Journal of Personality and Social Psychology1994, Vol. 66, No. 4,722-730 Copyright 1994 by the American Psychological Association Inc

0022-3514/94/S3.00

A Developmental Genetic Analysis of Adult Personality: Extraversionand Neuroticism From 18 to 59 Years of Age

Richard J. Viken, Richard J. Rose, Jaakko Kaprio, and Markku Koskenvuo

Developmental genetic analyses were conducted on Extraversion (E) and Neuroticism (N) scalescores from nearly 15,000 male and female Finnish twins, ages 18-53 at baseline, who were testedon 2 occasions, 6 years apart. Significant genetic effects on both traits were found, at all ages, in menand women, on each measurement occasion. For E, heritability was invariant across sex but de-creased from late adolescence to the late 20s, with a smaller additional decrease at about 50 years ofage. Heritabiiity for N also decreased from late adolescence to late 20s and remained stable thereaf-ter. For all ages after the early 20s, heritability of N was significantly higher among women. Meansfor E and N were sex-dependent and, apparently, influenced by cohort and time of assessment, aswell as by age. There was little evidence of new genetic contributions to individual differences afterage 30; in contrast, significant new environmental effects emerged at every age.

No dimensions of adult personality have been subjected togreater study than Extraversion and Neuroticism, the "BigTwo" in current five-factor models of personality. That fact isapparent in the research literature on twin studies: > 50,000 ad-olescent and adult twins have completed self-report measuresof Extraversion and Neuroticism, and substantial consistencyemerges from their data (Loehlin, 1992). There is compellingevidence of a strong genetic effect on both Extraversion andNeuroticism, from samples of both twins reared together andtwins reared apart (Eaves, Eysenck, & Martin, 1989; Pedersen,Plomin, McClearn, & Friberg, 1988; Rose, 1988; Rose, Kosken-vuo, Kaprio, Sarna, & Langinvainio, 1988; Tellegen, Lykken,Bouchard, Wilcox, Segal, & Rich, 1988). The genetic effects areconsistent across gender and culture and are found both in self-reports and in ratings by others (Heath, Neale, Kessler, Eaves,& Kendler, 1992). Common environmental effects on Extraver-sion and Neuroticism are less evident in twin data, althougheffects of shared experience can be found in contrasts of geneti-cally identical twin pairs who differ in frequency of their socialcontact; cotwins in more contact are more alike (Rose & Ka-prio, 1988). But effects of social contact are modest and cannotexplain consistent differences in behavioral resemblance be-tween monozygotic and dizygotic cotwins (Rose et al., 1988).

Richard J. Viken and Richard J. Rose, Department of Psychology,Indiana University Bloomington; Jaakko Kaprio, Department of PublicHealth, University of Helsinki, Helsinki, Finland; Markku Koskenvuo, De-partment of Public Health, University of Turku, Turku, Finland-

Data analyses and article preparation were supported by Grants AA-08315 and AA-07611 from the National Institute on Alcohol Abuseand Alcoholism (NIAAA). Formation of the older Finnish Twin Cohortand data collection in 1975 and 1981 were supported by the Council forTobacco Research. Richard J. Rose was supported by NIAAA ResearchScientist Award AA-00145, and Jaakko Kaprio was supported by theFinnish Academy of Science. We thank Kauko Heikkilla and Eila Voi-pio for their skilled assistance.

Correspondence concerning this article should be addressed to Rich-ard J. Viken, Department of Psychology, Indiana University, Blooming-ton, Indiana 47405.

Thus, twin studies of Extraversion and Neuroticism, the largesttwin studies in the behavior-genetic literature, affirmatively an-swer the basic question: Do genes contribute to individual vari-ation in personality? But more important questions remain,and this study addresses several: Does the genetic contributionto Extraversion and Neuroticism vary across ages of adulthood?Do genes contribute to consistency and change in Extraversionand Neuroticism during adult development? If so, what is therelative magnitude of genetic contributions to age-to-age con-sistency? Is the genetic contribution to age-to-age stability mod-ulated by age? By gender? Is the genetic stability different forExtraversion than for Neuroticism?

Despite many twin studies of Extraversion and Neuroticism,few have investigated age-related changes in genetic and envi-ronmental effects. Eaves et al. (1989) reanalyzed Swedish twindata reported by Floderus-Myrhed, Pedersen, and Rasmuson(1980), to evaluate heterogeneity of genetic effects on Extraver-sion and Neuroticism for three age cohorts; genetic variance forboth traits was highest in the youngest twin pairs. In a longitu-dinal follow-up (average interval 10 years) of twins who werelate adolescents or young adults at baseline, McGue, Bacon,and Lykken (1993) assessed age modulation of genetic effectson the Multidimensional Personality Questionnaire (MPQ) andfound a consistent tendency for heritabilities to be greater atbaseline, when the twins were younger.

Information on gender differences in genetic effects on per-sonality is also very limited, in part because large samples arenecessary to establish such effects. In their reanalysis of theFloderus-Myrhed et al. (1980) data from the Swedish Twin Reg-istry, Eaves et al. (1989) found evidence for greater genetic con-tributions to Extraversion and Neuroticism for female subjects.Similarly, Martin and Jardine (1986) reported significantlyhigher neuroticism heritability for women, but no significantgenetic gender modulation for Extraversion. Thus, there may begender differences in genetic effects on personality, particularlyfor Neuroticism.

