Traits and Metatraits: Their Reliability, Stability, and Shared Genetic Influence

Transcription

1 PERSONALITY PROCESSES AND INDIVIDUAL DIFFERENCES Traits and Metatraits: Their Reliability, Stability, and Shared Genetic Influence Scott L. Hershberger and Robert Plomin Pennsylvania State University Nancy L. Pedersen The Karolinska Institute Metatraits measure individual differences in construct relevancy, whereas traits measure individual differences in construct extremity. Twenty-four traits and metatraits were examined in this study using 57 pairs of identical twins reared together, 95 pairs of identical twins reared apart, pairs of fraternal twins reared together, and 8 pairs of fraternal twins reared apart obtained from the Swedish Adoption/Twin Study of Aging (see N. L. Pedersen et al., 99). Reliability and stability analyses of the metatraits revealed generally lower reliability and stability across time compared to traits. Quantitative genetic analyses of the relationship between traits and metatraits indicated that approximately 69% of the metatrait's genetic variance is shared with the trait, with 3% of its genetic variance unique to the metatrait. These results suggest that metatraits provide a useful additional view of personality. The most common measure of individual differences in personality research has undoubtedly been the total scale score obtained on an inventory or questionnaire. This total scale score, formed by the sum of item responses, serves as a measure of an individual's standing on a trait. Although the conclusion has been vigorously contested (e.g., Epstein, 983), some researchers have asserted that the use of total scale scores for the investigation of personality, particularly for the prediction of behavior, has been less than productive (e.g., Mischel, 968). Bern and Allen (974) suggested that the source of this predictive difficulty might rest with the relevancy of the personality trait for the individual's behavior. According to Bern and Allen, if a person is behaviorally consistent across situations, the personality trait responsible for the behavior may be more relevant to this person than to a behaviorally inconsistent person. As one measure of consistency, Bern and Allen calculated an ipsatized variance index, or the variance of item responses on one trait scale divided by the total variance across all items on several trait Scott L. Hershberger and Robert Plomin, Department of Psychology, Pennsylvania State University; Nancy L. Pedersen, The Karolinska Institute, Stockholm, Sweden. The Swedish Adoption/Twin Study of Aging (SATSA) is an ongoing study conducted at the Department of Epidemiology of the Institute for Environmental Medicine at the Karolinska Institute in Stockholm in collaboration with the Research Center for Developmental and Health Genetics at Pennsylvania State University. SATSA is supported in part by grants from the National Institute on Aging (AG-4563 and AG- IO 75), the Swedish Social Research Council, and the John D. and Catherine T. MacArthur Foundation Research Network on Successful Aging. Correspondence concerning this article should be addressed to Scott L. Hershberger, who is now at the Department of Psychology, University of Kansas, 46 Fraser Hall, Lawrence, Kansas scales. In their study, the ipsatized variance index proved successful in differentiating individuals who were cross-situationally consistent from those who were inconsistent in conscientiousness, but it was not successful for the trait of friendliness. Various efforts to reproduce and extend Bern and Allen's results have met with modest success (e.g., Chaplin & Goldberg, 985). Some researchers have pointed to statistical difficulties with the ipsatized index itself, among which is the confounding of item variance caused by the trait with the interitem variance of all other measured traits (e.g., Paunonen & Jackson, 985). Others have claimed that measures of consistency possess little discriminant validity from total scale scores (e.g., Burke, Kraut, & Dworkin, 984). Yet another problem exists with the correlation between the total scale score and ipsatized variance index, which is frequently not linear but curvilinear because of the restrictions placed on the extremity of the total scale scores of inconsistently responding persons (Paunonen, 988). Nonetheless, recent reviews of attempts to use behavioral consistency as a measure of trait relevancy have found some promise in the approach (Paunonen, 988; Tellegen, 988). Indeed, the work of Baumeister and Tice (988; Baumeister, 99; Tice, 989) has been particularly supportive of the usefulness of behavioral consistency as an index of trait relevancy. Baumeister and Tice (988) proposed the term metatrait to describe that "trait of having versus not having a particular trait." A metatrait is measured by the simple standard deviation of an individual's item responses, thus avoiding the problems attendant on the Bern and Allen ipsatized variance index. Although not without its critics (e.g.,. Paunonen, 988), the interitem standard deviation index has proved useful in many contexts (Tice, 989). Metatraits, as measured by the interitem standard deviation index, have been shown to mediate the relationship between total scale score (locus of control) and a behavioral criterion (persistence on a laboratory task) and, when the Journal of Personality and Social Psychology, 995, Vol. 69, No. 4, Copyright 995 by the American Psychological Association, Inc. -354/95/S3 673

2 674 S. HERSHBERGER, R. PLOMIN, AND N. PEDERSEN metatrait is self-esteem, to predict a behavioral criterion (advice seeking) directly (Baumeister & Tice, 988). Metatraits have also been shown to be stable across time and to be equally reliable for "traited" and "untraited" individuals (Baumeister, 99). However, the reliability estimates were restricted to test-retest correlations. Reliability estimates obtained from test-retest correlations commonly confound the reliability of a trait with its stability across time. Behavioral genetic studies of personality scale scores have revealed a substantial degree of genetic influence on personality, with heritabilities ranging typically between.5 and.5 (Bouchard & McGue, 99; Plomin, Chipuer, & Loehlin, 99). However, the heritabilities of metatraits are unknown. Waller and Reise (99) examined genetic influence on individual differences in scalability, a concept related to but distinct from the concept of metatraits. Scalability refers to the degree to which an individual's item response pattern conforms to a perfect Guttman scale. Although an extreme degree of item response inconsistency is incompatible with a perfect Guttman scale, two individuals with item patterns conforming identically to a Guttman scale may nonetheless demonstrate dramatically different degrees