Sex differences in the genetic and environmental influences on the development of antisocial behavior

Transcription

1 Development and Psychopathology, 14 (2002), Copyright 2002 Cambridge University Press Printed in the United States of America Sex differences in the genetic and environmental influences on the development of antisocial behavior KRISTEN C. JACOBSON, a,b CAROL A. PRESCOTT, a,b AND KENNETH S. KENDLER a,b a Virginia Institute for Psychiatric and Behavioral Genetics; and b Medical College of Virginia/Virginia Commonwealth University Abstract The present study uses a population-based sample of 6,806 adult twins from same-sex and opposite-sex twin pairs to examine sex differences in the underlying genetic and environmental architecture of the development of antisocial behavior (AB). Retrospective reports of AB during three different developmental periods were obtained: prior to age 15 years (childhood), age years (adolescent), and age 18 years and older (adult). Structural equation modeling analyses revealed that there was no evidence for sex-specific genetic or sex-specific shared family environmental influences on the development of AB; that is, the types of genetic and environmental influence were similar for males and females. For both sexes, a model that allowed for genetic influences on adolescent and adult AB that were not shared with childhood AB fit better than a model with a single genetic factor. In contrast, shared environmental influences on adolescent and adult AB overlapped entirely with shared environmental influences on childhood AB. Genetic factors played a larger role in variation in childhood AB among females, whereas shared environmental factors played a larger role among males. However, heritability of AB increased from childhood to adolescence and adulthood for both sexes, and the magnitude of genetic and environmental influences on adolescent and adult AB was approximately equal across sex. We speculate that sex differences in timing of puberty may account for the earlier presence of genetic effects among females. One of the more intriguing results from re- vergence of findings between studies of search on genetic and environmental influences childhood and adolescent AB (e.g., conduct on antisocial behavior (AB) is the di- disorder [CD] and delinquency) and studies of adult AB (e.g., antisocial personality disorder and criminality). Research using children and This work was supported by National Institutes of Health adolescents typically find that both shared Grants MH-40828, AA-09095, and MH/AA and a rearing environments and genetic factors are Research Scientist Award (MH-01277) to Dr. Kendler and the Rachel Brown Banks Endowment Fund. The Virginia a significant source of family resemblance for Twin Registry, now the Mid-Atlantic Twin Registry AB (Edelbrock, Rende, Plomin, & Thompson, (MATR), was established by Walter Nance, M.D., Ph.D., 1995; Eley, Lichtenstein, & Stevenson, 1999; and is maintained by Linda Corey, Ph.D. and Lenn Murre- Rowe, 1986). In contrast, studies of adult lle, Ph.D., and is supported by the U.S. National Institutes populations often find that genetic factors are of Health, the Carmen Trust, and the W. M. Keck, John Templeton, and Robert Wood Johnson Foundations. We the sole source of familiality (Cadoret, 1974; thank Sarah Woltz, M.A., Frank Butera, M.S., Lisa Halber- Cloninger & Gottesman, 1987; Crowe, 1974; stadt, M.S., Barbara Brooke, M.S.W., and Patsy Waring DiLalla & Gottesman, 1989; Mednick, Gabrifor their assistance with data collection and maintenance. elli, & Hutchings, 1984). A meta-analysis of Address correspondence and reprint requests to: Kris- 24 twin and adoption studies of aggression reten C. Jacobson, Ph.D., 800 E. Leigh Street, PO Box , Richmond, VA ; jacobson@ vealed that age of subjects was a significant physio.vipbg.vcu.edu. predictor of both heritability and shared envi- 395

2 396 K. C. Jacobson, C. A. Prescott, and K. S. Kendler ronmental influences (Miles & Carey, 1997). Segal, & Lykken, 1990). However, the 95% Specifically, genetic and shared environmental confidence intervals surrounding these estivariation influences contributed about equally to mates were large, and results from this small in aggression among studies of children and rather unusual sample were not replicated and adolescents, whereas, among studies in a second, large-scale study that used retro- of adults, the average heritability was substantial, spective reports of AB from a sample of 3,226 accounting for upwards of 50% of varia- male twin pairs from the Vietnam Twin Regspective tion, and no shared environmental influences istry (Lyons, True, Eisen, Goldberg, Meyer, were detected. Faraone, Eaves, & Tsuang, 1995). This second study supported the hypothesis of increasing genetic effects on AB. Genetic influences Developmental Changes in Genetic on juvenile antisocial traits were quite mod- Influence on Antisocial Behavior est, explaining only 7% of the variance, compared However, the drawing of conclusions about to 43% of variance explained by genetic developmental changes in the relative influ- factors for adult antisocial traits. Conversely, ence of genetic and shared environmental factors shared environmental influences explained on AB from cross-study comparisons is 31% of the variation in juvenile antisocial problematic, because studies use different traits, but only 5% of the variation in adult samples and different measures of AB. These antisocial traits. differences in methodologies may introduce In the Lyons et al. (1995) study, the corre- systematic biases. In addition, studies that fo- lation between juvenile and adult antisocial cus on only one time point (i.e., childhood, traits was.44, and genetic and shared environ- adolescence, or adulthood) cannot determine mental influences each accounted for approximately the source of potential increases in genetic influence. one-third of this correlation. This study For example, heritability of AB may also found that the same set of genetic and increase in adulthood because of the presence shared environmental factors accounted for of new genetic influences on adult AB that variation in both juvenile and adult antisocial are not shared with child and adolescent AB. traits; that is, there were no genetic or shared Alternatively, the same set of genetic factors environmental influences specific to either juvenile may influence AB at all time points, but the or adult antisocial traits. Thus, the magnitude of the genetic influence may in- greater heritability estimate for adult antisocial crease with age. Finally, genetic influences traits was due to an increase in the magnicrease might remain constant across time while envi- tude of the genetic influence on adult antisocial ronmental influences decrease in importance. traits, rather than to the presence of new This too would account for an increase in her- genetic influences. itability. Ideally, prospective, longitudinal, genetically However, this study had two important informative studies of AB are needed limitations: first, only two points in time were to disentangle these effects. Although several used with a cutoff point of age 15, so the mea- longitudinal twin and adoption studies are sure of adult antisocial traits included behaviors in progress (Hewitt et al., 1997; Plomin & during middle and late adolescence, DeFries, 1983), none currently have data from as well as adult behaviors. Because most adolescent both childhood and adulthood. Thus, a second twin siblings live together until age strategy is to collect retrospective reports of 18, combining behavior from ages 15 to 17 child and adult AB using the same adult with the behavior after age 18 may have over- sample. stated the continuity of shared environmental At least two prior studies have used this factors between juvenile and adult AB. Likemethod. A study of 32 monozygotic twin wise, combining adolescent and adult behaviors pairs reared apart reported heritability estimates may have obscured any new genetic inpairs of.41 and.28 for childhood and adult fluence that occurs after age 18. Second, antisocial personality disorder (APD), respectively because the Vietnam Twin Registry was re- (Grove, Eckert, Heston, Bouchard, stricted to male twins, sex differences could

