Generalist genes and the Internet generation: etiology of learning abilities by web testing at age 10

Abstract

A key translational issue for neuroscience is to understand how genes affect individual differences in brain function. Although it is reasonable to suppose that genetic effects on specific learning abilities, such as reading and mathematics, as well as general cognitive ability (g), will overlap very little, the counterintuitive finding emerging from multivariate genetic studies is that the same genes affect these diverse learning abilities: a Generalist Genes hypothesis. To conclusively test this hypothesis, we exploited the widespread access to inexpensive and fast Internet connections in the UK to assess 2541 pairs of 10-year-old twins for reading, mathematics and g, using a web-based test battery. Heritabilities were 0.38 for reading, 0.49 for mathematics and 0.44 for g. Multivariate genetic analysis showed substantial genetic correlations between learning abilities: 0.57 between reading and mathematics, 0.61 between reading and g, and 0.75 between mathematics and g, providing strong support for the Generalist Genes hypothesis. If genetic effects on cognition are so general, the effects of these genes on the brain are also likely to be general. In this way, generalist genes may prove invaluable in integrating top-down and bottom-up approaches to the systems biology of the brain.

Full text

Page 1

Genes, Brain and Behavior (2008) 7: 455–462
# 2007 The Authors
Journal compilation # 2007 Blackwell Publishing Ltd
Generalist genes and the Internet generation: etiology
of learning abilities by web testing at age 10
O. S. P. Davis*,†, Y. Kovas†, N. Harlaar†,
P. Busfield†, A. McMillan†, J. Frances†,
S. A. Petrill‡, P. S. Dale§and R. Plomin†
†Social, Genetic and Developmental Psychiatry Centre, Institute
of Psychiatry, King’s College London, London, United Kingdom,
‡Department of Human Development and Family Science, Ohio
State University, Columbus, OH, and§Department of Speech and
Hearing Sciences, University of New Mexico, Albuquerque, NM,
USA
*Corresponding author: O. S. P. Davis, Institute of Psychiatry
P080, De Crespigny Park, London, SE5 8AF, UK. E-mail: oliver.
davis@iop.kcl.ac.uk
A key translational issue for neuroscience is to under-
stand how genes affect individual differences in brain
function. Although it is reasonable to suppose that
genetic effects on specific learning abilities, such as
reading and mathematics, as well as general cognitive
ability (g), will overlap very little, the counterintuitive
finding emerging from multivariate genetic studies is
that the same genes affect these diverse learning abili-
ties: a Generalist Genes hypothesis. To conclusively test
this hypothesis, we exploited the widespread access to
inexpensive and fast Internet connections in the UK to
assess 2541 pairs of 10-year-old twins for reading, math-
ematics and g, using a web-based test battery. Heritabil-
ities were 0.38 for reading, 0.49 for mathematics and 0.44
for g. Multivariate genetic analysis showed substantial
genetic correlations between learning abilities: 0.57
between reading and mathematics, 0.61 between read-
ing and g, and 0.75 between mathematics and g, pro-
viding strong support for the Generalist Genes
hypothesis. If genetic effects on cognition are so general,
the effects of these genes on the brain are also likely to
be general. In this way, generalist genes may prove
invaluable in integrating top-down and bottom-up ap-
proaches to the systems biology of the brain.
Keywords: General cognitive ability, generalist genes, genetic
correlation, Internet, mathematics, reading, twins
Received 8 August 2007, revised 25 October 2007, accepted
for publication 31 October 2007
Naturally occurring genetic variation is a Rosetta Stone for
translating causal effects from genes to brain to cognition,
especially once specific genes are identified (Plomin et al. in
press). Although learning abilities and disabilities are highly
heritable, specific genes responsible for their heritability
have not yet been identified despite several promising
candidate genes for reading disability (Fisher & Francks
2006). Nonetheless, more can be gleaned from quantitative
genetic research, such as twin studies that compare identical
and fraternal twins, than the mere fact that they are heritable.
A major advance in quantitative genetics is multivariate
genetic analysis, which investigates not only the variance
of traits considered one at a time but also the covariance
among traits. In this way, it indicates the extent to which the
same or different genes affect several traits (Neale et al.
2006), using a statistic known as a genetic correlation
(Plomin et al. in press). These findings can constrain explan-
ations of brain processes that underlie the traits. For exam-
ple, it is reasonable to suppose that genetic effects will be
specific to the substantially different cognitive processes
involved in reading and mathematics. Such genetic specific-
ity would indicate the need to identify the genetically driven
differences in brain processes that underlie these cognitive
differences.
However, a very different result is emerging from multi-
variate genetic research on learning abilities and disabilities.
Most genetic effects appear to be general in that the same
genes affect different learning abilities and disabilities. A
review of multivariate genetic research on learning abilities
found that genetic correlations varied from 0.67 to 1.0
between reading and language (five studies), 0.47 to 0.98
between reading and mathematics (three studies) and 0.59 to
0.98 between language and mathematics (two studies)
(Plomin & Kovas 2005). The average genetic correlation was
about 0.70. Moreover, the general effects of genes appear to
extend beyond specific learning abilities such as reading and
mathematics to other cognitive abilities such as verbal
abilities (e.g. vocabulary and word fluency) and non-verbal
abilities (e.g. spatial and memory). The average genetic
correlation between specific learning abilities and general
cognitive ability (g), which encompasses these verbal and
non-verbal cognitive abilities, is about 0.60 (Plomin & Kovas
2005). These findings have led to a Generalist Genes hypoth-
esis (Plomin & Kovas 2005), which has far-reaching implica-
tions for cognitive neuroscience (Kovas & Plomin 2006).
Re-use of this article is permitted in accordance with the
Creative Commons Deed, Attribution 2.5, which does not
permit commercial exploitation.
doi: 10.1111/j.1601-183X.2007.00370.x
455

