Independent pathway model. Dataare shown as components of variance (95% confidence interval). The fit ofan ADE model is as follows: χ224 goodness of fit= 14.52, P = .93, AIC = −33.48. The fit ofan ACE model is as follows: χ224 goodness of fit= 19.23, P = .74, AIC = −28.77. A indicatesadditive genetic effects; AIC, Akaike Information Criterion; C in ACE, sharedenvironmental effects; D, nonadditive or dominance genetic effects; E, uniqueenvironment (which includes measurement error); Pt, parent report; S, adolescentself-report; and T, teacher report.
Common pathway model. Data areshown as components of variance (95% confidence interval). The variance ofthe latent variable, "P," was constrained to equal 1.00. The fit of an ADEmodel is as follows: χ228 goodness of fit = 16.81, P= .95, AIC = −39.19. The fit of an ACE model isas follows: χ228 goodness of fit = 19.52, P= .88, AIC = −36.48. A indicates additive geneticeffects; AIC, Akaike Information Criterion; C in ACE, shared environmentaleffects; D, nonadditive or dominance genetic effects; E, unique environment(which includes measurement error); P, conduct problems' phenotype commonto all raters; Pt, parent report; S, adolescent self-report; and T, teacherreport.
Simple multiple rater model. Dataare shown as components of variance (95% confidence interval). The varianceof the latent variable, "P," was constrained to equal 1.00. The fit of anADE model is as follows: χ234 goodness of fit =58.48, P = .006, AIC = −9.52. The fit of anACE model is as follows: χ234 goodness of fit =58.62, P = .005, AIC = −9.38. A indicates additivegenetic effects; AIC, Akaike Information Criterion; C in ACE, shared environmentaleffects; D, nonadditive or dominance genetic effects; E, unique environment(which includes measurement error); EPt, ES, and ET, error for parent rating, self-rating, and teacher rating, respectively;P, conduct problems' phenotype common to all raters; Pt, parent report; S,adolescent self-report; and T, teacher report.
Scourfield J, Van den Bree M, Martin N, McGuffin P. Conduct Problems in Children and AdolescentsA Twin Study. Arch Gen Psychiatry. 2004;61(5):489–496. doi:10.1001/archpsyc.61.5.489
Evidence supports a genetic influence on conduct problems as a continuous
measure of behavior and as a diagnostic category. However, there is a lack
of studies using a genetically informative design combined with several different
informants and different settings.
To examine genetic and environmental influences on conduct problems
rated by parent and teacher reports and self-reports and to determine whether
their ratings reflect a common underlying phenotype.
A twin study design was used to examine conduct problem scores from
ratings by teachers, parents, and twins themselves.
Twins aged 5 to 17 years participating in the Cardiff Study of All Wales
and North England Twins (CaStANET) project.
Main Outcome Measures
Conduct problem scale from the Strengths and Difficulties Questionnaire.
Conduct problem scores were significantly heritable based on parent
and teacher reports and self-reports. Combining data from all 3 informants
showed that they are rating a common underlying phenotype of pervasive conduct
problems that is entirely genetic, while teacher ratings show separate genetic
influences that are not shared with other raters.
Conduct problems are significantly heritable based on parent and teacher
reports and self-reports, and are also influenced by environmental effects
that impinge uniquely on children from the same family. There is a cross-situational
conduct problems' phenotype, underlying the behavior measured by all informants,
that is wholly genetic in origin. No significant influence of shared environmental
effects was found.