Many studies have established the relative stability of individ-ual differences in Extraversion and Neuroticism in adulthood

722

PERSONALITY FROM 18 TO 59 YEARS 723

(McCrae & Costa, 1990). Far less is known about age-relatedchanges in stability, although some evidence suggests greaterstability among older age cohorts (Finn, 1986). Finally, very lit-tle is known about the age-to-age stability of genetic and envi-ronmental effects on personality throughout adulthood. Genesare often thought of as static influences that remain stablethroughout life, but there may, in fact, be substantial changes ingenetic effects over time. Age-specific genetic variation has beenclearly identified during childhood (Cardon, Fulker, DeFries, &Plomin, 1992), but new genetic variance may emerge at anypoint in the life span. With genetically informative longitudinaldata, it is possible to estimate the stability of genetic and envi-ronmental effects, as reflected in the genetic (rg) and environ-mental (re) correlations across time. Genetic contributions toage-to-age stability are a function of both heritability (propor-tion of variance explained by genetic variation at any time indevelopment) and rg (the genetic correlation across periods ofdevelopment). Stable genetic effects can contribute little to sta-bility if heritabilities are low; conversely, if genetic stability islow, genes can contribute little to phenotypic stability, evenwhen heritabilities are high. Thus, estimation of rg and re allowsus to evaluate the extent to which observed phenotypic stabilityis determined by genetic and environmental effects. Despite theobvious interest and potential importance of such analyses,there is, to our knowledge, only one report on environmentaland genetic stabilities for adult personality, and that report islimited to young adulthood. In a longitudinal follow-up (aver-age interval 10 years) of late adolescents and young adults,McGue et al. (1993) found significant genetic stabilities for thehigher order MPQ factors of positive affectivity (.81) and nega-tive affectivity (.72), as well as other scales in the MPQ. Themagnitude of these estimates suggests that genetic effects weresubstantial contributors to personality consistency, but suggests,too, that new genetic variance, not present at baseline, was alsocontributing to personality change. Environmental stabilitieswere lower, suggesting that genes were the primary contributorto personality consistency during this developmental period.

As this brief review indicates, little is known about age-re-lated changes in genetic or environmental effects on personality,or about age-related change in the stability of genetic and envi-ronmental effects. In part, this is because the identification ofage-related effects is no simple task. Measures taken to evaluateage effects will always be subject to age effects, birth cohort (gen-erational) effects, and time of assessment (period) effects. Moststudies investigating age effects in personality have used cross-sectional designs. Cross-sectional studies confound age and co-hort effects, such that observed differences may be due to truedevelopmental effects or to differences in the historical periodswithin which each cohort lived. Schaie and Willis (1991) re-cently demonstrated how easily cohort effects on personalitycan be mistaken for age effects when researchers rely on cross-sectional designs. Longitudinal studies, which follow a cohortthrough time, offer an informative alternative. However, single-cohort longitudinal studies require a lifetime to provide data onlife-span development, and they have the disadvantage of con-founding age effects with effects of time of assessment.

A variety of more sophisticated developmental designs havebeen proposed to facilitate differentiating age effects from otherinfluences on stability and change (Schaie & Hertzog, 1982).

One of these is the cross-sequential design, in which multiplecohorts, differing in age at baseline, are assessed during two ormore time points. Because multiple age cohorts provide a start-ing place, longitudinal change can be observed over a muchlarger age range, without waiting for a single cohort to progressthrough the life span. More importantly, subjects can bematched on any one of the age, cohort, and time variables whileleaving them free to vary on the others. The cross-sequentialdesign is not without its own disadvantages (e.g., long-term"sleeper" effects of early conditions cannot be identified be-cause of the short follow-up), and no practical developmentalmodel provides a perfect solution. Age, cohort, and time arealways confounded. Each variable is a linear composite of theothers (e.g., age = time - cohort), so that any effect that appearsto be due to one of them can always be represented as an in-teraction of the other two. However, when one cross-sequentialeffect provides a simple and theoretically coherent explanationof data and the alternative is an improbable, complex interac-tion of the other two, there is support for the simpler model. Inthe absence of solutions that are both perfect and practical, it isnecessary to use the best design available for the data and com-pare results across multiple studies using multiple strategies.

In an earlier report, we (Rose et al., 1988) analyzed geneticand environmental effects on Extraversion and Neuroticismamong > 14,000 adult Finnish twins surveyed by postal ques-tionnaire in 1981. These twins, and their 1981 survey data, areincluded in the present analysis. The Finnish Twin Cohort hadalso been surveyed 6 years earlier, in 1975, with a questionnairethat included the same measures of Extraversion and Neuroti-cism. We now report results of longitudinal analyses of the twowaves of data collection on this very large, population-based co-hort of adult twins. These informative data permit us to extendearlier longitudinal twin studies of personality (Dworkin,Burke, Maher, & Gottesman, 1976, 1977; Pogue-Geile & Rose,1985) that, limited by small samples and brief follow-up, coulddemonstrate only that genes contribute to patterns of continuityand change over time. With a larger, more informative popula-tion-based twin registry, we now assess age and gender modula-tion of genetic and environmental influences on age-to-age sta-bilities over four decades of adult development.

MethodSample

The sample consisted of 6,828 twin brothers and 8,104 twin sisterswho were aged 18-53 years at the time of baseline testing in 1975. Thetwins are members of the (older) Finnish Twin Cohort, a population-based twin registry compiled from the Central Population Registry ofFinland (Kaprio, Sarna, Koskenvuo, & Rantasalo, 1978). Subjects re-sponded to a medical and psychosocial postal questionnaire in 1975 andagain in 1981. Zygosity of the twin pairs was determined by a validatedquestionnaire assessment of perceived similarity and "confusability" byothers. Zygosity of approximately 93% of the twin pairs could be classi-fied by the questionnaire items, with an expected error rate below 2%,confirmed by polymorphic blood markers (Sarna & Kaprio, 1980). Ourability to access updated information from the population registry, to-gether with extremely high compliance of Finnish twins to question-naire requests (Kaprio, Koskenvuo, Artimo, Sarna, & Rantasalo,1979), negligible migration, and very low attrition between 1975 and1981, provided us with an unusually representative longitudinal sam-

724 VIKEN, ROSE, KAPRIO, AND KOSKENVUO

AGE

12

<-> 56

18-23

1975

24-29

19811975

30-35

19811975

36-41

19811975

42-47

19811975

48-53

19811975

54-59

1981

Figure 1. Cross-sequential structure of the sample. Subjects can be matched by birth cohort (horizontallyacross age and time of assessment), age (vertically across cohort and time of assessment), and time of as-sessment (diagonally across age and cohort).

pie. Of pairs responding in 1975, only 1.8% were lost to the 1981 follow-up, due to death or inability to locate a current residential address. Ofthe remaining twins, pairwise response rates in 1981 were 87.5% formen and 96.7% for women. A preliminary analysis indicated that pairswithout complete 1981 data did not differ from pairs with completedata when compared on 1975 mean scores for Extraversion, r(7,618) =1.42, /J = ns, or Neuroticism, /(7,618) = -1.24, p = ns, or on additivegenetic, shared environment, and unshared environment estimates forExtraversion, x2(3) = 2.01, or Neuroticism, x2(3) = 2.88. Subjects notproviding complete data in 1981 were significantly older, ((7,618) =10.06, p<. 000.

