Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m

Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m
160 The Journal of Social Psychology, 2010, 150(2), 160–180 Copyright © Taylor & Francis Group, LLC VSOC 0022-4545 name, Vol. 150, No. 2, 1 2009: pp. 0–0 The Journal of Social Psychology Genetic Contributions to Antisocial Personality and Behavior: A Meta-Analytic Review From an Evolutionary Perspective The Journal of Social Psychology Ferguson CHRISTOPHER J. FERGUSON Texas A&M International University ABSTRACT. Evidence from behavioral genetics supports the conclusion that a significant amount of the variance in antisocial personality and behavior (APB) is due to genetic contributions. Many scientific fields such as psychology, medicine, and criminal justice struggle to incorporate this information with preexisting paradigms that focused exclu- sively on external or learned etiology of antisocial behavior. The current paper presents a meta-analytic review of behavioral genetic etiological studies of APB. Results indicated that 56% of the variance in APB can be explained through genetic influences, with 11% due to shared non-genetic influences, and 31% due to unique non-genetic influences. This data is discussed in relation to evolutionary psychological theory. Keywords: antisocial behavior, evolution, violence THE DEGREE WITH WHICH ANTISOCIAL, aggressive, and violent behavior can be explained through evolved genetic factors continues to be debated in the social sciences. This paper concerns itself with identifying the degree of variance in antisocial behavior that can be explained through genetic and non-genetic fac- tors in order to better inform this debate. It is my argument that the existing “Standard Social Science Model” (SSSM; Tooby & Cosmides, 1992) has, despite vaguely acknowledging the importance of “nature” in human behavior, persisted in focusing primarily on social or environmental explanations of behavior at the expense of genetic explanations of antisocial behavior. This may be, in part, a practical concern wherein genetics is difficult for most social scientists to con- trol. Yet this practical concern appears to filter down into theory and intervention priorities. Ultimately, social science may be experiencing a problem related to “levels of analysis” (Tinbergen, 1963). Specifically, the SSSM, due to practical Address correspondence to Christopher J. Ferguson, Department of Behavioral Sciences, Texas A&M International University, 5201 University Boulevard, Laredo, TX 78041, USA; [email protected] (e-mail). Ferguson 161 limitations, may have become accustomed to viewing behavior through “proxi- mal” or immediate causes or even mere correlates. For example, relationships between being abused as a child and acting violently as adults, though not always large, are commonly found in the literature (e.g. Ferguson et al., 2008). However, it remains unclear if this relationship between family violence and violent crime is due to learning/socialization factors or genetic factors or some combination of the two. Specific genes or socialization influences that contribute to behavior may be considered proximal or immediate causes of behavior. The ultimate cause of behavior, however, may more likely be the evolutionary process, which caused those specific genes to be selected due to specific environmental pres- sures, or which caused the ability to learn aggressive behavior to become an adaptive trait. It is the position of this article that understanding evolution as an ultimate cause of antisocial behavior is an important part of the discussion on antisocial behavior and violence that, as of yet, has not received adequate atten- tion. Thus, this paper will examine the evidence for genetic contributions to anti- social behavior and explain these contributions from an evolutionary framework. The Genetics of APB In the current article, I refer to antisocial personality and behavior (APB) as the main variable of interest. This variable is defined as including both the innate traits and motivation that direct individuals toward antisocial behavior (i.e., anti- social personality disorder, psychopathy) and antisocial behavior itself (i.e., aggression, violence, lying, stealing, etc.). Use of the term APB is intended to reflect the range of dependent measures used in relevant studies, some of which examine personality factors (i.e., Psychopathy Checklist), and others of which examine behaviors (i.e., arrest records, self-reported criminal activity, etc.) Evi- dence for a genetic (or partially genetic) etiology for APB comes from two main sources. Specifically, these are studies investigating the relationship of specific genes and APB, and behavioral genetics research, which attempts to determine the variance in APB that is due to genetic and non-genetic influences. Regarding specific genes, understanding of the human genome and its influ- ences on human behavior remains preliminary. However research on specific genes in human populations has begun to provide some evidence for how genes may contribute in the development of APB. For example, in one study of 240 children with attention deficit hyperactivity disorder, children with the valine/ methionine variant in the catechol O-methyltransferase (COMT) gene showed greater antisocial behaviors than those without this variant (Thapar et al., 2005). This variant of the COMT gene also appeared to interact with neonatal risk to increase APB. The COMT gene variant may have influenced the development of the prefrontal cortex, potentially reducing control over aggressive impulses. Caspi et al., (2002) used a longitudinal design to examine the impact of the MAOA gene located on the X-chromosome and its interaction with exposure to 162The Journal of Social Psychology maltreatment in the family. Results indicated that males with both a low-MAOA activity genotype and exposure to maltreatment were significantly more likely to exhibit child conduct disorder and adult antisocial behavior, including criminal arrests, than were high-MAOA activity genotype males who had been similarly maltreated. Although the low-MAOA genotype on its own did not increase APB, it appears that its presence places individuals at risk for APB, which can be triggered by maltreatment in the family. The presence of this gene on the X-chromosome may also help explain why males are more aggressive, on aver- age, than females, particularly if the low MAOA activity genotype is recessive in nature. This finding related to the low-MAOA genotype has been replicated in other studies (e.g., Nilsson et al., 2007; Kim-Cohen et al., 2006). Retz, Retz-Junginger, Supprian, Thome, and Rosler (2004) found a relation- ship between the serotonin transporter promoter gene (5-HTT) and impulsive violence in a forensic sample of 153 males. Specifically, a deletion/insertion polymorphism on this gene predicted impulsively violent behavior within this population. In and of itself, polymorphism on this gene was not able to explain the majority of violent behavior, but it appears that this gene may be one part of a larger genetic puzzle. It is important to point out that these single genes do not appear, in and of themselves, to deterministically cause APB in the same sense that the HD gene invariably leads to Huntington’s disease. Rather, these genes likely interact with each other in ways that remain poorly understood. Further, there are likely numerous other genes that are involved either directly or indirectly (i.e. via inter- actions) that have yet to be identified. Finally, these studies demonstrate that genetic vulnerability and exposure to family violence interact to lead to APB. These studies are unable to inform us regarding the relative contribution of genetic and non-genetic influences in explaining proportions of variance in the etiology of APB. For this purpose, behavioral genetic procedures such as twin and adoption studies have been employed as a means of gauging or estimating the relative contribution of genetic and non-genetic influences in APB etiology. Differences in Evolutionary and Behavioral Genetic Approaches to APB On the surface, evolutionary and behavioral genetic approaches to under- standing APB (and other phenomenon) would appear to be directly compatible. Specifically, for genes that promote certain behavior to exist within individuals, they must have developed through the process of natural selection. Similarly for behaviors to have evolutionary origins, they must be coded for in genes that are passed down through sexual reproduction. However, it has been observed (e.g. Horgan, 1999) that evolutionary psychologists and behavioral geneticists often- times distance themselves from one another. Both fields may appear to benefit from distancing themselves from unpopular elements of the other field. For instance, in downplaying evolution, behavioral geneticists can avoid the controversy of Ferguson 163 evolutionary theories. Some behavioral geneticists also may mollify critics by placing emphasis on gene-environment interactions or environmental effects (e.g. Moffit, 2005) rather than the direct effect of genes. Horgan (1999) argues that evolutionary psychologists also avoid specific genetic causal explanations for differences in behavior by focusing on commonalities in behavior rather than differences between individuals. The result of the chasm between evolutionary psychology and behavioral genetics is that theoretical differences in the explanation of behavior arise. Tinbergen (1963) suggested that there are multiple levels at which behavior can be explained. Arguably, behavioral geneticists prefer to focus on proximate explanations for behavior (i.e. specific genes), whereas evolutionary psycholo- gists look at ultimate causes of behavior (i.e., natural selection). It is not too much of a leap perhaps, to suggest that ultimate causes lead to proximal causes, which in turn lead to behavior. In other words, behavioral genetics research actu- ally provides evidence in support of the mechanism through which the natural selection of evolutionary psychology influences human behavior, both in regards to commonalities and differences. Behavioral Genetics and APB Stated briefly, behavioral genetics studies attempt to ascertain the relative contribution of genetic and non-genetic influences in explaining traits, character- istics, or patterns of behavior at the population/sample level. Given that identical (monozygotic, MZ) twins share all of their genetic material and fraternal (dizygotic, DZ) twins share approximately half of their genetic material, but (it is assumed) share environments to similar degrees, it is concluded that the corre- lated behavior of MZ twins should be twice that of DZ twins (Moffitt, 2005). Any variations from this observation can be attributed to non-genetic effects. Meta-analytic reviews of behavioral genetic studies for APB suggest that approximately 40%–50% of the variance can be explained through genetic influ- ences (Rhee & Waldman, 2002; Miles & Carey, 1997). These results suggest that genetic influences provide a significant contribution to APB. These reviews, while providing important evidence for the genetic basis are not without some limitations. Miles and Carey (1997), for instance, discuss laboratory measures of aggression that have not been properly validated in regards to APD (Tedeschi & Quigley, 1996; Ferguson, 2007). Rhee and Waldman (2002) include a wide range of studies, which is both a strength (for inclusion and breadth) and a weakness (as arguably more recent studies may have improved on deficiencies from previous studies). Neither of the previous meta-analyses attempts to understand the results from behavioral genetics studies of APB from a theoretical framework that would explain the development of genetic predisposition for APB and understand how genetic risk is influenced by environmental threats. Nonetheless, the behav- ioral genetics data provides evidence for proximal genetic causes of APB. Yet it 164The Journal of Social Psychology is important to understand proximal genetic traits by examining evolution as an ultimate cause of those genetic traits. The Controversy Over Evolutionary Explanations of Human Behavior Given that behavioral genetic studies of APB provide evidence for a signifi- cant genetic contribution, it is somewhat surprising to find considerable resis- tance to discussing or studying the evolution of APB or violence in the social sciences. Arguably, modern social science has focused on “pitfalls” of modern life such as media violence, toy guns, and Western values, although violence and homicide rates are found to be high among non-advanced cultures without access to these modern accruements (Buss & Shackelford, 1997). For instance, a bro- chure on youth violence available from the American Psychological Associa- tion’s Web site states that “There is no gene for violence. Violence is a learned behavior . . .” (APA, 1996). Although the brochure goes on to suggest that genet- ically influenced factors such as learning disabilities and impulsivity may interact with learned violence, this brochure inaccurately suggests that genetics do not directly contribute to violence. The brochure is outdated also in suggesting that youth violence is on the rise, when in fact it has been on a precipitous decline for 15 years. Similarly, it has been noted that the National Institutes of Health have de-emphasized genetic, evolutionary, or other biological studies of violence behavior (Enserink, 2000). Critiques of biological theories of aggression are per- haps epitomized by Berkowitz (1993), who claims that aggression has not been linked to biological storage areas for an aggressive instinct, and that aggression is normally provoked by external stimuli. From this perspective, aggression would seem to only be biological if it were univariate, purposeless, and unprovoked. Tooby and Cosmides (1992) argue that perspectives such as those of Berkowitz are indicative of the “Standard Social Science Model” (SSSM), which postulates the brain as a general-purpose learning device, devoid of content at birth, with behavior solely a product of subsequent learning. By contrast, evolutionary psychology views the brain as having evolved through countless generations of environmental pressures, wherein certain largely pre-wired brain “modules” give organisms in-born mechanisms for dealing with likely environmental stressors. As for the reluctance of the psychological discipline to embrace the contributions of evolutionary theory, three main reasons would appear to be relevant. The first of these is simply a matter of dogma and indoctrination. Prior to the mid-20 th century, genetic explanations for behavior were common. However, difficulties in conducting genetics research in humans and abuses of evolutionary theory, including eugenics and the belief in racial differences in intelligence (Kamin, 1974) decreased the appeal of evolutionary explanations for many. As environmental explanations (particularly those focused on learning) gained prominence in psychology in the 20 th century, genetic theories of human behavior were largely repudiated in favor of environmental theories. Arguably, this movement coincided with social changes in the United States Ferguson 165 at that time, such as the increased influence of feminist theories. Which tended to focus on the innate “sameness” of individuals rather than on innate differences. Secondly, misunderstandings about the nature of how evolutionary theory applies to human behavior may also increase resistance. Two common miscon- ceptions about evolutionary psychology involve the “naturalistic fallacy” and concerns with regards to biological hard determinism. Briefly, the naturalistic fallacy is the belief (or fear) that if something is demonstrated to be due to biology, then this provides moral justification for the behavior. Similarly, biolog- ical hard determinism would imply that human behavior is due only to genetic (or other biological) effects and is not influenced by the environment, nor open to the effects of agency. However, evolutionary psychologists have indicated clearly that they do not endorse either the naturalistic fallacy or biological hard determinism (see Wilson, Dietrich, & Clark, 2003 for a discussion). Finally, a qualitative difference exists between genetic causal influences and non-genetic causal influences. If non-genetic influences (whether biological or socialization) are demonstrated to cause a negative outcome, this observation provides an evident solution, namely that removing the causal influence will likely reduce the negative behavior. However, genetic influences that are shown to cause a negative outcome cannot logically or ethically be removed. This may be interpreted as suggesting that genetic causal influences offer no evident solu- tion for negative behavior reduction; individuals may thus be reluctant to endorse such research under this misimpression, as it appears to offer little hope for behavioral change. However, blinding oneself to the influence of genetic ele- ments on behavior, by necessity, blinds oneself also to gene/environment interac- tion effects, which may offer some solutions for the reduction of negative behavior. Understanding the genetic influences on behavior, and identifying these genetic risks within individuals, may result in treatments that theoretically could be targeted early and preventatively toward individuals who may have this genetic risk. Understandably, discussion of these possibilities raises considerable ethical concerns (see, for example, Williams, 1994). This is not to say that such techniques may not prove to be useful in the future, but that great care must be taken to ensure that any behavioral or medical interventions for violence preven- tion are undertaken only under strict ethical guidelines. Put directly, we enter problematic ethical territory once individuals are punished for crimes they have not yet committed. Ultimately, it may be decided that such approaches are either unpractical or wholly unethical. Nevertheless, even if that should be true, obser- vations regarding the limits of the “practicality” of genetic effects on violence should not be intertwined with discussions of the “truthfulness” of such effects. Understanding APB From an Evolutionary Framework APB may be understood as a bi-product of normal human aggression. Here, use of the term “aggression” is similar to that proposed by Baron and Richardson 166The Journal of Social Psychology (1994). Specifically, aggressive behavior is here defined as intentional behavior produced to cause physical harm or humiliation to another person who wishes to avoid the harm. Violent antisocial behavior, by extension, is here defined as intentional behavior intended to cause physical harm or humiliation to another person who contrastingly wishes to avoid the harm and the behavior is out of proportion with precipitating or provoking stimuli and carried out with disregard to the welfare or rights of others. Thus all violent antisocial behaviors are aggressive, but not all aggressive behaviors are necessarily antisocial. For example, acting in self-defense in response to a threatening individual would be considered aggressive behavior, but not antisocial behavior. Buss and Shackelford (1997) provide an excellent review of the underlying premises of an evolutionary understanding of aggressive behavior. Buss and Shackelford argue that aggression can be thought of as an adaptive response that can result in certain gains, such as co-opting or defending resources, increased mating options and mate fidelity and increased status. Aggression is not unitary or context-blind, but rather modular and context-specific, and one would expect aggression to be provoked by external stimuli. Our understanding of APB can be built upon several assumptions that are derived from evolutionary psychology models such as those provided by Buss and Shackelford. These are: 1. Human aggression is a normative and adaptive response that has a selec- tive advantage for individuals (note this does not imply that it is morally desirable). 