Page 2

However, the Generalist Genes hypothesis has yet to be
tested by direct cognitive test measures in a sample large
enough toconclusively establishthe magnitude of thegenetic
correlations between learning abilities. To address this prob-
lem, we developed an online test battery that includes
measures of reading, mathematics and g and used it to
assess a UK-representative population sample of 2541 pairs
of 10-year-old twins: by far the largest twin sample with
cognitive test data. The purpose of the present study was to
exploit the potential of web-based administration to provide
a powerful test of the Generalist Genes hypothesis.
Methods
Participants
The Twins Early Development Study (TEDS) recruited families of
twins born in England and Wales in 1994, 1995 and 1996: three annual
cohorts (Oliver & Plomin 2007; Trouton et al. 2002). The present paper
describes results at 10 years, where a two-cohort subsample of TEDS
(twins born in 1994 and 1995) was tested. Despite inevitable attrition
(Oliver & Plomin 2007), the sample remains representative of the UK
population (ascertained by comparison with census data from the
Office of National Statistics). Notably, because of the widespread
availability of fast Internet access, we found no evidence of sampling
bias in favor of higher socioeconomic status families. Informed
consent was obtained by post or online consent forms, and a test
administrator was then assigned who telephoned the family, sorted
out any problems with the Internet or testing and generally assisted
and encouraged the participating family. Ethical approval for TEDS has
been provided by the Institute of Psychiatry Ethics Committee,
reference number 05/Q0706/228.
We excluded from the analyses children with severe current
medical problems and children who had suffered severe problems
at birth or whose mothers had suffered severe problems during
pregnancy from the analyses. We also excluded twins whose
zygosity was unknown or uncertain or whose first language was
other than English. Finally, we included only twins whose parents
reported their ethnicity as ‘white’, which was 93% of this UK sample.
The present analyses are based on 2541 twin pairs [919 monozygotic
(MZ) pairs, 817 same-sex dizygotic (DZ) and 805 opposite-sex DZ].
Internet testing
Widespread access to inexpensive and fast Internet connections in
the UK has made online testing an attractive possibility for collecting
data on the substantial samples necessary for genetic research,
especially for multivariate genetic research. The advantages and
potential pitfalls of data collection over the Internet have been
reviewed in detail elsewhere (Birnbaum 2004). For older children,
most of whom are competent computer users, it is an interactive and
enjoyable medium. Through adaptive branching, it allows the use of
hundreds of items to test the full range of ability, while requiring
individual children to complete only a relatively small number of items
to ascertain their level of performance. In tests where it is appropriate,
streaming voiceovers can minimize the necessary reading. In addi-
tion, the tests can be completed over a period of several weeks,
allowing children to pace the activities themselves, although they are
not allowed to return to items previously administered. Finally, it is
possible to intersperse the activities with games. All of these factors
help maintain children’s engagement with the tests.
Measures
Reading was assessed by an adaptation of the Peabody Individual
Achievement Test (PIAT-Revised; Markwardt 1997) Reading Com-
prehension Scale, and mathematics by three subtests based on the
nferNelson Math 5-14 Series (nferNelson 2001): Understanding
Number, Non-Numerical Processes and Computation and Knowl-
edge. g was indexed by two verbal tests – WISC-III-PI multiple-choice
Vocabulary and Information (Wechsler 1992) – and two non-verbal
tests – WISC-III-UK Picture Completion (Wechsler 1992) and Raven’s
Standard Progressive Matrices (Raven et al. 1996). We created a g
score with equal weights for the four tests by summing their
standardized scores. This unit-weighted score correlated 0.99 with