MeasuresA short form of the Eysenck Personality Inventory developed by Flod-

erus (1974) and subsequently used in collaborative Swedish and Finnishtwin surveys was used to assess Extraversion and Neuroticism. Internalconsistency (alpha) was .73 for both the 10-item Extraversion scale andthe 9-item Neuroticism scale. Behavioral correlates of the short-formExtraversion and Neuroticism scales in Swedish and Finnish samplesoffer evidence of their construct validity; for example, Neuroticismscores are correlated with smoking, sedentary work and leisure habits,alcohol abuse, prescribed antihypertensive medication, self-reports ofimpaired sleep, and self-prescribed sleeping medications (Floderus,1974; Koskenvuo, Langinvainio, Kaprio, Rantasalo, & Sarna, 1979).

AnalysesTo evaluate the contributions of gender, age, birth cohort, and time of

assessment to the pattern of observed data, we organized the sample inthe form of a cross-sequential design as shown in Figure 1. Each row ofFigure 1 represents a 6-year birth cohort. Thus, the first cohort, theyoungest, comprises all subjects born 1952-1957. Moving horizontallyacross the rows, the figure illustrates that each cohort was assessed attwo time points, 1975 and 1981. The 6-year birth cohorts match the 6-year follow-up period, so that the age of each cohort at the 1981 follow-up is matched with the age of the next cohort at the 1975 baseline. Asan illustration, the first cohort was 24-29 years of age in 1981, whereasthe second cohort was 24-29 years of age in 1975. The seven resultingage groups are designated across the top of the figure. The sample wasalso divided by gender and zygosity (monozygotic [MZ] or dizygoticIDZ]). Thus, the basic data structure consisted of 24 groups (6 cohorts

X 2 genders X 2 zygosities), each of which was assessed for Extraversionand Neuroticism at two points in time.'

For each personality trait, genetic and environmental componentswere simultaneously fit to the 24-group covariance structure using max-imum likelihood estimation in LISREL (Joreskog & Sorbom, 1989).For each trait, we fit a general bivariate model (Heath, Neale, Hewitt,Eaves, & Fulker, 1989; Neale & Cardon, 1992), in which the covariancestructure of baseline and follow-up measures was attributed to a com-bination of additive genetic effects (A), common (shared) environmen-tal effects (C), and unshared environmental effects (E), or ACE, andrelationships between baseline and follow-up measures were attributedto longitudinal covariances of the ACE effects.

For both Extraversion and Neuroticism, a standard set of five modelswas estimated to evaluate the cross-sequential (CS) effects of age, co-hort, and time. The CS-invariant model assumed no heterogeneity ofACE effects across age, cohort, or time, and therefore constrained theseparameters to be equal across the 12 (6 cohorts X 2 times) cells in Figure1. The CS-invariant model would be true if there were no variation ingenetic or environmental effects associated with age, cohort, or time.The CS-free model assumed complete heterogeneity of ACE effectsacross the 12 cells and allowed all parameters to freely differ across allcells. The age model allowed differences among age groups, but con-strained estimates from age-matched subjects (subjects in the same col-umn of Figure 1) to be equal. The cohort model allowed differencesamong birth cohorts, but constrained estimates within each cohort tobe equal in 1975 and 1981. Finally, the time model allowed differencesacross the two occasions of assessment, but constrained the ACE esti-mates to be equal across cohorts.

The five models were compared by using chi-square difference tests.The goal of these comparisons was to identify the model that best rep-resented the data. Two criteria guided evaluation of alternative models:parsimony (reflected in a small number of free parameters and a corre-sponding high number of degrees of freedom) and fit (reflected in asmall chi-square value). When a more restrictive model (e.g., CS-invari-ant) is a nested submodel of a less restrictive model (e.g., CS-free), wecan make a direct comparison of fit by subtracting the CS-free degreesof freedom from the CS-invariant degrees of freedom and subtractingthe CS-free chi-square from the CS-invariant chi-square. The resultingdifference is itself distributed as chi-square. In general, a significant

' Correlation matrices for the 96 cohort-time-gender-zygosity vari-able units are available from us.

PERSONALITY FROM 18 TO 59 YEARS 725

difference chi-square indicates that the less restrictive model (e.g., theCS-free model) is preferred, because the added constraints of the morerestrictive model have resulted in a significant deterioration of fit. Forinstance, if constraining genetic variance to be equal at 1975 and 1981resulted in a significant deterioration in fit, then we would choose theless restrictive model allowing 1975 and 1981 parameters to differ. Ifthe difference chi-square is nonsignificant, the more restrictive model ispreferred, because it provides a more parsimonious representation ofthe data without resulting in a significant deterioration in fit.

In these analyses, we always compared the CS-free and CS-invariantmodels first. If this test indicated that there was significant heterogeneityin the data (i.e., CS-free was preferred), we evaluated the possibility thatthe age, cohort, or time models, which are less restrictive than the CS-invariant model but more restrictive than the CS-free model, could pro-vide a parsimonious compromise. For example, if birth cohort (eachrow of Figure 1) were the primary source of heterogeneity in ACE esti-mates, and if neither time of assessment nor longitudinal change in agewere important, then a model that allows estimates to differ across co-horts but constrains them to be invariant across time and longitudinalchange in age would provide the best fit to the data. If time of assessment(each diagonal of Figure 1) was the primary source of heterogeneity,then the best fitting model would allow estimates to differ across times(diagonals) but would assume invariance across cohorts and cross-sec-tional ages. If age (each column of Figure 1) was the primary source ofheterogeneity, then the best fitting model would allow age differences,but would assume invariance across cohort and time for age-matchedsubjects. As noted above, cross-sequential designs allow only a partialdecomposition of the age, cohort, and time effects. If the true model issome combination of or interaction of these factors, for instance, it willnot be possible to establish the correct model through formal statisticalcriteria. Nonetheless, it is possible to establish which models do not fitthe data, and therein lies the value of modeling.

