Developmental path between language and autistic-like impairments: a twin study

Infant and Child DevelopmentInf. Child Dev. 17: 121–136 (2008)

Published online 8 January 2008 in Wiley InterScience

(www.interscience.wiley.com) DOI: 10.1002/icd.536

Developmental Path BetweenLanguage and Autistic-likeImpairments: A Twin Study

Katharina Dworzynskia,*, Angelica Ronalda, Marianna E.Hayiou-Thomasb, Fiona McEwana, Francesca Happea,Patrick Boltona and Robert Plomina

a Institute of Psychiatry, King’s College London, London, UKb Department of Psychology, University of York, York, UK

Autism spectrum disorders (ASDs) are diagnosed when indivi-duals show impairments in three behavioural domains: commu-nication, social interactions, and repetitive, restrictive behavioursand interests (RRBIs). Recent data suggest that these three sets ofbehaviours are genetically heterogeneous. Early language delay isstrongly associated with ASD, but the basis for this associationand the relationship with individual sub-domains of ASD hasnot been systematically investigated. In the present study, datacame from a population-based twin sample with languagedevelopment data at 2–4 years, measured by the MacArthurCommunicative Development Inventory (MCDI), and data at 8years using the Childhood Asperger Syndrome Test (CAST). Forthe total CAST and the three subscales at 8 years, approximately300 same-sex twin pairs were selected as showing extremeautistic-like traits (ALTs), defined here as pairs in which at leastone member of the twin pair scored in the highest 5% of thedistribution. Phenotypic analyses indicated that children show-ing extreme social and communication ALTs (but not the RRBIsubscale) at 8 years were below average in language developmentat 2–4 years. A regression model for selected twin data suggestedthat genetic influences account for this overlap, but that theseeffects are only in part mediated by genes that are sharedbetween language and extreme autistic traits. Copyright # 2008John Wiley & Sons, Ltd.

Key words: autism spectrum disorders; early language development

��

*Correspondence to: K. Dworzynski, Institute of Psychiatry, The Newcomen Centre,St. Thomas Street, London SE1 9RT, UK. E-mail: [email protected]

Copyright # 2008 John Wiley & Sons, Ltd.

INTRODUCTION

Communication impairments, social impairments, and restrictive and repetitivebehaviours or interests (RRBIs) form the core triad of features characteristic ofautism spectrum disorders (ASDs; DSM-IV, American Psychiatric Association,1994; ICD-10, World Health Organization, 1994). As well as communicationproblems, however, a significant proportion of individuals with ASDs have adelay in, or total lack of, the development of spoken language. Indeed, thedistinguishing feature between autism and Asperger’s syndrome is whether ornot early language delay is present. The aim of the present study is to investigateto what degree early language problems share the same genetic and environ-mental risk factors as core autistic features.

Twin studies have shown that autism is highly heritable (Rutter, 2005).Furthermore, autistic traits have been found to be continuously distributed andindividually highly heritable (Constantino & Todd, 2000, 2003; Ronald, Happe, &Plomin, 2005). Recently, however, it has been shown that there is only modestgenetic overlap between communication impairments, social impairments andRRBIs, when assessed as traits in the population (Ronald et al., 2005; Ronald,Happe, Bolton, et al., 2006), as well as when assessed at the extreme (Ronald,Happe, Price, Baron-Cohen, & Plomin, 2006), providing evidence for geneticheterogeneity within the autistic triad. This has now been shown in two differentstudies using two different measures.

In light of these recent findings of genetic heterogeneity between communicationimpairments, social impairments and RRBIs, as well as their known associationwith language delay, it is now paramount to investigate the cause of the linkbetween each of these symptom domains individually with language delay.

In the present study a longitudinal analytical strategy was used which is calledretrodictive (Oliver, Dale, & Plomin, 2004). This strategy differs from bothprospective and retrospective analyses. A prospective study includes manychildren whose early problems are transient, whereas the targets of theretrodictive study are children with problems later on in development. Incontrast to the retrospective approach, rather than using data that are collectedretrospective, here the data are collected longitudinally, and can in this wayprovide important information about earlier precursors of later autistic-likeimpairments. This approach was employed to investigate to what degree earlylanguage delay shares the same genetic and environmental causes as each of thethree autistic symptom domains characteristic of autism, measured duringmiddle childhood.

The prediction is that language will show greatest phenotypic and geneticoverlap with the communication domain since a strong relationship existsbetween social non-verbal communication and the level of language develop-ment in autism (e.g. those that show better joint attention develop more receptivelanguage and at an earlier age, Mundy, Sigman, & Kasari, 1994). It wasalso expected that language problems might show an association withsocial impairments, since language is a vital tool in the social context. However,work by Bishop and colleagues has suggested that communication and languageproblems can exist in individuals who do not show any social impairments per se,such as individuals with specific language impairment (SLI) (Bishop, 2000).An association between RRBIs and language problems was not expectedbecause of previous evidence that language level is not related toRRBIs in children with autism or language-impaired children (Lord & Pickles,1996).