3 Sex differences in antisocial behavior 397 not be examined. It is possible that the genetic Bohman, & Knorring, 1982) Thus, evidence and environmental architecture underlying the for sex differences in the heritability of AB is development of AB differs for males and fe- equivocal. In addition, no published study has males. There is substantial evidence for sex investigated sex differences in the underlying differences in mean levels of both childhood genetic and environmental influences on the CD and adult ASP (Cohen, Cohen, Kasen, development of antisocial behavior over time. Velez, Hartmark, Johnson, Rojas, Brook, & Streuning, 1993; Kessler, McGonagle, Zhao, Nelson, Hughes, Eshleman, Wittchen, & Kendler, The Present Study 1994; Robins & Reiger, 1991; Rutter, The present study uses data on behavior from Giller, & Hagell, 1998; Simonoff, Pickles, three different developmental periods (prior to Meyer, Silberg, Maes, Loeber, Rutter, Hew- age 15, between the ages of 15 and 17, and itt, & Eaves, 1997), but it is less clear whether age 18 and older) to examine four primary sex differences in mean level are associated questions concerning sex differences in the with sex differences in the sources of individual genetic and environmental architecture under- differences in AB. lying the development of AB. First we test whether the genetic and shared environmental Sex Differences influences that impact AB are the same for males and females (a qualitative sex difference). To date, only a few studies have examined For example, it has been suggested that sex differences in genetic and environmental the greater prevalence of AB among males influences on AB. One of the first studies of may be due to genes related to sex-specific behavioral deviance using an unselected sam- hormones such as testosterone. If so, genetic ple of 13-year-old twins found stronger evidence factors influencing AB should overlap only for genetic influences among males partially, if at all, for males and females, be- than among females (Graham & Stevenson, cause these genes would be expressed only 1985). A more recent large-scale study using among males. Similarly, we can test whether two separate samples of adolescent twins the shared environmental factors that influ- found that the heritability of nonaggressive ence AB are the same for males and females. delinquent behavior was higher among females Second, we examine whether genetic and en- than among males, although no sex dif- vironmental factors exert similar magnitudes ferences emerged for aggressive delinquent of influence on AB for males and females (a behavior (Eley et al., 1999). Still other twin quantitative sex difference). For example, if studies have failed to find evidence for significant genes activated at puberty are an important sex differences in the heritability of AB influence on AB, genetic influences on early (e.g., Slutske, Heath, Dinwiddie, Madden, adolescent AB might be stronger for females Bucholz, Dunne, Statham, & Martin, 1997). than for males, given the earlier age of puberty Results from adoption studies are similarly experienced by females. Third, we in- conflicting. A recent adoption study found vestigate the underlying structure of genetic that CD among adopted males was predicted and shared environmental influences on the by adoptive family environment alone, while development of AB. Specifically, we test biological background and gene environment whether a single genetic factor, and/or a sin- interactions predicted CD among females, gle shared environmental factor, can account suggesting greater genetic influence on CD for the variation in AB at all three time points among females (Langbehn, Cadoret, Yates, and, conversely, whether genetic influences Troughton, & Stewart, 1998). However, results on antisocial behavior at the three time points from earlier adoption studies suggested are completely independent. Finally, we ex- that the same genetic factors predicted antisocial amine whether the same underlying genetic behavior in both males and females (e.g., and environmental structure can account for Baker, Mack, Moffitt, & Mednick, 1989; Cadoret & Cain, 1980; Sigvardsson, Cloninger, continuity and change in AB for both males and females.