2. Restraining aggression (i.e., impulse control) is also a normative and adaptive response that has a selective advantage for individuals. 3. Aggressive and aggressive impulse control instincts respond to environ- mental stimuli, or catalysts, which are cognitively processed in order to select the most adaptive response to an environmental stressor. 4. The human brain has evolved separate systems or “devices” to manage separate aggression and aggression-reduction impulse control drives. These devices may at times compete, particularly when environmental catalysts are ambiguous. From an evolutionary perspective, a behavior as ubiquitous as aggression is best understood as an adaptation to environmental pressures that provided a selective advantage to members of the species. That is to say, members of the species who possessed the genotype related to the production of aggressive behaviors were more likely to survive and produce viable offspring than mem- bers of the species with genotypes that were less likely to produce aggressive behaviors. The selective advantages provided by aggressive behaviors may be related to external pressures common to hominid species, such as the benefits of risk-taking during hunting, and fending off or attacking predatory species. Benefits to aggression may also include higher success regarding intraspecies Ferguson 167 competitive pressures such as mate selection and supply of resources (see Sagan & Druyan, 1992 for a discussion). Although aggression, in moderate amounts and in proportion to environmen- tal threats, may be beneficial, high levels of aggression may clearly be “too much of a good thing” phenotypically speaking. High levels of aggression may place an individual at extreme risk for harm or may result in social rejection and depriving the individual of the benefits of social groups, the development of which also likely contributes to the survivability of individual hominid organ- isms. Therefore, an individual may benefit not only from being aggressive, but knowing when to be aggressive and when to restrain aggressive impulses. Just as an aggressive instinct may provide a selective advantage under some circum- stances, so too an aggression-reduction instinct may provide a selective advantage under other circumstances. These ideas are consistent with Gray’s the- ories on behavioral approach and inhibition systems (Gray, 1990). The aggres- sion-reduction instinct may be synonymous with what is often referred to as “impulse control” or “executive functioning.” Deficits in portions of the brain (i.e., frontal lobes of the cortex) related to executive functioning have been demonstrated to predict overly aggressive (i.e., antisocial) behavior (Mercer & Selby, 2005; Donovan & Ferraro, 1999; Soderstrom et al., 2002). Neuroimaging studies document that frontal cortex lesions are associated with impulsive aggres- sion, and less-so with trait aggression (Leon-Carrion & Chacartegui-Ramos, in press). For example, Raine, Lencz, Bihrle, LaCasse, and Colletti (2000) found that individuals with APB have 11% less grey matter in the prefrontal cortex as compared to non-APB individuals. This was the case even in individuals without history of brain injury. Critchley et al., (2000) found similar results in violent individuals, as compared to non-violent controls, in the prefrontal cortex and in the amygdala and hippocampus. The numerous studies on frontal lobe function- ing and violence are too numerous to summarize here in full, although several excellent review sources exist (e.g., Davidson, Putnam & Larson, 2000; Hare, 1993; Leon-Carrion & Chacartegui-Ramos, in press). Thus it is reasonable to conclude that an “impulse control device,” such as would decrease the frequency of aggressive responses to ambiguous stimuli, may be located in the frontal lobes. This gene/environment interaction is important to emphasize. The antiso- cial genotype, as with any genotype, is unlikely to produce a static array of behaviors across all environmental situations. Rather, genotype produces a behavioral range or range of behaviors in order to allow the individual to adjust to differing environmental threats. Environments with low threat or strain are less likely to elicit antisocial behavioral responses than are environments with high threat or strain. Understanding what environmental situations are likely to produce antisocial behaviors from individuals high in antisocial personality may provide promising avenues for prevention and intervention, with interven- tions targeted at increasing the behavioral range of antisocial individuals to 168The Journal of Social Psychology include more non-aggressive behavioral options. From an evolutionary prospective, the way in which genes and environment interact renders the individual more flexible in dealing with a host of potential environmental threats. A more behaviorally flexible organism is inherently more adaptive than a behaviorally rigid one. The Current Study The current meta-analysis will seek to explain the etiological origin of genetic effects, as well as gene-environmental interactions from an evolutionary framework. As such, the present study will provide an ultimate causal explana- tion for the proximal influences of specific genetic traits. Consistent with both previous research and an evolutionary perspective, it is hypothesized that a significant amount of the variance, most likely the largest single variance compo- nent, will be explainable through genetic influences. It is further hypothesized that genetic influences will be found to be highest in children and to diminish slightly across age, due to individuals accumulating non-genetic biological and non-biological influences, experiences, and injuries across time. Method Study Selection and Categorization PsycINFO was searched for all articles published between the years of 1996 and 2006 (this criterion discussed below) that included the following search terms: (adopt* or twin or heritabil* or behavior genetic* or behavioral genetic*) and (violen* or violent crim* or crim* or aggress* or antisocial). The references of primary sources revealed in this search were also examined for studies that were not discovered during this initial search. Articles were judged relevant if they met the following criteria: (a) Articles had to have been published between the years of 1996–2006 (effectively an 11-year publication period). Limiting meta-analyses to more recent research allows for an examination of recent trends in the literature, in which methods may have improved over past years. (b) Outcome variables had to clearly measure some element of antisocial, violent, or aggressive behavior. Criteria were essentially identical to those Rhee and Waldman (2002) discuss in depth. (c) Articles had to include methodology (i.e., twin or adoption) for determining relative contributions of genetic and non-genetic influences in APB. A total of 38 published studies comprised of 53 separate observations (some studies broke down results by gender, some did not) were found that met the Ferguson 169 above criteria. The combined sample size for the included articles was 96,918. Articles in the current study were coded for the presence of several potential moderator variables, namely: (a) Age (below age twelve, twelve to eighteen and above eighteen), (b) Sex, (c) Whether the outcome variables clearly used measures of antisocial personality disorder according to DSM-IV criteria, indices of violent behavior such as police reports or self-reported violent criminal activ- ity, or broader measures of violent or aggressive behaviors as found in clinical measures of behavior disorders related to aggression. Calculating Effect Size Estimates and Statistical Analyses Pearson’s r, a flexible and easily interpreted index of effect size, was used as the effect size estimate in this study. Correlation coefficients were transformed to Fisher’s z, weighted, averaged and transformed back to a pooled r, denoted r +. The Comprehensive Meta-Analysis (CMA) software program was used to fit fixed effects models. Once the combined effect sizes were calculated for genetic, shared non-genetic and unique non-genetic variance components, these were transformed back to (h 2), (c 2) and (e 2) variance estimates. Due to rounding dur- ing the effect size calculation process it is likely that the final total variance may differ slightly from 100%. Table 1 presents all included studies, effect sizes for (h 2), (c 2) and (e 2) in terms of r and characterization on moderator variables. Results Table 2 presents the results from the meta-analysis on the entire group of studies broken down by genetic, shared non-genetic, and unique non-genetic variance components. As can be seen, genetic influences account for the largest component of the variance in APB, with 56% explained, shared non-genetic influences explaining 11% of the variance in APB, and unique non-genetic influ- ences explaining 31%. These results indicate that genetics is a significant con- tributor to APB, but that non-genetic influences also remain important. Tests of homogeneity indicate the presence of moderator variables. Analyses of Moderator Variables Bivariate correlations were used to examine the effect of moderator vari- ables on the effect size of the three variance components for APB. Results for genetic influences indicated that effect size was moderated by type of measure- ment used (r = .38, p < .01) with broader measure of aggression obtaining higher effect sizes than measures limited to DSM-IV criteria for antisocial personality disorder. Similarly, age was correlated (r = −.40, p < .01) with effect size, sug- gesting that genetics is a more powerful predictor of APB in younger individuals than in adults. Sex was not a significant moderator (r = .06, p > .05). 170The Journal of Social Psychology TABLE 1. Included Studies in Current Meta-Analysis StudyrAge Sex Outcome Arseneault (2003) h 2 .91 Child Mixed Externalizing c 2 .00 e 2 .42 Bartels (2003) 1 h 2 .83 Teen Male Externalizing c 2 .10 e 2 .55 2 h 2 .85 Teen Female c 2 .00 e 2 .53 Blonigen (2006) 1 h 2 .70 Adult Male Antisocial c 2 .00 e 2 .72 2h 2 .70 Adult Female c 2 .00 e 2 .72 Blonigen (2005) h 2 .70 Teen Mixed Externalizing c 2 .00 e 2 .71 Brendgen (2006) h 2 .64 Child Mixed Violent c 2 .00 e 2 .77 Brendgen (2005) h 2 .79 Child Mixed Violent c 2 .00 e 2 .62 Bullock (2006) h 2 .53 Child Mixed Externalizing c 2 .53 e 2 .66 Button (2005) h 2 .65 Child Mixed Externalizing c 2 .14 e 2 .75 Cleveland (2003) h 2 .64 Teen Mixed Violent c 2 .24 e 2 .73 Coccaro (1997) h 2 .69 Adult Male Violent c 2 .00 e 2 .73 Eley (2003) h 2 .80 Child Mixed Externalizing c 2 .56 e 2 .33 (Continued ) Ferguson 171 TABLE 1. (Continued ) StudyrAge Sex Outcome Eley (1999) 1 h 2 .83 Child Male Externalizing c 2 .23 e 2 .50 3h 2 .83 Child Female Externalizing c 2 .23 e 2 .50 Gelhorn (2006) h 2 .78 Teen Mixed Antisocial c 2 .00 e 2 .62 Gjone (1997) h 2 .85 Child Mixed Externalizing c 2 .00 e 2 .53 Goldstein (2001) h 2 .56 Adult Female Antisocial c 2 .31 e 2 .77 Hicks (2004) h 2 .89 Teen Mixed Antisocial c 2 .00 e 2 .45 Hudziak (1999) 1 h 2 .85 Child Male Externalizing c 2 .00 e 2 .52 2h 2 .83 Child Female c 2 .00 e 2 .56 Hudziak (2003) 1 h 2 .78 Child Male Externalizing c 2 .45 e 2 .43 2h 2 .78 Child Female c 2 .46 Jacobson (2000) h 2 .57 Adult Male Antisocial c 2 .56 e 2 .60 Jacobson (2002)1 h 2 .54 Adult Male Antisocial c 2 .39 e 2 .75 2h 2 .64 Adult Female c 2 .24 e 2 .75 (Continued ) 172The Journal of Social Psychology TABLE 1. (Continued ) StudyrAge Sex Outcome Jaffee (2002) h 2 .84 Child Mixed Externalizing c 2 .00 e 2 .54 Koenen (2006) 1 h 2 .88 Child Male Externalizing c 2 .00 e 2 .47 2h 2 .88 Child Female c 2 .00 e 2 .48 Larsson (2006) h 2 .79 Teen Mixed Antisocial c 2 .00 e 2 .69 Ligthart (2005) 1 h 2 .77 Child Male Externalizing c 2 .44 e 2 .46 2h 2 .79 Child Female c 2 .38 e 2 .47 Malone (2004) h 2 .57 Adult Male Antisocial c 2 .36 e 2 .74 McGue (2006) 1 h 2 .65 Teen Male Antisocial c 2 .41 e 2 .64 2h 2 .69 Teen Female c 2 .35 e 2 .63 O’Connor (1998) h 2 .79 Teen Mixed Externalizing c 2 .47 e 2 .40 Polderman (2006) h 2 .70 Child Mixed Externalizing c 2 .00 e 2 .71 Rushton (1996) 1 h 2 .86 Adult Male Antisocial c 2 .00 e 2 .51 2h 2 .00 Adult Female c 2 .80 e 2 .60 Slutske (2001) 1 h 2 .82 Adult Male Externalizing c 2 .00 e 2 .57 (Continued ) Ferguson 173 TABLE 1. (Continued ) StudyrAge Sex Outcome Taylor (2003) h 2 .62 Teen Male Antisocial c 2 .00 e 2 .78 Thapar (1996) h 2 .53 Child Mixed Antisocial c 2 .63 e 2 .57 Tuvblad (2006) 1 h 2 .24 Teen Male Antisocial c 2 .72 e 2 .65 2h 2 .77 Teen Female c 2 .41 e 2 .49 Tuvblad (2004) 1 h 2 .52 Teen Male Externalizing c 2 .65 e 2 .56 2h 2 .63 Teen Female c 2 .53 e 2 .57 van Beijsterveldt (2003) 1 h 2 .82 Child Mixed Externalizing c 2 .44 e 2 .37 2h 2 .74 Child Female c 2 .54 e 2 .41 Viding (2005) h 2 .75 Child Mixed Antisocial c 2 .41 e 2 .52 Vierikko (2003) 1 h 2 .52 Child Male Externalizing c 2 .84 e 2 .30 2h 2 .74 Child Female c 2 .56 e 2 .41 Vierikko (2005) 1 h 2 .52 Teen Male Externalizing c 2 .79 e 2 .32 2h 2 .84 Teen Female c 2 .42 e 2 .35 Note. All studies are listed by first author last name and year. Manuscripts with multiple analyses (including analyses separated by gender) are delineated numerically after the date. Mixed = mixed sex sample. 174The Journal of Social Psychology Results for shared non-genetic influences did not reveal significant moderat- ing effects for measure used (r = .06), age (r = −.05) or sex (r = .12). Results for unique non-genetic influences did reveal significant moderators. Not surpris- ingly, these effects were opposite those for genetic influences with a significant correlation with type of measurement used (r = −.54, p ≤ .01) and age (r = .59, p< .01). Sex was not a significant moderator (r = −.06, p ≥ .05). Thus unique non-genetic influences are higher on measurements that use stricter DSM-IV cri- teria for antisocial personality disorder and for older individuals. Results for the age moderator are particularly worthy of consideration. Figure 1 presents changes in effect size for genetic, shared non-genetic, and unique non-genetic effects across child, adolescent, and adult groups in terms of effect size r +. As can be seen, the influence of both genetic and shared non- genetic influences decreases into adulthood, whereas the influence of unique TABLE 2. Meta-Analytic Results for Genetic, Shared Non-Genetic, and Unique Non-Genetic Variance Components in Antisocial Personality and Behavior Component r + 95% C.I. Homogeneity Test Variance Genetic .75 (.74, .75) X 2(52) = 5813.8, p ≤ .001 56% Shared Non-Genetic .33 (.33, .34) X2(52) = 7575.0, p ≤ .001 11% Unique Non-Genetic .55 (.54, .55) X2(52) = 3550.8, p ≤ .001 31% Note. r + = pooled correlation coefficient; C.I. = Confidence intervals; Variance = proportion of variance explained. FIGURE 1. Age Trends in Genetic and Non-Genetic Influences in APB. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Child Adolescent Adult Age Effect Sizer Genetic Shared Non-Genetic Unique Non-Genetic Ferguson 175 non-genetic influences increases into adulthood. As such, the proportion of vari- ance in APB explained through genetic or shared non-genetic factors declines slightly across the lifespan. This is not surprising and likely reflects the gradual accumulation of non-genetic influences such as head injuries, infections, as well as socialization, and potentially increased agency (Bandura, 2006; Rychlak, 1999). It is possible as well that highly APB individuals are selected against by the environment. The likelihood of being killed or injured during an aggressive incident naturally increases across time, thus showing increased effect size for unique non-genetic effects. Thus it is important to understand that understanding the influence of genetic factors on APB depends somewhat at what point in the lifespan APB is examined. As unique non-genetic factors accumulated across time in the individual lifespan, naturally their influence on behavior increases. Genetic factors, being essentially unchanging, remain fairly static by comparison. Discussion Results from the current study highlight both the genetic effects and non- genetic effects on APB. As can be seen, genetic influences account for a consid- erable percentage of the variance. Understanding the development of these genetic factors from an evolutionary perspective allows us to understand the ultimate causal processes that lead to the development of APB. Shared non- genetic influences, which arguably are a reasonable indicator of family socializa- tion (although other environmental influences that cause direct biological insults may also be part of this variance) has the smallest influence on APB. It is worth mentioning however, that 11% of the variance in APB is still a considerable percentage and worthy of attention, particularly as this portion of the variance may be particularly amenable to prevention or intervention. Unique non-genetic influences accounted for about a third of the variance in APB. As this portion of the variance includes non-genetic biological influences such as head injuries and infections, non-family socialization processes, and potential agency effects (Bandura, 2006) this portion of the variance is most dif- ficult to interpret. As such, until studies undertake to measure specific potential influences on APB, this portion of the variance may best be considered “unknown” variance, aside from the fact that we know it is variance that is unique to individuals and not due to shared influences. Utilizing specific mea- sures of family violence exposure, family environment, peer relations, medical history, etc., as part of twin studies may help in elucidating the specific contribu- tions of each of these potential non-genetic influences in the development of APB. Several studies have already begun to adopt such procedures (e.g., Caspi et al., 2004). Results from this analysis are generally consistent with literature indicating that genetic contributions are an important influence in the development of APB (Larsson, Andershed, & Lichtenstein, 2006). As indicated earlier in the paper, 176The Journal of Social Psychology the presence of a significant genetic component to APB suggests evolutionary origins for this behavior. It would appear reasonable to conclude that aggressive behaviors have promoted the survivability of individual members in the species. Genes, thusly, are one predictor for individual variation (a component of natural selection) in this behavioral trait. Acknowledging this effect may be an impor- tant step in then examining specific gene-environment interactions that may pro- mote APB. An issue that bears mentioning is that of measurement standardization and validity. For instance, analyses have found that there is an inverse relationship between rigorous methodology and measurement in fields such as organiza- tional psychology (Terpstra, 1981) and media violence effects (Ferguson, 2007), with higher effects reported in studies using poorly standardized and poorly val- idated instruments than for better instruments. Given criticisms of behavioral- genetics research, it is imperative that future research takes great care to use methods of the highest rigor and measurements with demonstrated reliability and validity. Suggestions for future research involve including valid measures of non- genetic influences into twin study methodology to specifically examine the etiological contributions of these influences, once genetics has been controlled. Similarly, further research on gene/environment interactions would be helpful and likely offer the most positive route for intervention or prevention. Also, examining the catalytic impact of environmental strain on antisocial personalities and how the impact of this strain may be reduced would be a worthy avenue of research. Lastly, further research on the impulse control device theorized here may provide inroads in understanding how impulsive aggression, in particular, may be understood and treated. AUTHOR NOTE Christopher J. Ferguson is an associate professor of clinical forensic psychology at Texas A&M International University. He holds a PhD in clinical psychology from the University of Central Florida and is licensed as a psycholo- gist in Texas. His main research areas have focused on violent behavior and the positive and negative effects of playing violent video games. REFERENCES Note. *Article included in meta-analysis. American Psychological Association. (1996). An APA brochure on youth violence. Retrieved February 2, 2007, from http://www.apa.org/ppo/issues/pbviolence.html *Arseneault, L., Moffitt, T., Caspi, A., Taylor, A., Rijsdijk, F., & Jaffee, S., et al. (2003). Strong genetic effects on cross-situational antisocial behaviour among 5-year-old children according to mothers, teachers, examiner-observers, and twins’ self-reports. Journal of Child Psychology and Psychiatry, 44, 832–848. Ferguson 177 Bandura, A. (2006). Toward a psychology of human agency. Perspectives on Psychological Science, 1, 164–180. Baron, R., & Richardson, D. (1994). Human aggression. New York: Plenum Press. *Bartels, M., Hudziak, J., van den Oord, E., van Beijsterveldt, C., Rietveld, M., & Boomsma1, D. (2003). Co-occurrence of aggressive behavior and rule-breaking behav- ior at age 12: Multi-rater analyses. Behavior Genetics, 33, 607–621. Berkowitz, L. (1993). Aggression. New York: McGraw-Hill. Blonigen, D., Hicks, B., Krueger, R., Patrick, C., & Iacono, W. (2006). Continuity and change in psychopathic traits as measured via normal-range personality: A longitudinal- biometric study. Journal of Abnormal Psychology, 115, 85–95. Blonigen, D., Hicks, B., Krueger, R., Patrick, C., & Iacono, W. (2005). Psychopathic personality traits: Heritability and genetic overlap with internalizing and externalizing psychopathology. Psychological Medicine, 35, 637–648. *Brendgen, M., Vitaro, F., Boivin, M., Dionne, G., & Perusse, D. (2006). Examining genetic and environmental effects on reactive versus proactive aggression. Develop- mental Psychology, 42, 1299–1312. *Brendgen, M., Dionne, G., Girard, A., Boivin, M., Vitaro, F., & Perusse, D. (2005). Examining genetic and environmental effects on social aggression: A study of 6-year-old twins. Child Development, 76, 930–946. *Bullock, B., Deater-Deckard, K., & Leve, L. (2006). Deviant peer affiliation and prob- lem behavior: A test of genetic and environmental influences. Journal of Abnormal Child Psychology, 34, 29–41. Buss, D., & Shackelford, T. (1997). Human aggression in evolutionary psychological perspective. Clinical Psychology Review, 17, 605–619. *Button, T., Scourfield, J., Martin, N., Purcell, S., & McGuffin, P. (2005). Family dysfunc- tion interacts with genes in the causation of antisocial symptoms. Behavior Genetics, 35, 115–120. Caspi, A., McClay, J., Moffitt, T., Mill, J., Martin, J., Craig, I., et al. (2002). Role of geno- type in the cycle of violence in maltreated children. Science, 297, 851–854. Caspi, A., Moffitt, T., Morgan, J., Rutter, M., Taylor, A., Arseneault, L., et al. (2004). Maternal expressed emotion predicts children’s antisocial behavior problems: Using monozygotic twin differences to identify environmental effects on behavioral develop- ment. Developmental Psychology, 40, 149–161. Critchley, H.D., Simmons, A., Daly, E.M., Russell, A., van Amelsvoort, T., Robertson, D.M., et al. (2000). Prefrontal and medial temporal correlates of repetitive violence to self and others. Biological psychiatry, 47(10), 928–34. *Cleveland, H. (2003). Disadvantaged neighborhoods and adolescent aggression: Behavioral genetic evidence of contextual effects. Journal of Research on Adolescence, 14, 211–238. *Coccaro, E., Bergeman, C., Kavoussi, R., & Seroczynski, A. (1997). Heritability of aggression and irritability: A twin Study of the Buss-Durkee Aggression Scales in adult male subjects. Biological Psychiatry, 41, 273–284. Davidson, R., Putnam, K., & Larson, C. (2000). Dysfunction in the neural circuitry of emotion regulation—a possible prelude to violence. Science, 289, 591–594. Donovan, W., & Ferraro, R. (1999). Frontal lobe deficits in domestic violence offenders. Genetic, Social and General Psychology Monographs, 125, 71–102. *Eley, T., Lichtenstein, P., & Moffitt, T. (2003). A longitudinal behavioral genetic analy- sis of the etiology of aggressive and non-aggressive antisocial behavior. Development and Psychology, 15, 383–402. *Eley, T., Lichtenstein, P., & Stevenson, J. (1999). Sex differences in the etiology of aggressive and non-aggressive antisocial behavior: Results from two twin studies. Child Development, 70, 155–168. 178The Journal of Social Psychology Enserink, M. (2000). Searching for the mark of Cain. Science, 289, 575–579. Ferguson, C. J., Rueda. S., Cruz, A., Ferguson, D., Fritz, S., Smith, S. (2008). Violent video games and aggression: Causal relationship or byproduct of family violence and intrinsic violence motivation? Criminal Justice and Behavior, 35, 311–332. Ferguson, C.J. (2007). Evidence for publication bias in video game violence effects litera- ture: A meta-analytic review. Aggression and Violent Behavior, 12, 470–482. *Gelhorn, H., Stallings, M., Young, S., Corley, R., Rhee, S., Hopfer, C., et al. (2006). Common and specific gene influences on aggressive and non-aggressive conduct disorder domains. Journal of the American Academy of Child and Adolescent Psychia- try, 45, 570–577. *Gjone, H., & Stevenson, J. (1997). A longitudinal twin study of temperament and behavior problems: Common genetic or environmental influences. Journal of the Amer- ican Academy of Child and Adolescent Psychiatry, 36, 1448–1456. *Goldstein, R., Prescott, C., & Kendler, K. (2001). Genetic and environmental factors in conduct problems and adult antisocial behavior among adult female twins. Journal of Nervous and Mental Disease, 189, 201–209. Gray, J. (1990). Brain systems that mediate both emotion and cognition. Cognition and Emotion, 4, 269–288. Hare, R. (1993). Without conscience. New York, NY: Guilford Press. *Hicks, B., Krueger, R., Iacono, W., McGrue, M., & Patrick, C. (2004). Family transmission and heritability of externalizing disorders. Archives of General Psychiatry, 61, 922–928. Horgan, J. (1999). The undiscovered mind: How the human brain defies replication, medication and explanation. New York, NY: Simon and Schuster. *Hudziak, J., van Beijsterveldt, C., Bartels, M., Rietveld, M., Rettew, D., Derks, E., et al. (2003). Individual differences in aggression: Genetic analyses by age, gender, and informant in 3-, 7-, and 10-year-old Dutch twins. Behavior Genetics, 33, 575–589. *Hudziak, J., Rudiger, L., Neale, M., Heath, A., & Todd, R. (1999). A twin study of inattentive, aggressive, and anxious/depressed behaviors. Journal of the American Academy of Child and Adolescent Psychiatry , 39, 469–476. *Jacobson, K., Prescott, C., & Kendler, K. (2000). Genetic and environmental influences on juvenile antisocial behavior assessed on two occasions. Psychological Medicine, 30, 1315–1325. *Jacobson, K., Prescott, C., & Kendler, K. (2002). Sex differences in the genetic and envi- ronmental influences on the development of antisocial behavior. Development and Psychopathology, 14, 395–416. Jaffee, S., Moffitt, T., Caspi, A., Taylor, A., & Arseneault, L. (2002). Influence of adult domestic violence on children’s internalizing and externalizing problems: An environ- mentally informative twin study. Journal of the American Academy of Child and Adolescent Psychiatry, 41, 1095–1103. Kamin, L. (1974). The science and politics of IQ. Oxford: Lawrence Erlbaum. Kim-Cohen, J., Caspi, A., Taylor, A., Williams, B., Newcombe, R., Craig, I., et al. (2006). MAOA, maltreatment, and gene-environment interaction predicting children’s mental health: New evidence and a meta-analysis. Molecular Psychiatry, 11, 903–913. *Koenen, K., Caspi, A., Moffitt, T., Rijsdijk, F., & Taylor, A. (2006). Genetic influences on the overlap between low IQ and antisocial behavior in young children. Journal of Abnormal Psychology, 115, 787–797. *Larsson, H., Andershed, A., & Lichtenstein, P. (2006). A genetic factor explains most of the variation in the psychopathic personality. Journal of Abnormal Psychology, 115, 221–230. Leon-Carrion, J. & Chacartegui-Ramos, F. (in press). Brain injuries and violent crime. In C. Ferguson (Ed.), Violent crime: Clinical and social implications. Thousand Oaks: Sage. Ferguson 179 *Ligthart, L., Bartels, M., Hoekstra, R., Hudziak, J., & Boomsma, D. (2005). Genetic con- tributions to subtypes of aggression. Twin Research and Human Genetics, 8, 483–491. *Malone, S., Taylor, J., Marmorstein, N., McGue, M., & Iacono, W. (2004). Genetic and environmental influences on antisocial behavior and alcohol dependence from adoles- cence to early adulthood. Development and Psychopathology, 16, 943–966. *McGue, M., Iacono, W., & Krueger, R. (2006). The association of early adolescent prob- lem behavior and adult psychopathology: A multivariate behavioral genetic perspective. Behavioral Genetics, 36, 591–602. Mercer, K., & Selby, M. (2005). The effects of psychopathy, violence and drug use on neuropsychological functioning. American Journal of Forensic Psychology, 23, 65–86. Miles, D., & Carey, G. (1997). Genetic and environmental architecture on human aggres- sion. Journal of Personality and Social Psychology, 72, 207–217. Moffitt, T. (2005). The new look of behavioral genetics in developmental psychopathol- ogy: Gene-environment interplay in antisocial behaviors. Psychological Bulletin, 131, 533–554. Nilsson, K., Sjoberg, R., Wargelius, H., Leppert, J., Lindstrom, L., & Oreland, L. (2007). The monoamine oxidase A (MAO-A) gene, family function and maltreatment as predictors of destructive behaviour during male adolescent alcohol consumption. Addiction, 102, 389–398. *O’Connor, T., McGuire, S., Reiss, D., Mavis-Hetherington, E., & Plomin, R. (1998). Co-occurrence of depressive symptoms and antisocial behavior in adolescence: A common genetic liability. Journal of Abnormal Psychology, 107, 27–37. *Polderman, T., Posthuma, D., De Sonneville, L., Verhulst, F., & Boomsma, D. (2006). Genetic analysis of teacher ratings of problem behavior in 5-year-old twins. Twin Research and Human Genetics, 9, 122–130. Raine, A., Lencz, T., Bihrle, S., LaCasse, L., Colletti, P. (2000). Reduced prefrontal gray matter volume and reduced autonomic activity in antisocial personality disorder. Annals of General Psychiatry, 57(2), 119–27. Retz, W., Retz-Junginger, P., Supprian, T., Thome, J., & Rosler, M. (2004). Association of serotonin transporter promoter gene polymorphism with violence: Relation with personality disorders, impulsivity and childhood ADHD psychopathology. Behavioral Sciences and the Law, 22, 415–425. Rhee, S., & Waldman, I. (2002). Genetic and environmental influences on antisocial behavior: A meta-analysis of twin and adoption studies. Psychological Bulletin, 128, 490–529. *Rushton, J. (1996). Self-report delinquency and violence in adult twins. Psychiatric Genetics, 6, 87–89. Rychlak, J. (1999). Social constructionism, postmodernism, and the computer model: Searching for human agency in the right places. Journal of Mind and Behavior, 20, 379–389. Sagan, C., & Druyan, A. (1992). Shadows of forgotten ancestors: A search for who we are. New York, NY: Random House. *Slutske, W., Eisen, S., Xian, H., True, W., Lyons, M., Goldberg, J., et al. (2001). A twin study of the association between pathological gambling and antisocial personality disorder. Journal of Abnormal Psychology, 110, 297–308. Soderstrom, H., Hultin, L., Tullberg, M., Wikkelso, C., Ekholm, S., & Forsman, A. (2002). Reduced frontotemporal perfusion in psychopathic personality. Psychiatry Research: Neuroimaging, 114, 81–94. *Taylor, J., Loney, B., Bobadilla, L., Iacono, W., & McGue, M. (2003). Genetic and envi- ronmental influences on psychopathy: Trait dimensions in a community sample of male twins. Journal of Abnormal Child Psychology, 31, 633–645. 180The Journal of Social Psychology Tedeschi, J., & Quigley, B. (1996). Limitations of laboratory paradigms for studying aggression. Aggression & Violent Behavior, 2, 163–177. Terpstra, D. (1981). Relationship between methodological rigor and reported outcomes in organization development evaluation research. Journal of Applied Psychology, 66, 541–543. Thapar, A., Langley, K., Fowler, T., Rice, F., Turic, D., Whittinger, N., et al. (2005). Catechol O-Methyltransferase gene variant and birth weight predict early-onset antiso- cial behavior in children with Attention-Deficit/Hyperactivity Disorder. Archives of General Psychiatry, 62, 1275–1278. *Thapar, A., & McGuffin, P. (1996). A twin study of antisocial and neurotic symptoms in childhood. Psychological Medicine, 26, 1111–1118. Tinbergen, N. (1963). On aims and methods of ethology. Zeitschrift fur Tierpsychologie, 20, 410–433. Tooby, J., & Cosmides, L. (1992). The psychological foundations of culture. In J. Barkow, L. Cosmides & J. Tooby. (Eds.), The adapted mind: Evolutionary psychology and the generation of culture (pp. 19–36). New York: Oxford University Press. *Tuvblad, C., Eley, T., & Lichtenstein, P. (2005). The development of antisocial behavior from childhood to adolescence: A longitudinal twin study. European Journal of Child and Adolescent Psychiatry, 14, 216–225. *Tuvblad, C., Grann, M., & Lichtenstein, P. (2006). Heritability for adolescent antisocial behavior differs with socioeconomic status: Gene-environment interaction. Journal of Child Psychology and Psychiatry, 47, 734–743. *van Beijsterveldt, C., Bartels, M., Hudziak, J., & Boomsma, D. (2003). Causes of stability of aggression from early childhood to adolescence: A longitudinal genetic analysis in Dutch twins. Behavioral Genetics, 33, 591–605. *Viding, E., James, R., Blair, R., Moffitt, T., & Plomin, R. (2005). Evidence for substan- tial genetic risk for psychopathy in 7-year-olds. Journal of Child Psychology and Psychiatry, 46, 592–597. *Vierikko,E., Pulkkinen,L., Kaprio, J., Viken, R., & Rose, R. (2003). Sex differences in genetic and environmental effects on aggression. Aggressive Behavior, 29, 55–68. *Vierikko,E., Pulkkinen,L., Kaprio, J., & Rose, R. (2006). Genetic and environmental sources of continuity and change in teacher-rated aggression during early adolescence. Aggressive Behavior, 29, 308–320. Williams, J. (1994). Violence, genes and prejudice. Discover Magazine. Retrieved January 14, 2007, from http://discovermagazine.com/1994/nov/violencegenesand446 Wilson, D., Dietrich, E., & Clark, A. (2003). On the inappropriate use of the naturalistic fallacy in evolutionary psychology. Biology and Philosophy, 18, 669–682. Received July 23, 2007 Accepted June 25, 2008
Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m