In a recent government survey1 of themental health of children in the United Kingdom, conduct disorder emergedas the most common child psychiatric disorder and one of the most frequentreasons for referral to specialist services. Children with conduct disorderare more likely to have comorbid substance abuse, depression, and anxiety,2,3 and in adult life are much more likelyto develop antisocial personality disorder and other psychiatric disorders.4,5 Follow-up studies have found high ratesof premature death, often from violent causes,6 andup to 46% of those who attempt suicide have had conduct disorder.7 Even those without a frank psychiatric disorder inadulthood have pervasive social difficulties compared with their peers.8
There has been increasing awareness in recent years that genetic influenceshave a role in the development of conduct problems, whether measured as acontinuum of behavior or as a diagnostic category. A recent meta-analysis9 reported a heritability of 41% (additive and nonadditivegenetic effects) for antisocial behavior measured in various ways. It is unlikelythat there will turn out to be genes for these behaviors, in the sense thata specific genetic constellation would provide necessary and sufficient cause.Rather, the path from genes to behavior is likely to be a complex one involvingenvironmental stresses and individual differences in liability. There is evidencesupporting a genetic influence on conduct disorder as a diagnostic category10,11 and on continuous measures of conductproblems and antisocial behavior,12- 23 althoughfindings vary among studies.
Discrepant findings are likely to be the result of different samplecharacteristics and measurement strategies and the fact that these behaviorsprobably represent a heterogeneous group.24,25
Because children's behavior can vary in different settings, the mostvalid measurements are those that include data from more than one informantand from more than one setting. Behavioral genetic approaches can indicatethe extent to which different informants are rating the same substrate ofbehavior and the degree of unique perspective that different informants contribute.The present investigation examines genetic and environmental influences ona dimensional measure of conduct problems using parent and teacher reportsand self-reports from the Strengths and Difficulties Questionnaire (SDQ),26 a measure widely used in the United Kingdom clinicallyand epidemiologically.
The Cardiff Study of All Wales and North England Twins (CaStANET) isa population-based twin register of all twin births in Wales and greater Manchester,England. This investigation is based on the subsample of families from southWales: 1109 twin pairs aged 5 to 17 years, consisting of all twin births betweenJanuary 1, 1980, and December 31, 1991, in the former counties of Mid andSouth Glamorgan. The methods used to identify and trace twins have been describedelsewhere.27
Data collection was by mail. Parents were asked to complete a packetof questionnaires about their twins, including the SDQ.26 Threereminders were sent, and telephone reminders were used when numbers couldbe traced. Parental consent was obtained to contact a teacher, who was thenasked to complete the teacher version of the SDQ. Self-report versions weresent to twins 11 years and older. The SDQ is a brief behavioral screeningquestionnaire that can be completed in 5 minutes. Its brevity and its inclusionof strengths as well as difficulties make it particularly suitable for usein general population samples. It includes a continuous measure of behaviorproblems, the Conduct Problems Scale. The following 5 items contribute tothe Conduct Problems Scale: "Often has temper tantrums or hot tempers; generallyobedient, usually does what adults request; often fights with other childrenor bullies them; often lies or cheats; steals from home, school or elsewhere."Parents were asked to report behavior observed in the past 6 months, and itemsare scored "not true" (0), "somewhat true" (1), or "certainly true" (2), withreverse scoring when appropriate. This conduct scale correlates highly withthe conduct/externalizing scales of the longer-established Rutter Scales andthe Achenbach Child Behavior Checklist.26,28 Zygositywas assigned using a twin similarity questionnaire that is more than 90% accuratein distinguishing monozygotic (MZ) from dizygotic (DZ) twins.29,30
The influence of latent genetic factors (the sum of effects of individualalleles at a locus), shared environmental factors, and unique environmentalfactors on conduct problem scores from parents, teachers, and adolescentswere first estimated in univariate models. Because MZ twins share 100% oftheir genes and DZ twins share, on average, 50%, additive genetic effectsare reflected in a pattern of correlations by which the MZ correlation istwice the DZ correlation. When DZ correlations are less than half the MZ correlations,nonadditive or dominance genetic effects can be examined. This refers to interactioneffects between alleles at the same locus. Dizygotic twins share only 25%of nonadditive or dominance genetic effects, so these will be reflected ina DZ correlation that is less than half the MZ correlation. Alternatively,when DZ correlations are less than half the MZ correlation, a sibling interactionmodel can be tested; this model examines genetic and environmental influencesin the presence of sibling interactions. Shared environmental influences arethose that tend to make twins more similar, and will inflate DZ correlationsrelative to MZ correlations, while unique environmental influences are thosethat make twins different from one another. They are seen in the extent towhich the MZ correlation is less than 1.0. The overall acceptability of amodel is evaluated using a χ2 test (assessing goodness of fit)and Akaike Information Criterion (assessing parsimony), which has a more negativevalue the better the fit. The significance of individual parameters withina model is assessed by constraining their effects to 0 and then calculatingthe change in χ2 for the model. A significant deteriorationin the goodness of fit implies a significant influence from the constrainedparameter. A commercially available software program (SPSS for Windows) wasused31 to obtain descriptive statistics andthe covariance matrices for genetic analysis. A software package (Mx)32 was used for genetic model fitting. The methods andprinciples of the model-fitting procedure are described in detail by Nealeand Cardon.33 Before model fitting, scoreswere transformed to approximate normality using log(score + 1).