4 398 K. C. Jacobson, C. A. Prescott, and K. S. Kendler Previous analyses using the same-sex fe- least partly heritable (Plomin & McClearn, male twins from the present study found that 1993; Rowe, 1994). To the extent that these the heritability of CD symptoms up to age 18 characteristics also influence the development was.41, and shared environmental influences of AB, one might predict that a common set did not significantly differ from zero (Gold- of genes would influence AB at all three time stein, Prescott, & Kendler, 2001). In contrast, points. Thus, models constraining genetic in- a paper using a two-wave measurement model fluences on antisocial behavior to be independent to assess genetic and environmental influences across time periods should not fit the on CD among the same-sex male twins data well. In contrast to LCP antisocial behavior, in the present study reported that genes and AL delinquency is considered to be norin shared environments both explained approxi- mative, particularly among males, and may be mately one third of the variation in CD, and most strongly influenced by factors such as both estimates were significantly greater than peer group composition, parental discipline zero (Jacobson, Prescott, & Kendler, 2000). and monitoring, and structural factors at the However, it should be noted that these reports family, school, and community level (Moffitt, were independent analyses of the same-sex 1993). This would suggest that environmental pairs only, and thus cannot answer the ques- influences, both shared and nonshared, would tion of sex differences directly. be more important for variation in child and Results from the present study may have adolescent AB than for adult AB. implications for developmental taxonomies of However, it should be noted that the design antisocial behavior such as those proposed by of the present study does not directly test Moffitt (1993) and DiLalla and Gottesman hypotheses concerning developmental taxo- (1989). Longitudinal research has shown that nomies. Predictions pertaining to different genetic although the majority of individuals diagnosed and environmental etiologies of LCP with adult APD met criteria for CD in and AL individuals require either person-centered childhood and adolescence, most individuals analyses that focus on group differences diagnosed with CD do not go on to become or growth curve analyses. Thus, although results antisocial adults (Robins, 1978). This unidirectional from this study might be consistent with effect, coupled with the finding that predictions based on developmental typologies, delinquent behavior in adolescence is a nearly they do not test them directly. Instead, the universal phenomenon (e.g., West & Farring- present study focuses on a different developmental ton, 1973) has led a number of researchers to issue, namely, estimating the timing propose the existence of two major types of and relative importance of genetic and environmental delinquents: transitory or adolescent-limited influences on the development of (AL) delinquents, whose antisocial be- AB and testing whether these influences are havior is limited to the adolescent years, versus similar for males and females. continuous or life-course-persistent (LCP) delinquents, whose antisocial behavior begins at a younger age and continues from Methods adolescence into adulthood (DiLalla & Gottesman, 1989; Moffit, 1993). In addition to Sample and procedure showing different patterns of AB across the Data are from two longitudinal studies of psychiatric life span, these two groups of individuals are disorders in adult twins: a four-wave further surmised to have different genetic and longitudinal study of female female twins environmental etiologies (Rutter, MacDonald, (the FF study; Kendler, Neale, Kessler, Le Couteur, Harrington, Bolton, & Bailey, Heath, & Eaves, 1992) and an ongoing threewave 1990; Wilson & Herrnstein, 1985). study of male male and male female In particular, the antisocial behavior of twins (the MMMF study; Kendler & Prescott, LCP individuals is thought to be influenced, in part, by relatively stable cognitive and personality characteristics, many of which are at 1999). Twins were ascertained via the Virginia Twin Registry (VTR, now part of the Mid-Atlantic Twin Registry). The VTR was

5 Sex differences in antisocial behavior 399 formed by a systematic search of all Virginia In total, we received SRQs from 6,823 individual birth certificates since Twins were eligible twins. Eight twins were eliminated for participation if one or both twins because of missing information concerning could be successfully matched to state records, zygosity, and 9 twins were eliminated because they were Caucasian, and they were of missing data for all three measures of anti- born between 1934 (FF) or 1940 (MMMF) social behavior. The remaining 6,806 twins and Inclusion in the FF study also re- included the following: 2,580 twin pairs in quired that both twins in a pair return a mailed which both twins had data on antisocial be- questionnaire, whereas MMMF participants havior (346 monozygotic female female were first recruited by a telephone interview. [MZF] pairs, 212 dizygotic female female Initial response rates were 64% (FF) and 73% [DZF] pairs, 635 monozygotic male male (MMMF). Both studies were approved by the [MZM] pairs, 432 dizygotic male male local Institutional Review Board, and subjects [DZM] pairs, and 955 dizygotic male female were informed about the goals of the study [DZOS] pairs), 8 sets of triplets (creating 1 and provided verbal consent prior to phone in- DZF, 9 MZM, 4 DZM, and 10 DZOS pairs), terviews and written consent prior to in-person and 1,622 twins whose cotwins did not have interviews and collection of DNA sam- valid data on antisocial behavior (206 MZF, ples. 169 DZF, 299 MZM, 322 DZM, and 626 Data for the present study come from self- DZOS). Twins had an average of 13.5 (SD = report questionnaires (SRQ) that were part of 2.7, males) and 14.0 (SD = 2.5, females) years the wave 4 (FF) or wave 2 (MMMF) data col- of education at the time of the SRQ. Female lection. The FF study includes 2,164 twins twins were slightly younger than male twins originally interviewed at the first wave of data (M female = 36.7, SD = 8.7; M male = 37.1, SD = collection in and 275 twins who 9.1), F (1, 6,792) = 6.5, p <.05, and MZ twins were ascertained and studied subsequently. Of were younger than DZ twins (M MZ = 36.1, these 2,439 twins, 1,934 (79%) were successfully SD = 8.9; M DZ = 37.4 SD = 8.9), F (1, 6,792) = reinterviewed via telephone at the fourth 36.2, p <.001. However, the interaction be- wave of data collection in Wave tween sex and zygosity for age was not signif- 4 participants were also sent an SRQ and were icant, F (1, 6,792) = 0.02, p >.50, indicating asked to fill it out and mail it back later. We that the difference in age between MZ and DZ received SRQs from 1,497 (77%) twins inter- twins was similar for males and females. viewed at wave 4. The majority of SRQs Thus, our results concerning sex differences (85%) were returned within 3 months of the in estimates of heritability and shared environmental wave 4 interview. A minority of SRQs (1%) influences should not be biased by were answered orally. the slight age differences between males and The MMMF study includes 6,847 twins females. originally interviewed in Eightythree percent (N = 5,651) completed a second wave interview in , and 5,326 Measures (94%) of those interviewed at wave 2 also completed an SRQ. The majority (80%) of Zygosity. Zygosity of same-sex twin pairs wave 2 interviews were conducted face-toface, who both participated at the initial assessment with the SRQ filled out during the mid- was determined by a combination of twins dle of the interview. An additional 1.8% completed responses to standard questions regarding twin the SRQ prior to the wave 2 interview similarity, photographs, and DNA typing. As- ([M] = 8.6 days, SD = 19.7), and 18.2% returned signment of zygosity for twins whose samemonths, it after the wave 2 interview (M = 3.53 sex cotwins did not cooperate at wave 1 was SD = 8.9). In approximately one- done using a discriminant function analysis of quarter of cases, the SRQ was answered orally, most often because the wave 2 interview was given over the phone. items regarding physical similarity and twin self-report of zygosity, with DNA-typed twins as the comparison group.