K. Dworzynski et al.122

Copyright # 2008 John Wiley & Sons, Ltd. Inf. Child Dev. 17: 121–136 (2008)DOI: 10.1002/icd

A related earlier study by our group, based on the same cohort as used in thepresent study, investigated the link between language ability and autistic traits in thegeneral population (Dworzynski et al., 2007). This study found that language abilitywas linked, both phenotypically and genetically, with both social impairments andcommunication impairments, but not RRBIs. The present study differs from ourprevious study in that it investigates the link between language in early childhood (2–4 years) and autistic symptoms in middle childhood (7–9 years) at the extreme of thedistribution. It employs only a sub-sample of the original sample, that is those childrenat the high end of the distribution who show autistic traits in the top 5% in middlechildhood. t also differs from previous studies because it uses a different method.Instead of an individual differences approach with structural equation models, weused a means-based regression model to assess the role of genetic and environmentalfactors at the extremes. In this regression model, the scores of cotwins are predictedfrom the proband scores and a score that estimates the genetic relationship betweentwins (1 for monozygotic pairs and 0.5 for dizygotic pairs). This is important because itis possible that the aetiology of the association between language performance andautistic traits at the extremes could differ from the aetiology of the association betweenlanguage and autistic-like traits (ALTs) in the normal range.

METHODS

Participants

Data came from TEDS, a longitudinal, UK-based population study of twins bornin 1994–1996 (Oliver & Plomin, 2007; Trouton, Spinath, & Plomin, 2002). Afterchecking for infant mortality, all families identified by the UK Office for NationalStatistics (ONS) as having twins born in these years were invited to participate inTEDS when the twins were about 18 months old. At ages 2, 3 and 4 years, 6285,6056 and 8148 families participated which is 55.4%, 53.4% and 48.5% (age 4contains the third cohort for the first time hence the larger number but lowerpercentage) of the entire sample. There were a total number of 9974 families whosupplied data on at least one of 2, 3 and 4 years questionnaires (and also had firstcontact data available) corresponding to 59.3% of the entire sample. For thelanguage data, families were excluded if English was not the only languagespoken at home (N ¼ 518 families). In 2002–2004 when the twins were aged 8,families were contacted again and 6771 families (55.0% of active contact families)returned questionnaires. The following medical exclusions were made: childrenwith a specific medical syndrome (not including ASDs) such as Down’ssyndrome, Fragile X, chromosomal anomalies, including cystic fibrosis, andcerebral palsy (there were 405 individual cases who were excluded on thesegrounds). Twins were also excluded when zygosity data were unavailable(N ¼ 83 families). Further exclusions were made for those families who had notsigned consent (N ¼ 57 families), where the birth order had been incorrectlyrecorded (N ¼ 207 families at age 8) and overall those who had no first contactdata available (N ¼ 763 families). At age 8 the remaining sample, after exclusion,consisted of 6087 families.

Language Measures

When the twins were 2, 3 and 4 years of age, parents completed age-appropriate, reduced versions of the MacArthur Communicative Development

Autistic-like Impairments and Early Language 123


Inventory which were adapted for the UK (MCDI: UKSF; Dionne, Dale, Boivin, &Plomin, 2003). At age 2, a measure of expressive vocabulary was used which consistedof a possible maximum of 100 items of words that the child could say at the time. Bothfor ages 3 and 4 a combined expressive vocabulary and grammar measures wereanalysed. This consisted of the combination of number of spoken words (out of 100and 48 items of an upward extension of the MCDI for ages 3 and 4, respectively) aswell as points on a grammar complexity scores. Expressive grammar was measuredusing an anglicized abbreviation of the MCDI grammar scales. At age 3 this consistedof 13 items in which parents had to choose from a pair of word combinations the typeof utterance that best characterized what their child can say. These scores are thenrescaled into a 5-point composite (Dionne et al., 2003). Sentence complexity wasmeasured at age 4 where parents had to rate more global aspects of their children’sadvancing grammar (six questions). At both ages 3 and 4, a verbal composite wascalculated by averaging the standardized vocabulary and grammar scores.

Language Factor Score (LFS)

From these three language measures for the ages 2, 3 and 4 for each child a factorscore was calculated. This was carried out for two reasons. Firstly, a factor score fromthe three measured language scores would represent stable language performanceand secondly a factor score can be calculated even for cases with missing values atsome time points and therefore the sample size is increased which in turn allows formore clinically relevant cut-off points to be used. This procedure was carried outusing a specifically designed program for this purpose (the latent variable scorecalculator is publicly available at http://statsgen.iop.kcl.ac.uk/lsc). This is astructural equation model fitting approach to calculate scores and in this case thecovariance matrix and factor loadings are required as input files. These two fileswere produced with the MX program which handles incomplete data by using rawdata maximum likelihood. Since a script with three measured variables and a one-factor solution has as many parameters as the saturated model the goodness of fit isdifficult to determine. A comparative factor analysis using the SPSS package (whichhandles missing cases differently to the MX software) showed that the one-factorsolution accounted for more than 70% of the variance. Since the method of analysisrequires the relationship between the two measures to have positive (i.e. above 0)values, LFSs were reversed, to aid interpretation of results.

Autistic-like Traits

The CAST (Scott, Baron-Cohen, Bolton, & Brayne, 2002) questionnaire data werecollected when the twins were between 7 and 9 years old. It is a screening instrumentfor autism spectrum conditions, which is completed by parents, and is designed fornon-clinical samples. It consists of 37 items overall, of which 31 contribute to a child’stotal score. The remaining six items are control items (one such control item pertinentto the current study is ‘Was s/he speaking by 2 years old?,’ which is neither includedin the total rating and communication subscale scores}see Scott et al., 2002, for detailson all control items). All questionnaire items are answered ‘yes’ or ‘no’ and responsesare scored additively with a score above 14 indicating a cut off for children at risk forASDs. In case of missing data a total score is calculated by summing all items andconverting this into a mean of total possible scores given the number of itemscompleted (with the constraint that more than half of items have to be answered).