the corresponding factor score. Further information about the meas-
ures is available elsewhere (Oliver & Plomin 2007). We have shown
that the web-based tests are reliable, stable and valid (Haworth et al.
2007). As an index of reliability, Cronbach’s alpha was 0.95 for PIAT
subtests, 0.78–0.93 for the math subtests and 0.74–0.91 for the g
subtests (n ¼ 2569–2924). In terms of stability and validity, scores on
the online version of our reading and mathematics tests correlate
highly with traditional in-person versions administered in person 1–
3 months later: r ¼ 0.80 for PIAT and 0.92 for mathematics (n ¼ 30).
Download time, a proxy for computer performance, accounted for
less than 2% of the variance in the PIAT and less than 0.5% of the
variance in the other tests.
Statistical analyses
According to the quantitative genetic model (Plomin et al. in press;
Rijsdijk & Sham 2002), same-sex twins reared together resemble
each other because of the additive effects of shared genes (A) or
shared (common) environmental factors (C). For identical or MZ twins,
the correlation between their genes is 1.00, whereas for non-identical
or DZ twins, the correlation is 0.50 because DZ twins on average
share half of their segregating alleles. The correlation between twins
for shared environment is, by definition, 1.00 for both MZ and DZ
twins growing up in the same family, while non-shared environmental
influences (E) are uncorrelated and contribute to differences between
twins. For the twin analyses, standardized residuals correcting for age
and sex were used because the age of twins is perfectly correlated
across pairs, which means that, unless corrected, variation within
each age group at the time of testing would contribute to the
correlation between twins and be misrepresented as shared environ-
mental influence (Eaves et al. 1989). The same applies to the sex of
the twins because MZ twins are always of the same sex. Likewise,
download time, a proxy for computer performance, was also re-
gressed out of the twins’ test scores. The assumptions of the
classical twin model, and their validity, have been discussed in detail
elsewhere (Boomsma et al. 2002; Martin et al. 1997; Rijsdijk & Sham
2002).
As well as examining twin correlations, we used standard ACE
model-fitting analysis in MX1.7.01 (Neale et al. 2006) where ACE
stands for additive genetic influences (A), shared or common envi-
ronmental influences (C) and non-shared environmental influences
(E), as above. Model-fitting analysis specifies a correlational structure
(a model) using matrix algebra. This model is a hypothesis about the
structure of the dataset and is derived from what we know about how
MZ and DZ twins are related to each other (see above). By fitting the
model to the data using an iteration process, we can derive its
‘goodness of fit’ and parameter estimates for the contributions of A, C
and E. Before embarking on our multivariate analysis, we initially
examined sex differences in the genetic and environmental parameter
estimates by comparing the fit of three full univariate ACE models
(one each for reading, mathematics and g) with that of various nested
models, dividing the twin pairs into five groups: MZ male, MZ female,
DZ male, DZ female and DZ opposite-sex pairs. These sex-limitation
models (Eley 2005) allowed us to estimate qualitative and quantitative
etiological differences between the sexes (Galsworthy et al. 2000).
Finally, we used multivariate genetic model fitting to investigate
the genetic and environmental etiology of the covariation between
learning abilities. Figure 1 shows the phenotypic Cholesky decom-
position, which partitions variance into a universal factor influencing
all three traits, a factor influencing reading and mathematics
independent of g and a factor influencing mathematics independent
of reading and g. As shown on the left of Fig. 2, the genetic
Cholesky decomposition partitions this variance further into genetic,
shared environmental and non-shared environmental components.
456
Genes, Brain and Behavior (2008) 7: 455–462
Davis et al.