of response variability. One goal of the present study was to examine genetic and environmental influences on metatraits. A second goal was to address the important issues of the reliability and stability of metatraits. Baumeister's (99) study addressed the two latter issues with the correlation between metatrait scores assessed at two time periods several weeks apart. The current study includes three periods of measurement spanning 6 years, thus allowing for a more refined approach to reliability and stability estimation. A third goal was to examine the correlation between trait and metatrait scores. If, as Burke et al. (984) claim, metatraits essentially provide no information in addition to that provided by the level of a trait, the metatrait and trait level should be highly correlated. If this is the case, most of the genetic influence affecting the metatrait should be shared with the trait, with little genetic influence affecting the metatrait uniquely. If metatraits capture an aspect of personality not captured by the trait level, a significant proportion of the genetic variance affecting the metatrait should be unique to the metatrait and not shared with the trait. Participants Method The data for this study were obtained from the Swedish Adoption/ Twin Study of Aging (SATSA; see Pedersen et al., 99), a longitudinal study of the cognitive, personality, and health characteristics of a sample of twins. The twins were identified from the Swedish Twin Registry, which includes nearly 5, pairs of same-sex twins born in Sweden between 886 and 958 (Cederlof & Lorich, 978). The twins reared together were matched to the twins reared apart on the basis of age, gender, and county of birth. For the twins reared apart, the average age of separation was.8 years; 48% were separated before their first birthday and 8% before their fifth. None of the twin pairs were separated after the age of. The first wave of testing was conducted in two parts in 984. Responses to the first questionnaire were received from 74 intact twin pairs: 99 identical twins reared apart (MZA), 66 identical twins reared together (MZT), 38 fraternal twins reared apart (DZA), and fraternal twins reared together (DZT). A second questionnaire was mailed directly after receipt of the first; responses were received from 576 intact twin pairs: 83 MZA, 36 MZT, 8 DZA, and 75 DZT. The second wave of testing was conducted in 987, with a single questionnaire sent to all individuals who had not requested discontinuation of participation. Responses were received from 53 intact twin pairs: 7 MZA, 7 MZT, 78 DZA, and 55 DZT. An identical questionnaire was used for the third wave of testing in 99. Responses were received from 474 intact twin pairs: 6 MZA, 4 MZT, 53 DZA, and 45 DZT. The average age of the twins was 58.6 years at the first wave of testing, ranging from 6 to 87 years. Forty percent of the sample was male and 6% female. Initially, zygosity status was based on the physical similarity reported at the first wave. All twins older than 5 years, both of whom had responded to the first wave's questionnaires, were approached for participation in a later health interview; during this health interview, in which approximately 5% of the first-wave twin pairs participated, all had their zygosity status determined by comparing serum proteins and red cell enzymes from blood assays. Zygosity was changed for 8% of the sample on the basis of the assay results. Subsequent analyses comparing behavioral genetic results before and after the zygosity changes showed little difference in results. Further information concerning SATSA may be found in Pedersen et al. (99). Measures Each of the measures used in this study was translated into Swedish by a professional translator and back-translated into English to ensure that the meaning of the original items was retained. Short form of the Eysenck Personality Inventory. The short form of the Eysenck Personality Inventory (Floderus, 974) was used to measure extraversion and neuroticism. Each scale consists of nine items, with a yes or no response format. Internal consistency as measured by Cronbach's a was.66 for extraversion and.75 for neuroticism (Pedersen, Plomin, McClearn, & Friberg, 988). EAS Temperament Survey. The EAS Temperament Survey for adults (Buss & Plomin, 984) assesses the temperament traits of distress, fear, anger (the three composing an emotionality factor), activity, and sociability. Each trait is assessed with four items answered on a 5- point Likert scale. Buss and Plomin reported average -week test-retest reliability estimates of.8 for emotionality,.8 for activity, and.85 for sociability. Karolinska scales oflmpulsivity, Monotony Avoidance, and Inhibition of Aggression. Three scales were selected from the Karolinska Scales of Personality (Schalling, Edman, & Asberg, 983) impulsivity, monotony avoidance, and inhibition of aggression each measured on a 5-point Likert scale. The impulsivity scale measures the individual's propensity to act on the spur of the moment (e.g., "I tend to speak first and think afterwards"). The monotony avoidance scale measures the individual's desire to seek out novelty and excitement (e.g., "I am drawn to places where exciting things happen"). Cronbach's a is.7 for impulsivity and.76 for monotony avoidance (Pedersen et al., 988). The inhibition of aggression scale is more accurately conceived of as a measure of assertiveness rather than inhibited aggression, with items such as, "If I'm treated badly at a restaurant I don't like to complain" and "I find it difficult to refuse when someone asks me to do something for them even if I don't feel like doing it." Cronbach's a for inhibition of aggression is.69 (Pedersen, Lichtenstein etal., 989). Framingham hard-driving. Five items, responded to on a 5-point Likert scale, were adopted from the Framingham scale of Type A Behavior (Haynes, Levine, Scotch, Feinleib, & Kanner, 978) to assess the trait of hard-driving. The hard-driving items measure the demands that people put on themselves in their interactions with others (e.g., "I can be described as domineering"). Cronbach's a for the hard-driving scale is.7 (Pedersen, Lichtenstein, et al., 989)., life direction, and responsibility. Twenty-three forced-choice items from the Rotter (966) Locus of Control Scale were converted to