6 400 K. C. Jacobson, C. A. Prescott, and K. S. Kendler Child and adolescent antisocial behavior. sent (0) if none of the criteria were met. Items concerning antisocial behavior were Symptoms were then summed to create a conidentical in the FF and MMMF studies. Three tinuous measure, with a possible range of measures of antisocial behavior (AB) were 0 9. used: AB prior to age 15 years (child), AB For all three measures of AB, if a respondent age years (adolescent), and AB age 18 had missing data for a particular item, years and older (adult). For AB prior to age then he or she was given a score of 0 for that 15 years, twins were asked to report how often particular item. Twins with 50% or more items they had engaged in 11 specific antisocial missing for a particular scale were given behaviors, corresponding to 11 of the 13 scores of missing for that scale (0.91% for symptoms used to establish conduct disorder child AB, 1.03% for adolescent AB, and in the DSM-III-R (American Psychiatric Association 0.31% for adult AB). Ninety-five percent of [APA], 1987; the DSM-III-R item the sample (N = 6,464) did not have missing regarding forced sex was eliminated, given its data for any of the items. Only four individu- low prevalence in other samples and potential als had more than two missing items for any offensiveness, and two of the DSM-III-R of the three scales. items regarding theft behavior were combined Table 1 presents the proportion of twins into a single item). Responses ranged from 0 with each symptom count at each of the three (never) to3(6 or more times [or often]). Nine age points, separately by zygosity and sex. of the 11 items were repeated for AB age The prevalence of AB at all three time points years. (Two of the items, frequency was similar across zygosity, although a with which the respondent lied and started greater proportion of males than females re- fights, were only asked for the period prior ported one or more AB symptoms. Given the to age 15 years.) A computer algorithm was small proportion of cases with more than five applied to the frequencies for each item to in- childhood AB symptoms, more than four adolescent dicate whether a given symptom was present symptoms, and more that six adult (1) or absent (0). The algorithm was designed symptoms, categories were combined so that to match the wording of the DSM-III-R criterion the number of symptoms ranged from 0 to 5 as closely as possible. Summary scores (childhood), 0 to 4 (adolescent), and 0 to 6 of symptom counts were then calculated, and (adult); all variables were treated as ordinal in there was a possible range of 0 11 for AB the structural equation modeling; and thresholds prior to age 15 years and 0 9 for AB ages years. corresponding to each category were estiprior mated. Adult antisocial behavior. The SRQ also included Statistical analysis. Structural equation modtoms 17 items relating to 9 of the 10 symp- eling analyses were conducted using the sta- for adult antisocial personality disorder tistical package Mx (Neale, 1999). Models (ASPD; APA, 1987), such as frequency of be- were fit to correlation matrices (shown in Appendix ing irresponsible at work, frequency of arrest, A) created from the raw data. 1 The full and frequency of fighting. (Questions relating sex-limitation trivariate Cholesky model to the 10th adult symptom, failure to establish (Neale & Cardon, 1992) is shown in Figure 1. a monogamous relationship for at least 1 year, The diagram is shown for DZOS twin pairs, were not included.) All items were asked for with the parameters for male twins on the left the period age 18 years and older, and the (designated by the subscript m), and those for scale for each item ranged from 0 (never) to female twins on the right (designated by the 3(often). For adult AB, a computer algorithm subscript f). The model allows for 3 underlywas used that combined these 17 items into the 9 possible symptoms. A given symptom 1. Models were fit to correlation matrices, rather than to was coded as present (1) if respondents met covariance matrices, because the variables were ordicriteria for one or more of the behaviors that nal, and it is assumed that ordinal variables have variance represented that particular symptom and ab- = 1.0.

7 Table 1. Prevalence of AB symptoms Childhood AB a Adolescent AB b Adult AB c Females Males Females Males Females Males MZ DZ DZOS MZ DZ DZOS MZ DZ DZOS MZ DZ DZOS MZ DZ DZOS MZ DZ DZOS N ,317 1,569 1,183 1, ,318 1,565 1,180 1, ,320 1,574 1,185 1, Note: MZ, monozygotic twin; DZ, dizygotic twin from same-sex twin pair; DZOS, dizygotic twin from opposite-sex twin pair. a Defined as behavior prior to age 15 years. b Defined as behavior at ages years. c Defined as behavior at age 18 years and older. 401

8 402 Figure 1. The trivariate Cholesky model. The parameters for female twins are shown on the right; those for male twins are shown on the left. Child, AB before age 15 years; Adol, AB age years; Adult, AB 18 years and older; A, additive genetic influences; C, shared environmental influences; E, nonshared environmental influences.