In a recent validation study (Williams et al., 2005) the questionnaire was shownto have excellent sensitivity and specificity (the proportion of children with a



diagnosis scoring over the cut-off was 100% and the proportion of childrenwithout a diagnosis scoring under the cut-off was 95–97%). It was developed as aquantitative scale and is based on a dimensional concept of ASDs. The 31 itemsderived from behavioural descriptions from ICD-10 and DSM-IV (AmericanPsychiatric Association, 1994; World Health Organization, 1994) assess the threecore features of ASD (social impairments, communication abnormalities, andRRBIs). The present sample showed good internal consistency with a Cronbachalpha value of 0.69.

In addition to the total CAST score, three subscales were derived from thequestionnaire using the same method as Ronald, Happe, Price et al. (2006). Thedivision is made on theoretical grounds (questions are divided according to thethree sub-domains of ASD as reflected, for example, in DSM-IV) and missing dataare handled as for the total CAST. Examples of questionnaire items selected for thedifferent subscales are: Social subscale}‘Is it important to him/her to fit in with thepeer group?’ (reverse scored), Communication subscale}‘Can s/he keep a two-wayconversation going?’(reverse scored), and RRBI subscale}‘Does s/he try to imposeroutines on him/herself, or on others, in such a way that it causes problems?’ For allthe CAST items and their division into subscales refer to the Appendix.

A top 5% cut-off is a standardized z-score of 1.82 in the current sample which isequivalent to a raw total CAST score of 12 or above (for the sub-domains z-score cut-offs of 1.89 for social, 1.97 for communication and 1.82 for RRBI scales were used).

In the TEDS sample all parents of children who had been either reported bytheir parents to have an autism spectrum condition, or those who were above thescreening cut-off for CAST questionnaires were contacted. A telephone interviewwas then administered using a specifically designed module of the Developmentand Well-being Assessment (Goodman, Ford, Richards, Gatward, & Meltzer,2000). Two clinicians, specialized in ASDs, reviewed the data and reachedagreement to establish the numbers of children with an autism spectrumcondition within the TEDS sample.

Table 1 shows the proportion of children with an autism spectrum condition inthose selected as probands and those with early language difficulties.

Logistic Regressions Analysis

The categorical status of the three sub-domains (falling into the lowest 5% or not)is to be predicted from the continuous LFS variable. This then establisheswhether the odds are higher to fall into the ALT impaired group when there wasa lower prior language score. Continuous rather than categorical language scoresare used since this makes it comparable to the genetic analysis which also usescontinuous language scores. All twins were entered into logistic regressionanalyses with standard errors robust against non-independence of observationsfrom individuals within families (robust clusters) and against departures fromnormal assumptions and were carried out in STATA (version 9.2; Stata Corp.,College Station, TX).

Bivariate DeFries and Fulker (DF) Extremes Analysis: Do the Same Genes thatCause Autistic-like Impairments also Cause Early Language Delay?

Bivariate DF extremes analysis is a technique in which probands are identified onthe basis of low scores on one-dimensional measure (X}here ALTs) and thencompare their co-twins scores with another dimensional measure (Y}LFSs).



If extreme ALTs are linked genetically with LFSs then the quantitative trait scorefor the language score of probands’ co-twins should be more similar to thepopulation mean for DZ than MZ twins. Figure 1 illustrates the differentialregression to population means utilized by this analytic method.

The first statistic to consider is the phenotypic group correlation (PGC) whichis calculated by dividing the probands’ mean standardized z-score on theunselected variable (i.e. the composite language measure) by the probands’ meanstandardized z-score on the selected variable (i.e. the autism trait score). It is thusan index of the magnitude of the phenotypic relationship at the extreme betweentwo variables (i.e. this is within probands rather than across twin pairs). Thus, itis an indication of the magnitude of comorbidity of consistent early languagescores and ALTs.

When probands are selected for scores at extreme ends of the distributions on avariable of interest (in this case the top 5% of ALTs) both the MZ and DZ cotwinswill regress towards the mean of the unselected population (DeFries & Fulker,1985). If poor performance depended on chance factors specific to the individual,mean scores of co-twins would be expected to regress back to the populationmean. In comparison, if environmental factors common to both twins affectscores, then twins and co-twins should resemble one another, regardless ofzygosity. If genes play an important role in a condition then DZ cotwin scoresshould regress farther back to the population mean than MZ cotwin scores. Inother words, DF methodology addresses the genetic and environmental originsof the mean difference on the quantitative trait between probands and thepopulation. The basic multiple regression model for the analysis of selected twindata is as follows:

C ¼ b1Pþ b2Rþ A

where C is the cotwin’s predicted score, P is the proband’s score, R is thecoefficient of relationship, A is the regression constant. b1 is the partial regressionof the cotwin’s score on the proband’s score and is a measure of twin resemblanceindependent of zygosity, and b2 is the partial regression of the cotwin’s score on

Table 1. Confirmed proportions of children with autism spectrum conditions in theoverall sample as well as in the 5% cut-off groups of children with early languagedifficulties and those with later autistic-like trait impairments

Sample Proportion ofconfirmed autismspectrum conditions(number of cases/total number)

95%confidenceinterval

z scoredifference inproportions(compared tooverallsample)

p

Overall 1.0% (136/12 884) 0.87–1.23%Children with earlylanguage difficulties

4.3% (23/526) 2.62–6.12% 6.89 p50:0001

Children social autistictrait impairments

16.1% (97/601) 13.2–19.1% 27.77 p50:0001

Children withcommunication autistictrait impairments

15.9% (96/605) 13.0–18.8% 27.39 p50:0001

Children with RRBIautistic traitimpairments

12.9% (78/603) 10.2–15.6% 22.82 p50:0001

Note: Children were selected for assessment from CAST questionnaire responses.