To evaluate gender differences, we first estimated the five CS modelsunder an assumption of gender invariance, with ACE values for menand women constrained to be equal, and then under an assumption ofgender dependence, with male and female ACE values allowed to differ.Model comparisons by difference chi-square were used to identify sig-nificant gender differences. Because the older Finnish Cohort, fromwhich the current sample was obtained, includes only same-sex twinpairs, analysis of gender dependence was limited to genetic and environ-mental effects common to men and women.

If there are significant differences in phenotypic variances across gen-der, age, cohort, or time, the unstandardized estimates of ACE compo-nents may differ even if the proportions (standardized estimates) remaininvariant. For this reason, a second set of models was applied to allof the patterns of CS and gender constraints. The effect of these scalarmultiple models was to constrain the standardized estimates of ACE tobe equal across groups, while allowing the unstandardized estimates todiffer. Results of scalar multiple models are of little intrinsic interest.However, comparing the more restrictive scalar multiple model to theless restrictive variance component model of the same type makes itpossible to rule out the possibility that ACE differences are but reflec-tions of differences in phenotypic variances.

Genetic and environmental contributions to personality stabilitywere estimated by fitting a longitudinal correlation between geneticeffects at baseline and follow-up (rg) and longitudinal correlations be-tween environmental effects at baseline and follow-up (rc andrc; Heathet al., 1989; Kaprio, Viken, Koskenvuo, Romanov, & Rose, 1992). Be-cause the current data include assessments at only two time points, es-timation of stabilities must be conducted in a reduced form of the CSmodel. The basic unit in this reduced form is the age-cohort group (rowsof Figure 1), so that age and cohort effects can no longer be distin-guished. There is no time effect, because both times (1975 and 1981)are necessary to define one stability.

In addition to the usual genetic analysis of covariance structure, weevaluated the structure of sample means. By including sample means inmodel estimation, we can test the equivalence of mean scores for MZand DZ pairs and describe mean structure across gender, age, time, andcohort. It should be noted that power to identify mean differences isconsiderably higher than power to detect variance differences, so that, ina sample of nearly 15,000 individuals, mean differences of no practicalsignificance may be statistically significant.

Separately for Extraversion and Neuroticism, the first step was toevaluate which parameters were necessary in a model with free esti-mates across age, cohort, time, and gender. All parameters not signifi-cantly different from 0.0 or 1.0 were fixed to those values for subsequentanalyses. Then, constrained analyses of variance structure were con-ducted, with mean structure free to vary across age, cohort, time, andgender. Constrained analyses of means were conducted, with variancestructure free to vary across age, cohort, time, and gender.

Before we conducted analyses, twins' absolute intrapair differenceswere regressed on twin pair means to assess mean-variance relation-ships in the data. Linear, quadratic, and cubic regressions of pair differ-ences on pair means were small for Extraversion (.06 to - . 10) and Neu-roticism (.10 to -.03) and were not reduced by standard transforma-tions; accordingly, we used untransformed data in all reported analyses.

ResultsThere was significant additive genetic variance (A) for both

traits and both genders in all cohorts at both times of assess-ment, minimum x2(l) = 9.84. No shared environment effects(C) were significantly different from 0 for either trait, so we setC effects to 0.0 for the remaining analyses.2 Of the 24 geneticstabilities, 18, which were not significantly different from 1.0,were set to unity for all remaining analyses. The scalar multiplemodels did not provide a good fit to the data and will not bediscussed further.3

The 6-year phenotypic correlations for Extraversion andNeuroticism, computed for all individual subjects without re-gard to twin status, are shown in Table 1, which also providesthe number of individual twins in each of the six cohorts.

ExtroversionThe process of interpreting the first set of results is described

in detail to illustrate our analyses. Once the logic of model eval-

2 Given the absence of significant C effects, it would have been possi-ble to estimate nonadditive genetic effects for each separate cohort-time-gender sample. Small nonadditive effects were present for somesamples, but they were not consistent across age, cohort, time, or gender.Current versions of LISREL do not allow the types of constraints nec-essary for a cross-sequential analysis of highly correlated additive andnonadditive genetic parameters, and longitudinal analyses were pre-cluded, because significant nonadditive effects were not present at bothassessments for the same cohort. Accordingly, our analyses consider ad-ditive effects only. However, as in previous investigations (Eaves,Eysenck, & Martin, 1989), nonadditive effects were present, and itshould be recognized that broad heritability estimates (which includenonadditive variance) for all cohort-gender-time combinations wouldbe larger than are the narrow heritabilities on which our analyses focus.

3 When heterogeneity in A and E effects was identified in the currentanalyses, the two components always changed in opposite directions.Thus, the scalar multiple models' assumption of proportional changewas always incorrect. Interested readers can obtain details of the scalarmultiple models from us.

726 VIKEN, ROSE, KAPRIO, AND KOSKENVUO

Table 1Phenotypic 6- Year Correlations and Sample Size for Six Age Cohorts

Agecohortat baseline

18-2324-2930-3536-4142-4748-53

Phenotypic correlations

Extraversion

Men

.58

.68

.72

.72

.71

.69

Women

.60

.69

.70

.74

.70

.71

Neuroticism

Men

.48

.56

.59

.63

.62

.67

Women

.60

.57

.62

.63

.65

.65

MZ brothers

674500358272194144

Number of individuals

DZ sisters I

14041126770616476294

vlZ brothers DZ sisters

928660452304238198

17401234826620520384

Note. MZ = monozygotic; DZ = dizygotic.

uation is established, we describe only the main conclusions.Results of model fitting of additive genetic and unshared envi-ronmental effects (A and E) for Extraversion are shown in Table2. CS comparisons, depicted in the rows of Table 2, evaluateheterogeneity across the CS effects of age, cohort, and time ofassessment. Comparisons of the columns of Table 2 evaluategender differences.

Considering the gender-invariant column of the CS factor, themost basic comparison is between the CS-free model, whichallows the variance components to differ across any of the CScomponents, and the CS-invariant model, which constrains es-timates to be equal across all CS components. This comparisontells us whether there is significant heterogeneity in the AE esti-mates across age, time, or cohort. The comparison, x2(22) =78.6, p < .001, is highly significant. Thus, the invariance as-sumption provides parsimony (22 df), but it results in a signifi-cant, and therefore unacceptable, deterioration in fit. As a re-sult, we reject the invariance assumption and infer heterogene-ity in the AE estimates.