9 Sex differences in antisocial behavior 403 ing genetic (A 1,A 2,A 3 ), shared environmental DZ twins (r g1 r g3 in Figure 1) because identical (C 1,C 2,C 3 ), and nonshared environmental factors twins share 100% of their genes, and fra- (E 1,E 2,E 3 ) that represent influences that ternal twins, like nontwin siblings, share 50% appear during childhood, adolescence, and of their segregating genes (on average). Non- adulthood, respectively. 2 In this full model, shared environmental influences, by definition, earlier influences are assumed to persist into are not correlated across twins. The later stages, although this is a testable as- model shown in Figure 1 is sex limited in two sumption. ways. First, the model estimates each of the Variation in scores at a given time point is individual parameters separately for males calculated as the sum of the squared parameters and females. Thus, the absolute magnitude of that point to that particular time point. For genetic and environmental influences on AB example, variation in childhood AB is a can vary across sex (a quantitative sex differ- c e 2 1, and the heritability of childhood AB ence). 3 Second, this model can estimate the is simply a 2 1. For adolescent and adult AB degree to which genetic influences are shared however, variation arises both from influ- across males and females (a qualitative sex ences specific to that time point as well from difference) by allowing each of the r g coeffi- influences that are shared with previous time cients to vary from 0.5 among DZOS twins. points. For example, the total variation in Similarly, the degree to which shared environ- adult AB is a a a c c c mental influences are the same for males and e e e 2 6, and the heritability of adult AB females is tested or by allowing each of the r c is a a a 2 6. Thus, the heritability of adult coefficients to vary from 1.0 among DZOS AB can be decomposed into new genetic influences twins. Because of constraints imposed by the specific to adult AB (a 2 6 ), genetic in- twin design, qualitative sex differences in ge- fluences that are common to adolescent but netic and shared environmental influences not childhood AB (a 2 5 ), and genetic influences cannot be tested simultaneously. that are common to AB at all three time points The absolute fit of the model shown in Fig- (a 2 3 ). Total shared environmental influences ure 1 is obtained by comparing the likelihood are calculated by c c c 2 6 and can be of this model to the likelihood of a model that similarly decomposed. The hypothesis that a fits the raw data perfectly (i.e., a saturated single set of genetic factors influences variation model), using the likelihood ratio test statistic in behavior at all three time points is (LRC), which is calculated as twice the differ- tested by fitting a nested submodel that constrains ence in log-likelihoods (Neale & Cardon, the paths a 4, a 5, and a 6 to zero, thereby 1992). The LRC is distributed as a chi-square eliminating any genetic influence on variation value and is an indication of model fit, and a in adolescent or adult AB that is not shared nonsignificant LRC indicates that the model with the genetic influence on child AB. Likewise, fits the raw data well. Similarly, the relative the hypothesis that there is a single set fit of nested submodels can be obtained by of shared environmental factors that influence calculating the LRC from the difference in AB is tested by a submodel constraining the log-likelihoods between the full model shown paths c 4, c 5, and c 6 to be zero. In contrast, the in Figure 1 and the particular submodel. When hypothesis that genetic influences on AB are two competing, nonnested submodels both completely independent across time is tested have nonsignificant LRCs, Akaike s Informa- by constraining the paths a 2, a 3, and a 4 to zero. Among same-sex twin pairs, each of the 3. Because of the inclusion of DZOS twin pairs, the pathree shared environment factors is correlated rameters specific to each of the three time points (i.e., 1.0 across twins (r c1 r c3 in Figure 1), regardto those with the subscripts 1, 4, and 6) were constrained less of zygosity. Each of the three genetic faceters be nonnegative to avoid a situation in which param- were estimated as negative for one sex but positors is correlated 1.0 for MZ twins and 0.5 for tive for the other. It should be noted that this constraint still allows genetic and environmental influences on 2. All latent factors were constrained to have a variance covariation between time points to be negative in one of unity. sex and nonnegative in the other.