Page 3

As in the phenotypic Cholesky, variance attributable to genetic
influences is divided into a universal genetic factor influencing g,
reading and mathematics (A1); a genetic factor specific to reading
and mathematics (A2) and a genetic factor unique to mathematics
(A3). The shared environmental influences (C1, C2, C3) and non-
shared environmental influences (E1, E2, E3) are partitioned in the
same way.
The correlated factors model (on the right) can be derived from the
Cholesky model. It specifies three latent genetic factors, one for each
trait, and calculates the correlation between them. Once again, the
same is true for the shared and non-shared environments. These
statistics, the genetic, shared environmental and non-shared environ-
mental correlations, are unique to multivariate analysis. Along with
bivariate heritability and environmentality (the proportion of the
phenotypic correlation accounted for by genes or the environment),
they give us an essential insight into how genetic factors and
environments are shared in the etiology of learning abilities. It is
important to note that the genetic and environmental correlations are
independent of the heritability or environmentality of the traits. For
example, two traits with very little heritability can nevertheless be
highly correlated genetically (i.e. share the same genetic influence),
and two highly heritable traits can be genetically uncorrelated
(independent genetic influences).
In Fig. 2, the squared path coefficients influencing each measured
variable in the correlated factors model can be derived from the
corresponding squared paths in the Cholesky model. Finally, individual
Cholesky pathways were dropped one at a time, and the fit was
compared with the full model. This tests the statistical significance of
the influence of each latent factor on g, reading and mathematics and
indicates the most parsimonious model.
Results
Phenotypic analyses
As shown in Table 1, the measures exhibited small mean
differences for sex (boys higher for mathematics and g) and
zygosity (DZs higher for reading and g), although the differ-
ences are statistically significant, given the large sample
sizes. Altogether, sex and zygosity accounted for less than
1% of the variance in all measures.
Phenotypic correlations for the age-, sex- and download-
time-regressed indicators were as follows (using one member
of each twin pair): mathematics–reading, r ¼ 0.51, n ¼ 2457,
P ¼ 0.000; mathematics–g, r ¼ 0.63, n ¼ 2342, P ¼ 0.000;
reading–g, r ¼ 0.54, n ¼ 2333, P ¼ 0.000.
Univariate genetic analyses
Intraclass correlations (Shrout & Fleiss 1979; twin similarity
coefficients) are shown in Table 2 for the total group of MZ,
DZ same-sex and DZ opposite-sex twins, as well as for the
male and female subgroups among the same-sex twin pairs.
Correlations between MZ twins were consistently higher
than those between DZ twins, suggesting a genetic contri-
bution to reading, mathematics and g. As a first estimate,
doubling the difference between the MZ and same-sex DZ
correlations yields moderate heritability estimates of 40% for
mathematics, 40% for reading and 36% for g. Shared
environmental influences are also moderate, estimated as
the extent to which MZ resemblance exceeds heritability:
31% for mathematics, 24% for reading and 35% for g. The
remainder of the variance is attributed to non-shared environ-
mental influences (plus error of measurement): 29% for
mathematics, 36% for reading and 29% for g.
As shown in Table 3, ACE model-fitting results are consis-
tent with estimates based on the twin correlations in Table 2.
For mathematics, reading and g, genetic influence is moder-
ate (49%, 38% and 44%, respectively, for the best-fitting
model). Shared and non-shared environmental influences are
more modest.
Moreover, across zygosity, correlations within male and
female pairs (Table 2) were similar, suggesting similar ACE
estimates for boys and girls. In addition, correlations between
same-sex DZ twins were similar to those between opposite-
sex DZ twins, suggesting no qualitative sex differences.
Sex-limitation model fitting (Eley 2005) confirmed these
expectations, yielding no significant sex differences in ACE
parameter estimates or in comparisons between same-sex
and opposite-sex DZ twins. The best-fitting model from the
sex-limitation analyses was the null model, which includes no
quantitative or qualitative differences in etiology between
males and females: likelihood ratio w2test with Ddf compared
with full sex-limitation model for mathematics, reading and
g, respectively: 4.17, 3, P ¼ 0.24; 0.18, 3, P ¼ 0.98; 3.79, 3,
P ¼ 0.29. For this reason, and to maximize power, our multi-
variate genetic analyses combined sexes.
Multivariate genetic analysis
Cross-trait twin correlations (e.g. twin 1 reading versus twin 2
mathematics) are the essence of multivariate genetic analy-
sis. Table 4 shows that cross-trait twin correlations are
consistently greater for MZ than for DZ twins; the MZ
cross-trait twin correlations are nearly as great as the pheno-
typic correlations for the same individual, shown in the third
column of Table 4. Doubling the difference between the MZ
and DZ cross-trait correlations estimates the genetic contri-
bution to the phenotypic correlations (0.18, 0.32 and 0.32,
respectively, for the three rows of Table 4), and dividing these
Figure 1: Path diagram for the full phenotypic Cholesky
decomposition model. The Cholesky partitions variance into a
universal factor influencing all three traits (latent factor 1), a factor
influencing reading and mathematics (latent factor 2) and a factor
influencing mathematics alone (latent factor 3).
Genes, Brain and Behavior (2008) 7: 455–462
457
Generalist genes and the Internet generation