We next evaluate the possibility that age, cohort, or time

Table 2Results of Model Fitting for Extraversion

Constraint type

CS freeCS invariantAgeCohortTime

CS freeCS invariantAgeCohortTime

Gender invariant

df x2

Gender dependent

df

Variance structure constraints274296284286294

289.06352.66306.33325.53332.91

250294270274290

Mean structure constraints262273267268272

368.64508.07375.18448.64479.64

250272260262270

X2

258.03345.76292.00303.56324.76

258.03432.59280.38366.04391.34

Note- CS = cross-sequential.

models, which are less restrictive than the CS-invariant modelbut more restrictive than the CS-free model, provide a parsimo-nious compromise. The age model, which assumes that onlyage-related differences are important, provided a significantlybetter fit than the CS-invariant model, x2(12) = 46.36, indicat-ing that there were age-related differences in the data. The non-significant comparison of the age and CS-free models, x2(10) =17.27, indicated that, once age effects were considered, residualcohort and time effects could be constrained to zero without asignificant deterioration in fit. The overall fit of the age modelwas excellent, x2(284) = 306.33, p = . 173, GFI = .98. Althoughboth the cohort, x

2(10) = 27.13, and the time, x2(2) = 19.75,models provided a better fit than the CS-invariant model, theyalso provided a significantly worse fit, x2(12) = 36.6; x2(20) =43.85, than the CS-free model. These results indicate that therewas significant heterogeneity associated with cohort and time,but that neither could explain all of the significant variability:Both longitudinal and cross-sectional age differences must beconsidered to account for variability in the AE estimates. Incontrast, there were no significant residual time or cohort effectsonce age had been considered.

The same comparisons can be made under the gender-depen-dent condition, where estimates are not constrained to be equalfor men and women. Under this condition, the age, cohort, andtime models are all superior to the CS-invariant model, butnone is superior to the CS-free model. When comparing acrossgender conditions, the gender-invariant age model is superior tothe gender-dependent, CS-free, CS-invariant, and age models.Thus, the gender-invariant age model is the preferred represen-tation of the data.

Heritability estimates for Age 1 (ages 18-23 years) throughAge 7 (ages 54-59 years) were .52, .45, .42, .46, .43, .41, and.41. These estimates suggest a gradual decrease in heritability ofExtraversion through the life span. Just as model fitting allowedus to identify age as the significant source of heterogeneity inthe general model, we can fit more restrictive age models to eval-uate the significance of individual age differences. In the case ofExtraversion, a restricted model with three age-related levelsand invariance within levels provided a fit no worse than thegeneral age model, x2(8) = 9.97. Estimates from this model,with elements constrained to be equal underlined, were: .52,.44, .44, .44, .44, .41, and .41. Separate fo\\ow-up of A and E

PERSONALITY FROM 18 TO 59 YEARS 727

Table 3Genetic and Environmental Stability Over 6 Years

Agecohort

at baseline

18-2324-2930-3536-4142-4748-53

18-2324-2930-3536-4142-4748-53

Men

.881.001.001.001.001.00

Extroversion

.33

.46

.56

.53

.47

.57

.831.00.80

1.001.001.00

Neuroticism

.25

.40

.50

.43

.45

.54

Women

.861.001.001.001.001.00

.801.00.88

1.001.001.00

.36

.42

.45

.46

.58

.40

.25

.30

.43

.28

.52

.40

ternative with no deterioration in fit over the CS-free model,X2(5) = 2.73. A more restrictive age model constraining the lastfour ages to be equal provided a fit as good as the general agemodel, x2(3) = 7.01, /> < .1, although the comparison ap-proached significance. Estimates across the seven ages, with el-ements constrained to be equal underlined, were 4.74, 4.43,4.26,4.16,4.16,4.16, and 4.16. For women, the age model pro-vided a significantly poorer fit than the CS-free model, x2(5) =19.62, indicating that age effects alone could not explain thepattern of means for women. Women did show a general de-crease in Extraversion associated with cross-sectional age and,in the first and last cohorts, a decrease in Extraversion associ-ated with longitudinal increase in age. In the middle cohorts,however, the samples tended to show an increase in Extraversionbetween 1975 and 1981. None of the basic restricted modelsthat can be evaluated in the current design can account for thispattern. One effect that is clear is the significant drop in meanExtraversion from the ages of 18-23 years to the ages of 24-29years in both cross-sectional, x2(l) = 31.91, and longitudinal,X2(l)= 19.67, comparisons.

effects indicated that the decreasing heritabilities were associ-ated with a significant increase in E from the first to the laterage levels, x2(2) = 23.65, and a nonsignificant decrease in Aacross all three age levels, x2(2) = 5.65.

Estimates of genetic and environmental stabilities for Extra-version are shown in Table 3. For both men and women, thecorrelation between genetic effects at 1975 and 1981 was sig-nificantly different from 1.0 in the first cohort only. The genderdifference in rg was not significant, x2( 1) = 0.15. Thus, we foundno evidence of significant new genetic influences on Extraver-sion after the age of 29 years and no evidence of gender modu-lation of genetic stability. There was significant heterogeneity inre across both cohort, x2( 10) = 37.08, and gender, x2(6) = 14.18.With the exception of lower estimates of re in the first cohort,however, the pattern of differences in re was not readily inter-pretable.

There was no significant zygosity difference in mean Extra-version, x2(24) = 34.59. Results of the model fitting for CS andgender contributions to mean structure of Extraversion arefound in Table 2, and means computed on subjects as individu-als are found in Table 4. Note that the AE effects were left freeto vary by gender, age, cohort, and time in all of the models onmeans, just as means were free to vary by gender, age, cohort,and time in the variance component models. For this reason,the gender-dependent CS-free conditions are identical for thetwo groups of models. Modeling revealed a substantial genderdifference in mean Extraversion, x2(12) = 110.34. Evaluationof chi-square difference tests indicated that the gender-depen-dent age model fit significantly better than the CS-invariantmodel. However, the comparison with the gender-dependentCS-free model was significant, x2(10) = 22.35, indicating thatthere were significant influences on means after age and genderwere considered. None of the other models provided a reason-able fit to the data.