10 404 K. C. Jacobson, C. A. Prescott, and K. S. Kendler tion Criteria (AIC; Akaike, 1987) can be used. to assess AB, should be lower in magnitude The AIC indicates the balance of goodness of than the correlations between child and adofit and parsimony (Williams & Holahan, lescent AB, which use the same items. 1994), and models with more negative AIC values are preferred. Twin correlations Results Inspection of the twin correlations presented in Appendix A can provide some expectation Reliability of what results from the structural equation modeling analyses may be. 4 For example, if A subset of twins (N = 127 from the FF study; MZ correlations are greater than DZ correla- N = 172 from the MMMF study) completed a tions, genetic influences are suggested. To the second SRQ an average of 28 days after the extent that DZ correlations are greater than initial SRQ. The intraclass correlations for one-half the MZ correlation, shared environmental AB prior to age 15, age 15 17, and age 18 influences are suggested. Finally, a years and older were.77,.71, and.67, respec- comparison of the same-sex and opposite-sex tively, among the FF sample and.72,.71, and twin correlations can give some indication of.69, respectively, among the MMMF sample. the likelihood of significant sex differences. The age of the twins used for reliability varied Examination of the correlations suggests a from 21 to 57. Therefore, we were able to ex- few general patterns. First, DZOS correlations amine whether age affected short-term reliability are not substantially lower than the same-sex by taking the absolute difference be- DZ correlations, suggesting that there are no tween the number of symptoms obtained from qualitative differences in genetic or shared en- the original SRQ and the number of symp- vironmental influences on AB. To take the toms obtained from the reliability SRQ and most extreme example, if the genetic and en- regressing this difference score onto age. This vironmental influences on AB were completely was done for each of the three variables, separately different for males and females, correwas by gender. Age did not significantly lations among DZOS twins would be zero. predict any of the absolute difference scores Instead, correlations among DZOS twins for either males or females (t range = 1.22 range from approximately.10 to.25. However, 0.20; all p >.20). there is some suggestion of quantitative sex differences (i.e., sex differences in the Within-person correlations magnitudes of genetic and environmental influences on AB). Most notably, for child AB, Within-person polychoric correlations were the MZF correlation (.39) is substantially calculated using SAS v The correlation between childhood and adolescent AB was.56 for males and.51 for females. Correla- 4. Formal structural equation modeling programs such as Mx are preferred to a visual inspection of correlations tions between childhood and adult AB were for three primary reasons. First, standard errors on cor-.39 (males) and.42 (females), and the correla- relations can vary widely across zygosity group, making tions between adolescent and adult AB were determinations of statistically significant differtions.56 and.49 for males and females, respecwhen ences difficult. Mx takes into account sample size estimating parameters and calculates confidence tively. The fact that the correlations between intervals around the parameters. Relatedly, Mx proadolescent and adult AB were nearly identical vides fit statistics that indicate the goodness of fit of to the correlations between child and adoles- each model and allows for specific hypothesis testing. cent AB demonstrates the cross-time validity Finally, in multivariate models, both cross-twin, of our measures of AB. Specifically, if meatwin, within-trait (e.g., Child_1 with Child_2) and crosssures cross-trait (e.g., Child_1 with Adol_2) correla- of adult AB were tapping a different tions are taken into account simultaneously when esticonstruct than that measured by child and admating parameters. Thus, simple visual comparisons of olescent AB, the correlations between adoles- pairs of correlations can sometimes yield misleading cent and adult AB, which use different scales results.

11 Sex differences in antisocial behavior 405 greater than the DZF correlation (.09), indi- in fit compared to Model 1 (LRCs < 1.0, df = cating strong genetic influence and no shared 1, all p >.50). We ran a similar series of mod- environmental influence among females. In els estimating the r c among DZOS twins, and contrast, the MZM correlation (.34) is only again, none of the models offered a signifi- slightly greater than the DZM correlation cant improvement in fit compared to Model 1 (.30), suggesting that the primary sources of (LRC < 1.00, df = 1, all p >.50, results not familial resemblance for child AB among shown). Thus, there was no evidence for qualitative males are shared environmental influences. sex differences in genetic and environmales Second, especially among male twins, mental influences on the development of AB. MZ correlations increase from childhood to In contrast, Model 2, which tested the hypothadulthood, while the DZ twin correlations are esis that there were no quantitative sex differences, similar across developmental periods. This did fit the data significantly more suggests that the heritability of AB increases poorly than Model 1 (LRC = 33.91, df = 15, with age. Finally, a comparison of MZ and p <.003), indicating that the magnitude of the DZ cross-twin, cross-trait correlations (e.g., genetic and environmental influences on the Child_1 with Adol_2) shows that MZ crosstwin, development of AB varied significantly cross-trait correlations are uniformly across males and females. Thus, parameters higher than their DZ counterparts, indicating were allowed to vary across males and fe- some overlap of genetic influence. males in all subsequent models. Models 3 6 tested whether the development Trivariate Cholesky analyses of AB could be explained by a single set of genetic factors (Models 3 and 5) and/or Prior to beginning our primary analyses, we by a single set of shared environmental facexamined whether thresholds were signifi- tors (Models 4 and 6). These analyses were cantly different across Twin 1 and Twin 2 first conducted separately by sex. Among (same-sex pairs only), across zygosity, or males, the hypothesis that there was a single across sex. Based on the LRC statistic, neither set of genetic factors that influenced variation equating thresholds across Twin 1 and Twin 2 in AB at all three time points could be rejected among same-sex pairs nor equating thresholds (Model 3) because the fit of this model across zygosity within sex resulted in a signif- was significantly worse than Model 1 (LRC = icant deterioration in fit (LRC = 42.85, df = 8.63, df = 3, p <.05). In contrast, the model 60, p =.95; LRC = , df = 120, p =.72, that allowed for a single set of shared environmental respectively). In contrast, equating thresholds influences on AB among males across sex resulted in a highly significant de- (Model 4) did not fit the data significantly terioration in fit (LRC = , df = 135, more poorly than Model 1 (LRC = 3.86, df = p <.001), indicating that the prevalence of 3, p =.28). Among females, neither the model AB varied significantly across sex. Therefore, with a single genetic factor (Model 5) nor the thresholds were constrained to be equal model with a single shared environmental factor within genders but were allowed to vary (Model 6) fit the data significantly more across sex for all subsequent analyses. poorly than Model 1 (LRC = 6.07, df = 3, Results from the primary model-fitting p =.11, Model 5; LRC = 4.50, df = 3, p =.21, analyses are presented in Table 2. Model 1 is Model 6). However, a model simultaneously the full trivariate Cholesky shown in Figure testing for a single set of genetic and a single 1, with parameters allowed to vary across sex, set of shared environmental factors (Model 7) but with r g = 0.5 and r c = 1.0 among the did fit the data significantly more poorly DZOS twins. This model fit the data very well (LRC = 28.18, df = 6, p <.001), indicating (p =.82). Next we ran three models estimating that there was some familial influence on adothat each of the three r g parameters in DZOS lescent and/or adult AB among females twins (data not shown). For all three models, the r g was estimated at close to 0.50, and none of the models was a significant improvement was not shared with the familial influence on childhood AB. Model 6 had a more negative AIC value, indicating that the model with a