Data from all informants were subsequently examined together in theadolescent sample, in which 3 sources of data were potentially available.For these analyses, scores were first standardized for age and sex so thatany influence of age or sex would not bias parameter estimates. To maximizethe available data, a raw data approach in the software package (Mx) was usedat this stage. This removes the need to summarize data in covariance matricesacross all informants, which would result in the exclusion of any twin pairsin which data from one informant were missing. This raw maximum likelihoodapproach generates a value (−2 × log likelihood), which is thencompared with a saturated model to generate a goodness-of-fit χ2.32
First, an independent pathway model was tested (Figure 1). This tests the extent to which genetic and environmentalinfluences are shared across informants, while also allowing for discretegenetic and environmental influences on ratings from each informant. Next,a more stringent model, described as a psychometric or common pathway model,33 was tested (Figure2). This type of model implies that parents, teachers, and adolescentsare all rating the same underlying conduct problems' phenotype (with contributinggenetic and environmental influences), but allows for rater-specific estimatesof genetic and environmental influences. The arrows from the latent phenotype,"P," represent the factor loadings of the latent phenotype on each of theobserved measures. This type of model was chosen because it best representsthe sources of data herein, with each rater basing his or her appraisal ondifferent circumstances (ie, teachers rating behavior in school, parents ratingbehavior primarily at home, and adolescents rating behavior across settings).Finally, a simple multiple rater model was tested (Figure 3). This again allows the 3 observed measures (parent, teacher,and adolescent ratings) to be influenced by a latent conduct problems' phenotype.It also allows for variation in the extent to which different raters contributeto the latent phenotype and for variation in the error associated with eachrater, but it does not allow for discrete genetic and shared environmentalinfluences on each observed measure.
Replies were received from 682 families, a response rate of 61.5% froman initial sample of 1109. There were 12 families in which either the parentor one of the twins failed to return a questionnaire, leaving 670 familieswho returned parent- and twin-rated measures. Of these families, permissionto contact a teacher was given by 561 (83.7%), and 443 teachers replied (79.0%of those families giving consent). Adolescent self-report packs were sentto 570 twin pairs aged 11 to 17 years, and were returned by 286 twin pairs,a response rate of 50.2%.
The age of the twins ranged from 5 to 17 years, with a mean for bothsexes of 10.6 years. There were 14 pairs in which zygosity could not be determinedbecause of missing or inconsistent data.
A comparison of responders and nonresponders using Townsend Scores,34 which are based on National Census data for eachelectoral ward, revealed no significant (P = .10)sociodemographic differences between the 2 groups, suggesting that the sampleis representative of the local population in sociodemographic terms. Furthermore,a comparison with population norms for the SDQ showed that mean conduct problemscores for this sample are not significantly different from those in the generalpopulation. For twins, the mean (95% confidence interval) scores were as follows:parent report, 1.83 (1.39-2.27); teacher report, 0.98 (0.51-1.45); and self-report,2.41 (1.85-2.97). For singletons, the mean (95% confidence interval) scoreswere as follows: parent report, 1.60 (1.35-1.85); teacher report, 0.90 (0.64-1.16);and self-report, 2.20 (1.88-2.52).