3 METATRAITS Likert-type items; of these, items (4 for each) were selected to measure the traits of, Life Direction, and. These three scales previously were identified by Klockars and Varnum (975) in a factor analysis of Rotter's scale. The items assess the individual's belief in the importance of luck in determining life events (e.g., "Most people don't realize to what extent their lives are ruled by coincidences"). The Life Direction items assess the individual's belief in personal control over life events (e.g., "When I make plans I'm almost certain that I can follow them through"). The items are concerned with personal responsibility for not having a successful career or friendships ("People who are disliked don't know how to get on with others"). Cronbach's as for, Life Direction, and are.55,.49, and.66, respectively (Pedersen, Gatz, Plomin, Nesselroade, &McClearn,989). NEO openness. A shortened version of the to Experience scale from the NEO Personality Inventory (Costa & McCrae, 985) was used. Of the original 48 items, 6 were retained. The validity of the shortened to Experience scale was assessed by its correlation with an adjective checklist, a method employed by McCrae and Costa (985) for the original to Experience scale. The correlations for the short and long versions of the scale did not differ significantly. Cronbach's a is.77 for the to Experience scale (Bergeman et al., 993). OARS depression, somatic symptom, and alienation. The Short Psychiatric Evaluation Schedule from the OARS (Duke University, 978) comprises three scales depression, somatic symptom, and alienation each assessed with five yes or no response-type items. The depression items are concerned with mood and sleep patterns; the somatic-symptom items, with perceived healthfulness; and the alienation items, with difficulty in interpersonal relations. Gatz, Pedersen, and Harris (987) cite internal consistency (Cronbach's a) estimates of.57 for depression,.64 for somatic symptom, and.7 for alienation. Cook-Medley hostility. The original Cook-Medley Scale (Cook & Medley, 954) contains 5 items; however, items were retained on the basis of a factor analysis in which 5 items loaded highly on a Paranoid factor, and 5 items loaded on a Cynicism factor. These items were combined to create a hostility scale, each with a 5- point Likert-scale format. Examples include, for Paranoid, "I have a feeling that I've often been punished unjustly," and for Cynicism, "I think most people would lie to get ahead." Cronbach's a for the hostility scale is (Pedersen, Lichtenstein, et al., 989).. was measured by 3 items, scored on a 5-point Likert-type scale, taken from Wood, Wylie, and Sheafor (969). One example is, "Most of my expectations have been filled." Cronbach's a for this scale is.8. State anxiety. Three scales, adapted from Spielberger (979), were used to measure state anxiety: a 5-item state anxiety scale referring to positive affective states (e.g., "I feel relaxed"); a 5-item state anxiety scale referring to negative affective states (e.g., "I feel nervous"); and a total state anxiety scale that sums the positive and negative scales. Responses were obtained on a 5-point Likert-type scale. Metatrait scales. The metatrait scales corresponding to each of the 3 traits described above were constructed by calculating the standard deviation of item responses for each individual. In addition, a metatrait factor was created by conducting a principal components analysis of of the metatrait scales (leaving out state anxiety positive and negative because of their dependency on the state anxiety total score) and retaining the first component. This first principal component accounted for % of the variance at the first occasion, 5% of the variance at the second occasion, and 6% of the variance at the third occasion. The metatrait factor represents the general behavioral consistency of the individual. However, given the great heterogeneity in the personality scales used in this study, a comparable component for the trait scale has no meaning and was not created. Figure. Reliability and stability model across three occasions of measurement for metatraits and trait levels, where X r = the true score; X o = the observed score; r» x = the square root of the reliability coefficient; j8 = the stability coefficient; and e - random error. Model Fitting Two models were evaluated in the present study: The first model assesses the reliability and stability of trait and metatrait scores; the second was a bivariate behavioral genetic model that indicates the degree to which the observed correlation between traits and metatraits is due to shared genetic and environmental influences. Reliability and stability model. The model used to estimate the reliability (r xx ) of the trait and metatrait scores is shown in Figure. In this model, adapted from Heise (969), at each of the three occasions, the observed score (X o ) is a linear function of the true score (X T ) and random measurement error (e). The relationship between the true scores across time is represented by ft the coefficient of stability. This model assumes, in accordance with the classical test theory model (e.g., Allen & Yen, 979), that random measurement error is uncorrelated with true scores and uncorrelated with the random measurement error from other occasions. Although /3, and fe may be unequal, the reliability of the scores is assumed constant across time. In addition, the disturbances from the second and third occasions are also assumed to be equal. The model may then be tested with one (fe # /8 3 ) or two (/3 = /3 3 ) degrees of freedom. The model of Figure was evaluated by the model-fitting program LISREL 8 (Joreskog & Sorbom, 993), which provides maximum-likelihood estimates of the parameters. Bivariate behavioral genetic model. The linear model underlying quantitative genetic theory posits that the variance in a phenotype may be partitioned into genetic and environmental sources (Plomin, De- Fries, & McClearn, 99). When two or more phenotypes are analyzed simultaneously, the degree of shared genetic and environmental sources among the phenotypes can be determined as well. In this study, a bivariate model-fitting analysis was conducted on the covariance between trait and metatrait scores to determine (a) the proportion of variance attributable uniquely to genetic and environmental sources for metatraits, and (b) the extent of shared genetic and environmental sources of variance between trait and metatrait scores. To conduct a quantitative genetic analysis, the correlations between relatives of varying genetic relatedness are required. When the relatives have been exposed to differing environments, further information of the effect of the environment is obtained. With MZ and DZ twins reared apart and together, the SATS A study provides a powerful design to assess genetic and environmental parameters. The first parameter, additive genetic effects (A), represents the additive (or summative) influence of the multiple genes affecting the phenotype. Identical twins share all their genetic effects and have a unit correlation for additive genetic effects; fraternal twins share, on average, 5% of their genes and are correlated.5 for additive genetic effects. The second parameter, nonadditive genetic effects, or dominance