the coefficient of relationship. b2 equals twice the difference between MZ and DZcotwin means after covariance adjustment for any differences between MZ andDZ proband scores. By standardizing and transforming the scores, i.e. expressingeach score as a deviation from the control mean (standardizing) and thendividing by the difference between the proband and control means prior tomultiple regression (transforming), b2 provides a direct estimate of groupheritability}h2:x: The current analysis makes use of the bivariate extension of thebasic multiple regression model for the analysis of selected twin data (Knopik,Alarcon, & DeFries, 1997; Light and DeFries, 1995; Purcell et al., 2001; Stevenson,Pennington, Gilger, DeFries, & Gillis, 1993). This method considers whether thesame genes are implicated in two heritable traits. In the bivariate extension, theprobands are selected on the ‘selection variable’, x (in this case ALTs), but thecotwin’s score is investigated on the ‘unselected variable’, y (in this case earlylanguage performance). The bivariate DF regression equation is now:

Cy ¼ b1Px þ b2Rþ A

where Cy is the cotwin’s predicted score on the unselected (language scores)variable, Px is the proband’s score on the selected variable (each ALT domain), Ris the coefficient of relationship, and A is the intercept. The following

0 0.2 0.4 0.6 0.8 1

DZ

MZ

DZ

MZU

nsel

ecte

d va

riabl

e (la

ngua

gefa

ctor

sco

re)

Sel

ecte

d va

riabl

e (a

utis

tic-li

ketr

ait s

core

)

probandco-twin

bivariate .5 h2g.xy

univariate .5 h2g.x

Figure 1. Illustration of bivariate DF extremes analysis. Data are transformed by dividingproband scores on the selected variable (in this case the autistic-like trait scores or X) andproband scores on the unselected variable (here the composite language scores or Y) aswell as co-twin selected (autistic-like trait) and co-twin unselected (language factor) scoresby the proband zygosity specific mean of the selected variable (autistic-like trait). Therefore,the proband score of the selected variable has a population mean ¼ 0 and a probandmean ¼ 1: The proband’s transformed score on the selected variable (autistic-like traits)can then be used in a regression model which also includes a coefficient of relatedness (i.e.1 for MZ and 0.5 for DZ twins) to predict co-twin’s scores on the unselected variable(language factor score). The term h2

g:xy then assesses the extent to which genetic factors areresponsible for the lowered Y scores of probands with low X. See text for further details.Note: The h2

g:xy estimate is logically constrained to be equal to, or lower than, the MZtransformed cotwin mean.



transformation of the data is carried out prior to regression. If probands areselected on variable x then each score is expressed as a deviation from therespective x and y mean of the unselected population and then divided by thedifference between the proband and the control mean in x scores (see for exampleKnopik et al., 1997). When the data are transformed in this manner prior tomultiple regression analysis, b2 equals hxhyrgðxyÞ which is an index of the extent towhich the proband x deficit is due to genetic factors that also influence ydisability (from now on referred to as h2:xy). The ratio of h2:xy to the PGC providesa measure of the proportion of their covariance that can be attributed to geneticfactors. These analyses are bidirectional which means that selecting probands fordisorder x or disorder y are separate studies that can yield different results. Inother words, results can vary depending on proband selection (either x or y), thereason being that even though some of the same individuals are selected for bothx and y many of the x probands are not y probands and vice versa. The focus ofthis paper is the analysis of language once probands are selected for high ALTs atage 8, but the h2:yx for the reverse relationship is also analysed in order tocalculate genetic correlations (Knopik et al., 1997), since genetic correlations arederived using the following formula (which requires h2ðxyÞ as well as h2ðyxÞ):

rgðxyÞ ¼ ððh2:xyÞ*ðh2:yxÞ=ðh2:xÞ* ðh2:yÞÞ

The genetic correlation is an index of the degree to which mean differences inthe two measures between probands and the population are due to the samegenetic influences.

h2 estimates should not be higher than the MZ transformed cotwin mean, butthis can sometimes occur when data show a non-additive genetic pattern ofeffects (that is, when DZ transformed cotwin means are less than half the MZcotwin means). In these circumstances, h2 are constrained to the value of the MZtransformed cotwin mean.

RESULTS

We first report phenotypic relationships between the CAST subscales in middlechildhood (7–9 years) and language scores in early childhood (2–4 years) usinglogistic regression and PGCs. We then use bivariate DF extremes analysis toestimate genetic and environmental sources of these phenotypic relationships.

Phenotypic Logistic Regression

Odds ratios are shown in Figure 2 for the prediction of membership in the top 5%of ALTs in middle childhood from language performance in early childhood. Allodds ratios are significantly larger than 1.0 indicating that children in the top 5% ofALTs in middle childhood are more likely to have lower language scores in earlychildhood for all three ALTs (communication subscale: Wald chi-square ¼ 95:56;df ¼ 1; p50:0001; OR 1.92 CI 1.49–1.89; RRBI subscale: Wald chi-square ¼ 15:83;df ¼ 1; p50:0001; OR 1.32 CI 1.15–1.51; social subscale: Wald chi-square ¼ 73:61;df ¼ 1; p50:0001; OR 1.68 CI 1.49–1.89). However, Figure 2 shows that earlylanguage scores are most predictive of communication ALTs in middle childhood.The confidence intervals indicate that the odds ratio is significantly greater for thecommunication subscale than the RRBI subscale.