Genetic and Environmental Influences on Antisocial Behavior: A Meta-Analysis of Twin and Adoption Studies Soo Hyun Rhee and Irwin D. Waldman Emory University A meta-analysis of 51 twin and adoption studies was conducted to estimate the magnitude of genetic and environmental influences on antisocial behavior. The best fitting model included moderate proportions of variance due to additive genetic influences (.32), nonadditive genetic influences (.09), shared environmental influences (.16), and nonshared environmental influences (.43). The magnitude of familial influences (i.e., both genetic and shared environmental influences) was lower in parent– offspring adoption studies than in both twin studies and sibling adoption studies. Operationalization, assessment method, zygosity determination method, and age were significant moderators of the magnitude of genetic and environmental influences on antisocial behavior, but there were no significant differences in the magnitude of genetic and environmental influences for males and females. Considerable research has focused on the goal of explaining the etiology of antisocial behavior. In particular, the role of familial influences on antisocial behavior has been studied extensively. Dysfunctional familial influences such as psychopathology in the parents (e.g., Robins, 1966), coercive parenting styles (e.g., Patter- son, Reid, & Dishion, 1992), physical abuse (Dodge, Bates, & Pettit, 1990), and family conflict (e.g., Norland, Shover, Thornton, & James, 1979) have been shown to be significantly related to antisocial behavior. Often, these variables are considered environ- mental influences, and the possibility that they may also reflect genetic influences is not considered. This is unfortunate because disentangling the influences of nature and nurture is the first step toward the goal of eventually explaining the etiology of antisocial behavior. Also, estimating the relative magnitude of genetic and environmental influences on antisocial behavior is an important step toward the search for specific candidate genes and environ- mental risk factors underlying antisocial behavior. Although it is not possible to disentangle genetic from environmental influences in family studies because genetic and environmental influences areconfounded in nuclear families, twin and adoption studies have the unique ability to disentangle genetic and environmental influences and to estimate the magnitude of both simultaneously. More than a hundred twin and adoption studies of antisocial behavior have been published. Nonetheless, it is difficult to draw clear conclusions regarding the magnitude of genetic and environ- mental influences on antisocial behavior given the current litera- ture. The main reason for this difficulty is the considerable heter- ogeneity of the results in this area of research, with published heritability estimates (i.e., the magnitude of genetic influences) ranging from very low (e.g., .00; Plomin, Foch, & Rowe, 1981) to very high (e.g., .71; Slutske, Heath, et al., 1997). Various hypoth- eses have been proposed to explain these heterogeneous results across studies, including differences in the age of the sample (e.g., Cloninger & Gottesman, 1987), the age of onset of antisocial behavior (e.g., Moffitt, 1993), and the measurement of antisocial behavior (e.g., Plomin, Nitz, & Rowe, 1990). We conducted a meta-analysis of twin and adoption studies in order to provide a clearer and more comprehensive picture of the magnitude of genetic and environmental influences on antisocial behavior. Given previous hypotheses proposed to explain the het- erogeneity in the results, we examined the possible moderating effects of three study characteristics (i.e., the operationalization of antisocial behavior, assessment method, and zygosity determina- tion method) and two participant characteristics (i.e., the age and sex of the participants) on the magnitude of genetic and environ- mental influences on antisocial behavior. We examined the opera- tionalization of antisocial behavior given the evidence that anti- social personality disorder (ASPD), conduct disorder (CD), crim- inality, and aggression are distinct but related constructs (e.g., Robins & Regier, 1991). We examined assessment method and zygosity determination because of evidence suggesting that these are potential methodological confounders (e.g., McCartney, Har- ris, & Bernieri, 1990; Plomin, 1981). Sex was examined given the consistent evidence that antisocial behavior is more prevalent in males than females (e.g., Hyde, 1984; J. Q. Wilson & Herrnstein, 1985). Age was examined because of the potential to test an interesting hypothesis regarding the development of antisocial Soo Hyun Rhee and Irwin D. Waldman, Department of Psychology, Emory University. Earlier versions of this article were presented at the meeting of the American Society of Criminology in Chicago, Illinois, November 1996, and the meeting of the Behavior Genetics Association in Toronto, Ontario, Canada, July 1997. This work was supported in part by National Institute on Drug Abuse Grant DA-13956 and National Institute of Mental Health Grant MH-01818. We thank the authors who made data from unpublished studies available through personal communication. We also thank Deborah Finkel, Jenae Neiderhiser, Wendy Slutske, and Edwin van den Oord for making the data from their studies available before their publication, and we thank Scott O. Lilienfeld, Kim Wallen, and Terrie E. Moffitt for helpful comments on earlier versions of this article. Correspondence concerning this article should be addressed to Soo Hyun Rhee, who is now at the Institute for Behavioral Genetics, University of Colorado at Boulder, Campus Box 447, Boulder, Colorado 80309. E-mail: [email protected] Psychological BulletinCopyright 2002 by the American Psychological Association, Inc. 2002, Vol. 128, No. 3, 490 –5290033-2909/02/$5.00 DOI: 10.1037//0033-2909.128.3.490 490 behavior. L. F. DiLalla and Gottesman (1989) and Moffitt (1993) have suggested that individuals who engage in antisocial behavior can be divided into a smaller group whose antisocial behavior is persistent throughout the life course and influenced predominantly by genetics and a larger group whose antisocial behavior is limited to adolescence and influenced predominantly by environment. If their hypothesis is correct, the magnitude of genetic influences on antisocial behavior should be lower in adolescence than in child- hood or adulthood. Previous Reviews Examining Behavior Genetic Studies of Antisocial Behavior A number of traditional literature reviews (e.g., Carey, 1994; Gottesman & Goldsmith, 1994; Plomin et al., 1990) of twin and adoption studies of antisocial behavior have been published, and most researchers in this area have concluded that both genetic and environmental influences are important contributors to indi- vidual differences in antisocial behavior. Although these reviews are informative, they did not provide a quantitative estimate of the genetic and environmental influences on individual differences in antisocial behavior across studies. Three previous meta-analyses have provided such an estimate. Walters (1992) examined 11 family studies, 14 twin studies, and 13 adoption studies of crimi- nality and found genetic influences on crime that were low to moderate in magnitude (i.e., a mean unweighted phi coefficient of .25 and a mean weighted phi coefficient of .09). Mason and Frick (1994) examined 12 twin studies and 3 adoption studies of anti- social behavior and attributed approximately 50% of the variance in measures of antisocial behavior to genetic influences. They examined several moderating variables and found that effect sizes did not vary across the type of antisocial behavior (i.e., criminality, aggression, or antisocial personality), demographic variables (i.e., sex, age, and racial composition), and two methodological vari- ables (i.e., sample size and zygosity determination), but they found larger estimates of genetic influences for severe antisocial behav- ior, antisocial behavior in clinic-referred samples, and studies with optimal blinding(i.e., assessment of antisocial behavior that is blind to the relatives’level of antisocial behavior). Miles and Carey (1997) examined 20 twin studies and 4 adoption studies of aggression and concluded that genetic influences account for up to 50% of the variance. They also tested several potential moderators of genetic and environmental influences on aggression, including sex, age, and assessment method. The heritability estimate for males was higher than that for females, and the heritability esti- mate for younger samples was lower than that for older samples. Studies using parent reports yielded a lower heritability estimate and a higher estimate for the magnitude of shared environmental influences than those using self-reports. Walters’s (1992) and Mason and Frick’s (1994) meta-analyses have some methodological problems that make the interpretation of their results difficult. Mason and Frick (1994) provided a detailed description of the methodological problems (e.g., inclu- sion of nonindependent samples) in Walters’s meta-analysis. A serious concern about Mason and Frick’s meta-analysis is the effect size they chose to report. They used an effect size ofdfor both adoption studies and twin studies, subtracting the dizygotic (DZ) correlation from the monozygotic (MZ) correlation in twin studies and subtracting the adoptee–adoptive parent correlationfrom the adoptee–biological parent correlation in adoption studies. This effect size,d,is not appropriate because the difference be- tween the MZ correlation and the DZ correlation is not comparable to the difference between the adoptee–biological parent correla- tion and the adoptee–adoptive parent correlation. Heritability is estimated in twin studies by doubling the difference between the MZ correlation and the DZ correlation, whereas heritability is estimated in adoption studies by doubling only the adoptee– biological parent correlation. Another methodological problem in Mason and Frick’s study is that their effect size,d,included the difference between the concordances of MZ and DZ twins as well as the difference between the correlations of MZ and DZ twins. Concordances vary according to the base rate, such that the same concordances with different base rates are associated with different correlations (A. Heath, personal communication, March 1994). There are several important differences between the present meta-analysis and the previous three meta-analyses. First, the present study is more comprehensive, examining 10 independent adoption samples and 42 independent twin samples from 51 stud- ies (two separate samples were examined in Eley, Lichtenstein, & Stevenson, 1999). Second, we adopted a broader conceptualization of antisocial behavior, examining relevant diagnoses, criminality, aggression, and antisocial behavior (i.e., a composite index of delinquency and aggression). Third, as in Mason and Frick (1994), nonindependent samples were not treated as independent. Fourth, as in Miles and Carey (1997), direct analysis of the data was conducted. Fifth, more potential moderators were examined, in- cluding operationalization, assessment method, zygosity determi- nation method, sex, and age. Sixth, the present meta-analysis entailed a direct comparison between the results of twin and adoption studies. Seventh, the present meta-analysis also addresses several issues that could not be examined quantitatively in the meta-analysis because not enough studies in the literature exam- ined them. These issues include the role of genotype–environment interaction on antisocial behavior, longitudinal studies of antisocial behavior, and specific environmental influences on antisocial behavior. Operationalization as a Moderator The operationalizations of antisocial behavior can be divided into three major categories (Plomin et al., 1990). First, antisocial behavior has been examined in terms of psychiatric diagnoses, such as ASPD and CD. Second, antisocial behavior has been operationalized in terms of the violation of legal or social norms, that is, as criminality and delinquency. Third, antisocial behavior has been operationalized as aggressive behavior. TheDiagnostic and Statistical Manual of Mental Disorders(4th ed.;DSM–IV;American Psychiatric Association, 1994) described the essential features of ASPD as“a pervasive pattern of disregard for, and violation of, the rights of others that begins in childhood or early adolescence and continues into adulthood”(p. 645). A diagnosis of ASPD requires a history of CD before the age of 15 and three or more of the following criteria: failure to conform to social norms with respect to lawful behaviors (i.e., as indicated by repeatedly performing acts that are grounds for arrest), deceitful- ness, impulsivity, irritability and aggressiveness, reckless disre- gard for safety, consistent irresponsibility, and lack of remorse. CD, a criterion for the diagnosis of ASPD, is described by the 491 ANTISOCIAL BEHAVIOR DSM–IVas“a repetitive and persistent pattern of behavior in which the basic rights of others or major age-appropriate societal norms or rules are violated”(American Psychiatric Association, 1994, p. 90). It usually occurs in childhood or early adolescence and is manifested as aggression toward people and animals, de- struction of property, deceitfulness or theft, and serious violations of rules. Criminalityhas been defined as an unlawful act that leads to arrest, conviction, or incarceration, whereasdelinquencyhas been defined as unlawful acts committed as a juvenile. In addition to official records, past researchers also have assessed delinquency with anonymous self-reports of criminal activity that has not led to arrest, conviction, or incarceration. Aggression is usually studied as a personality characteristic and assessed with measures such as the Adjective Check List (Gough & Heilbrun, 1972) and the Multidimensional Personality Questionnaire (Tellegen, 1982, as cited in Tellegen et al., 1988). The operationalization of aggression has been very heterogeneous in the past, ranging from negative affect (Partanen, Bruun, & Markkanen, 1966) to the number of hits to a Bobo doll (Plomin et al., 1981). For the present review, the operationalization of aggression was restricted to the type of behavioral aggression described in theDSMcriteria for CD (e.g., bullying, initiating physical fights, and using a weapon that can cause serious physical harm). In deciding which studies to include in the present review, an important question had to be considered. Do diagnoses, criminality–delinquency, and aggression reflect the same con- struct? It is clear that the three operationalizations are related. CD and criminality are criteria for ASPD, whereas aggression and delinquency are criteria for CD. Past research shows moderate correlations between self-report measures of aggression and ASPD in a sample of individuals engaging in substance abuse (Mutaner et al., 1990) and a significant relation between criminality and ASPD in a sample of individuals engaging in criminal activity (e.g., number of prior arrests was significantly related to presence of antisocial disorder; Abram, 1989). In addition, childhood aggres- sion was found to predict adult criminality (e.g., boys rated by peers and teachers as highly aggressive at age 8 had more than five arrests on average at age 26 compared with less than two arrests in comparison boys; Pulkkinen & Pitka¨nen, 1993). On the other hand, the three operationalizations of antisocial behavior are not synonymous. The Epidemiologic Catchment Area study led by Robins and Regier (1991) reports that only 27% of boys and 21% of girls with three or more CD symptoms will be diagnosed with ASPD in adulthood, whereas 49% of boys and 33% of girls with six or more CD symptoms will be diagnosed with ASPD. Delinquency before age 15 predicted later ASPD in 29% of males and 13% of females. Also, whereas 40% of male criminals and 18% of female criminals qualify for an ASPD diagnosis, 55% of males with ASPD and 17% of females with ASPD are criminals. Although there are no definitive conclusions regarding the re- lations among the diagnoses of ASPD and CD, criminality– delinquency, and aggression other than that they are moderately overlapping constructs, studies examining all three operationaliza- tions of antisocial behavior were included in the present review for the following reasons. First, past reviews have focused on only one operationalization (e.g., aggression in Miles & Carey, 1997) or reviewed the results of studies using different operationalizationsseparately (e.g., Plomin et al., 1990). This is understandable, as conclusive evidence showing that the different operationalizations reflect the same construct is lacking, and the magnitude of genetic and environmental influences on antisocial behavior may differ across operationalizations (Plomin et al., 1990). Thus, in the present meta-analysis, studies examining all three operationaliza- tions were included to conduct a quantitative test of this issue. Second, age was examined as a possible moderator to examine potential developmental shifts in the relative magnitudes of genetic and environmental influences on antisocial behavior. In order to do so, studies using different operationalizations of antisocial behav- ior had to be included because antisocial behavior is expressed differently by children and adults and therefore is defined differ- ently for them. Third, adopting broader inclusion criteria increases the power of the meta-analysis, given that it is based on a greater number of studies. In addition to clinical diagnoses, criminality, and aggression, “antisocial behavior,”an omnibus operationalization that includes aggression and delinquency items, was examined. Some research- ers (e.g., Rowe, 1983) conducted twin studies of delinquency, although the measures used in these studies also included aggres- sion items. Also, some studies used measures including both aggression and delinquency items, such as the externalizing scale from the Child Behavior Checklist (Achenbach & Edelbrock, 1983). In addition to the general moderator analyses of operation- alization, we examined possible differences between ASPD and CD in a more focal analysis. Assessment Method as a Moderator Researchers have shown that assessment method can influence the results of behavior genetic studies. For example, McCartney et al. (1990) compared parent and self-reports of sociability and found that parent reports resulted in higher correlations than self- reports in MZ twins but resulted in lower correlations than self- reports in DZ twins. They also found that for activity–impulsivity, parent reports resulted in higher correlations than self-reports in both MZ and DZ twins. In contrast, Miles and Carey (1997) found that behavior genetic studies of aggression using parent reports resulted in a lower heritability estimate when compared with those using self-reports. Researchers studying temperament have found that parent re- ports tend to yield DZ correlations that are very low or even negative. This may be the result of parents’exaggerating the differences between their DZ twins, which has been described as arater contrast effect(Loehlin, 1992a). One example of such a finding emerged from the MacArthur longitudinal twin study (Emde et al., 1992). No resemblance of DZ twins on measures of behavioral inhibition and shyness was found using parent reports, but significant DZ resemblance was found using observational measures of the same constructs. Plomin’s (1981) review of twin studies examining personality concluded that objectively assessed behavior yielded lower heritabilities than self-reports and parent reports. Similarly, Miles and Carey’s (1997) meta-analysis of behavior genetic studies of aggression concluded that two studies using an objective method found little evidence of genetic influ- ences on aggression, in contrast to studies using self-report or parent report. 