4 676 S. HERSHBERGER, R. PLOMIN, AND N. PEDERSEN MZ= DZ=.5 Trait Level Score Metatrait Score Metatrait Score Trait Level Score MZ= DZ=.5 Figure. Bivariate path model showing common and unique factors for genetic and environmental sources of variance and covariance for trait levels and metatraits. Latent variables Ens, A, D, and Es refer to nonshared environmental, additive genetic, nonadditive genetic, and shared rearing environmental effects, respectively. All primed (') symbols refer to the co-twin. (D), refers to the nonlinear interaction between alleles at the same locus. Identical twins are again correlated unity for nonadditive dominance effects, whereas fraternal twins are, on average, correlated.5 for nonadditive dominance effects. In general, significant genetic effects on a phenotype are indicated when the identical twin correlation significantly exceeds the fraternal twin correlation. When the concern is with the genetic mediation between two phenotypes, a greater identical twin cross-correlation (the correlation between one twin's metatrait score and the co-twin's trait scale score) is indicative of significant shared genetic effects. The third parameter, shared-rearing environmental effects (Es), refers to environmental factors that produce similarities in siblings reared together. By definition, twins reared together, MZT and DZT, are correlated unity for shared-rearing environmental effects; twins reared apart, MZA and DZA, are correlated zero. Thus, if the shared-rearing environment is significant, the average of the MZT and DZT correlations should exceed the average of the MZA and DZA correlations. Analogously, if the average of the MZT and DZT cross-correlations exceeds the average of the MZA and DZA cross-correlations, the correlation between the two phenotypes is mediated by significant shared-rearing environmental effects. The fourth parameter, nonshared environmental effects (Ens), refers to environmental effects that contribute to a lack of twin resemblance. The correlation for each of the twin types attributable to nonshared environmental effects is therefore zero. Unsystematic measurement error (unreliability), itself an effect that cannot contribute to the correlation between the twins, must by necessity inflate the estimate of the nonshared environment when no information is included in the analysis concerning reliability of measurement. Figure presents the bivariate model of trait and metatrait scores. The bivariate (or Cholesky) model is commonly used in quantitative genetic analyses (Neale & Cardon, 99) and allows for the decomposition of the correlation between two phenotypes into components of variance shared between the phenotypes and unique to one of the phenotypes. It should be noted that only sources of variance unique to one phenotype are identifiable in the Cholesky model with the four twin types; the metatraits were selected for the identification of unique effects because of the emphasis of this study o"n the properties of metatraits. Each of the four parameters discussed previously is included in the model. Path coefficients subscripted with represent effects unique to the metatrait scores. Path coefficients subscripted with or represent effects emanating from factors shared by the two phenotypes. Thus, the model of Figure implies the following variances of the traits and metatraits, as well as their covariance: Var, Var Ttai t = n u + a u + dn + s n. () = «l + fl i + ^ + * + «+ <Z + ^ +.S - () CV T + S,,S I. (3) The model of Figure also implies the following twin covariance traits (Equations 4 through 7) and twin covariances for the metatraits (Equations 8 through ): Cov M a,, + d n Cov DZTTniii =.5a n d n (4) (5)

5 METATRAITS 677 COV DZAM, = an + d x s l + a d + s.5a i + s l +.5fl + -5d + s = a l + d, + a + d.5a.5d t.5a.5d +d u d l +s n s i. Cov DZTltailJiltuli>it =.5a, ifl i + -5 d,, d x i +d u d l.5d u d l. (7) (8) (9) () () Specific to bivariate models are the cross-twin covariances between the two phenotypes: () (3) (4) (5) The model of Figure was also evaluated with LISREL 8. Covariances between the twin pairs, rather than correlations, were used in the model-fitting analyses. Comprehensive summaries of the univariate analyses of the personality trait scores may be found in Pedersen et al. (99) and Plomin and McClearn (99). Results Reliability and Stability of Metatraits To conduct the reliability and stability analyses of the metatraits, the covariances among the metatrait scores across the three occasions of measurement were required. The variances of the metatrait scores at each occasion, as well as the correlations among the occasions, are shown in Table. Comparable information is given for the traits in Table. The variances and covariances are based on the entire SATSA sample, regardless of whether the twin had a co-twin responding to the personality inventory. In each case, the average of the across-occasion trait correlations exceeds the average of the across-occasion metatrait correlations. However, on average, the metatrait stability correlations are at least moderate, with the metatrait factor exhibiting the highest level of stability. Using LISREL 8, tests were conducted to assess the equality of the variances and covariances of the traits and metatraits across the three occasions. As shown in Table, for of the metatraits, the variance of one occasion differed significantly from that of the other two occasions. In most cases, the largest variance occurred at the first occasion. For the trait variances in Table, all of the 4 traits had one occasion variance that differed significantly from the other two occasion variances, but again, the largest variance appeared most frequently at the first occasion. Not unexpectedly, as shown in Tables and, a majority of the covariances between the first and third occasions were significantly smaller than the two covariances between adjacent occasions. In a few instances, one of the covariances between adjacent occasions significantly exceeded the other two covariances; it is important to note, however, that in no instance was the covariance between the first and third occasions significantly greater than the other occasion covariances. Thus, the data conform to a simplex structure, which stipulates that the correlation between occasions closer together in time must be greater than between occasions further apart in time. Table 3 presents the results of the reliability and stability Table Variances and Intercorrelations Among Metatrait Scores Across Three Occasions Measure N o, Variances o o 3 O, Correlations O 3 o 3 Metatrait factor *.3**.55**.7*.76** **.9** 9 5** ** * **.5..3** ** ** ** * ** ** *.4.56**.533*.374**.4.4**.456**.8.373** **.56**.64** Note. A correlation noted as significant differed significantly from adjacent correlations. *p<. **p<.l.