Phenotypic Group Correlations (PGCs)

As shown in Table 2, PGCs yield similar results. All PGCs are rather modest withthe greatest relationship being between communication subscale scores and LFSs(0.18), somewhat less for those between social subscale and early language (0.14),and clearly less for RRBI subscale and language (0.07).

As can be seen from the means and S.D., children in the top 5% of the CASTcommunication score at 8 years are nearly 0.5 of a S.D. below the mean on the LFSin early childhood, which is a moderate effect size (Cohen, 1988).

Bivariate DF Extremes Analysis

In order to estimate the genetic correlation between early language and laterALTs, univariate DF results are needed for all measures. These are not the focusof the present study and will be presented only briefly here. Univariate h2x DFextremes estimates for ALTs were 0.68 for selection on total CAST, 0.65 for thesocial subscale, 0.76 for the communication subscale, 0.66 for the RRBI subscale.These univariate estimates are consistent with previous results from TEDS}seeRonald, Happe, et al., 2006}slight variations in estimates are due to the largersample size in the current study). For low language in early childhood, the groupheritability estimate is 0.64, which is also consistent with previous reports fromour group (Spinath, Price, Dale, & Plomin, et al., 2004}again variation due toslightly different sample and language variable).

As indicated in the Methods section, bivariate heritability in this case restson the mean LFSs of co-twins of probands selected for high ALTs. As shown in

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2

Social subscale

Communication subscale

RRBI subscale

Au

tist

ic-l

ike

trai

t im

pai

rmen

t

Odds ratios

*

Figure 2. Odds ratios (and 95% confidence intervals) for the prediction of groupmembership to the top 5% of autistic-like traits at age 8 from stable language performanceduring ages 2–4. All OR significant p50:001: Note: Difference between odds ratiosindicated by n significant p50:05: All individual odds ratios significantly ðp50:0001Þlarger than 1.



Table 2, for all total CAST and subscale probands, there was a differentialregression of the MZ and DZ cotwin means for language scores. MZ cotwins, onaverage, do not regress more than 0.03 S.D. units for total CAST as well as thesubscales towards the unselected population mean for LFSs. DZ cotwins, onaverage, regress between 0.22 (RRBI probands) and 0.34 (total CAST scoreprobands) S.D. units for language scores. This differential regression indicatesthat there is genetic influence on the relationship between extreme ALTs andlanguage delay. The pattern of these results corresponds to the phenotypic resultsof the logistic regression (Figure 2) and the PGCs (Table 2) in that the total CASTand communication subscale showed the largest effects.

h2:xy estimates are also given in Table 3, which index the extent to which theproband extreme ALTs are due to genetic factors that also influence earlylanguage performance. These h2:xy estimates range from 0.12 to 0.24 with thelargest estimates for selections on the total CAST score and communication sub-domain. The ratios of the h2:xy estimates to the observed PGC (ratios are: totalCAST 0.24/0.19, social 0.12/0.14, communication 0.23/0.18, RRBI 0.13/0.07)indicate that the entire covariance is attributable to genetic factors. Ratios largerthan 1 are sometimes possible particularly when there is a pattern of non-additivity in the data (DZ transformed cotwin means less than half of MZtransformed cotwin means). However, the h2:xy estimate confidence intervals allinclude the respective PGC values.

Looking at the confidence intervals of these for the subscales, it is clear thath2:xy estimates for the total CAST and communication subscale are significant butthose for the Social and RRBI subscale probands are only marginally so.

The formula to calculate genetic correlations also requires the reverse bivariateheritabilities (i.e. those selected for language difficulties). h2:xy estimates forbivariate language selection: �0:19; �0:05; �0:27 and �0:03 for total CAST, social,communication and RRBI, respectively. The last column of Table 3 gives theresults of the genetic correlations (bivariate rg) for the relationships that areanalysed in each panel of the table. It can be seen in Table 3 (final column) thatthe largest genetic correlation of 0.36 was between the extreme communicationsubscale probands and the LFS. This shows a modest overlap indicating thatearly stable language delay appear to be due to some of the same geneticinfluences as those that play a role in later communication problemscharacteristic of ASDs. Only very modest genetic overlap was seen for both the

Table 2. Phenotypic group correlations (PGCs) for probands who fell within the top 5%of autistic like traits

Measure}cut off threshold (top 5%) Mean (SD) N PGC

Total CAST 2.82 (1.03) 600LFS 0.54 (1.14) 453 0.19Social subscale 2.77 (1.00) 603LFS 0.39 (1.06) 486 0.14Communication subscale 2.71 (0.73) 601LFS 0.50 (1.10) 453 0.18RRBI subscale 2.47 (0.62) 598LFS 0.19 (1.14) 452 0.07

Note: These are the ratios of the probands’ mean score on the unselected variable and the probands’mean score on the selected variable. LFS, language factor score (from ages 2 to 4 years); RRBI,repetitive, restrictive behaviours and interests; S.D., standard deviation; N; number of probands.



social and RRBI subscale selections (0.12 and 0.10, respectively). The relationshipbetween these different statistics and the implication for the phenotypicassociation between these measures is depicted in Figure 3. All analyses werere-run excluding those children who were diagnosed as having an ASD to seewhether patterns of effects were highly influenced by these cases. Results of theseanalyses were only minutely different from those including all children and aretherefore not reported here.

DISCUSSION

The present study set out to analyse developmental pathways between earlylanguage performance for those children who have been selected for specificimpairments in social, communication and RRBI domains of ALTs in middlechildhood.