492 RHEE AND WALDMAN In addition to self-report, report by others (i.e., parent and teacher report), and objective measures, antisocial behavior has been assessed by two other methods. Criminality has been assessed with official records, and aggression has been assessed by exam- ining reactions to aggressive material (e.g., whether one finds aggressive humor to be funny or not; G. D. Wilson, Rust, & Kasriel, 1977). In the present review, assessment method was examined as a moderator, comparing self-report, report by others (i.e., parent and teacher report), official records, objective mea- sures, and reactions to aggressive material. Zygosity Determination Method as a Moderator Zygosity determination method was examined as a possible moderator of genetic and environmental influences on individual differences in antisocial behavior. Zygosity determination methods used in twin studies of antisocial behavior include blood grouping, questionnaires, and a combination of the two methods. The inac- curacy of blood grouping in determining the zygosity of twin pairs is less than 1% (e.g., Smith & Penrose, 1955). Questionnaire methods of determining zygosity, which involve asking about the physical similarity of the twin pairs, have been found to agree highly with zygosity diagnosis by blood grouping. For example, Kasriel and Eaves (1976) found that if all twin pairs who agree that they were confused in childhood and are alike in appearance are determined to be MZ, only 3.9% of the sample would be diagnosed incorrectly. Nevertheless, estimates of the magnitude of genetic and environmental influences may be affected by the zygosity determination method. McCartney et al. (1990) predicted that studies that used blood grouping would have higher effect sizes for MZ twins and lower effect sizes for DZ twins because use of blood grouping in zygosity determination would purify the MZ and DZ samples. They found that studies using blood grouping did have higher effect sizes for MZ twins, but that the zygosity determina- tion method did not moderate effect sizes for DZ twins. In the present review, studies using blood groupings, questionnaires, and a combination of the two methods (i.e., studies using the question- naire method for the whole sample and the blood grouping method for a subset of the sample) are compared. Age as a Moderator It is important to investigate age as a possible moderator of genetic and environmental influences on human behavior in gen- eral and on antisocial behavior in particular. In the behavior genetics literature, there is a general finding for a variety of traits that as age increases, the magnitude of genetic and nonshared environmental influences increases, whereas the magnitude of shared environmental influences decreases (Loehlin, 1992a; Plo- min, 1986). One example of such a finding is Matheny’s (1989) longitudinal study of temperament. Over 12 to 30 months of age, MZ twins became more concordant than DZ twins for age-to-age changes in temperament measures of emotional tone, fearfulness, and approach. McCartney et al. (1990) conducted a meta-analysis of develop- mental changes in genetic and environmental influences on intel- ligence and several personality variables. They reported correla- tions between the components of variance (i.e., heritability, shared environment, and nonshared environment) and age for the threevariables examined in the most studies (i.e., intelligence, sociabil- ity, and activity–impulsivity). In general, the correlations for both MZ and DZ twin pairs decreased as age increased, and this finding also applied to the eight studies examining aggression. The results were inconsistent, and the authors cautioned that they may not be reliable because they are based on few data points. Also, it should be noted that researchers conducting another meta-analysis that examined genetic and environmental influences on intelligence (Devlin, Daniels, & Roeder, 1997) concluded that an age-effects model, which allowed the heritability of IQ to increase with age, failed to fit the data better than a simpler model. In Miles and Carey’s (1997) meta-analysis of behavior genetic studies examin- ing aggression, the magnitude of shared environmental influences decreased and the magnitude of genetic influences increased from childhood to adulthood. There appears to be conflicting evidence regarding age as a moderator of genetic and environmental influences on criminality. In five early twin studies examining juvenile delinquency, the weighted average of concordance rates for MZ and DZ twins was .87 and .72, respectively (Cloninger & Gottesman, 1987). In com- parison, in seven early twin studies examining adult criminality, the weighted average of concordance rates for MZ and DZ twins was .51 and .23, respectively (Cloninger & Gottesman, 1987). These results suggest that juvenile delinquency during adoles- cence, unlike criminality during adulthood, is only moderately affected by genetic influences but is very strongly affected by shared environmental influences. Given these results, researchers have theorized that genetic influences on individual differences in delinquency may be minimal because the base rate for delinquency is very high (L. F. DiLalla & Gottesman, 1989) or because envi- ronmental influences such as peer pressure are particularly strong in adolescence (Raine & Venables, 1992). Pertinent to these hy- potheses, Lyons et al. (1995) assessed juvenile and adult ASPD symptoms in the same participants using retrospective self-report. They found that the heritability for the adult antisocial traits (h 2 .43) was higher than that for the juvenile antisocial traits (h2 .07), supporting Cloninger and Gottesman’s conclusions. In contrast, Rowe (1983) examined anonymous self-reports of delinquent acts and found that both genetic and environmental influences are substantial for juvenile delinquency. Some review- ers have attributed Rowe’s contradictory finding to his use of self-report and have suggested that the finding of genetic influ- ences is reflecting the response to questionnaires rather than the construct of juvenile delinquency (e.g., L. F. DiLalla & Gottes- man, 1991). They have also noted that the finding of genetic influences may be a function of including items that assess ag- gression rather than delinquency. Other limitations of this study include a low response rate, which raises issues regarding sam- pling biases, and the use of a mailed questionnaire, which raises the possibility of nonindependent responses. On the other hand, Rowe and Rodgers (1989) asserted that it is premature to conclude that genetic influences are not important for delinquency, as the early twin studies had many methodological problems (e.g., hap- hazard sampling, small sample size, and variance in zygosity determination method). Carey (1994) admitted that the methodol- ogy of the early twin studies was generally poor but noted that similar methodological problems did not prevent finding genetic influences on adult criminality. 493 ANTISOCIAL BEHAVIOR It is not possible to conclude from the information provided by the traditional literature reviews (e.g., Cloninger & Gottesman, 1987; L. F. DiLalla & Gottesman, 1989) whether age is an impor- tant moderator of genetic and environmental influences on antiso- cial behavior or criminality. In the present meta-analysis, we used participants’age as a moderator in order to examine this issue, comparing results for children (below age 13), adolescents (ages 13–18), and adults (above age 18). The significance of age of onset and the continuity of antisocial behavior is discussed in several traditional literature reviews (e.g., Cloninger & Reich, 1983; L. F. DiLalla & Gottesman, 1989; Gottesman & Goldsmith, 1994). In particular, L. F. DiLalla and Gottesman (1989) hypothesized that there are three different types of offenders: continuous antisocials (i.e., those are who are delin- quent as youths and continue to be criminal as adults), transitory delinquents (i.e., youths who are delinquent but not criminal as adults), and late bloomers (i.e., adults who are criminal but were not delinquent as adolescents). They accepted the conclusion of the early twin studies (e.g., Cloninger & Gottesman, 1987) that genetic influences are minimal for juvenile delinquency, and they hypoth- esized that delinquency is in many cases transitory and primarily affected by peer pressure. A review by Moffitt (1993) concurs with L. F. DiLalla and Gottesman’s (1989) hypothesis. Moffitt noted that although anti- social behavior shows impressive continuity over age, the preva- lence of antisocial behavior increases almost 10-fold during ado- lescence. She also suggested a subtype hypothesis for antisocial behavior, with the first subtype comprising a small group of members who are antisocial from an early age and who continue to be antisocial during adulthood, and the second subtype being a much larger group whose members have a later age of onset for antisocial behavior and are only antisocial during adolescence. She hypothesized that the correlates and causes of persistent crime or antisocial psychopathology (e.g., genetic influences) may not be found in those who engage in juvenile delinquency. Two recent twin studies have yielded data that are relevant to the issues of age of onset and continuity of antisocial behavior. First, Slutske, Lyons, et al. (1997) found that antisocial behavior that is earlier in onset is no more heritable than later-onset anti- social behaviors, but they also found that antisocial behavior that is persistent across the life span is more heritable than antisocial behavior that is limited to either childhood or adulthood. Slutske, Lyons, et al. cautioned that the use of retrospective reports may be a limitation of their study. Second, Waldman, Levy, and Hay (1997) examined the etiology of four types of antisocial behavior (i.e., oppositionality, aggression, property violations, and status violations) that vary monotonically in their median age of onset from 6 years old (oppositionality) to 9 years old (status violations). They found that antisocial behavior with an earlier age of onset is more heritable and shows a lesser magnitude of shared environ- mental influences than antisocial behavior with a later age of onset. Given that so few twin studies have addressed the issue of age of onset or continuity of antisocial behavior, the present review cannot provide conclusive evidence for or against L. F. DiLalla and Gottesman’s (1989) hypothesis. If one assumes, however, that antisocial behavior in adolescents is more transitory in general (although adolescents with continuous and transitory antisocial behavior are not distinguished), the results should indicate that themagnitude of genetic influences on antisocial behavior should be lowest in adolescence. Sex as a Moderator No matter how antisocial behavior is operationalized or as- sessed, it is more prevalent in males than females (e.g., Hyde, 1984; J. Q. Wilson & Herrnstein, 1985). Given this sex difference in prevalence, it is important to consider whether the magnitude of genetic and environmental influences differs in males and females. Therefore, the present meta-analysis examined whether sex is a significant moderator of the results of behavior genetic studies of antisocial behavior by comparing the results for males, females, and both sexes (i.e., studies reporting results for a combined sample of males and females or studies reporting results for opposite-sex twin pairs). Past literature reviews (e.g., Widom & Ames, 1988) have suggested that the magnitude of genetic and environmental influences on antisocial behavior is equal for the two sexes, whereas Miles and Carey (1997) found that the mag- nitude of genetic influences on aggression was slightly higher for males than for females. One confusion in this area has to be addressed. The polygenic multiple threshold model attempts to explain the sex difference in prevalence by suggesting that the less affected sex needs a greater liability to manifest the disorder. There has been substantial sup- port for this model in the area of antisocial behavior (e.g., Med- nick, Gabrielli, & Hutchings, 1983; Sigvardsson, Cloninger, Boh- man, & von Knorring, 1982). Raine and Venables (1992) sug- gested that such support for the polygenic multiple threshold model conflicts with the evidence that heritability is equal in males and females (Widom & Ames, 1988). The polygenic multiple threshold model makes a prediction about the degree of liability (both genetic and environmental) needed to express a disorder, rather than the magnitude of genetic or environmental influences on within-sex individual differences in antisocial behavior, how- ever. The fact that females may need more liability (either genetic or environmental) to express antisocial behavior does not mean that genetic influences are of greater magnitude in females than males. Confounding Among Moderators In examining age, operationalization, and assessment method as moderators, the potential confounding among these variables must be investigated. Antisocial behavior is operationalized and as- sessed differently for children, adolescents, and adults (e.g., CD assessed by means of parent report in children vs. ASPD assessed by means of self-report in adults). Also, certain operationalizations of antisocial behavior are most frequently or readily assessed using certain methods (e.g., criminality by means of official records). Therefore, the age of the participants, the operationalization, and the assessment method may be all highly correlated across the studies of antisocial behavior. Given these concerns, we assessed the potential confounding among these moderators in the studies included in the present meta-analysis. Tables 1 and 2, which show the number of studies at each level of the moderators, demonstrate the problem of con- founding between the following pairs of moderators: age and operationalization, age and assessment method, and operational- 494 RHEE AND WALDMAN ization and assessment method. If there was no confounding be- tween the potential moderators, the numbers of studies in these tables would be distributed equally throughout the tables. For example, males and females were nearly equally distributed across the four types of operationalization. In an extreme example, all of the studies using the assessment method of official records were studies examining the operationalization of criminality. In a less extreme example, the studies using the assessment method of report by others tended to be those examining antisocial behavior in childhood. This type of confounding can make the interpretation of results difficult in two ways. First, if two confounded moderators are both found to be significant, it is possible that the second moderator is significant only because of its confounding with the first moder- ator. Fortunately, this problem can be assessed in the present meta-analysis. Each of the three moderators in question, opera- tionalization, assessment, and age, was tested for significance after the other two moderators were controlled for statistically. Second, if one level of a moderator is completely confounded with a level of another moderator (e.g., all studies examining criminality being assessed by records), it is unclear whether the results reflect the first or second moderator. Unfortunately, we cannot resolve this problem in the present review. This problem can be addressed in future research, however, by diversifying the pairings among op- erationalization, assessment method, and age (e.g., by conducting more studies of criminality using a variety of assessment methods, rather than criminal records alone). Tables 1–2, and the corre- sponding tests of moderators in the meta-analysis, thus serve as aguide to fruitful directions for future behavior genetic studies of antisocial behavior. Comparisons Between Twin and Adoption Studies The results of twin and adoption studies were directly compared in the present meta-analysis. Twin and adoption studies have unique assumptions or biases that can make interpretations of their results difficult. Comparing the results of twin and adoption stud- ies can help determine whether the results of behavior genetic studies have been influenced by these unique assumptions or biases. To the degree that the results of twin and adoption studies are similar, it is more likely that the results reflect the true magnitude of genetic and environmental influences. One cannot rule out the possibility, however, that the results of twin and adoption studies are similar because they share similar biases to some extent that influence their results in the same direction. Therefore, the following assumptions and biases always should be considered when interpreting the results of behavior genetic studies. In twin studies comparing the correlations between MZ and DZ twin pairs, one has to make the equal environments assumption, or the assumption that the environmental influences on the trait being examined are no more or less similar for MZ twins than for DZ twins. It is possible that the environmental influences on MZ twins are more similar because they are treated more similarly given their similar appearance. This bias could result in the overestima- tion of genetic influences. Another factor to consider in the equal environments assumption is that approximately two thirds of MZ twin pairs are monochorionic (i.e., share the same chorion), whereas one third