Because of the strong gender dependence in the mean struc-ture, the age model was evaluated separately for men andwomen. In men, the age model provided a parsimonious al-

NeuroticismResults of model fitting of genetic and environmental effects

for Neuroticism are shown in Table 5. The best fitting modelwas the gender-dependent age model, which was superior toboth the CS-free and the CS-invariant models and provided anexcellent fit to the data, X2(268) = 294.60, p = . 127, GFI = .97.Heritability estimates for men across the seven ages were .54,.29, .35, .35, .29, .31, and .17; follow-up of the age effects formen indicated that a restricted model specifying only two agelevels of heritability (.54 for Age 1 and .31 for all other ages) fit

Table 4Mean Extraversion and Neuroticism Scores at 1975 and 1981

Age cohortat baseline

18-2324-2930-3536-4142-4748-53

All 18-53 MAll 18-53 SD

18-2324-2930-3536-4142-4748-53

All 18-53 3 /All 18-53 SD

1975

4.734.444.344.184.184.19

4.442.45

3.973.943.934.194.103.98

4.002.51

Men

1981

Extraversion

4.434.214.294.054.294.14

4.272.51

Neuroticism

3.953.743.754.124.143.81

3.902.45

Women

1975

4.304.013.853.873.713.70

4.012.51

5.104.674.674.494.334.15

4.722.42

1981

4.044.034.023.933.723.64

3.962.55

4.534.244.264.154.134.03

4.302.38

728 VIKEN, ROSE, KAPRIO, AND KOSKENVUO

Table 5Results of Model Fitting for Neuroticism

Constraint type

CS FreeCS InvariantAgeCohortTime

CS FreeCS InvariantAgeCohortTime

Gender invariant

df X2

Gender dependent

df

Variance structure constraints272294282284292

324.27407.77350.57366.42378.50

248292268272288

Mean structure constraints260271265266270

572.90870.20651.23822.21643.88

248270258260268

X2

264.78367.42294.60321.83333.92

264.70672.20351.77585.29388.44

Note. CS = cross-sequential.

no worse than the general age model, x2( 10) = 10.7. Heritabilityestimates for women were .53, .44, .39, .44, .43, .34, and .40; atwo-level age model for women (.53 for Age 1 and .42 for allother ages) also fit as well as the general age model, x2(10) =10.87. Follow-up of the gender dependency indicated that esti-mates for men and women were significantly different only inthe older age level, x2(2) = 38.96. Separate follow-up of A andE effects indicated that there was a significant increase in E withage for both men, x2(l) = 31.39, and women, x

2(l) = 14.18,and a significant decrease in A with age for both men, x2(l) =32.56, and women, x2(0 = 7.65.

Genetic and environmental stabilities for Neuroticism areshown in Table 3. Genetic stabilities were significantly differentfrom 1.0 only in the first, x

2(2) = 31.37, and third, X2(2) = 9.71,cohorts, and the difference between men and women was notsignificant, x2(2) = 1.74. Environmental stabilities were hetero-geneous across cohorts within gender, x2(10) = 49.24, andacross gender, x2(6) = 12.93. As was the case for Extraversion,however, the pattern of environmental stabilities was not readilyinterpretable.

There was a significant zygosity difference in mean Neuroti-cism scores, x2(24) = 46.25, with the DZ mean .16 higher thanthe MZ mean. Although the difference is significant with thislarge sample, it is very small (0.06 SD) and does not indicate thepresence of substantial zygosity differences. Results of modelfitting for gender and CS effects on mean Neuroticism scores areshown in Table 4. Means based on subjects treated as individu-als are shown in Table 5. Although the gender-dependent age,cohort, and time models each provided a substantial improve-ment over the CS-invariant models, none came close to being anacceptable alternative to the gender-dependent CS-free model.Thus, there was significant heterogeneity of means left afterconsidering each of the CS components, and no simple modelcould account for the data. There were substantial genderdifferences, x2(12) = 308.20. Differences in means across thelife span for men were small: The 12 estimates were all within0.1 SD of the joint mean. For women, there was a consistent

decrease in Neuroticism associated with increasing cross-sec-tional age and with longitudinal increase in age. In the youngercohorts, however, the decrease associated with the 6-year follow-up period was greater than the decrease associated with thesame change in cross-sectional age. Thus, when samples werematched on age, the sample assessed in 1981 had a lower meanscore.

DiscussionOur analysis provides, for the first time, a comprehensive

view of genetic effects on extraversion and neuroticism througha 40-year span of adult life. For the first time in any behaviorgenetic study, the modulating effects of age and gender wereevaluated in a cross-sequential model to allow an evaluation ofcompeting cohort and time-related explanations for genetic andenvironmental heterogeneity. There was significant heritabilityof both traits at all ages, cohorts, and times, consistent with abroad range of previous work (Loehlin, 1992). And, consistentalso with prior investigations in which shared environmentaleffects are modeled but not measured (Rose, Kaprio, Williams,Viken, & Obremski, 1990), we found no evidence of them. ForExtraversion, heritabilities were gender invariant, but modelingdemonstrated significant age modulation of both genetic andenvironmental effects for Extraversion, with no significant het-erogeneity of effects, once age was considered. The best fittingmodel indicated a decrease in heritability from the late teensand early 20s to the late 20s and a small additional decreaseafter the late 40s. Separate estimates of A and E effects indicateda nonsignificant decrease in A with age and a significant in-crease in E from the late teens and early 20s to later ages. Theincrease in E may well reflect increased contributions from un-shared environments, as cotwins leave a childhood home,marry, and establish separate lives. McGue et al. (1993) havereported a similar longitudinal increase in unshared environ-mental effects for the related construct of Positive Emotionalityfrom the MPQ, and, as found here for Extraversion, they foundthat the increase in unshared environmental effects was associ-ated with a decrease in heritability.