6 678 S. HERSHBERGER, R. PLOMIN, AND N. PEDERSEN Table Variances and Intercorrelations Among Trait Scores Across Three Occasions Measure N o, Variances o 5.49* * * 5..83*'" 8.784*'*.73.6*"*.6.634*'* *" * *" ' " * *' 3.84* * * ** * " o ** ** ** * ** ** Note. A correlation noted as significant differed significantly from adjacent correlations. *p<. **p<.l. Correlations o **.673*.738**.58*.6*.64**.77** O,,.65**.685**.586**.65**.65** ** *.679** **.45*.479 Table 3 Reliability and Stability of Trait and Metatrait Scores Across Time Metatrait scores Trait scores Measure. ' XX X df r XX /3 X c V Metatrait factor I : >

7 METATRAITS 679 analyses of the trait and metatrait scores. The reliability of a variable is a function of the intercorrelations across the three occasions and is given by ( r ^ ^ / r u, for example, in the case of the neuroticism metatrait, ( X.53)/.44 =.57. The reliabilities of the metatraits vary widely, ranging from.8 for responsibility to for the metatrait factor. The trait reliabilities vary less widely, from.58 for luck to for hostility. For each variable, the reliability of the trait scores exceeded the reliability of the metatrait scores. The stability of the variables is, as in the case of reliability, a function of the correlations between occasions. For example, the stability of the neuroticism metatrait between the first two occasions (/8 i) is given by r 3 /r 3 and between the second and third occasions ( f e ) by r, 3 /r,, or.44/.53 = and.44/ =.9, respectively. The average of these two stability coefficients is equal to.88, the average of the maximum-likelihood estimates of the stability coefficients obtained from the model fitting. For a majority of the variables, the stabilities of the metatrait scores were within points of the stabilities of the trait scores, the trait stabilities generally exceeding the metatrait stabilities. Notably, the metatrait stabilities of life satisfaction, hard-driving, openness, state anxiety positive, and state anxiety negative were more than points lower than the comparable trait stabilities. All of the models fit from Table 3 had nonsignificant chi-squares, indicating that the reliability-stability model fit the data acceptably for all variables. In conclusion, although some of the test-retest reliabilities of the metatraits were barely acceptable, others were close to the generally more reliable trait scale scores. On the other hand, traits and metatraits differ less in stability, with the stability of metatraits on the whole only slightly less than that of traits. Correlation Between Traits and Metatraits One goal of the present analysis was to determine to what extent the correlation between traits and metatraits is mediated genetically. But a question that arises is this: Does the (linear) correlation between traits and metatraits capture adequately the relationship between the two? Past research suggested a strong nonlinear component to the correlation, resulting from the improbability of receiving a high metatrait score at the extremes of the trait level. To the degree that this is so, the linear correlation may underestimate the true relationship, possibly leading to the underestimation of shared genetic influence. Thus, to determine if a significant nonlinear component underlies the relationship between traits and metatraits, hierarchical multiple regressions were conducted, with the metatrait score first regressed onto the trait score alone and, in a second step, regressed onto both the trait score and the square of the trait score to detect the presence of a significant quadratic relationship. The cube of the trait score was added in the equation in a third step to detect a significant cubic relationship. Results from the regression analyses are presented in correlational form in Table 4. The correlations presented in the table are the semipartial correlations between the traits and metatraits from the step at which a term entered the equation. It is important to note that the linear relationship between the trait and metatrait scores has been partialed from the quadratic correlation and that the linear and quadratic have both been partialed from the cubic Table 4 Semipartial Correlations Between Trait and Metatrait Scores Measure 'Quadratic tubic N.54.Ola a a.8a a a.a.4a.a.6a -.a a.4a a.8a -a.7a.7a a Note, p < for all correlations unless indicated otherwise, a = not significant. correlation. This is important because, with the exception of luck, which did not have a significant linear, quadratic, or cubic correlation, each of the quadratic correlations and several of the cubic correlations were significant. Indeed, in most cases, the quadratic term not only added significantly to the regression equation, it also was significantly greater than the value of the linear term. Furthermore, three scales extraversion, anger, and responsibility had a significant quadratic but not a significant linear correlation. For these three variables, the relationship between traits and metatraits resembled an inverted U- shaped function, illustrated most dramatically by extraversion. Figure 3 shows the plot of extraversion metatrait scores against trait scores. The general shape of the function for variables with an additional, significant cubic term did not differ from those with a significant quadratic term, illustrated in Figure 4 by depression. Given the existence of significant nonlinear correlations for nearly all the scales, the question remained whether the magnitude of the shared genetic influence would be underestimated if the linear correlation were used for the analysis. To examine this, modeling analyses were conducted on both the linear and quadratic correlations for extraversion, a trait for which the linear correlation did not differ from zero and the quadratic correlation was near unity, and for life direction, for which the correlations were approximately equal. Behavioral Genetic Model-Fitting Analyses Model-fitting analyses of the model depicted in Figure were conducted with the covariances between twins, computed from a double-entry data file from the first wave of data collection, shown in Table 5. All twin data were corrected for the effects of

8 68 S. HERSHBERGER, R. PLOMIN, AND N. PEDERSEN.6 H I Trait Level i Figure 3. The relationship between extroversion metatrait scores and trait level scores. age, gender, and the Age X Gender interaction within each twin group separately. The standardized maximum-likelihood path coefficients from the modeling analyses are presented in Table 6. The chisquare goodness-of-fit test indicated a good fit of the model for 7 of the 4 variables. Note that the metatrait factor was analyzed under a univariate model. To calculate the proportion of genetic and environmental variance attributable uniquely to the metatrait scores, each of the paths leading to the metatrait were squared and summed. For example, for neuroticism, the proportion of variance.6- caused by additive genetic effects is a i + # ( =.3); the proportion of variance caused by nonadditive effects is cfei + fi? ( =.3); the proportion of variance caused by the shared-rearing environment is 5 i + s (.3 + ) = ; and the proportion of variance caused by nonshared environmental effects is «i + «( =.6). The results of these univariate calculations for the 4 metatraits are presented in Table 7. On average, 5% of the variance in the metatrait scores was attributable to genetic influence (6% to additive effects and 9% to nonadditive effects); 5% to shared-rearing environmental in Trait Level Figure 4. The relationship between depression metatrait scores and trait level scores.

9 METATRAITS 68 Table 5 Twin Intraclass Correlations for the Metatraits Metatrait Metatrait factor MZT (9-57) DZT (9-) MZA (6-95) DZA (3-8) Note The sample sizes refer to twin pairs. Also, the sample sizes for the metatrait factor were somewhat lower in comparison to the other variables: MZT (58); DZT (8); MZA (35); DZA (68). Table 6 Standardized Path Coefficients From Bivariate Analysis of Trait and Metatrait Scores Common factors Unique factors Measure «n On dn S\\ " «4 i S ««dn * X Metatrait factor * 53.44** * * ** 77.7*** * Note. All df= except for the Metatrait factor, for which df= 4. */><. **p<.. ***p<.