It was shown that the relatively small phenotypic associations were entirelymediated by genetic factors that were, however, only partially shared betweenALTs at the extremes of the distribution with early language. When comparingthese results to those from the analysis of overall individual differences of thewhole population (Dworzynski et al., 2007) close similarities are evident. In both

Table 3. DeFries–Fulker bivariate extremes model fitting results

Zygosity N Proband’sLFS score

Cotwin h2g:xy rg

Total CAST probandsMZ 130 Stand 0.64 0.65DZ 155 Stand 0.53 0.19MZ 130 Trans 0.23 0:24� 0:08DZ 155 Trans 0.19 0:07� 0:06 0:24� 0:08 0.33

Socially impaired probandsMZ 128 Stand 0.36 0.33DZ 174 Stand 0.38 0.12MZ 128 Trans 0.13 0:12� 0:08DZ 174 Trans 0.14 0:04� 0:05 0:12� 0:08 0.12

Communication impaired probandsMZ 146 Stand 0.64 0.61DZ 156 Stand 0.47 0.25MZ 146 Trans 0.24 0:23� 0:09DZ 156 Trans 0.17 0:09� 0:06 0:23� 0:09 0.36

RRBI probandsMZ 135 Stand 0.29 0.32DZ 149 Stand 0.24 0.02MZ 135 Trans 0.12 0:13� 0:10DZ 149 Trans 0.10 0:01� 0:07 0:13� 0:10 0.10

Note: h2g:xy estimates constrained to be equal or lower than MZ transformed cotwin mean which is a

logical upper bound for the estimate. LFS, language factor score; N; number of probands.Standardized scores (Stand) expressed as S.D. units from the mean of whole unselected sample (i.e.z scores). For transformed scores (Trans), see text. 95% confidence intervals provided, calculated usingstandard errors corrected for double entry. Mean transformed proband language scores on thetransformed language measure are shown for interest, but do not feature in the DF analysis (i.e.probands’ autistic-like trait scores are used to predict cotwins’s language scores).



analyses communication items were those with slightly higher associations toearlier language performance. However, these were still relatively weak and onlypartially due to genetic influences common to both. One conclusion to draw fromthese similarities is that there may not really be a differentiation between thosewho score within the normal range on ALT dimensions and those scoring above acertain cut-off. Another explanation for these similarities is a strictly distribu-tional effect. Since the CAST yields skewed data, ACED analysis (as used in therelated individual differences paper) will be especially influenced by the highextremes.

It was predicted that the strongest genetic relationship with early languagedevelopment would be for those probands selected on the communicationdomain and to a lesser degree those with social problems. Results of the presentstudy supported this prediction, but with smaller effects than was anticipated. Itwas shown that the proportion of children who later were diagnosed with anautism spectrum condition was significantly higher in the early language delaygroup than in the general population. This corresponds to other research findings(Conti-Ramsden, Simkin, & Botting, 2006) as well as reported experiences inclinical practice (as highlighted by the accounts of Rapin, 1996; Rutter, 1967). Allof the modest phenotypic overlap between all three autistic triad impairmentsand early language was due to genetic influences. However, these genetic effectswere mostly not shared between autistic traits and early language. The geneticoverlap was largest between language performance and communicationimpairments (but only just over one-third), but very small between socialimpairments and early language, and between RRBIs and language scores. Thegenetic correlation of 0.36 suggests that language and communication deficits are,in part, due to the same genetic influences.

Limitations of the current work include reliance on parental report and thefact that CAST data are positively skewed. This could be resolved bytransformation of the data but recent evidence by Bishop (2005) has shownthat the DF procedure (at least in univariate cases) is fairly robust to thesedeviations from normal distributions. Another drawback concerns the causalinferences from this work. To draw conclusions regarding effects of onevariable on the other (language and ALTs) ideally both measures would be

Predominantlynot the samegenes (rg .33,.12, .36 and .10)

Entirelymediated by geneticinfluences

Small phenotypicoverlap (PGCs .19, .14,.18 and .07)

Probands scoring high on Total, Social, Communication or RRBI autistic-like traits

Early languageperformance

Figure 3. Relationship between phenotypic group correlations, the amount of covarianceaccounted for by genetic influences and genetic correlations.



obtained at both ages. With the present study autistic traits were only available atage 8 and not at earlier ages.

In terms of interpreting these findings, results highlight that there are bothoverlap and differences between language as defined by lexical and semanticcomponents and those communicative aspects of language relating to pragmaticsthat are impaired in children with ASDs. A genetic correlation of 0.36, as reportedhere, highlights that a large proportion of the relationship is due to specific genesthat operate at each age and for each trait. Communication ALTs encompassedmainly pragmatic skills which underpin communication in the social context.Thus, they were quite distinct from those skills measured in the MCDI (lexicaland grammatical abilities). Our findings therefore provide evidence that somechildren who have a good level of early vocabulary and syntax can nonethelesshave later difficulties with pragmatic skills such as for example turn-taking orkeeping a conversation going.

The genetic correlation between language scores and later autistic-likeimpairments showed that even though genes mediate all of this relationship,the majority of genetic influences are not shared with the highest level of geneticoverlap being only just over one-third. Such levels of genetic overlap areconsistent with molecular linkage studies of autism and SLI, which haveon the whole not corresponded to the same regions of the genome butwhich have revealed some genomic areas where there are overlaps in linkagesignal. This raises the possibility that these specific regions may harboursusceptibility genes for those autistic traits corresponding to communicationdifficulties and SLI}such as the 13p region reported in a pedigree studyof SLI (Bartlett et al., 2002). It should, however, be highlighted that having a highCAST score is not synonymous with being on the autism spectrum. In fact,the original CAST reference (Scott et al., 2002) pointed out that a number ofthose children scoring in the screening cut-off did in fact have languageimpairments rather than autism. Therefore, the current results could simplyreflect the overlap between early language and later language impairments.Since language impairment status is not systematically collected for all twins inthe TEDS sample it is hard to know how many of those high scoring childrendid in fact have diagnoses other than ASDs. If it were the case that this is simply areflection from earlier to later language then taking those children with diagnosesof ASDs out of the analyses should strengthen the current associations. Thiswas not the case. A rerun of the current analyses without diagnosed childrenrevealed no significant differences for overall CAST scores and all threesubscales.