Mean conduct problem scores, standard deviations, correlations, andfinal sample sizes of twin pairs with complete data are shown in Table 1. Conduct problem scores were higherin male than female twins, in keeping with the existing literature, for parentand teacher reports and self-reports (parent-rated t1300 = 3.3, P<.001; teacher-rated t831 = 3.4, P<.001;and adolescent-rated t530 = 2.6, P = .01). There were no significant differences in meanscores between MZ and DZ twins for any of the measures (parent-rated t1273 = 0.65, P = .50;teacher-rated t809 = 1.6, P = .10; and adolescent-rated t519 =−0.2, P = .90). The F ratio of variance inMZ/DZ pairs was not significant for any of the measures (parent-rated F =1.4, P = .20; teacher-rated F = 2.4, P = .20; and adolescent-rated F = .02, P =.90). The correlation between conduct problem scores and age was significantfor parent report (Spearman ρ = −0.09, P =.02), but not for other informants. The distribution of conduct scores waspositively skewed so scores were transformed to approximate normality beforemodel fitting using log(score + 1). After transformation, skewness and kurtosisindices were between −1.0 and 1.0, implying that not much distortionis to be expected.35
Table 2 presents the resultsof univariate model fitting for the 3 informants. For results from differentinformants to be comparable (because self-reports are only available for 11-to 17-year-old subjects), results for children (those aged <11 years) andadolescents (those aged 11-17 years) are presented separately.
Complete data were available from 621 twin pairs for these analyses.The pattern of correlations (0.63 MZ/0.16 DZ in male pairs and 0.53 MZ/0.33DZ in female pairs) suggested genetic influence on the variance of scores,and this was supported by model fitting. The best-fitting models show thatmore than half the variance was accounted for by additive genetic influences,while the remainder was influenced by unique environmental factors. Sharedenvironmental influences and nonadditive genetic influences were insignificant.
The same model-fitting procedure was applied to teacher-rated SDQ conductproblems. There were 39 twin pairs in which the twins were rated by differentteachers, and these were excluded from the present analyses. Complete datawere available from 387 twin pairs. The pattern of correlations (0.69 MZ/0.13DZ in male pairs and0.75 MZ/0.26 DZ in female pairs) suggested genetic influences.The results of model fitting are summarized in Table 2, and show that, in adolescents, almost three quarters ofthe variance in scores was attributable to genetic influences, with the remainderaccounted for by unique environmental factors. In children, the same proportionof variance was due to genetic influence, but in this case the genetic influencewas nonadditive. A sibling interaction model was also tested, but fitted thedata less well than the nonadditive or dominance genetic effects and uniqueenvironmental effects model. Again, the influence of shared environment wasinsignificant.
Complete data were available from 254 twin pairs for these analyses.The pattern of correlations (0.39 MZ/0.05 DZ in male pairs and 0.53 MZ/0.17DZ in female pairs) suggested genetic influence. The same model-fitting procedurewas applied, and results are summarized in Table 2. Of the variance in scores, 35% was accounted for by additivegenetic influences, with the remainder due to influences of the twins' uniqueenvironment. Again, the influences of the shared environment and the nonadditivegenetic effects were insignificant.
Results for the independent pathway model are shown in Figure 1. A model including dominance genetic effects is shown,but a model including additive genetic effects, shared environmental effects,and unique environmental effects was also tested; fit statistics for bothmodels are given in the legend to Figure 1. The model including additive genetic effects, nonadditive or dominancegenetic effects, and unique environmental effects provides a slightly betterfit.