10 68 S. HERSHBERGER, R. PLOMIN, AND N. PEDERSEN Table 7 Proportions of Variance Attributable to Genetic and Environmental Sources for Metatraits Metatrait Metatrait factor A D Es 5 Is) Ens Note. A = additive genetic variance; D = dominance variance; E s = shared environmental variance; E m = nonshared environmental variance. fluence; and 8% to nonshared environmental influence. Although the value of 5% for the average degree of variance explained by genetic sources may appear to be on the low side, genetic influence varied dramatically among the metatraits, ranging from a total of 36% for life satisfaction to % for distress, sociability, and openness to experience. Thus, whereas genetic influence may be relatively unimportant for some metatraits (as is the case for some trait scale scores), it is quite important for others. Shared-rearing environment was on the average low, similar to its magnitude for the trait scale score, although the value of % for the sociability metatrait is nearly twice the value found for the sociability scale score (cf. Plomin, Pedersen, McClearn, Nesselroade, & Bergeman, 988). Not unexpectedly, the proportion of variance caused by nonshared environmental effects is high throughout, in part caused by unreliability of measurement. To calculate the proportion of genetic influence that is common to traits and metatraits, and the proportion of genetic variance that is unique to the metatrait, the genetic path coefficients that lead to the metatrait (e.g., a i, a, d l, d ) were squared and summed. This again represents the total amount of genetic variance affecting the metatrait. The square of fhe path coefficients a and d, paths that are unique to the metatrait, is an estimate of the metatrait's genetic variance that is not shared with the trait. Therefore, (a + d )/(a + a i + d + d ) is an estimate of the proportion of genetic variance that is unique to the metatrait. The remainder of the genetic variance is due to genetic influence shared with the trait. For example, the sum of the four squared genetic path coefficients leading to impulsivity (a + a + d, + d or OO ) is.3 (the heritability of the impulsivity metatrait). The square of the genetic paths unique to impulsivity {a + d or ) is.4. Then.4/.3, or 48% of the variance affecting the impulsivity metatrait is unique to the metatrait. The residual, 5%, is the proportion of genetic variance the metatrait shares with the trait. The results of these calculations for the common and unique genetic proportions of variance are shown in Table 8. According to the results of Table 8, an average of approximately 3% of the genetic variance is unique to the metatrait scores; 69% of the genetic variance is held in common with the trait scale scores. The genetic variance unique to the metatrait scores appears to be less than half the magnitude of the genetic variance shared with the trait scores. However, a number of the proportions in Table 8, particularly several % proportions under the column for common variance, are not significant on the basis of the results of chi-square difference tests removing the relevant parameters from the model. For example, although % of the genetic variance is held in common for activity and openness to experience, the total amount of genetic variance is extremely small, 3% and nearly %, respectively. If only the variables with significant common genetic variance are averaged, the average rises from 69% to 88%. On the other hand, if a comparable average is calculated only for variables with significant unique genetic variance, the average rises from 39% to 84%. Notably, only one variable, impulsivity, had both significant common and significant unique genetic variance. Separate modeling analyses were conducted with the covariance between extraversion and life direction metatrait scores and squared trait scale scores (with the linear component par- Table 8 Proportion of Genetic Influence Common to Trait and Metatrait Scores and Unique to Metatrait Scores Measure Common * * * * 66* 7* * * * * 4 86* 78* Note. Significant proportions based on Xdartest, df=. *p<. Unique * 4 8* 6 48* * 86* 9* 4

11 METATRAITS 683 tialed from each). The common path coefficients from the analysis of extraversion were n i = -.8, a i = -, d x = -.49, and S =.4; and the unique coefficients were «= -.8, =, d =., and s =. The proportion of genetic variance unique to the extraversion metatrait is now %, and the proportion of genetic variance held in common with the extraversion scale score is now %, a complete reversal of the results found from analyzing the linear covariance. For life direction, the common path coefficients were n l = -.7, a i = -., d x = -, and s i = -.8; and the unique coefficients were n =.88, a =, d =, and s =.4. Here, % of the genetic variance held is shared between the life direction trait and metatrait, identical to the results obtained from performing the model-fitting analyses with the linear covariances. Thus, using the linear covariance when there is a significant quadratic component does not necessarily lead to the underestimation of the magnitude of the genetic variance common to the trait and metatrait. An extreme example (extraversion) in which the linear correlation did not differ from zero was deliberately selected to emphasize the caution that must be exercised in interpreting the results of linear correlations between traits and metatraits. If both the linear and quadratic components contributing to the association between the trait and metatrait level are of equivalent magnitude, no bias toward underestimating the genetic variance held in common is introduced. It could also be argued that the failure of seven of the bivariate models reported in Table 6 to fit the data could be due to the presence of a significant nonlinear component (either a quadratic or cubic or both) to the correlation between the trait and metatrait. When model-fitting analyses were performed on the quadratic covariances for these seven variables, only two models, neuroticism {\ =.8, ns) and extraversion (X =, ns), now fit the data well. In no case among the seven variables did modeling the cubic covariances result in a goodfitting model. Therefore, the failure of the seven bivariate models to fit the data using the linear covariances cannot be largely attributed to the presence of nonlinearity. Discussion Metatraits have received increasing attention from personality researchers since publication of the seminal paper by Bern and Allen (974). Yet a number of aspects of metatraits remain relatively unexamined. This is certainly not the case for traits, the total scale score on an inventory; studies of traits are endemic to the field of personality psychology, in which their measurement characteristics have been well described. Within behavioral genetic research also, the focus has almost exclusively been on the genetic and environmental sources of individual differences in traits. This study sought to add to the growing information on metatraits by exploring their reliability and stability over time and by exploring their genetic and environmental structure, independent of traits. The test-retest reliabilities of the 4 