Our results provide further evidence in support of genetic heterogeneity ofALTs and have implications for both molecular and intervention studies. Infuture it will be important to clarify developmental and aetiological pathways tothe various facets of the syndrome as this might lead to more targeted treatmentapproaches.

APPENDIX A. SUBSCALES IN THE CHILDHOOD ASPERGERSYNDROME TEST ðCASTÞn

The following are all items in the original questionnaire (Scott et al., 2002: theCambridge Autism Research Centre details regarding scoring can be found athttp://www.autismresearchcentre.com/tests/default.asp).

Instructions to parents:



Please read the following questions carefully, and circle the appropriateanswer. All responses are confidential.

1Social Does s/he join in playing games with other children? Yes No2Social Does s/he come up to you spontaneously for a chat? Yes No3 Was s/he speaking by 2 years old? Yes No4 Does s/he enjoy sports? Yes No5Social Is it important to him/her to fit in with the peer group? Yes No6RRBI Does s/he appear to notice unusual details that others miss? Yes No7Com Does s/he tend to take things literally? Yes No8Social When s/he was 3 years old, did s/he spend a lot of time

pretending (e.g. play-acting being a superhero, or holdingteddy’s tea parties)?

Yes No

9RRBI Does s/he like to do things over and over again, in the sameway all the time?

Yes No

10Com Does s/he find it easy to interact with other children Yes No11Com Can s/he keep a two-way conversation going? Yes No12 Does s/he read appropriately for his/her age? Yes No13RRBI Does s/he mostly have the same interests as his/her peers? Yes No14RRBI Does s/he have an interest which takes up so much time that

s/he does little else?Yes No

15Social

Does s/he have friends, rather than just acquaintances? Yes No

16Social

Does s/he often bring you things s/he is interested in toshow you?

Yes No

17Com Does s/he enjoy joking around? Yes No18Com Does s/he have difficulty understanding the rules for polite

behaviour?Yes No

19RRBI Does s/he appear to have an unusual memory for details? Yes No20Com Is his/her voice unusual (e.g. overly adult, flat, or very

monotonous)Yes No

21Social

Are people important to him/her? Yes No

22 Can s/he dress him/herself? Yes No23Com Is s/he good at turn-taking in conversation? Yes No24Social

Does s/he play imaginatively with other children, andengage in role-play?

Yes No

25Com Does s/he often do or say things that are tactless or sociallyinappropriate?

Yes No

26 Can s/he count to 50 without leaving out any numbers? Yes No27Social

Does s/he make normal eye-contact? Yes No

28RRBI Does s/he have any unusual and repetitive movements? Yes No29Social

Is his/her social behaviour very one-sided and always onhis/her own terms?

Yes No

30Com Does s/he sometimes say ‘you’ or s/he’ when s/he means‘I’?

Yes No

31Social

Does s/he prefer imaginative activities such as play-actingor story-telling, rather than numbers or lists of facts?

Yes No

32Com Does s/he sometimes lose the listener because of notexplaining what s/he is talking about?

Yes No



33 Can s/he ride a bicycle (even without stabilizers)? Yes No34RRBI Does s/he try to impose routines on him/herself, or on

others, in such a way that it causes problems?Yes No

35Social

Does s/he care how s/he is perceived by the rest of thegroup?

Yes No

36Com Does s/he often turn conversations to his/her favouritesubject rather than following what the other person wants totalk about?

Yes No

37Com Does s/he have odd or unusual phrases? Yes No

nSuperscripts indicate Social, Communication (Com) and Repetitive or RestrictiveBehaviours and Interests (RRBI) subscale items. Those without superscripts arecontrol items (3, 4, 12, 22, 26 and 33) and are not included in subscale division.

Subscales were designed and previously used by Dr. Ronald (see Ronald et al.,2006; Ronald, Happe, et al., 2006).

In each subscale more than half the items are required, otherwise it is classed asa missing score.

ACKNOWLEDGEMENTS

The first author is funded by a National Alliance for Autism Research projectgrant to Patrick Bolton. TEDS is funded by an MRC program grant G0500079. Wethank the parents of the TEDS twins for their participation.

REFERENCES

American Psychiatric Association. (1994). Diagnostic and statistical manual of mental disorders(DSM-IV ed.) Washington, DC: APA.

Bartlett, C. W., Flax, J. F., Logue, M. W., Vieland, V. J., Bassett, A. S., Tallal, P., et al. (2002).A major susceptibility locus for specific language impairment is located on chromosome13q21. American Journal of Human Genetics, 71, 45–55.

Bishop, D. V. M. (2000). Pragmatic language impairment: A correlate of SLI, a distinctsubgroup, or part of the autistic continuum? In D. V. M. Bishop & L. B. Leonard (Eds.),Speech and language impairments in children: Causes, characteristics, intervention and outcome(pp. 99–113). Hove, UK: Psychology Press.

Bishop, D. V. M. (2005). DeFries–Fulker analysis of twin data with skewed distributions:Cautions and recommendations from a study of children’s use of verb inflections.Behavior Genetics, 35, 479–490.

Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Hillsdale, NJ:Erlbaum.

Constantino, J. N., & Todd, R. D. (2000). Genetic structure of reciprocal social behavior.American Journal of Psychiatry, 157, 2043–2045.

Constantino, J. N., & Todd, R. D. (2003). Autistic traits in the general population: A twinstudy. Archives of General Psychiatry, 60, 524–530.

Conti-Ramsden, G., Simkin, Z., & Botting, N. (2006). The prevalence of autistic spectrumdisorders in adolescents with a history of specific language impairment (SLI). Journal ofChild Psychology and Psychiatry and Allied Disciplines, 47, 621–628.

DeFries, J. C., & Fulker, D. W. (1985). Multiple regression of twin data. Behavior Genetics, 15,467–473.

Dionne, G., Dale, P. S., Boivin, M., & Plomin, R. (2003). Genetic evidence for bidirectionaleffects of early lexical and grammatical development. Child Development, 74, 394–412.



Dworzynski, K., Ronald, A., Hayiou-Thomas, M., Rijskijk, F., Happe, F., Bolton, P.,& Plomin, R. (2007). Aetiological relationship between language performance andautistic-like traits in childhood: A twin study. International Journal of Language andCommunication Disorders, 42(3), 273–292.

Goodman, R., Ford, T., Richards, H., Gatward, R., & Meltzer H. (2000). The developmentand well-being assessment: Description and initial validation of an integratedassessment of child and adolescent psychopathology. Journal of Child Psychology andPsychiatry, 41, 645–655.

Knopik, V. S., Alarcon, M., & DeFries, J. C. (1997). Comorbidity of mathematics andreading deficits: Evidence for a genetic etiology. Behavior Genetics, 27, 447–453.

Light, J. G., & DeFries, J. C. (1995). Comorbidity of reading and mathematics disabilities:Genetic and environmental etiologies. Journal of Learning Disabilities, 28, 96–106.

Lord, C., & Pickles, A. (1996). Language level and nonverbal social-communicativebehaviors in autistic and language-delayed children. Journal of the American Academy ofChild and Adolescent Psychiatry, 35, 1542–1550.

Mundy, P., Sigman, M., & Kasari, C. (1994). Joint attention, developmental level, andsymptom presentation in autism. Development and Psychopathology, 6, 389–401.

Oliver, B., Dale, P. S., & Plomin, R. (2004). Verbal and nonverbal predictors of earlylanguage problems: an analysis of twins in early childhood and infancy. Journal of ChildLanguage, 31, 609–631.

Oliver, B. R., & Plomin, R. (2007). Twins early development study (TEDS): A multivariate,longitudinal genetic investigation of language, cognition and behavior problems fromchildhood through adolescence. Twin Research and Human Genetics, 10(1), 96–105.

Purcell, S., Eley, T. C., Dale, P. S., Oliver, B., Petrill, S. A., Price, T., et al. (2001). Comorbiditybetween verbal and non-verbal cognitive delays in 2-year-olds: A bivariate twinanalysis. Developmental Science, 4, 195–208.

Rapin, I. (1996). Historical data. In I. Rapin (Ed.), Preschool children with inadequatecommunication: Developmental language disorder, autism, low IQ. Clinics in developmentalmedicine no. 139. (pp. 58–97). London: Mac Keith Press.

Ronald, A., Happe, F., Bolton, P., Butcher, L. M., Price, T., Wheelwright, S., et al. (2006).Genetic heterogeneity between the three components of the autism spectrum: A twinstudy. Journal of the American Academy of Child and Adolescent Psychiatry, 45, 691–699.

Ronald, A., Happe, F., & Plomin, R. (2005). The genetic relationship between individualdifferences in social and nonsocial behaviours characteristic of autism. DevelopmentalScience, 8, 444–458.

Ronald, A., Happe, F., Price, T. S., Baron-Cohen, S., & Plomin, R. (2006). Phenotypic andgenetic overlap between autistic traits at the extremes of the general population. Journalof the American Academy of Child and Adolescent Psychiatry, 45, 1206–1214.

Rutter, M. (1967). Concepts of autism: A review of research. Journal of Child Psychology andPsychiatry, 9, 1–25.

Rutter, M. (2005). Aetiology of autism: Findings and questions. Journal of IntellectualDisability Research, 49, 231–238.

Scott, F. J., Baron-Cohen, S., Bolton, P., & Brayne, C. (2002). The CAST (childhood aspergersyndrome test): Preliminary development of a UK screen for mainstream primary-school-age children. Autism, 6, 9–31.

Spinath, F. M., Price, T., Dale, P. S., & Plomin, R. (2004). The genetic and environmentalorigins of language disability and ability. Child Development, 75, 445–454.

Stevenson, J., Pennington, B. F., Gilger, J. W., DeFries, J. C., & Gillis, J. J. (1993).Hyperactivity and spelling disability: Testing for shared genetic aetiology. Journal ofChild Psychology and Psychiatry, 34, 1137–1152.

Trouton, A., Spinath, F. M., & Plomin, R. (2002). Twins early development study (TEDS): Amultivariate, longitudinal genetic investigation of language, cognition and behaviorproblems in childhood. Twin Research, 5, 444–448.

Williams, J., Scott, F. J., Stott, C. M., Allison, C., Bolton, P., Baron-Cohen, S., et al. (2005). TheCAST (childhood asperger syndrome test): Test accuracy. Autism, 9, 45–68.

World Health Organization. (1994). International classification of diseases. (ICD 10 ed.).Geneva: World Health Organization.



Documents

Developmental path between language and autistic-like impairments: a twin study