Results for the common pathway model are shown in Figure 2. Again, a model including dominance genetic effects isshown, and fit statistics for the model including additive genetic effects,nonadditive or dominance genetic effects, and unique environmental effectsand the model including additive genetic effects, shared environmental effects,and unique environmental effects are given in the legend to Figure 2; the model including additive genetic effects, nonadditiveor dominance genetic effects, and unique environmental effects gives a slightlybetter fit with the data. A model including additive genetic effects, sharedenvironmental effects, and unique environmental effects estimated all sharedenvironmental influences at 0. The common pathway model, which tests whethera shared underlying phenotype accounts for similarities in ratings from differentobservers, is more psychologically meaningful and a more stringent representationof the data. It can be compared with the more general independent pathwaymodel (change in χ24 = 2.29, P = .68), revealing a nonsignificant change of fit. This suggests thatthe common pathway model is a better explanation of the data than the independentpathway model. The common pathway model indicated that a wholly heritablecommon phenotype (genetic influence = additive genetic effects + nonadditiveor dominance genetic effects = 1.00) underlies cross-situational conduct problemsthat are rated by different informants in different settings. In addition,the common pathway model suggests discrete genetic effects on teacher ratings(additive and nonadditive) and on self-ratings (nonadditive). When these discreteeffects were constrained to 0 and the fit compared with a full model includingadditive genetic effects, nonadditive or dominance genetic effects, and uniqueenvironmental effects, the genetic effects on self-ratings were not significant(change in χ22 = 5.37, P =.07), but discrete genetic effects on teacher ratings were significant (changein χ22 = 16.02, P<.001).
Results for the simple multiple rater model are shown in Figure 3. This model provides a poor fit with the data. The multiplerater model is more stringent than the other 2 models because it offers fewerparameters and, in particular, does not allow for any separate influences,other than error, on each measure. This restriction results in a significantdeterioration of fit compared with the common pathway model (change in χ26 = 41.67, P<.001), indicatingthat the presence of discrete rater influences provides a better explanationof the data.
The results of this investigation lend further support to existing evidencesuggesting that conduct problems in children and adolescents have a heritablecomponent. In this community sample of twins, a substantial genetic influencehas been shown using a dimensional measure that is widely used in clinicalpractice and epidemiology.26 Parent, teacher,and adolescent reports all provided significant heritability estimates. Therewas no significant effect of the shared environment in any of these models,and only in the teacher-report scores for adolescents were the confidenceintervals fairly wide around this estimate (0%-32%), suggesting that a largersample might have detected a significant effect.
When conduct scores from all informants were combined in a single model,the best results were achieved for a model that included an underlying phenotypethat was common to all informants and separate genetic and environmental influencesacting on each measure. This suggested that all informants were, to a considerableextent, rating a common underlying phenotype that was wholly influenced bygenetic factors, additive and nonadditive. This highly genetic phenotype refersonly to aspects of behavior that are common to all informants and, therefore,pervasive across settings. However, ratings by teachers also showed a significantspecific genetic influence, suggesting that different genes may be influencingthe behavior that teachers were rating. Conduct problems occurring in schoolseem to be different from those that are pervasive across settings. This isa phenomenon that has also been observed in ratings of antisocial behaviorin younger children36 and in childhood hyperactivity.37,38 It supports the value of SDQ datafrom teachers because they are clearly contributing a measure of behaviorthat is not tapped by parent reports or self-reports. It also has implicationsfor phenotype definition in molecular genetic research, implying that differentgenes are influencing measures from different informants.
Another approach to data from multiple informants is to examine theextent of rater bias, but this approach is better suited to multiple ratersof behavior in one setting (eg, mothers and fathers rating behavior at home)and does not easily generalize to multiple raters in different settings. Giventhat children may behave differently at home and at school, it is possiblethat parents and teachers provide valid ratings of behavior in their respectivedomains, and to what extent this represents rater bias or true phenotypicdifferences is difficult to disentangle.