of MZ twin pairs and all DZ twin pairs are dichorionic (Melnick, Myrianthopoulos, & Christian, 1978). Fail- ure to account for the effect of sharing a chorion may bias esti- mates of genetic and environmental influences if prenatal environ- ment influences the trait being examined (Prescott, Johnson, & McArdle, 1999). Several studies have found that monochorionic MZ twins are more similar than dichorionic MZ twins in person- ality (e.g., Reed, Carmelli, & Rosenman, 1991; Sokol et al., 1995) and cognitive ability (e.g., Rose, Uchida, & Christian, 1981), although others have failed to find significant differences between the two types of MZ twins (e.g., in temperament, Riese, 1999; in cognitive ability, Sokol et al., 1995). Also, sharing a chorion actually may lead to decreased similarity in monochorionic MZ twin pairs because of competition for resources within a twin pair as evidenced by greater similarity in birth weight for dichorionic MZ twins than for monochorionic MZ twins (Corey, Nance, Kang, & Christian, 1979; Vlietinck et al., 1989). Table 1 Confounding Between Assessment Method and Operationalization: Number of Samples at Each Level of Moderator OperationalizationAssessment method Self Others Records Reaction Objective Diagnosis 11 2 Criminality 5 Aggression 7 4 2 1 Antisocial behavior 4 8 Table 2 Confounding Between Age and Operationalization–Assessment Method: Number of Samples at Each Level of Moderator ModeratorAge Children Adolescents Adults Operationalization Diagnosis 2 5 4 Criminality 3 Aggression 6 7 Antisocial behavior 7 5 2 Assessment method Self 1 7 12 Others 12 2 Records 3 Reactive 1 1 Objective 1 495 ANTISOCIAL BEHAVIOR Another assumption of studies examining twins reared together is the assumption that the genetic variance is primarily additive and that there is no epistasis (i.e., interaction between alleles in different loci). The violation of this assumption may lead to overestimation of heritability and underestimation of the magni- tude of shared environmental influences (Grayson, 1989). Al- though twin studies examine models including dominance (i.e., interaction between alleles in the same locus), they do not examine models including epistasis. The coefficient for genetic relationship for epistatic interactions depends on the number of loci involved and the type of interaction (Falconer & Mackay, 1996). The coefficient for genetic relationship for dominance is equal to the coefficient for genetic relationship for epistatic interactions only for Additive Additive interactions between two loci. Eaves (1988) pointed out that in many behavior genetic studies, the difference between MZ and DZ correlations is much bigger than that predicted under an additive genetic model or a model includ- ing dominance (i.e., interaction between alleles in the same locus) alone. He also demonstrated that duplicate gene interactions be- tween pairs of moderately frequent alleles at polygenic loci pro- duce very small genetic correlations (approximately .12) between siblings compared with a genetic correlation of .50 for additive genetic influences and .25 for dominance genetic influences. Another issue to consider when interpreting the results of twin studies is the generalizability of the findings. First, volunteers in social science studies tend to be above average in socioeconomic status (SES), and this would pertain to twin studies just as it does for other studies. Second, pre- and perinatal complications are more common in twin pairs than in singletons. Twins are born 3 to 4 weeks premature on average, are 30% lighter at birth, and tend to have delayed language development (Plomin, DeFries, McClearn, & Rutter, 1997). Given this concern, several research- ers have compared the prevalence of antisocial behavior in twins and singletons and reached differing conclusions. Gjone and Nøvik (1995; Norwegian twins) and van den Oord, Koot, Boomsma, Verhulst, and Orlebeke (1995; Dutch twins) found that the level of antisocial behavior in twins is similar to that of singletons. On the other hand, Gau, Silberg, Erickson, and Hewitt (1992; Virginia twins) found small but consistent differences between the level of antisocial behavior in twins and singletons. Twins had higher levels of antisocial behavior than singletons in both older and younger children. They also found tentative support for the relation between increased perinatal complications and increased child- hood behavior problems in twins. If the range of environmental influences is restricted in twin samples for any reason (e.g., higher SES in volunteers; more pre- and perinatal complications), the magnitude of genetic influences may be overestimated. Adoption studies also have several selection or sampling biases that make interpretation of their results difficult. First, it may be difficult to generalize results of adoption studies because adoptees have a higher rate of antisocial behavior compared with the general population. This finding has been replicated in adoptees in several countries—for example, New Zealand (Fergusson, Lynskey, & Horwood, 1995), the Netherlands (Verhulst, Versluis-den Bieman, van der Ende, Berden, & Sanders-Woudstra, 1990), and the United States (Sharma, McGue, & Benson, 1998). Second, the range of the adoptee’s adoptive home environment is restricted. For example, Fergusson et al. (1995) found that adoptees had sev- eral advantages over children in the general population infamily stability, educational opportunities, standards of health care, material living standards, and mother–child interactions. Although one can ensure that the SES of the adoptive families is similar to that of the control families (e.g., Scarr & Weinberg, 1978), genetic influences may be overestimated and shared environmental influences may be underestimated when the sam- ple’s range of environments is restricted (Stoolmiller, 1999). Third, selective placement (viz., matching the environmental characteristics of the biological parents’home and the adoptive parents’home) often occurs in adoptions. Clerget-Darpoux, Goldin, and Gershon (1986) demonstrated how a genetic effect is simulated in adoption studies when there is a positive corre- lation between the adoptive and biological parents for an etio- logic environmental variable. Two types of adoption studies were included in the present meta-analysis: (a) parent–offspring adoption studies (i.e., compar- ing the correlation between adoptees and their adoptive parents with the correlation between adoptees and their biological parents) and (b) sibling adoption studies (i.e., comparing the correlation between adoptive siblings with the correlation between biological siblings). When parent–offspring data are interpreted, it is impor- tant to consider the possibility that the correlations between the parents and the offspring may be reduced by the age difference between the two generations and that the magnitude of familial (i.e., genetic and shared environmental) influences may be under- estimated. Genetic influences on a trait may differ from one generation to another because the genes affecting the same trait may differ in their expression across age because of genotype– environment interaction. For example, genetic influences in the younger generation may be increased because of environmental facilitation of antisocial behavior—for example, by means of sec- ular increases in substance use and less stringent parenting prac- tices (e.g., Lykken, 1997). Also, there may be cohort-specific shared environmental influences other than the cultural transmis- sion from parents to offspring. Unfortunately, the parent–offspring adoption studies included in the present meta-analysis did not provide enough information to address these possibilities. There- fore, each type of adoption study was compared with the twin studies separately, and the parent–offspring adoption studies were compared with the rest of the studies combined (i.e., twin studies and sibling adoption studies). Behavior genetic studies also make the assumption of random mating. The studies included in the present meta-analysis did not control for the effects of assortative, or nonrandom, mating. A significant correlation between the phenotypes of couples is evi- dence of assortative mating, and Krueger, Moffitt, Caspi, Bleske, and Silva (1998) found evidence of substantial positive assortative mating for antisocial behavior in their sample of 360 couples from New Zealand. Although assortative mating for personality traits related to antisocial behavior was low (r .15), assortative mating for self-reports of antisocial behavior and tendency to associate with peers who engage in antisocial behavior was high (r .54). Positive assortative mating leadsh 2estimates to be biased down- ward andc 2estimates to be biased upward in twin studies because the genetic resemblance between DZ twins is increased. In adop- tion studies, theh 2estimate is biased upward because the genetic resemblance between the adopted away offspring and the biolog- ical parent, as well as between biological siblings, is increased in the presence of positive assortative mating. As Krueger et al. 496 RHEE AND WALDMAN suggested, future behavior genetic studies examining antisocial behavior should attempt to control for the effects of assortative mating in order to obtain unbiased estimates of the magnitude of genetic and environmental influences. Method Search Strategy We began our search for twin and adoption studies of antisocial behavior by examining the PsycINFO andMedlinedatabases. Appendix A shows the search terms used in this process. The references from the research studies and review articles found through this method were examined for any additional studies that might have been missed or published before the databases were established. Also, information about relevant unpublished manuscripts or manuscripts in press was obtained by examining pertinent reviews and the abstracts of the 1995, 1996, 1997, and 1998 Behavior Genetics Association meetings and searching theDissertation Abstracts InternationalandEducational Resources Information Centerdatabases. Authors of 14 manuscripts provided unpublished data; four of these manu- scripts were published subsequently. One hundred forty-one twin and adoption studies examining antisocial behavior were identified. After we excluded unsuitable studies according to the criteria described below (i.e., construct validity, inability to calculate tetrachoric or intraclass correlations, and assessment of related disor- ders), 96 studies remained. After we addressed the problem of noninde- pendence in these studies, 51 studies (i.e., 10 independent adoption sam- ples and 42 independent twin samples [two separate samples were examined in Eley et al., 1999]) remained. Tables 3 and 4 list the 26 adoption studies and 70 twin studies that met the first three inclusion criteria, respectively. The tables are grouped by the 10 independent adoption samples and the 42 independent twin samples in the meta-analysis, and the inclusion–exclusion column indicates which studies were included and which studies were excluded. Tables 3 and 4 also indicate the operationalization examined in the study, the method of assessment, the method of zygosity determination, the mean or midpoint age, the sex of the sample, the number of pairs, the relationship of the pairs, and the effect sizes. Inclusion Criteria for Studies in the Meta-Analysis Construct Validity General issues.Thirteen studies were excluded from the meta-analysis because of inadequate construct validity. The validity of the measures used in the studies considered for the meta-analysis was an important issue in deciding whether to include or exclude a study. Only studies examining antisocial behavior were included, and those examining related constructs such as anger and hostility were excluded. The included studies met one of the following qualifications. First, a study was included if it was clearly evident that it examined ASPD, CD, criminality, or aggression. Examples include studies assessing criminality with official records and ASPD with DSMcriteria. Second, a study was included if there was empirical evidence that the measure of antisocial behavior used successfully discriminated between an antisocial group and a control group or if the measure was significantly related to a more established operationalization of antisocial behavior. We discuss the validity issues in more detail below for each operationalization of antisocial behavior. Clinical diagnoses.As mentioned above, studies that usedDSMcri- teria to assess ASPD or CD were included. It was not as clear whether studies examining psychopathy should be included in the operationaliza- tion of clinical diagnoses. TheDSM–IV(American Psychiatric Associa- tion, 1994) states,[The pattern of ASPD] has also been referred to as psychopathy, sociopathy, or dyssocial personality disorder. Because deceit and manipulation are central features of ASPD, it may be especially helpful to integrate information acquired from systematic clinical assessment with information collected from collateral sources. (pp. 645–646) However, some researchers have emphasized the difference between the DSMcriteria and the traditional concept of psychopathy, noting that the DSMcriteria for ASPD focus on antisocial behavior whereas the traditional concept of psychopathy focuses on personality traits (e.g., Hare, Hart, & Harpur, 1991). The two personality measures used most often in assessing psychopathy are the Minnesota Multiphasic Personality Inventory (MMPI; Hathaway & McKinley, 1942) Psychopathic Deviate (Pd) scale and the California Psychological Inventory (CPI; Gough, 1969) Socialization (So) scale. The MMPI was constructed empirically to distinguish nonpsychopathological from psychopathological populations, whereas the CPI was justified the- oretically to describe variation within the general population. Approxi- mately one third of the items on the CPI were derived from the MMPI, however. Given the evidence that psychopathy measures and theDSM criteria are related (e.g., Cooney, Kadden, & Litt, 1990), psychopathy measures were included as an operationalization of diagnosis. Nonetheless, given the concern that psychopathy and ASPD are not synonymous (e.g., Hare et al., 1991), the meta-analysis was repeated after excluding studies examining psychopathy (eight samples; Brandon & Rose, 1995; D. L. DiLalla, Carey, Gottesman, & Bouchard, 1996; Gottesman, 1963, 1965; Loehlin & Nichols, 1976; Loehlin, Willerman, & Horn, 1987; Taylor, McGue, Iacono, & Lykken, 2000; Torgersen, Skre, Onstad, Edvardsen, & Kringlen, 1993) to examine the sensitivity of the results to such studies. Criminality and delinquency.All studies examining criminality used the assessment method of official records of arrests or convictions and were therefore included in the meta-analysis. Aggression.A study examining aggression was included if it examined behavioral aggression (e.g., physical fighting, cruelty to animals, and bullying). For studies that did not meet this criterion, several issues regarding validity had to be resolved. First, 12 studies that examined other related variables such as anger, hostility, or impulsivity were not included. These studies were excluded because it was not clear whether they exam- ined aggression or some related but distinct trait. Second, Partanen et al. (1966) was excluded because although it reported that it examined aggres- sion, the aggression items examined in this study (e.g.,“Are you readily insulted?”and“Do you easily become unhappy about even small things?”) suggest that negative affect or anger, rather than aggression, was being assessed. Third, some studies examining aggression used measures with questionable validity (i.e., lack of evidence or inconclusive evidence re- garding validity). For example, the Missouri Children’s Picture Series (Sines, Pauker, & Sines, 1966, as cited in Owen & Sines, 1970) used by Owen and Sines distinguished institutionalized aggressive boys from boys from the general population (Defilippis, 1979) but did not distinguish teacher-referred children with behavior problems versus learning problems (Ollendick & Woodward, 1982). The meta-analysis was repeated after excluding the studies using measures with questionable validity (2 sam- ples; Owen & Sines, 1970; G. D. Wilson et al., 1977) to assess the sensitivity of the obtained results to inclusion of such measures. Antisocial behavior.A fourth operationalization, antisocial behavior, was included because several studies clearly examined antisocial behavior without specifically examining ASPD, CD, criminality, delinquency, or aggression (e.g., Rowe, 1983; Stevenson & Graham, 1988; Waldman, McGue, Pickens, & Svikis, in press). All of these studies examined a combination of delinquency and aggression items (e.g., the externalizing scale from the Child Behavior Checklist; Achenbach & Edelbrock, 1983). Many of the individual items examined in these studies are criteria for CD, (text continues on page 508) 497 ANTISOCIAL BEHAVIOR Table 3 Effect Sizes for Adoption Studies of Antisocial Behavior Study Operationalization Assessment Age SexNRelationship Effect size Inclusion–exclusion Texas adoptees Loehlin et al. (1985) Psychopathy Self-report both-m 52 a-bf .16 Excluded both-fm 53 a-bm .06 Excluded both-m 241 a-af .03 Excluded both-fm 253 a-am .02 Included—adoption Loehlin et al. (1987) Psychopathy Self-report both-m 81 a-bf .12 Included—averaged both-fm 81 a-bm .07 Included—averaged both-m 180 a-af .07 Excluded both-fm 177 a-am .01 Excluded St. Louis adoptees Cadoret et al. (1975) ASB Records–parent report both-both 76 a-bp .23 Excluded Cunningham et al. (1975) ASB Records–parent report both-both 77 a-bp .16 Included—largest Dutch adoptees van den Oord et al. (1994) ASB Parent report 12.40 fm-fm 35 b sibs .42 Included—independent 48 a sibs .37 Included—independent m-m 30 b sibs .46 Included—independent 44 a sibs .19 Included—independent m-f 46 b sibs .52 Included—independent 129 a sibs .11 Included—independent U.S. adoptees (CO, IL, MN, WI) McGue et al. (1996) ASB (externalizing) Self-report 15.00 both-both 194 a sibs (a-a) .17 Included—averaged 72 a sibs (a-b) .06 Included—averaged ASB (antisocial) 194 a sibs (a-a) .05 Included—averaged 72 a sibs (a-b) .34 Included—averaged Iowa adoptees Cadoret (1978) ASB/criminality–ASP/CD Records–self- or m-both 81 a-bp .10 Excluded parent report fm-both 56 a-bp .42 Excluded Cadoret et al. (1985) ASB–ASP Records–self- or parent