Page 4

Figure 2: Multivariate analysis. The top panel gives the estimates for the additive genetic (A) component of the variance, with the
Cholesky solution on the left and the correlated factors solution on the right. The middle panel does the same for the shared environment
(C) and the bottom panel gives the estimates for non-shared environmental effects (E). In the Cholesky diagrams, line weights and
intensities represent strength of association. Dotted paths can be dropped individually without a significant (P > 0.05) decrement
in model fit. In the correlated factors diagrams, the curved arrows represent correlations between the latent factors. R, reading;
M, mathematics.
458
Genes, Brain and Behavior (2008) 7: 455–462
Davis et al.

Page 5

estimates by the phenotypic correlations for the same
individual (third column of Table 4) indicates the proportional
contribution of genetic influences to the phenotypic correla-
tion: the bivariate heritability (35%, 51% and 59%). As
indicated in the Methods, the genetic correlation, unlike
bivariate heritability, is independent of heritability. The genetic
correlation can be estimated by dividing the genetic contribu-
tion to the phenotypic correlation by the product of the square
roots of the heritabilities of the two traits (Plomin & DeFries
1979). These rough estimates of genetic correlations are
substantial: 0.40 for reading and mathematics, 0.73 for
reading and g, and 0.68 for mathematics and g.
The results of multivariate model-fitting analyses echo this
simple analysis based on the cross-trait twin correlations.
The Cholesky decomposition model is shown in Fig. 1.
Tables 5 and 6 show parameter estimates and confidence
intervals, with the results summarized visually in Fig. 2. As
shown in Table 5 and Fig. 2 (curved arrows, top right),
genetic correlations are substantial (0.57 for reading and
mathematics, 0.61 for reading and g, and 0.75 for mathe-
matics and g), providing evidence in support of the Gener-
alist Genes hypothesis. Bivariate heritabilities (Table 5) are
about 50%, indicating that about half of the phenotypic
correlations across g, reading and mathematics are medi-
ated genetically. The Cholesky analysis (Table 6 and left side
of Fig. 2) also indicates that, independent of g, there is no
residual genetic overlap between reading and mathematics
and that there are significant genetic influences specific to
reading and mathematics.
Shared environment also shows substantial overlap
between reading, mathematics and g, contributing almost
as much as genetics to their phenotypic correlations and
yielding correlations of 0.89–0.94. Non-shared environment,
which includes error of measurement, is the chief contributor