In univariate twin designs, effects attributed to E includeboth unshared environmental effects and error variance, so it ispossible that our results suggesting increased E might reflectincreased error of measurement. However, both the internalconsistencies (Tarkkonen, Koskenvuo, Kaprio, Langinvainio,& Floderus-Myrhed, 1981) and the 6-year test-retest corre-lations of the current measures remain stable or increase withage, suggesting that increased error of measurement cannot ex-plain the increase in E. The correlations between environmen-tal effects at baseline and follow-up were modest, indicating thatenvironmental effects contribute to both change and stability inindividual differences in extraversion across the life span. Thelowest values of rc were observed in the youngest cohort, re-flecting the dramatic change in unshared environmental influ-ences after cotwins leave home, serve in the military, and beginadult lifestyles.

In contrast to the presence of new environmental variance ineach age group, there was no evidence of significant new geneticvariance after the age of 30 years, and the primary longitudinalcontribution of genetic effects was to maintain stability of indi-

PERSONALITY FROM 18 TO 59 YEARS 729

vidual differences in Extraversion. At all ages, genetic effectsmade the most important contribution to stable variation inExtraversion. The reduced (< 1.0) genetic stability found for theyoungest cohort is consistent with the longitudinal results ofMcGue et al. (1993) in a young twin sample. Younger twinsalso exhibited reduced genetic stability in age-to-age patterns ofalcohol consumption (Kaprio et al., 1992).

Mean Extraversion scores were gender-dependent, with higherExtraversion observed among men. A simple age model pro-vided a satisfactory explanation of heterogeneity of means onlyfor men, who showed a substantial decrease in Extraversionfrom late teens to late 20s and a more gradual decrease thereaf-ter. Women showed a significant decrease of similar magnitudefrom late teens to late 20s and inconsistent changes in lateryears. For both men and women, mean changes after age 30were quite small. A similar pattern of mean changes for Extra-version and Neuroticism has been reported for singleton Fin-nish adults, who were administered a Finnish translation of theEysenck Personality Questionnaire (Haapasalo, 1990).

Genetic and environmental effects on Neuroticism scoreswere age modulated and gender dependent. Heritability forNeuroticism scores dropped from the first age to all later ages.The decrease in heritability with age was associated with sig-nificant decreases in A and significant increases in E for bothmen and women. McGue et al. (1993) reported a similar de-crease in heritability with longitudinal age, associated with adecrease in genetic variance but stable E. In our analyses, thegender differences were significant only at the later ages, withlower heritabilities for men. Increased heritability for women isconsistent with work by Martin and Jardine (1986) and theEaves et al. (1989) reanalysis of the data from the Swedish TwinCohort. That gender differences were more substantial amongolder twins is consistent with the data presented by Eaves et al.(1989), although they did not explicitly test this effect, and itmay explain the absence of gender differences in the Loehlinand Nichols (1976) sample, whose members were approxi-mately the age of our younger twins.

As was found for Extraversion, correlations for Neuroticismbetween environmental effects at baseline and follow-up weremodest, indicating that environmental effects were contributingto both stability and change in individual differences for Neu-roticism. Modeling indicated that significant new genetic vari-ance was contributing to Neuroticism as late as the age of 40years. In general, however, genetic effects contributed to stabil-ity in individual differences rather than to change. For women,genetic effects were the primary source of age-to-age stabilityin Neuroticism. For older men, modest heritabilities and highgenetic stabilities provided about the same contribution to phe-notypic stability as did strong environmental effects coupledwith modest environmental stabilities.

Mean structure for Neuroticism was also gender dependent,with higher scores for women. Neuroticism scores for men werethe only measures that did not show a decrease from the firstto later ages. Although there was heterogeneity of Neuroticismscore means for men, the means did not follow any simple pat-tern, and there were no clear age effects. The pattern of meansfor women showed a decrease with age, but in younger cohorts,longitudinal increase in age (from 1975 to 1981) was associatedwith greater change in Neuroticism than cross-sectional age

differences. A reasonable explanation for this pattern wouldcombine an age-modulated decrease in Neuroticism across thelife span with a secular time-related effect (e.g., public educa-tion or attitudes) that influenced younger women between base-line and follow-up. However, a combination of age and cohorteffects (e.g., accelerated developmental influences in cohortsborn after 1945) could also explain this pattern. Unfortunately,limitations in the current design do not allow us to make a for-mal comparison of these (or other) alternative models.

The results of our analyses of means and longitudinal stabilityare consistent with broad evidence that Extraversion and Neu-roticism are relatively stable after the age of 30 years (McCrae& Costa, 1990). Changes in sample means after the age of 30years were small, and the substantial 6-year retest correlationsreflected both stability (primarily genetic) and instability (pri-marily environmental) in individual differences. However, wenote that the self-assessment scales we used in this research weredesigned to assess stable dispositions; larger age-related effectsmight well be found with indexes designed to assess character-istics that should be sensitive to developmental influences (e.g.,Whitbourne, Zuschlag, Elliot, & Waterman, 1992). Large-sam-ple longitudinal twin studies and emerging analytic techniques(Williams, Viken, & Rose, 1992) offer robust methods for pur-suing developmental issues in adult personality, and we antici-pate that this report will be followed by many more.

ReferencesCardon, L. R., Fulker, D. W., DeFries, J. C, & Plomin, R. (1992). Con-

tinuity and change in general cognitive ability from 1 to 7 years ofage. Developmental Psychology, 28, 64-73.

Dworkin, R. H., Burke, B. W., Maher, B. A., & Gottesman, 1.1. (1976).A longitudinal study of the genetics of personality. Journal of Person-ality and Social Psychology, 34, 510-518.

Dworkin, R. H., Burke, B. W., Maher, B. A., & Gottesman, 1.1. (1977).Genetic influences on the organization and development of personal-ity. Developmental Psychology, 13, 164-165.

Eaves, L. J., Eysenck, H. J., & Martin, N. G. (1989). Genes, culture, andpersonality: An empirical approach. San Diego, CA: Academic Press.

Finn, S. (1986). Stability of personality self-ratings over 30 years: Evi-dence for an age/cohort interaction. Journal of Personality and SocialPsychology, 50, 813-818.

Floderus, B. (1974). Psycho-social factors in relation to coronary heartdisease and associated risk factors [Supplement 6]. Nordisk Hygie-niskTidskrift.