reportm-both fm-both127 87a-br a-br.43 .20Excluded Excluded Cadoret et al. (1986) ASB–ASP Records–self-report both-both 443 a-bp .40 Included—largest Cadoret et al. (1987) ASB–ASP Records–self-report m-both 160 a-bp .37 Excluded Cadoret et al. (1990) Criminality/delinquency-ASP Records–self- or parent reportm-both 286 a-bp .47 Excluded Cadoret & Stewart (1991) Criminality/delinquency–ASP Records–self-report m-both 283 a-bp .47 Excluded Iowa adoptees (1990s—males) Cadoret et al. (1995) ASPD–aggressivity Records–parent report m-both 95 a-bp .53 Included—independent Iowa adoptees (1990s—females) Cadoret et al. (1996) ASP–CD Records–self-report fm-both 102 a-bp .49 Included—independent Danish adoptees Hutchings & Mednick (1971) Criminality Records m-m 971 a-bf .19 Excluded Mednick et al. (1983) Criminality Records m-both 3,722 a-bp .20 Excluded Gabrielli & Mednick (1984) Criminality Records m-both 2,592 a-bp .17 Excluded Baker (1986) Criminality Records m-fm 2,532 a-bm .17 Excluded m-m a-bf .04 Excluded m-fm a-am .06 Excluded m-m a-af .06 Excluded 498 RHEE AND WALDMAN Study Operationalization Assessment Age SexNRelationship Effect size Inclusion–exclusion Danish adoptees (continued) Baker et al. (1989) Criminality Records fm-fm 7,065 a-bm .15 fm-m a-bf .12 fm-both a-bp .14 Included—largest fm fm-fm a-am .02 fm-m a-af .05 fm-both a-ap .01 Included—adoption m-fm 6,129 a-bm .20 m-m a-bf .14 m-both a-bp .17 Included—largest m m-fm a-am .06 m-m a-af .11 m-both a-ap .09 Included—adoption Swedish adoptees Bohman (1978) Criminality Records m-m 892 a-bf .01 Excluded m-f 1,077 a-bm .00 Included—largest m fm-both 1,988 a-bp .12 Included—largest fm Bohman et al. (1982) Criminality Records m-both 702 a-bp .11 Excluded Sigvardsson et al. (1982) Criminality Records fm-both 407 a-bp .34 Excluded 457 a-ap .23 Excluded Colorado adoptees Deater-Deckard & Plomin Aggression Parent report 9.5 both-both 78 a sibs .26 Excluded (1999)94 b sibs .39 Excluded Teacher report 78 a sibs .06 Excluded 94 b sibs .25 Excluded Delinquency Parent report 78 a sibs .22 Excluded 94 b sibs .43 Excluded Teacher report 78 a sibs .14 Excluded 94 b sibs .24 Excluded Young et al. (1997, CD–dimensional (adoptive sample) Self-report 18.60 both-both 56 a sibs .10 Excluded personal communication) CD–dimensional (control sample) 18.10 42 b sibs .17 Excluded CD–dimensional (treatment sample) 17.05 144 b sibs .04 Excluded CD–threshold (adoptive sample) 18.60 56 a sibs .00 Included—largest CD–threshold (control treatment) 17.58 186 b sibs .31 Included—largest Young et al. (1996, CD (adoptive sample) Self-report m-m 43 a-af .15 Excluded personal communication)m-fm 57 a-am .02 Excluded CD (control sample) m-m 96 b-bf .17 Excluded m-fm 87 b-bm .16 Excluded CD (treatment sample) m-m 34 b-bf .21 Excluded m-fm 86 b-bm .28 Excluded Parker et al., (1989, as cited in Aggression Parent report 4.00 both-both 66 b sibs .42 Excluded Carey, 1994)45 a sibs .54 Excluded 7.00 19 b sibs .55 Excluded 17 a sibs .28 Excluded Note.Information within parentheses indicates whether data were obtained from personal communication or another publication. both both male and female; m male; a-bf adoptee-biological father; Excluded excluded from the meta-analysis because another study examined the same sample that was larger, more unique in assessment method, or better described; fm female; a-bm adoptee-biological mother; a-af adoptee-adoptive father; a-am adoptee-adoptive mother; Included—adoption included only for analysis comparing parent–offspring studies with other types of studies; Included—averaged included after using the averaging method of dealing with nonindependence; a-bp adoptee–biological parent; ASB antisocial behavior; Included—largest included after using the largest sample method of dealing with nonindependence; b sibs biological siblings; Included—independent included because the study did not have a nonindependence problem; a sibs adoptive siblings; CO Colorado; IL Illinois; MN Minnesota; WI Wisconsin; a-a adoptive—adoptive sibling pair; a-b adoptive—biological sibling pair; ASP antisocial personality; CD conduct disorder; a-br adoptee–biological relative; ASPD antisocial personality disorder; a-ap adoptee–adoptive parent; b-bf biological child–biological father; b-bm biological child–biological mother. 499 ANTISOCIAL BEHAVIOR Table 3 (continued) Table 4 Effect Sizes for Twin Studies of Antisocial Behavior Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion Midwest twins Cates et al. (1993) Aggression (assault) Self-report Blood grouping–42.50 fm-fm 77 MZ .07 Included—averaged questionnaire 21 DZ .41 Included—averaged Aggression (verbal)77 MZ .41 Included—averaged 21 DZ .06 Included—averaged Aggression (indirect)77 MZ .40 Included—averaged 21 DZ .01 Included—averaged NAS-NRC twins Centerwall & Robinette (1989) Criminality Records Blood grouping–36.50 m-m 5,933 MZ .74 Included—largest questionnaire– fingerprinting7,554 DZ .29 Included—largest Horn et al. (1976) Psychopathy Self-report Blood grouping m-m 99 MZ .43 Excluded 99 DZ .25 Excluded Maudsley twins (psychiatric sample) Coid et al. (1993) Criminality Records Blood grouping–45.90 both-both 92 MZ .70 Included—independent questionnaire 109 DZ .80 Included—independent California twins Ghodsian-Carpey & Baker (1987) Aggression Parent report Questionnaire 5.20 both-both 21 MZ .78 Included—largest (CBCL–A) 17 DZ .31 Included—largest Aggression 17 MZ .65 Excluded (MOCL) 15 DZ .35 Excluded Danish twins Carey (1992) Criminality Records Blood grouping–Lifetime m-m 365 MZ .74 Included—largest questionnaire 700 DZ .47 Included—largest fm-fm 347 MZ .74 Included—largest 690 DZ .46 Included—largest m-fm 2,073 DZ .23 Included—largest Christiansen (1973) Criminality Records Blood grouping–Lifetime m-m 325 MZ .70 Excluded questionnaire 604 DZ .29 Excluded Christiansen (1974) Criminality Records Blood grouping–Lifetime m-m 132 MZ .02 Excluded questionnaire 191 DZ .41 Excluded 132 MZ .45 Excluded 191 DZ .03 Excluded 132 MZ .48 Excluded 191 DZ .18 Excluded Christiansen (1977a) Criminality Records Blood grouping–Lifetime m-m 325 MZ .70 Excluded questionnaire 611 DZ .28 Excluded fm-fm 328 MZ .72 Excluded 593 DZ .41 Excluded m-fm 1,547 DZ .01 Excluded Cloninger et al. (1978) Criminality Records Blood grouping–Lifetime m-m 338 MZ .69 Excluded questionnaire 637 DZ .36 Excluded fm-fm 323 MZ .72 Excluded 615 DZ .43 Excluded m-fm 2,053 DZ .24 Excluded 500 RHEE AND WALDMAN Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion London twins (children) Stevenson & Graham (1988) ASB Parent report Blood grouping–13.00 m-m 46 MZ .61 Included—independent questionnaire–48 DZ .40 Included—independent fingerprinting fm-fm 53 MZ .29 Included—independent 58 DZ .49 Included—independent London twins (adults—1970s) G. D. Wilson et al. (1977) Aggression Reaction to 30.50 both-both 49 MZ .59 Included—independent stimuli 52 DZ .34 Included—independent London twins (adults—1980s) Rushton et al. (1986) Aggression Self-report Blood grouping–30.00 m-m 90 MZ .33 Included—independent questionnaire 46 DZ .16 Included—independent fm-fm 206 MZ .43 Included—independent 133 DZ .00 Included—independent m-fm 98 DZ .12 Included—independent Rushton (1996) Delinquency Self-report 50.00 m-m 42 MZ .77 Excluded 27 DZ .44 Excluded fm-fm 126 MZ .73 Excluded 79 DZ .47 Excluded Violence m-m 42 MZ .53 Excluded 27 DZ .06 Excluded fm-fm 126 MZ .27 Excluded 79 DZ .05 Excluded Minnesota twins (reared apart) D. L. DiLalla et al. (1996) Psychopathy Self-report Blood grouping 40.40 both-both 66 MZ ra .62 Included—largest 45.10 54 DZ ra .14 Included—largest Grove et al. (1990) Adult ASP Self-report Blood grouping 43.00 both-both 32 MZ ra .41 Excluded Child ASP.28 Excluded Bouchard & McGue (1990) Psychopathy Self-report Blood grouping 41.50 both-both 45 MZ ra .53 Excluded 26 DZ ra .39 Excluded Gottesman et al. (1984, as cited Psychopathy Self-report both-both 51 MZ ra .64 Excluded in Carey, 1994)25 DZ ra .34 Excluded Tellegen et al. (1988) Aggression Self-report Blood grouping 40.90 both-both 44 MZ ra .46 Excluded 27 DZ ra .06 Excluded Minnesota twins (reared together— 1970s) Tellegen et al. (1988) Aggression Self-report Blood grouping 21.65 both-both 217 MZ .43 Included—exception 114 DZ .14 Included—exception Lykken et al. (1978) Aggression Self-report m-m 88 MZ .66 Excluded 46 DZ .06 Excluded fm-fm 174 MZ .43 Excluded 92 DZ .24 Excluded McGue et al. (1993) Aggression Self-report Blood grouping 19.80 both-both 79 MZ .61 Excluded 48 DZ .09 Excluded Aggression—follow-up 29.60 79 MZ .58 Excluded 48 DZ .14 Excluded (table continues) 501 ANTISOCIAL BEHAVIOR Table 4 (continued) Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion Minnesota twins (1960s—high school sample) Gottesman (1963, as cited in Psychopathy Self-report Blood grouping 16.00 both-both 34 MZ .57 Included—largest Gottesman & Goldsmith, 1994)34 DZ .18 Included—largest Dworkin et al. (1976) Psychopathy (MMPI) Self-report Blood grouping 15.90 both-both 25 MZ .49 Excluded 17 DZ .30 Excluded Psychopathy (CPI)25 MZ .51 Excluded 17 DZ .31 Excluded Psychopathy—follow-up 27.90 25 MZ .44 Excluded (MMPI)17 DZ .34 Excluded Psychopathy—follow-up 25 MZ .69 Excluded (CPI)17 DZ .46 Excluded Minnesota twins (1990s—adolescents) Taylor et al. (2000, personal Psychopathy Self-report Blood grouping–17.00 m-m 145 MZ .52 Included—averaged communication) questionnaire 77 DZ .15 Included—averaged fm-fm 107 MZ .48 Included—averaged 52 DZ .37 Included—averaged Delinquency m-m 145 MZ .51 Included—averaged 77 DZ .29 Included—averaged fm-fm 107 MZ .60 Included—averaged 52 DZ .38 Included—averaged Hershberger et al. (1995, personal CD Self-report Blood grouping–17.00 m-m 138 MZ .72 Excluded communication) questionnaire 67 DZ .38 Excluded Minnesota twins (1990s—adults) Finkel & McGue (1997) Aggression Self-report Blood grouping–37.76 m-m 220 MZ .37 Included—independent questionnaire 165 DZ .12 Included—independent fm-fm 406 MZ .39 Included—independent 352 DZ .14 Included—independent m-fm 114 DZ .12 Included—independent Minnesota twins (sample with alcoholism) Waldman et al. (in press, personal ASB Self-report Blood grouping–34.50 m-m 92 MZ .37 Included—independent communication) questionnaire 104 DZ .31 Included—independent fm-fm 50 MZ .50 Included—independent 46 DZ .34 Included—independent m-fm 74 DZ .30 Included—independent fm-m 26 DZ .43 Included—independent Boston twins (adolescents) Gottesman (1965, as cited in Psychopathy Self-report Blood grouping 16.00 both-both 80 MZ .46 Included—largest Gottesman & Goldsmith, 1994)68 DZ .25 Included—largest Gottesman (1966, as cited in Psychopathy Self-report Blood grouping 16.00 m-m 34 MZ .32 Excluded Carey, 1994)32 DZ .06 Excluded fm-fm 45 MZ .52 Excluded 36 DZ .26 Excluded Boston twins (children) Scarr (1966) Aggression Parent report Blood grouping 8.08 fm-fm 24 MZ .35 Included—independent 28 DZ .08 Included—independent 502 RHEE AND WALDMAN Table 4 (continued) Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion Vancouver twins Livesley et al. (1993) ASB Self-report Questionnaire 28.68 both-both 90 MZ .52 Included—independent 85 DZ .52 Included—independent National Merit Scholarship twins Loehlin & Nichols (1976) Psychopathy Self-report Questionnaire 18.00 m-m 202 MZ .52 Included—independent 124 DZ .15 Included—independent fm-fm 288 MZ .55 Included—independent 193 DZ .48 Included—independent Calgary twins Lytton et al. (1988) ASB Teacher report 9.50 m-m 15 MZ .85 Included—largest 22 DZ .47 Included—largest Mother report 13 MZ .87 Excluded 22 DZ .67 Excluded Father report 12 MZ .96 Excluded 16 DZ .96 Excluded Philadelphia twins Meininger et al. (1988) Aggression Teacher report Blood grouping 8.50 both-both 61 MZ .67 Included—independent 34 DZ .11 Included—independent Missouri twins Owen & Sines (1970) Aggression Reaction to Blood grouping 10.00 m-m 10 MZ .09 Included—independent stimuli 11 DZ .24 Included—independent fm-fm 11 MZ .58 Included—independent 13 DZ .22 Included—independent Colorado twins (1980s) M. O’Connor et al. (1980) Aggression Parent report Questionnaire 7.60 both-both 54 MZ .72 Excluded 33 DZ .42 Excluded Plomin & Foch (1980) Aggression (no. of hits) Objective test Questionnaire 7.60 both-both 42 MZ .44 Excluded 29 DZ .42 Excluded Aggression (intensity43 MZ .38 Excluded of hits)28 DZ .48 Excluded Aggression (no. of40 MZ .22 Excluded quadrants)28 DZ .44 Excluded Plomin (1981) Aggression (no. of hits) Objective test Questionnaire 7.60 both-both 53 MZ .42 Included—exception 32 DZ .42 Included—exception Aggression (intensity53 MZ .39 Excluded of hits)31 DZ .47 Excluded Aggression (no. of53 MZ .23 Excluded quadrants)31 DZ .41 Excluded Colorado twins (1990s) Zahn-Waxler et al. (1996) ASB Mother report Blood typing–5.00 both-both 200 MZ .87 Excluded questionnaire 200 DZ .59 Excluded Father report both-both 200 MZ .85 Excluded 200 DZ .60 Excluded Schmitz et al. (1994) ASB Parent report Questionnaire 2.75 m-m 32 MZ .55 Excluded 26 DZ .46 Excluded fm-fm 37 MZ .71 Excluded 45 DZ .39 Excluded m-fm 89 DZ .45 Excluded (table continues) 503 ANTISOCIAL BEHAVIOR Table 4 (continued) Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion Colorado twins (1990s) (continued) Schmitz et al. (1995) ASB Parent report Questionnaire 2.83 both-both 154 MZ .70 Included—largest 366 DZ .44 Included—largest Colorado twins (sample with LD and control sample) Willcutt et al. (1995, personal CD–sample with LD Parent report 13.00 both-both 80 MZ .76 Included—independent communication)118 DZ .57 Included—independent CD–control sample40 MZ .63 Included—independent 45 DZ .46 Included—independent Ohio twins Rowe (1983) ASB Self-report Blood grouping–17.50 m-m 61 MZ .66 Included—independent questionnaire 38 DZ .48 Included—independent fm-fm 107 MZ .74 Included—independent 59 DZ .47 Included—independent Norwegian twins (psychiatric sample) Torgersen et al. (1993) Psychopathy Self-report both-both 24 MZ .22 Included—independent 28 DZ .20 Included—independent California twins Rahe et al. (1978) Aggression Self-report Blood grouping 48.00 m-m 82 MZ .31 Included—independent 79 DZ .21 Included—independent Virginia twins Eaves et al. (1997, only largest CD Self-report Blood grouping–12.00 m-m 289 MZ .36 Included—largest sample reported) questionnaire 177 DZ .13 Included—largest fm-fm 380 MZ .24 Included—largest 185 DZ .19 Included—largest m-fm 283 DZ .10 Included—largest Silberg et al. (1994) ASB Parent report Questionnaire 12.00 m-m 242 MZ .85 Excluded 253 DZ .65 Excluded fm-fm 272 MZ .78 Excluded 233 DZ .64 Excluded m-fm 271 DZ .70 Excluded Silberg et al. (1996) ASB Parent report Blood grouping–14.00 m-m 106 MZ .68 Excluded questionnaire 82 DZ .36 Excluded fm-fm 162 MZ .68 Excluded 77 DZ .53 Excluded m-fm 130 DZ .41 Excluded 9.5 m-m 159 MZ .70 Excluded 81 DZ .33 Excluded fm-fm 185 MZ .69 Excluded 83 DZ .49 Excluded m-fm 132 DZ .40 Excluded Simonoff et al. (1995) Aggression–ASB Mother report Blood grouping–m-m 169 MZ .81 Excluded questionnaire 113 DZ .68 Excluded Father report 169 MZ .77 Excluded 113 DZ .58 Excluded Self-report 169 MZ .60 Excluded 113 DZ .41 Excluded 504 RHEE AND WALDMAN Table 4 (continued) Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion Virginia twins (continued) Simonoff et al. (1995) (continued) ODD and CD Mother report 169 MZ .65 Excluded 113 DZ .37 Excluded Father report 169 MZ .75 Excluded 113 DZ .28 Excluded Self-report 169 MZ .42 Excluded 113 DZ .11 Excluded Indiana twins Pogue-Geile & Rose (1985) Psychopathy Self-report Blood grouping 20.20 both-both 101 MZ .47 Excluded 102 DZ .15 Excluded 24.55 both-both 71 MZ .23 Excluded 62 DZ .20 Excluded Rose (1988, as cited in Psychopathy Self-report Blood grouping 24.00 both-both 228 MZ .47 Excluded Gottesman & Goldsmith, 1994)182 DZ .23 Excluded Brandon & Rose (1995, personal Psychopathy Self-report Blood grouping–20.35 both-both 289 MZ .48 Included—largest communication) questionnaire 228 DZ .27 Included—largest Western Reserve twins Edelbrock et al. (1995) ASB Parent report Blood grouping–11.00 both-both 99 MZ .79 Included—independent questionnaire 82 DZ .53 Included—independent British Columbia twins Blanchard et al. (1995, personal Aggression Self-report 36.18 both-both 96 MZ .59 Included—independent communication)48 DZ .34 Included—independent Australian twins (children) Waldman et al. (1995, personal CD Parent report Questionnaire 8.66 m-m 437 MZ .84 Included—independent communication)278 DZ .66 Included—independent fm-fm 495 MZ .89 Included—independent 225 DZ .61 Included—independent m-fm 437 DZ .56 Included—independent Australian twins (adults) Slutske, Heath, et al. (1997) CD Self-report Questionnaire 43.70 m-m 401 MZ .71 Included—independent 236 DZ .35 Included—independent fm-fm 940 MZ .68 Included—independent 540 DZ .47 Included—independent m-fm 604 DZ .32 Included—independent Dutch twins van den Oord et al. (1996) Aggression Parent report Blood grouping–3.00 m-m 210 MZ .81 Included—independent questionnaire 265 DZ .49 Included—independent fm-fm 236 MZ .83 Included—independent 238 DZ .49 Included—independent m-fm 409 DZ .45 Included—independent VET twins Lyons et al. (1995) ASPD (juvenile traits) Self-report Blood grouping–44.60 m-m 1,788 MZ .39 Included—averaged questionnaire 1,438 DZ .33 Included—averaged ASPD (adult traits) 1,788 MZ .47 Included—averaged 1,438 DZ .27 Included—averaged (table continues) 505 ANTISOCIAL BEHAVIOR Table 4 (continued) Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion VET twins (continued) Coccaro et al. (1997) Aggression (direct) Self-report Blood grouping–44.13 m-m 182 MZ .50 Excluded questionnaire 118 DZ .19 Excluded Aggression (indirect) 182 MZ .42 Excluded 118 DZ .02 Excluded Aggression (verbal)182 MZ .28 Excluded 118 DZ .07 Excluded NEAD twins Neiderhiser et al. (1998; results—ASB Multiple methods Questionnaire 13.50 m-m 43 MZ .77 Excluded personal communication)46 DZ .71 Excluded fm-fm 46 MZ .87 Excluded 46 DZ .57 Excluded T. G. O’Connor, McGuire, et al. ASB Multiple methods Questionnaire 13.70 both-both 93 MZ .81 Included—averaged (1998)99 DZ .68 Included—averaged T. G. O’Connor, Neiderhiser, ASB Multiple methods Questionnaire 16.70 both-both 63 MZ .75 Excluded et al. (1998)75 DZ .36 Excluded Deater-Deckard & Dodge (1997) ASB Mother report Questionnaire 13.70 both-both 93 MZ .84 Included—averaged (Wave 1) 99 DZ .59 Included—averaged Father report 93 MZ .77 Included—averaged (Wave 1) 99 DZ .59 Included—averaged Mother report 16.70 63 MZ .66 Excluded (Wave 2) 73 DZ .43 Excluded Father report 63 MZ .80 Excluded (Wave 2) 73 DZ .47 Excluded Pike et al. (1996) ASB Multiple methods