metatraits examined in this study were decidedly modest, with the exception of a metatrait factor, which had a reliability of. The test-retest reliabilities of the traits were uniformly higher than their metatrait counterparts, a result consistent with that of Baumeister (99): The average reliability of the traits was.7; the average reliability of the metatraits was.49. However, Baumeister's observation that the metatrait reliabilities were as close to the trait reliabilities as the trait reliabilities were to was not borne out. Indeed, Baumeister assumes from this relationship between the reliabilities of the traits and metatraits that the metatrait scores are as reliable as the trait scores, an assumption that is based on the unproven claim that the reliability of the trait serves as a "ceiling" for the reliability of the metatrait. Although in some cases this relationship between the reliabilities was found in this study, in other cases it was not. For example, the reliability of the extraversion trait was.75 (a.5 difference from ), and the reliability of the extraversion metatrait was.45 (a.3 difference from.75). Thus, the relative similarity mentioned by Baumeister obtains in this case. However, the reliability of the state anxiety total trait was.68 (a.3 difference from ), and the reliability of the state anxiety total metatrait was (a.3 difference from.78). We cannot conclude, as Baumeister did, that metatrait scores are as reliable as trait scores, but for several metatraits, the reliabilities are reasonable. On the whole, the stabilities of the metatraits fared better than their reliabilities, with an average stability of.85; the average trait stability was.9. Nonetheless, the metatrait stabilities covered a broad range of values, from.44 (for life satisfaction) to (depression). Some metatraits are extremely stable, whereas others are much less so. On the other hand, more uniformity existed among the values of the trait stabilities, which ranged from (for alienation) to. (for hard-driving). It can be inferred, if only for metatraits showing high stability, that individual differences in metatraits remain relatively constant over time. Low metatrait (inconsistent) people tend to remain low metatrait people across time, and high metatrait (consistent) people tend to remain high metatrait people across time. Thus, there is evidence for the predictability of inconsistency as well as consistency from one occasion to the next. Quantitative genetic analyses of the metatraits showed, on the average, 5% of the phenotypic variance to be attributable to genetic influence, 5% to shared-rearing environmental influence, and 8% to nonshared environmental influences. Although the magnitude of the shared-rearing environment is consistent with past quantitative genetic analyses of trait scale scores, the magnitude of the genetic effects is smaller, and the magnitude of the nonshared environment is larger. One possible source of this discrepancy might lie with the lower reliabilities of the metatraits. Unreliability of measurement contributes to the estimate of the nonshared environment, to the proportional decrement of the genetic effects. Bivariate quantitative genetic analyses were conducted on the trait and metatrait scores with the primary purpose of assessing how much of the metatrait's genetic influence could be uniquely attributed to the metatrait and how much was held in common with the trait. One difficulty with this analysis stemmed from the nonlinearity of the correlations between the traits and metatraits. Linear correlations cannot capture the complete magnitude of the association between two nonlinearly related variables. One possible result of using a linear correlation is the underestimation of the amount of genetic influence common to both variables. Using linear correlations, it was found that on the average, 3% of the genetic variance affecting the metatrait was unique to the metatrait, with the remaining 69% shared with the trait. Recognizing only variables with significant common and unique genetic variance yielded values of 88% and

12 684 S. HERSHBERGER, R. PLOMIN, AND N. PEDERSEN 84%, respectively. Bivariate analyses of two traits using quadratic correlations with the linear component removed showed that if the quadratic component contributing to the relationship between the variables was about equal to the linear component, the quantitative genetic results did not differ substantially. However, if the quadratic component was substantially larger, a reduction in the estimated amount of shared genetic influence between the traits and metatraits occurred, with an equivalent enhancement in the magnitude of the metatrait's unique genetic influence. In this study, the quadratic correlation between the traits and metatraits was substantially larger than the linear correlation in a minority of the cases. Thus, the analysis of the linear correlation provides an accurate description of the amount of genetic influence held in common between trait and metatraits. Two important implications follow from the results of this study. One is that the reliability and stability of metatraits can be respectable, and depending on the phenotype, researchers can use them with as much confidence as trait scale scores. The second implication is that the concept of a metatrait provides a viable, additional view of personality. This is supported not only by the high stability of some of the metatraits across time, but also by the unique genetic influence, independent of the trait levels, found for the metatraits. The common finding of a significant association between traits and metatraits has led some researchers (Burke et al., 984; Tellegen, 988) to dismiss the interitem standard deviation as an appropriate method of measuring metatraits. Although better methods may be possible, the existence of a significant correlation between traits and metatraits does not negate the unique information brought to the investigation of personality metatraits. As long as the correlation falls below unity, metatraits have unique, reliable variance to contribute to the prediction of individual differences in personality, beyond the variance contributed by the level of behavior. A useful future direction for metatrait research to take concerns identifying sources of influence common to two or more metatraits. Multivariate behavioral genetic models could partition the observed correlation between metatraits into shared genetic and environmental sources of variance, as the present study did for the observed correlation between a trait and metatrait. Undoubtedly, substantial correlations exist among metatraits, as evidenced by the metatrait factor used in this study, which accounted for an average of 4% of the variance. Multivariate behavioral genetic modeling has successfully uncovered the genetic and environmental correlations underlying the association among many traits; there is no reason to assume the same could not be accomplished for metatraits. References Allen, M. J., & Yen, W. M. (979). Introduction to measurement theory. Belmont, CA: Wadsworth. Baumeister, R. E. (99). On the stability of variability: Retest