to differences between abilities, accounting for only about
10% of the phenotypic correlations and yielding correlations
of 0.15–0.22.
Discussion
Using direct tests of cognitive ability, reading and mathema-
tics in the largest representative twin sample to date, the
Table 1: Measure means (M) and standard deviations (SDs) by sex and zygosity
Measure
MZMDZM DZOM MZFDZFDZOF
ANOVA P values
MSDMSDMSDMSDMSDM SDSexZygosity
Sex ?
Zygosity R2
n
Mathematics
Reading
g
0.04 1.02 0.15 1.01 0.03 1.01 ?0.11 0.96 ?0.05 1.05 0.00 0.94 0.00** 0.11
?0.06 1.06 0.04 1.04 0.04 1.02 ?0.06 0.96
0.06 0.97 0.14 0.96 0.05 1.02 ?0.11 1.03 ?0.08 1.03 0.00 0.95 0.00** 0.04*
0.49
0.90
0.58
0.01 4922
0.00 5351
0.01 4684
0.02 1.00 0.05 0.93 0.92 0.00**
DZF, DZ females; DZM, DZ males; DZOF, females in DZ opposite-sex pairs; DZOM, males in DZ opposite-sex pairs; MZF, MZ females; MZM, MZ
males.
*P values significant at the <0.05 level.
**P values significant at the <0.01 level.
Table 2: Twin similarity coefficients (intraclass correlations) for mathematics, reading and g
MeasureMZDZS DZODZall MZMMZFDZM DZF
Mathematics0.71 (0.68–0.75),
(n ¼ 863)
0.64 (0.60–0.67),
(n ¼ 919)
0.71 (0.67–0.74),
(n ¼ 833)
0.51 (0.45–0.56),
(n ¼ 762)
0.44 (0.39–0.50),
(n ¼ 817)
0.53 (0.47–0.58),
(n ¼ 728)
0.43 (0.37–0.49),
(n ¼ 751)
0.42 (0.36–0.47),
(n ¼ 805)
0.44 (0.38–0.50),
(n ¼ 709)
0.47 (0.43–0.51),
(n ¼ 1513)
0.43 (0.39–0.47),
(n ¼ 1622)
0.48 (0.44–0.52),
(n ¼ 1437)
0.69 (0.63–0.74),
(n ¼ 356)
0.64 (0.58–0.70),
(n ¼ 383)
0.71 (0.65–0.76),
(n ¼ 342)
0.73 (0.68–0.77),
(n ¼ 507)
0.63 (0.58–0.68),
(n ¼ 536)
0.71 (0.66–0.75),
(n ¼ 491)
0.47 (0.38–0.55),
(n ¼ 343)
0.43 (0.34–0.51),
(n ¼ 373)
0.50 (0.42–0.58),
(n ¼ 328)
0.54 (0.47–0.60),
(n ¼ 419)
0.46 (0.38–0.53),
(n ¼ 444)
0.54 (0.47–0.61),
(n ¼ 400)
Reading
g
All similarity coefficients are based on age-, sex- and download-time-corrected scores.
95% confidence intervals in parentheses.
Dzall, all DZ pairs; DZO, opposite-sex DZ pairs; DZS, same-sex DZ pairs; MZ, MZ pairs; MZF, MZ female pairs; MZM, MZ male pairs; n, number of
twin pairs.
Table 3: Parameter estimates for mathematics, reading and g
MeasureACE
Mathematics
Reading
g
0.49 (0.40–0.57)
0.38 (0.29–0.48)
0.44 (0.36–0.53)
0.23 (0.15–0.30)
0.25 (0.17–0.33)
0.27 (0.19–0.35)
0.29 (0.26–0.31)
0.37 (0.33–0.40)
0.28 (0.26–0.31)
These estimates are based on the best-fitting submodel of the full
sex-limitation model, the null model, indicating no quantitative or
qualitative differences in etiology between males and females.
95% confidence intervals in parentheses.
A, additive genetic influence; C, shared environmental influence;
E, non-shared environmental influence.
Genes, Brain and Behavior (2008) 7: 455–462
459
Generalist genes and the Internet generation