SCA to explore the genetic architecture of SCA. This would be most profitable if these analyses also included g, as discussed below (SCA.g).
4.2. Developmental changes in SCA heritability
One of the most interesting findings about g is that its heritability increases linearly from 20% in infancy to 40% in childhood to 60% in adulthood. SCA show average decreases in heritability from childhood to later life (column 1 in Fig. 4). Although several CHC domains show increases from early childhood (0–6 years) to middle childhood (7–11 years), this seems likely to be due at least in part to difficulties in reliably assessing cognitive abilities in the first few years of life.
It is puzzling that heritability increases developmentally for g but not for SCA because g represents what is in common among SCA. A previous meta-analysis that investigated cognitive aging found that the heritability of verbal ability, spatial ability and perceptual speed decreased after the age of around 60 (Reynolds & Finkel, 2015). While we did not find evidence for this for any of the SCA domains, we did observe the general trend of decreasing heritability for reading and writing (Grw) and visual processing (Gv) from middle childhood onwards.
We hoped to investigate the environmental hypothesis proposed by Kovas et al. (2013) to account for their finding that the heritability of literacy and numeracy SCA were consistent throughout the school years (∼65%), whereas the heritability of g increased from 38% age 7 to 49% at age 12 (Kovas et al., 2013). They hypothesised that universal education for basic literacy and numeracy skills in the early school years reduces environmental disparities, which leads to higher heritability as compared to g, which is not a skill taught in schools.
We hoped to test this hypothesis by comparing SCA that are central to educational curricula and those that are not. For example, reading and writing (Grw), quantitative knowledge (Gq) and comprehension-knowledge (Gc) are central to all curricula, whereas other SCA are not explicitly taught in schools, such as auditory processing (Ga), fluid reasoning (Gf), processing speed (Gs), short-term memory (Gsm) and reaction and decision speed (Gt). Congruent with the Kovas et al. hypothesis, Grw, Gq and Gc yield high and stable heritabilities of about 60% during the school years. However, too few twin comparisons are available to test whether Ga, Gf, Gs, Gsm and Gt show increasing heritability during the school years.
4.3. SCA independent of g (SCA.g)
Although few SCA.g data are available, they suggest another surprising finding. In these studies, SCA independent of g are substantially heritable, 53%, very similar to the heritability estimate of about 50% for SCA uncorrected for g. This finding is surprising because a quarter of the variance of SCA is lost when SCA are corrected for g. More SCA.g data are needed to assess SCA issues raised in our review about the influence of g in differential heritability and developmental changes in heritability.
Although more data on SCA.g are needed, our preliminary results are encouraging in suggesting that genetic influence on SCA does not merely reflect genetic influence on g. Although g drives much of the predictive power of cognitive abilities, it should not overshadow the potential for SCA to predict profiles of cognitive strengths and weaknesses independent of g.
An exciting aspect of these findings is their implication for research that aims to identify specific inherited DNA differences responsible for the heritability of SCA and especially SCA.g. Genome-wide association (GWA) methods can be used to assess correlations across millions of DNA variants in the genome with any trait and these data can be used to create a polygenic score for the trait that aggregates these weighted associations into a single score for each individual (Plomin, 2018). The most powerful polygenic scores in the behavioural sciences are derived from GWA analyses for the general cognitive traits of g (Savage et al., 2018) and educational attainment (Lee et al., 2018; Okbay et al., 2022). It is possible to use these genomic data for g and educational attainment to explore the extent to which they can predict SCA independent of g and educational attainment even when SCA were not directly measured in GWA analyses, an approach called GWAS-by-subtraction (Demange et al., 2021), which uses genomic structural equation modeling (Grotzinger et al., 2019). We are also employing a simpler approach using polygenic scores for g and educational attainment corrected for g, which we call GPS-by-subtraction (Procopio et al., 2021).
Ultimately, we need GWA studies that directly assess SCA and especially SCA.g. Ideally, multiple measures of each SCA domain would be used and a general factor extracted rather than relying on a single test of the domain. The problem is that GWA requires huge samples to detect the miniscule associations between thousands of DNA variants and complex traits known to contribute to their heritabilities. The power of the polygenic scores for g and educational attainment comes from their GWA sample sizes of >250,000 for g and more than three million for educational attainment. It is daunting to think about creating GWA samples of this size for tested SCA as well as g in order to investigate SCA.g. However, a cost-effective solution is to create brief but psychometrically valid measures of SCA that can be administered to the millions of people participating in ongoing biobanks for whom genomic data are available. For example, a gamified 15-min test has been created for this purpose to assess verbal ability, nonverbal ability and g (Malanchini et al., 2021). This approach could be extended to assess other SCA and SCA.g.
We conclude that SCA.g are reasonable targets for genome-wide association studies, which could enable polygenic score predictions of profiles of specific cognitive strengths and weaknesses independent of g (Plomin, 2018). For example, SCA.g polygenic scores could predict, from birth, aptitude for STEM subjects independent of g. Polygenic score profiles for SCA.g could be used to maximise children's cognitive strengths and minimise their weaknesses. Rather than waiting for problems to develop, SCA.g polygenic scores could be used to intervene to attenuate problems before they occur and help children reach their full potential.