12 Table 2. Model-fitting results Absolute Model Fit Relative Model Fit Model 2 LL df p Value a LRC b df p Value AIC 1 Full model 40, , All parameters: males = females 40, , Single genetic factor (males) 40, , Single shared environmental factor (males) 40, , Single genetic factor (females) 40, , Single shared environmental factor (females) 40, , Single genetic factor, single shared environmental factor (females) 40, , Single shared environmental factor (both sexes) 40, , Note: LRC, Likelihood ratio chi-square; AIC, Akaike s Information Criterion. The best-fitting model is indicated in bold. a The significance with the model is based on a comparison with the saturated model. The fit of the saturated model was 2 LL= 40,149.06, df = 20,112. b The LRC is obtained from a comparison with the full model (Model 1). 406

13 Sex differences in antisocial behavior 407 single shared environmental factor among series of post hoc analyses equating the heritabilities females was a better fit than the model of a and shared environmental estimates single set of genetic factors among females. at each time point across sex. In this case, a Thus, the next model (Model 8) tested significant LRC in comparison with Model 8 whether a single set of shared environmental indicates that estimates cannot be equated influences could be used to explain the development across sex. The heritability of childhood AB of AB for both males and females si- was significantly different for males and fe- multaneously. This model fit the data as well males (LRC = 4.09, df = 1, p <.05), although as Model 1 (LRC = 5.01, df = 6, p =.54), and heritabilities for adolescent and adult AB based on the AIC criteria it was the best-fitting, were not (LRC = 0.71, df = 1, p =.40 for most parsimonious model. adolescents; LRC = 0.15, df = 1, p =.70 for Finally, we tested whether we could constrain adults). Shared environmental estimates difpendent the genetic influences on AB to be inde- fered significantly across sex for both child (i.e., uncorrelated) across time periods. and adult AB (LRC = 4.27, df = 1, p <.05; Compared with Model 8, eliminating LRC = 6.16, df = 1, p <.01, respectively), but genetic influence on covariation across time not for adolescent AB (LRC = 0.03, df = 1, resulted in a significant deterioration in fit for p =.82). Thus, post hoc analyses confirmed both males (LRC = 18.30, df = 3, p <.001) that genetic influences were in fact stronger and females (LRC = 9.20, df = 3, p <.03; among females than among males for childhood results not shown), indicating that genetic factors AB and that shared environmental influresults did account for at least some of the stabil- ences on childhood AB were stronger for ity of AB from childhood to adolescence to males than for females. adulthood. Figure 2 presents the standardized parameter estimates from Model 8. Parameters for Equal environments males are shown on the left; those for females The equal environments (EE) assumption in are shown on the right. Table 3 presents the twin studies is that MZ and DZ twins are heritability and estimates of shared and non- equally correlated in their exposure to environmental shared environmental influences at each of the influences that impact the behavior three time points based on the parameters or trait in question. If this assumption is violated, shown in Figure 2. As can be seen in Table higher correlations among MZ twins 3, heritability estimates for both sexes were may be due to environmental factors, rather substantially lower for childhood AB than for than genetic factors, and heritability may be adolescent or adult AB (.06 vs..41 and.40 overstated. To examine whether the higher for males;.29 vs..50 and.42 for females). In heritability of childhood AB among females addition, the heritability estimate for child- may be due to violations of the EE assumption, hood AB was greater for females (.29 [95% we used multiple regression to examine confidence interval {CI} =.10;.34]) than for whether similarity of childhood environment males (.06 [95% CI =.00;.24]). In contrast, predicted within-pair differences in childhood the estimate of shared environmental influ- AB once zygosity was controlled for. The EE ence on childhood AB was greater among variable was a composite of four standard males (.28 [95% CI =.09;.38]) than among questions asking how often while growing females (.09 [95% CI =.02;.26]). Nonshared up... :... did you share a room,... environmental influences on AB ranged from did you have the same classroom at school,.43 to.66, indicating that approximately onehalf... did you have the same friends, and... of the variation in AB at each of the three did you dress alike. These questions were time points was explained by nonshared environmental asked during the wave 1 interview (FF) or factors. This was true for both sexes. wave 2 (MMMF) interview. Possible compos- To test the significance of the sex differences ite scores ranged from 4 to 16, with higher in the heritability and shared environ- scores indicating less equal environments. mental influences on childhood AB, we ran a Scores were averaged across twins to create a

14 408 Figure 2. The standardized parameter estimates from the best-fitting trivariate model. The parameters for the female twins are shown on the right; those for male twins on the left. Child, AB before age 15 years; Adol, AB age years; Adult, AB 18 years and older; A, additive genetic influences; C, shared environmental influences; E, nonshared environmental influences.