The genetic literature on conduct/externalizing problems in childrenand adolescents is unusual insofar as more than any other behavioral trait,it has shown a significant effect of shared environment.24 However,the lack of shared environmental influence shown herein is a far from uniquefinding. These results are in agreement with those of Rowe,14 whoused a self-report delinquency scale in adolescents; Schmitz et al,18 who used the delinquency subscale of the Child BehaviorChecklist in 4- to 18-year-old subjects; McGuffin and Thapar,20 whoused the Olweus adolescent self-report scale; Slutske et al,11 whomade retrospective DSM-III-R conduct disorder diagnosesin adults; Eaves et al,10 who used interview-baseddiagnoses of conduct disorder and oppositional defiant disorder in childrenand adolescents; and Miles et al,23 who usedinterview-derived symptom counts. The only other twin study36 weare aware of that has used multiple informants in different settings alsofound a highly heritable shared latent phenotype with no significant influencefrom the shared environment. Arsenault et al36 useddata from mothers, teachers, observers, and children themselves, and showedthat heritability was higher (82%) for a shared latent phenotype than forcorresponding single measures of antisocial behavior, as found in this investigation.In their sample of 5-year-old twins, a common pathway model was the best explanationof the data, as found herein. Their results and ours support the notion thatconduct problems that are pervasive across settings are a highly genetic phenotypein young children and adolescents. This is in keeping with a theory of antisocialdevelopment39 that suggests that pervasiveantisocial behavior that develops in childhood is a result of heritable neurodevelopmentalabnormalities and is likely to persist into adolescence and adulthood, althoughlongitudinal cross-situational data are required to examine this questionfurther.
The lack of significant shared environmental influences in this investigationwarrants some consideration. When shared environmental influences depend ongenotype (as in gene × environment interactions), then their effectsare subsumed in the estimate of genetic influences in twin analyses. Thereis some evidence40,41 suggestingboth gene × environment interaction and correlation in conduct problemsand related behaviors, and this may have contributed to the lack of any significantshared environmental influence in the present investigation.
The estimates for unique environmental effects represent those environmentalinfluences that are not shared and tend to make individuals different fromone another, while also including measurement error. Unique environmentalinfluences emerged for each informant, but not for the shared underlying phenotype.This is in keeping with another twin study19 thatused data from mothers, fathers, and twins, and found that the influence ofthe nonshared environment reduced and disappeared when comparing single withmultiple-rater models. This observation is likely to reflect a reduction inerror variance when data from multiple informants are combined. For observedphenotypes, nonshared environment estimates include all measurement error,but with a latent phenotype, as used in the multiple-rater models herein,much of this error is taken up elsewhere. Although the present investigationdoes not measure which specific environmental influences are at work for eachinformant, a significant role for unique environmental influences is in keepingwith other literature. There is evidence42,43 suggestingthat children, while growing up in the same family, can nevertheless experiencequite different environments, and Reiss et al44 haveshown a strong association between child-specific parental conflict withinfamilies and children's behavior problems.
The finding that conduct problems are highly heritable is not in anyway a suggestion that these behaviors are wholly genetically determined and,therefore, not amenable to intervention. Environmental influences of the kindthat tend to make twins different from one another had a significant influencein this investigation, and appreciating genetic influence is only a smallstep in understanding the complex pathway from genes to behavior. Complexgene-environment correlation and interaction suggests that individuals witha genetic propensity to certain behaviors will seek out environments thatexacerbate the behavior. Nature and nurture are inextricably entwined.
Three limitations of the present study warrant discussion. First, thefindings presented are based on questionnaire data; diagnostic interviewsare obviously preferable, but are expensive and were not a feasible optionfor this investigation. Second, another limitation is the response rate: 61.5%of a population-based sample of 1109 twin pairs. However, comparison of responderswith nonresponders revealed no significant differences in sociodemographicmeasures reflecting neighborhood social deprivation. Because these are thetype of environmental stresses associated with conduct problems, it was assumedthat the sample adequately represents the general population and there wasno systematic bias with regard to conduct problems. In addition, a comparisonof mean conduct problem scores with population norms revealed no significantdifference. However, we cannot completely discount the possibility of a responsebias. Third, the main measure of conduct problems26,28 usedis a relatively new and brief measure, with 5 items that generate the ConductProblems Scale. Despite this, its brevity and its inclusion of strengths aswell as difficulties were seen as an advantage in a general population postalsurvey of this sort.
Corresponding author: Jane Scourfield, PhD, MRCPsych, Departmentof Psychological Medicine, University of Wales College of Medicine, FourthFloor, Heath Park, Cardiff CF14 4XN, Wales (e-mail: email@example.com).
Submitted for publication August 11, 2003; final revision received December30, 2003; accepted January 18, 2004.
This study was supported by the Medical Research Council, London, England.