Floderus-Myrhed, B., Pedersen, N. L., & Rasmuson, I. (1980). Assess-ment of heritability for personality, based on a short form of theEysenck Personality Inventory: A study of 12,898 twin pairs. Behav-ior Genetics, 10, 153-162.

Haapasalo, J. (1990). The Eysenck Personality Questionnaire andZuckerman's Sensation Seeking Scale (Form V) in Finland: Agedifferences. Personality and Individual Differences, 11, 503-508.

Heath, A. G, Neale, M. C, Hewitt, J. K., Eaves, L. J., & Fulker, D. W.(1989). Testing structural equation models for twin data usingLISREL. Behavior Genetics, 19, 9-35.

Heath, A. G, Neale, M. C, Kessler, R. C, Eaves, L. J., & Kendler, K. S.(1992). Evidence for genetic influences on personality from self-re-ports and informant ratings. Journal of Personality and Social Psy-chology, 63, 85-96.

Joreskog, K. G., & Sorbom, D. (1989). LISREL 7 users reference guide.Mooresville, IN: Scientific Software.

Kaprio, J., Koskenvuo, M., Artimo, M., Sarna, S., & Rantasalo, I.

730 VIKEN, ROSE, KAPRIO, AND KOSKENVUO

(1979). Baseline characteristics of the Finnish Twin Registry: SectionI: Materials, methods, representativeness, and results for variablesspecial to twin studies. Publications of the Department of PublicHealth Science, M47, University of Helsinki, Finland.

Kaprio, J., Sarna, S., Koskenvuo, M., & Rantasalo, I. (1978). FinnishTwin Registry: Formation and compilation, questionnaire study, zy-gosity determination procedures and research program. Progress inClinical and Biological Research, 24B, 179-184.

Kaprio, J., Viken, R., Koskenvuo, M., Romanov, K., & Rose, R. J.(1992). Consistency and change in patterns of social drinking: A 6-year follow-up of the Finnish Twin Cohort. Alcoholism: Clinical andExperimental Research, 16, 234-240.

Koskenvuo, M., Langinvainio, H., Kaprio, J., Rantasalo, I., & Sarna,S. (1979). The Finnish Twin Registry: Baseline characteristics. Sec-tion III. Occupational and psychosocial factors. Publications of theDepartment of Public Health Science, M49, University of Helsinki,Finland.

Loehlin, J. C. (1992). Genes and environment in personality develop-ment. Newbury Park, CA: Sage.

Loehlin, J. C, & Nichols, R. C. (1976). Heredity, environment, and per-sonality. Austin, TX: University of Texas Press.

Martin, N. G., & Jardine, R. (1986). Eysenck's contribution to behaviorgenetics. In S. Modgil & C. Modgil (Eds.), Hans Eysenck: Consensusand controversy (pp. 13-47). Lewes, England: Falmer Press.

McCrae, R. R., & Costa, P. T. (1990). Personality in adulthood. NewYork: Guilford Press.

McGue, M., Bacon, S., & Lykken, D. T. (1993). Personality stabilityand change in early adulthood: A behavioral genetic analysis. Devel-opmental Psychology, 29, 96-109.

Neale, M. C, & Cardon, L. R. (1992). Methodology for genetic studiesof twins and families. Norwell, MA: Kluwer Academic Publishers.

Pedersen, N. L., Plomin, R., McClearn, G. E., & Friberg, L. (1988).Neuroticism, extraversion, and related traits in adult twins rearedapart and reared together. Journal of Personality and Social Psychol-ogy, 55, 950-957.

Pogue-Geile, M. F., & Rose, R. J. (1985). Developmental genetic studiesof adult personality. Developmental Psychology, 21, 547-557.

Rose, R. J. (1988). Genetic and environmental variance in content di-

mensions of the MMPI. Journal of Personality and Social Psychology,55,302-311.

Rose, R. J., & Kaprio, J. (1988). Frequency of social contact and intra-pair resemblance of adult MZ co-twins: Does shared experience in-fluence personality after all? Behavior Genetics, 18, 309-328.

Rose, R. J., Koskenvuo, M., Kaprio, J., Sarna, S., & Langinvainio, H.(1988). Shared genes, shared experiences, and similarity of personal-ity: Data from 14,288 adult Finnish co-twins. Journal of Personalityand Social Psychology, 54, 161-171.

Rose, R. J., Kaprio, J., Williams, C. J., Viken, R., & Obremski, K.(1990). Sibling contact and sibling similarity: Facts, issues, and redherrings. Behavior Genetics, 20, 763-778.

Sarna, S., & Kaprio, J. (1980). Use of multiple logistic analysis in twinzygosity diagnosis. Human Heredity, 30, 71-80.

Schaie, K. W., & Hertzog, C. (1982). Longitudinal methods. In B. B.Wofman (Ed.), Handbook of developmental psychology (pp. 91-115).Englewood Cliffs, NJ: Prentice-Hall.

Schaie, K. W., & Willis, S. L. (1991). Adult personality and psychomo-tor performance: Cross-sectional and longitudinal analyses. Journalof Gerontology: Psychological Sciences, 46, P275-P284.

Tarkkonen, L., Koskenvuo, M., Kaprio, J., Langinvainio, H., & Flode-rus-Myrhed, B. (1981). Cross validation of the Eysenck extroversionand neuroticism scales in Finland and Sweden. Publications of theDepartment of Public Health Science, M62, University of Helsinki,Finland.

Tellegen, A., Lykken, D. X, Bouchard, T. J., Wilcox, K., Segal, N., &Rich, S. (1988). Personality similarity in twins reared apart and to-gether. Journal of Personality and Social Psychology, 54, 1031 -1039.

Whitbourne, S. K., Zuschlag, M. K., Elliot, L. B., & Waterman, A. S.(1992). Psychosocial development in adulthood: A 22-year sequentialstudy. Journal of Personality and Social Psycholgy, 63, 260-261.

Williams, C. J., Viken, R., & Rose, R. J. (1992). Likelihood-based anal-yses of longitudinal twin and family data: Experiences with pedigree-based approaches. Behavior Genetics, 22, 215-223.

Received December 28, 1992Revision received August 25, 1993

Accepted August 28, 1993 •