4.4. Other issues
An interesting finding from our review is that SCA.g scores in which SCA are corrected phenotypically for g by creating residualised scores from the regression of g on SCA yield substantially higher estimates of heritability than SCA.g derived from Cholesky analyses.
We suspect that the difference is that regression-derived SCA.g scores remove phenotypic covariance with g, thus removing environmental as well as genetic variance associated with g. In contrast, Cholesky-derived estimates of the heritability of SCA independent of g are calibrated to the total variance of SCA, not to the phenotypic variance of SCA after g is controlled. Regardless of the reason for the lower Cholesky-derived estimates of the heritability of g as compared to regression-derived SCA.g scores, regression-derived phenotypic scores of SCA.g are likely the way that phenotypic measures of SCA will be used in phenotypic and genomic analyses. Instead, the Cholesky models involve latent variables that cannot be converted to phenotypic scores for SCA.g.
Another finding from our review is that heritability appears to be due to additive genetic factors. The average weighted MZ and DZ correlations across the 11 CHC domains for which twin comparisons were available were 0.72 and 0.44, respectively. This pattern of twin correlations, which is similar to that seen across all SCA as well as g, is consistent with the hypothesis that genetic influence on cognitive abilities is additive (Knopik et al., 2017). Additive genetic variance involves genetic effects that add up according to genetic relationships so that if heritability were 100%, MZ twins would correlate 1.0 and DZ twins would correlate 0.5 as dictated by their genetic relatedness. In contrast, if genetic effects operated in a non-additive way, the correlation between DZ twins would be less than half the correlation between MZ twins. Because MZ twins are identical in their inherited DNA sequence, only MZ twins capture the entirety of non-additive interactions among DNA variants. In other words, the hallmark of non-additive genetic variance for a trait is that the DZ correlation is less than half the MZ correlation. None of the SCA show this pattern of results (Fig. 3), suggesting that genetic effects on SCA are additive.
Finding that genetic effects on SCA are additive is important for genomic research because GWA models identify the additive effects of each DNA variant and polygenic scores sum these additive effects (Plomin, 2018). If genetic effects were non-additive, it would be much more difficult to detect associations between DNA variants and SCA. The additivity of genetic effects on cognitive abilities is in part responsible for the finding that the strongest polygenic scores in the behavioural sciences are for cognitive abilities (Allegrini et al., 2019) (Cheesman et al., 2017) (Plomin et al., 2013).
4.5. Limitations
The usual limitations of the twin method apply, although it should be noted that twin results in the cognitive domain are supported by adoption studies (Knopik et al., 2017) and by genomic analyses (Plomin & von Stumm, 2018).
As noted earlier, a general limitation is that some CHC categories have too few studies to include in meta-analyses. This is especially the case in the developmental analyses. Power is diminished by dividing the twin comparisons into five age categories. In addition, different measures are used at different ages; even when measures with the same label are used across ages, they might measure different things. Finally, the developmental results are primarily driven by cross-sectional results from different studies. Nonetheless, longitudinal comparisons within the same study have also found no developmental change in heritability estimates for some SCA (Kovas et al., 2013).
Another limitation of this study is that there might be disagreement concerning the CHC categories to which we assigned tests. We reiterate that we used the CHC model merely as a heuristic to make some sense of the welter of tests that have been used in twin studies, not as a definitive assignment of cognitive tests to CHC categories. We hope that Supplementary Table S-3 with details about the studies and measures will allow researchers to categorise the tests differently or to focus on particular tests. This limitation is also a strength of our review in that it points to SCA for which more twin research is needed. The same could be said for other limitations of SCA twin research such as the use of different measures across studies and the absence of any twin research at some ages.
A specific limitation of SCA.g is that removing all phenotypic covariance with g might remove too much variance of SCA, as mentioned in the Introduction. A case could be made that bi-factor models (Murray & Johnson, 2013) or other multivariate genetic models (Rijsdijk, Vernon, & Boomsma, 2002) would provide a more equitable distribution of variance between SCA and g indexed as a latent variable representing what is in common among SCA. However, the use of bifactor models is not straightforward (Decker, 2021). Moreover, phenotypic and genomic analyses of SCA.g are likely to use regression-derived SCA.g scores because bifactor models, like Cholesky models, involve latent variables that cannot be converted to phenotypic scores for SCA.g.
Finally, in this paper we did not investigate covariates such as average birth year of the cohort, or country of origin, nor did we examine sex differences in differential heritability or in developmental changes in heritability or SCA.g. Opposite-sex DZ twins provide a special opportunity to investigate sex differences. We have investigated these group differences in follow-up analyses (Zhou, Procopio, Rimfeld, Malanchini, & Plomin, 2022).
4.6. Directions for future research
SCA is a rich territory to be explored in future research. At the most general level, no data at all are available for five of the 16 CHC broad categories. Only two of the 16 CHC categories have data across the lifespan.
More specifically, the findings from our review pose key questions for future research. Why are some SCA significantly and substantially more heritable than others? How is it possible that SCA.g are as heritable as SCA? How is it possible that the heritability of g increases linearly across the lifespan, but SCA show no clear developmental trends?
Stepping back from these specific findings, for us the most far-reaching issue is how we can foster GWA studies of SCA.g so that we can eventually have polygenic scores that predict genetic profiles of cognitive abilities and disabilities that can help to foster children's strengths and minimise their weaknesses.
Posts
No comments:
Post a Comment