15 Sex differences in antisocial behavior 409 Table 3. Estimates of genetic and environmental influences from best model h 2 c 2 e 2 Males Females Males Females Males Females Child % CI.00;.24.10;.34.09;.38.02;.26.62;.72.61;.75 Adolescent % CI.31;.53.34;.55.01;.15.00;.19.46;.59.30;.58 Adult % CI.28;.52.33;.50.05;.21.00;.65.46;.53.56;.66 Note: h 2, heritability; c 2, estimate of shared environmental influence; e 2, estimate of nonshared environmental influence. single score, and analyses were restricted to and (b) shared environmental influences on same-sex twin pairs with complete data on AB are most important during childhood. both the EE variable and childhood AB (N Among males, heritability increased from.06 pairs = 298 MZF, 199 DZF, 642 MZM, 433 to approximately.40; for females, heritability DZM). increased from.28 to Among males, MZ twins did report more similar childhood shared environmental influences accounted environments than DZ twins (M MZ = for over one-quarter of the variation in child- 7.70, SD = 1.8; M DZ = 8.75 SD = 1.8), F (1, hood AB but only about 10% of the variation 1,571) = , p <.001, and females re- in adolescent and adult AB. Shared environmental ported more similar childhood environments influences among females were weak than males (M female = 7.88, SD = 1.9; M male = overall, explaining less that 10% of the 8.23, SD =1.9), F (1, 1,571) = 11.92, p <.001. variance in childhood and adolescence and However, the interaction between zygosity less than 1% of the variation in adult AB. and sex was not significant, F (1, 1,571) = This pattern of increasing heritability and de- 0.25, p =.62, indicating that the difference creasing shared environmental influences supports between MZ and DZ twins was similar for the conclusions drawn from previous males and females. Moreover, when the cross-sectional studies of juvenile and adult within-pair, absolute difference score for AB (e.g., Cadoret, 1974; Cloninger & Gottes- childhood AB was regressed onto the EE vari- man, 1987; Crowe, 1974; DiLalla & Gottesman, able and zygosity, EE did not predict twin 1989; Edelbrock et al., 1995; Eley et al., pair differences for either females (t = 0.75, 1999; Mednick et al., 1984; Rowe, 1986) and p =.45) or males (t = 0.47, p =.64). Thus, is consistent with results both from a meta- the higher heritability of childhood AB analysis (Miles & Carey, 1997) and a prior among females was not due to violations of retrospective study of adult male twins the EE assumption among females. (Lyons et al., 1995). Although the sample contained nearly Discussion 1,000 DZOS twin pairs, we could find no evidence that the specific genes and shared environments Our results demonstrate both similarities and that influence AB were qualitatively differences across sex in the genetic and envi- different across sex. This is consistent with ronmental architecture underlying the development evidence that the psychosocial risks associchildhood of antisocial behavior (AB) from ated with problem behavior are qualitatively to adulthood. For both sexes, this similar for males and females (Rowe, Vaz- study supports two primary hypotheses: (a) sonyi, & Flannery, 1994). In addition, our results genetic factors increase in relative importance indicate that there are similarities across from childhood to adolescence and adulthood, sex in the underlying structure of genetic and

16 410 K. C. Jacobson, C. A. Prescott, and K. S. Kendler environmental influences on the development (1989) have suggested that these late bloomers of AB. For both males and females, the bestfitting may have an even higher heritability of model allowed for unique genetic in- AB than the LCP delinquents. Thus, the new fluences on adolescent and adult AB, in addition genetic influence on adult AB may be related to those that persist from child AB, and to this late bloomer effect. In addition, ge- for a single set of shared environmental influences. netic influences on many phenotypes turn on These results are partly consistent with and off throughout the life span, and to date, the prior study of male twins from the Vietnam little is known about the mechanisms that may Registry, which found that both the ge- responsible for age-related genetic effects. netic and shared environmental factors that influenced Finally, our results may be supportive of adult AB overlapped completely the concept of LCP antisocial behavior be- with those factors that influenced juvenile AB cause a single set of genetic factors did influ- (Lyons et al., 1995). ence the development of AB across time, and a model suggesting completely independent genetic factors was rejected. Thus, there is evidence Genetic and environmental influences on the that certain genetically influenced development of antisocial behavior characteristics are related to antisocial behav- ior during both adolescence and adulthood. In our study, the finding of unique genetic influences Possible characteristics include both physiowith on adolescent AB that are not shared logical factors and personality characteristics, child AB may reflect the influence of genetically such as impulsivity and sensation seeking, influenced biological processes that both of which are to some degree heritable are first activated at puberty. For example, (Zuckerman, 1994). Nevertheless, it should be there is evidence from animal and human reiterated that the present study cannot directly studies that hormone levels, such as testosterone, address hypotheses driven from develstudies are related to aggression among males opmental theories such as those expounded by (Albert, Jonik, Watson, Gorzalka, & Walsh, Moffitt (1993) and DiLalla and Gottesman 1990; Brooks & Reddon, 1996; Dabbs & (1989) because this study focused on changes Morris, 1990; Olweus, Mattson, Schalling, & in genetic and environmental influences over Löw, 1988; Wagner, Beuving, & Hutchinson, time, not on how genetic and environmental 1979), although it should be noted that studies factors may vary across different typologies of hormonal effects on aggression among hu- of antisocial individuals. Such questions are mans are inconsistent (see Archer, 1991; Jacobson of considerable interest, however, and work is & Rowe, 2000, for review). A second currently under way to test these hypotheses explanation is that by middle to late adolescence, using more appropriate statistical methods. adolescents have greater latitude in se- lecting environments, such as peer groups, that are more consistent with their genetically Sex differences in the development influenced characteristics (Scarr & McCart- of antisocial behavior ney, 1983). Because there is evidence that peer selection in adolescence is heritable Although the present study suggests that the (Rowe, 1989), the new genetic influence on underlying structure of genetic and environadolescent AB may be related to these genetic mental influences on AB is similar across sex influences on peer selection. and that genetic and environmental factors The explanation for the unique genetic in- that influence AB are not qualitatively different fluences on adult AB is somewhat less certain. among males and females, there was evifluences There is evidence, both in our sample dence for sex differences in the magnitudes of and others, that some antisocial adults do not genetic and environmental influences on the report childhood or adolescent AB. Although these individuals are not discussed in Moffitt s (1993) typology, DiLalla and Gottesman development of AB (i.e., a quantitative sex difference). Based on post hoc analyses, the primary sex difference was in the magnitude