reliability of metatraits. Personality and Social Psychology Bulletin, 7, Baumeister, R. E., & Tice, D. M. (988). Metatraits. Journal ofpersonality, 56, Bern, D. J., & Allen, A. (974). On predicting some of the people some of the time: The search for cross-situational consistencies in behavior. Psychology Review, 8, Bergeman, C. S., Chipuer, H. M., Plomin, R., Pedersen, N. L., McClearn, G. E., Nesselroade, J. R., Costa, P. T., Jr., & McCrae, R. R. (993). Genetic and environmental effects on openness to experience, agreeableness, and conscientiousness: An adoption/twin study. Journal of Personality, 6, Bouchard, T. J., Jr., & McGue, M. (99). Genetic and rearing environmental influence on adult personality: An analysis of adopted twins reared apart. Journal of Personality, 58, Burke, P. A., Kraut, R. E., & Dworkin, R. H. (984). Traits, consistency, and self-schemata: What do our methods measure? Journal of Personality and Social Psychology, 47, Buss, A. H., & Plomin, R. (984). Temperament: Early developing personality traits. Hillsdale, NJ: Erlbaum. Cederlof, R., & Lorich, U. (978). The Swedish twin registry. In W. E. Nance, G. Allen, & P. Parisi (Eds.), Twin research: Part C. Biology and epidemiology (pp ). New York: Alan R. Liss. Chaplin, W. E, & Goldberg, L. R. (985). A failure to replicate the Bern and Allen study of individual differences in cross-situational consistency. Journal of Personality and Social Psychology, 47, Cook, W. W., & Medley, D. M. (954). Proposed hostility and pharisaic-virtue scales for the MMPI. Journal of Applied Psychology, 38, Costa, P. T, Jr., & McCrae, R. R. (985). The NEO Personality Inventory Manual. Odessa, FL: Psychological Assessment Resources. Duke University Center for the Study of Aging and Human Development. (978). Multidimensional functional assessment: The OARS methodology. Durham, NC: Duke University Medical Center. Epstein, S. (983). Aggregation and beyond: Some basic issues in the prediction of behavior. Journal of Personality, 5, Floderus, B. (974). Psycho-social factors in relation to coronary heart disease and associated risk factors. Nordisk Hygienisk Tidskrift [Monograph, Suppl. 6], -48. Gatz, M., Pedersen, N. L., & Harris, J. (987). Measurement characteristics of the mental health scale from the OARS. Journal of Gerontology, 4, Haynes, G. S., Levine, S., Scotch, N., Feinleib, M., & Kanner, W. B. (978). The relationship of psychosocial factors to coronary heart disease in the Framingham study. American Journal of Epidemiology, 7, Heise, D. R. (969). Separating reliability and stability in test-retest correlation. American Journal of Sociology, 75, 93-. Joreskog, K. G., & Sorbom, D. (993). LISREL 8 user's reference. Chicago: Scientific Software. Klockars, A. J., & Varnum, S. W. (975). A test of the dimensionality assumptions of Rotter's Internal-External scale. Journal of Personality Assessment, 39, McCrae, R. R., & Costa, P. T., Jr. (985). to experience. In R. Hogan & W. H. Jones (Eds.), Perspectives in personality (Vol., pp. 45-7). Greenwich, CT: JAI Press. Neale, M. C, & Cardon, L. R. (99). Methodology for genetic studies of twins and families. Dordrocht, The Netherlands: Kluwer. Mischel, W. (968). Personality and assessment. London: Wiley. Paunonen, S. V. (988). Trait relevance and the differential predictability of behavior. Journal of Personality, 56, Paunonen, S. V., & Jackson, D. N. (985). Idiographic measurement strategies for personality and prediction: Some unredeemed promissory notes. Psychological Review, 9, Pedersen, N. L., Gatz, M., Plomin, R., Nesselroade, J. R., & McClearn, G. E. (989). Individual differences in locus of control during the second half of the lifespan for identical and fraternal twins reared apart and reared together. Journal of Gerontology: Psychological Sciences, 44, -5. Pedersen, N. L., Lichtenstein, P., Plomin, R., DeFaire, U., McClearn, G. E., & Matthews, K. A. (989). Genetic and environmental influ-

13 METATRAITS 685 ences for Type A-like and related traits: A study of twins reared together and twins reared apart. Psychosomatic Medicine, 5, Pedersen, N. L., McClearn, G. E., Plomin, R., Nesselroade, J. R., Berg, S., & DeFaire, U. (99). The Swedish Adoption/Twin Study of Aging: An update. Ada Geneticae Medicae et Gemmellologiae, 4, 7-. Pedersen, N. L., Plomin, R., McClearn, G. E., & Friberg, L. (988)., extraversion, and related traits in adult twins reared apart and reared together. Journal of Personality and Social Psychology, 55, Plomin, R., Chipuer, H. M., & Loehlin, J. C. (99). Behavioral genetics and personality. In L. A. Pervin (Ed.), Handbook of personality: Theory and research (pp. 5-43). New York: Guilford Press. Plomin, R., DeFries, J. C, & McClearn, G. E. (99). Behavior genetics: A primer. New York: Freeman. Plomin, R., & McClearn, G. E. (99). Human behavioral genetics of aging. In J. E. Birren & K. W. Schaie (Eds.), Handbook of the psychology of aging (pp ). New York: Academic Press. Plomin, R., Pedersen, N. L., McClearn, G. E., Nesselroade, J. R., & Bergeman, C. S. (988). EAS temperaments during the last half of the life span: Twins reared apart and twins reared together. Psychology and Aging, 3, Rotter, J. B. (966). Generalized expectancies for internal versus external control of reinforcement. Psychological Monographs, 8 (, Whole No. 69). Schalling, D., Edman, G., & Asberg, M. (983). Impulsive cognitive style and inability to tolerate boredom: Psychological studies of temperamental vulnerability. In M. Zuckerman (Ed.), Biological bases of sensation seeking, impulsivity and anxiety. Hillsdale, NJ: Erlbaum. Spielberger, C. D. (979). Preliminary manual for the State-Trait Personality Inventory (STPI). Tampa: University of Florida. Tellegen, A. (988). The analysis of consistency in personality assessment. Journal of Personality, 56, Tice, D. M. (989). Metatraits: Interitem variance as personality assessment. In D. M. Buss & N. Cantor (Eds.), Personality psychology: Recent trends and emerging directions (pp. 94-). New York: Springer-Verlag. Waller, N. G., & Reise, S. P. (99). Genetic and environmental influences on item response pattern scalability. Behavior Genetics,, Wood, V., Wylie, M. L., & Sheafor, B. (969). An analysis of a short selfreport measure of life satisfaction: Correlation with rater judgment. Journal of Gerontology, 4, Received January 7,993 Revision received February 8,995 Accepted February 4, APA Convention Call for Programs The Call for Programs for the 996 APA annual convention appears in the September issue of the APA Monitor. The 996 convention will be held in Toronto, Ontario, Canada, from August 9 through August 3. The deadline for receipt of program and presentation proposals is December, 995. Additional copies of the Call are available from the APA Convention Office, effective in September. As a reminder, agreement to participate in the APA convention is now presumed to convey permission for the presentation to be audiotaped if selected for taping. Any speaker or participant who does not wish his or her presentation to be audiotaped must notify the person submitting the program either at the time the invitation is extended or before the December deadline for proposal receipt.