Questionnaire 14.50 both-both 93 MZ .80 Excluded 98 DZ .65 Excluded Indian twins Nathawat & Puri (1995) Socialization Self-report 24.00 both-both 15 MZ .67 Included—averaged 15 DZ .47 Included—averaged Indirect aggression15 MZ .56 Included—averaged 15 DZ .01 Included—averaged Verbal aggression15 MZ .10 Included—averaged 15 DZ .03 Included—averaged Cardiff twins Thapar & McGuffin (1996) ASB Parent report Questionnaire 12.00 m-m 44 MZ .62 Included—independent 45 DZ .60 Included—independent fm-fm 69 MZ .62 Included—independent 40 DZ .60 Included—independent Swedish twins (adults) Gustavsson et al. (1996) Aggression (indirect) Self-report both-both 15 MZ ra .22 Included—averaged 26 MZ .41 Included—averaged 29 DZ .27 Included—averaged Aggression (verbal)15 MZ ra .03 Included—averaged 26 MZ .22 Included—averaged 29 DZ .23 Included—averaged 506 RHEE AND WALDMAN Table 4 (continued) Study Operationalization Assessment Zygosity Age SexNRelationshipEffect size Inclusion–exclusion Swedish twins (children) Eley et al. (1999) Aggression Parent report Questionnaire 8.50 m-m 176 MZ .72 Included—averaged 182 DZ .41 Included—averaged fm-fm 160 MZ .82 Included—averaged 194 DZ .45 Included—averaged m-fm 310 DZ .41 Included—averaged Delinquency m-m 176 MZ .71 Included—averaged 182 DZ .59 Included—averaged fm-fm 160 MZ .78 Included—averaged 194 DZ .60 Included—averaged m-fm 310 DZ .46 Included—averaged British twins Eley et al. (1999) Aggression Parent report Questionnaire 12.00 m-m 99 MZ .68 Included—averaged 93 DZ .45 Included—averaged fm-fm 124 MZ .77 Included—averaged 80 DZ .44 Included—averaged m-fm 95 DZ .27 Included—averaged Delinquency m-m 99 MZ .65 Included—averaged 93 DZ .66 Included—averaged fm-fm 124 MZ .75 Included—averaged 80 DZ .57 Included—averaged m-fm 95 DZ .32 Included—averaged New York twins Seelig & Brandon (1997, personal ASB Mother report 15.00 both-both 45 MZ .75 Included—averaged communication)38 DZ/b sibs.31 Included—averaged Father report 45 MZ .79 Included—averaged 38 DZ/b sibs.75 Included—averaged Self-report 45 MZ .44 Included—averaged 38 DZ/b sibs.62 Included—averaged Note.Information within parentheses indicates whether data were obtained from personal communication or another publication. fm female; MZ monozygotic twin pairs; Included—averaged included after using the averaging method of dealing with nonindependence; DZ dizygotic twin pairs; m male; NAS-NRC National Academy of Sciences–National Research Council; Included—largest included after using the largest sample method of dealing with nonindependence; Excluded excluded from the meta-analysis because another study examined the same sample that was larger, more unique in assessment method, or better described; both both male and female; Included—independent included because the study does not have a nonindependence problem; CBCL–A Child Behavior Check List—Aggression; MOCL Mothers’Observations Check List; ASB antisocial behavior; ra reared apart; ASP antisocial personality; Included—exception one of two exceptions in which the largest sample was not included; MMPI Minnesota Multiphasic Personality Inventory; CPI California Psychological Inventory; CD conduct disorder; LD learning disability; ODD oppositional defiant disorder; VET Vietnam era twins; ASPD antisocial personality disorder; NEAD nonshared environment in adolescent development; DZ/b dizygotic twin pairs/biological sibling pairs; sibs siblings. 507 ANTISOCIAL BEHAVIOR Table 4 (continued) which involves delinquency and aggression, but the operationalization of CD was reserved for studies that assessed the actualDSMcriteria. Inability to Calculate Tetrachoric or Intraclass Correlations The effect sizes used in this meta-analysis were the Pearson product– moment or intraclass correlations that were reported in the studies or the tetrachoric correlations that were estimated from the concordances or percentages reported in the studies. These effect sizes were analyzed in model-fitting programs that estimate the relative contribution of genetic and environmental influences and test alternative etiologic models. Twenty-six studies were excluded from the meta-analysis because effect sizes were not reported or because there was not enough information reported to calculate the effect sizes. Four frequently cited adoption studies were excluded for this reason. First, Crowe (1972, 1974, 1975) found that adopted away offspring of female criminal offenders were more likely to be criminal and have antisocial personality than adopted away offspring of controls, thus yield- ing evidence for genetic influences on antisocial behavior. Crowe counted the 52 adoptees of 41 biological mothers as individual cases rather than counting the 41 biological mothers as individual cases, creating a problem of nonindependence in mother–adopted away offspring pairs. Second, Jary and Stewart (1985) found that biological fathers of children with aggres- sive CD were more likely to have antisocial personality than adoptive fathers of children with aggressive CDs. They did not report similar information regarding the parents of a control group without aggressive CD, however. A comparison between the aggressive group and the control group is necessary for the estimation of a tetrachoric correlation, as the control group would provide the base rate of antisocial behavior in this sample. Three early twin studies were excluded because they did not permit adequate effect size estimation (Dalgard & Kringlen, 1976; Hayashi, 1967; Rosanoff, Handy, & Rosanoff, 1934). These studies located twin pairs with at least one affected member (i.e., the proband), then compared the risk to the cotwin in MZ twin pairs and DZ twin pairs. The risk was estimated using either the pairwise concordance or the probandwise concordance. Studies using this method do not include the base rate of the variable of interest (i.e., how prevalent the condition is in the sample being studied), which is necessary for the estimation of tetrachoric correlations. In other words, these studies reported the number of twin pairs that are concordant for being affected and the number of twin pairs that are discordant but did not report the number of twin pairs that are concordant for being unaf- fected. There are two related problems that this entails. First, the estimation of the effect sizes used in this meta-analysis, tetrachoric or intraclass correlations, is impossible without the base rate. Second, concordances themselves may be misleading because their interpretation varies according to the base rate (A. Heath, personal communication, March 1991). One possible way to include these early studies in the meta-analysis would be to use the base rate for antisocial behavior in the country from which the sample was drawn and include this information to estimate tetrachoric correlations. One unpublished meta-analysis of antisocial be- havior (Ridenour & Heath’s, 1997, meta-analysis of categorically defined antisocial behavior) took such an approach. We decided against this ap- proach for the following reason. Even if the base rate for antisocial behavior were found for the specific countries of interest, it may not be appropriate for the specific operationalization used by the studies, the year the studies were published, and many other specific factors that can make the sample examined by the study quite different from a random sample from the population for which the base rate was derived. The early twin studies excluded from this meta-analysis because of failure to provide the appropriate base rate have been discussed in many traditional literature reviews (e.g., Christiansen, 1977b; Cloninger & Gottesman, 1987). With the exception of Dalgard and Kringlen (1976),who found only slightly higher concordances for criminality in MZ twins than in DZ twins, the early twin studies found genetic influences to be of substantial magnitude for criminality, but not for juvenile delinquency. Nineteen recent twin and adoption studies examining antisocial behavior also were excluded because they did not provide enough information for the calculation of effect sizes. For all of these studies, the information needed for the meta-analysis was found in other publications that analyzed data from the same sample. These excluded studies usually examined more complex issues (e.g., Cadoret, Cain, & Crowe, 1983, genotype–en- vironment interaction; Langbehn, Cadoret, Yates, Troughton, & Stewart, 1998, relationship between CD and oppositional defiant disorder symp- toms and adult antisocial behavior; Reiss et al., 1995, parenting style). Assessment of Related Disorders In several studies, another variable related to antisocial behavior (e.g., alcoholism, somatization disorder, or other personality disorders) was studied in addition to antisocial behavior. For example, one adoption study (Schulsinger, 1972) examined the aggregate risk for psychopathy, crimi- nality, alcoholism, drug abuse, or mental illness in adoptees of biological parents with psychopathy and biological parents who did not have psy- chopathy. This means that some adoptees who do not engage in antisocial behavior could have been counted as“affected”because of their problems with alcohol or drug abuse (i.e., variables outside the scope of this meta- analysis). Such studies were not included because the assessment of other disorders interfered with the assessment of antisocial behavior (e.g., alco- holism or drug abuse being counted as antisocial behavior). Six studies were excluded from the meta-analysis because of assessment of related disorders. Nonindependent Samples Another justification for exclusion from the meta-analysis was noninde- pendent sampling. Several effect sizes from studies in the original refer- ence list were from nonindependent samples as a result of several factors. Some authors published the same data in two different sources (e.g., Mednick et al., 1983; Mednick, Gabrielli, & Hutchings, 1984). In such cases, we only considered one of the studies for the meta-analysis. The other three factors leading to nonindependent samples are more compli- cated. First, some authors of a single publication examined more than one dependent measure of antisocial behavior in their sample (e.g., Ghodsian- Carpey & Baker, 1987). Second, several publications were a collection of follow-up data of the same sample (e.g., Cadoret, 1978; Cadoret, Trough- ton, O’Gorman, & Heywood, 1986). Third, several authors (in different publications) examined different dependent measures in the same sample (e.g., Grove et al., 1990; Tellegen et al., 1988). Experts on meta-analysis have several suggestions for dealing with nonindependent samples (Mullen, 1989; Rosenthal, 1991). For example, Mullen gave four options for dealing with this problem: choosing the best dependent measure, averaging the effect sizes of the different dependent measures, conducting separate meta-analyses for each of the dependent measures, or using nonindependent samples as if they were independent samples (the least recommended approach). We did not follow the option of choosing the best dependent measure, unless one of the dependent measures did not fulfill the inclusion criteria described above, making the decision easy. This option was not taken in order to avoid making subjec- tive choices, because we were aware of the effect sizes associated with each of the dependent measures. The option of conducting separate meta- analyses for each of the dependent measures was not chosen simply as a practical matter, because there were a large number of effect sizes from nonindependent samples. Therefore, the most viable option was to average the effect sizes from nonindependent samples. In model-fitting analyses, the sample size must be indicated. Therefore, the option of averaging multiple effect sizes was used in cases in which the 508 RHEE AND WALDMAN sample size was identical across the nonindependent samples. If the sample size was not identical across the nonindependent samples, the effect size from the largest sample was used. More specifically, in cases of noninde- pendence in which the same dependent measure was used in the same sample multiple times (e.g., in follow-up analyses), the effect size esti- mated from the largest sample was chosen. In cases of nonindependence in which different dependent measures were used in the same sample (e.g., the author of one publication examining more than one dependent measure or authors of different publications examining different dependent mea- sures in one sample), the effect sizes were averaged if the sample size was the same across the nonindependent samples, and the effect size from the largest sample was used if the sample size differed across the nonindepen- dent samples. When choosing the effect size from the largest sample, we made this decision without regard to other factors with two exceptions. M. O’Connor, Foch, Sherry, and Plomin (1980) and Plomin et al. (1981) studied the same Colorado twin sample. Although O’Connor et al.’s (1980) sample was larger by 2 more twin pairs, Plomin et al. (1981) was included in the meta-analysis instead. Plomin et al.’s (1981) study is the only study to examine an objective measure of aggression (except for Plomin and Foch, 1980, which also used the same sample), so it was important to include the study in the examination of the potential moderating effect of the assess- ment method. Tellegen et al. (1988) and Lykken, Tellegen, and DeRubeis (1978) reported results for the same aggression measure on the same sample. Although Lykken et al.’s sample was larger, Tellegen et al. was included instead, as Lykken et al. focused on the methodological issue of volunteer sampling and did not report information regarding two potential moderating variables (i.e., zygosity determination method and age), whereas this information was included in Tellegen et al. In several cases, it was unclear whether several studies reported results from the same sample (e.g., the Minnesota Twin Family Study). Several pieces of information, including the year of the publication, the age of the sample, and the description of the sample were used to decide whether two studies actually used the same sample. In some studies (e.g., Parker, 1989, as cited in Carey, 1994), this decision was impossible to make because a description of the sample was not reported. The assumption of noninde- pendent sampling was made for these studies. In Tables 3 and 4, the studies using the same samples are grouped together. The Inclusion–exclusion column indicates whether the study’s effect sizes were included or excluded.“Included—averaged”indicates an effect size that was included in the meta-analysis after using the averaging method (i.e., averaging effect sizes with the same associated sample size) of dealing with nonindependence.“Included—largest”indicates an effect size that was included in the meta-analysis after using the largest sample method (i.e., simply choosing the effect size associated with the largest sample size) of dealing with nonindependence.“Included—independent” indicates an effect size that was included in the meta-analysis because the study does not have a nonindependence problem.“Excluded”indicates an effect size that was excluded from the meta-analysis because the same sample was examined in another study that was larger, more unique in assessment method, or better described. Analyses Determination of the Effect Size Some adoption and twin studies used a continuous variable to measure antisocial behavior and reported either Pearson product–moment or intra- class correlations, which were the effect sizes used from these studies in the meta-analysis. In other studies, a dichotomous variable was used, and concordances, percentages, or a contingency table (including the number of twin pairs with both members affected, one member affected, and neither member affected) was reported. The information from the concordances or percentages was transformed into a contingency table, which was then used to estimate the tetrachoric correlation (i.e., the correlation between thelatent continuous variables that are assumed to underlie the observed dichotomous variables). For these studies, the tetrachoric correlation was the effect size used in the meta-analysis. For some studies, we directly estimated the tetrachoric correlation from the raw data either because we had access to the data (Slutske, Heath, et al., 1997; Waldman, Levy, & Hay, 1995; Waldman et al., in press) or because the tetrachoric correlation had to be estimated from contingency tables. For these studies, we were also able to estimate the weight matrix (i.e., the asymptotic covariance matrix of the correlation matrix). If the weight matrix can be estimated, it is possible to use weighted least squares (WLS) estimation, which is more appropriate for non-normally distributed vari- ables like diagnoses of CD or ASPD, rather than maximum-likelihood (ML) estimation, in the model-fitting analyses. One assumption of model-fitting analyses is that the variable being analyzed is normally distributed. Although we do not have access to the distributions of the variables being examined in the studies included in the meta-analysis, violation of the normal distribution assumption in studies examining antisocial behavior is often a problem. Typically, the distribu- tion is positively skewed (i.e., inverseJ-shaped) because the majority of the population exhibits little or no antisocial behavior. WLS estimation is preferable to ML estimation for obtaining asymptotically correct standard errors of parameter estimates and chi-square goodness-of-fit tests when the normal distribution assumption cannot be met or when correlations (rather than covariances) are analyzed (Neale & Cardon, 1992). For most of the studies included in the meta-analysis, however, we did not have access to the raw data and were limited to the published information. This meant that for these studies, we were limited to analyzing Pearson product–moment or intraclass correlations, and ML rather than WLS estimation had to be used. Model-Fitting Analyses The magnitude of additive genetic influences (a 2) and that of nonaddi- tive genetic influences (d 2) constitute the amount of variance in the liability for antisocial behavior that is due to genetic differences among individuals. If genetic influences are additive, this means that the effects of alleles from different loci are independent and“add up”to influence the liability for a trait. If genetic influences are nonadditive, this means that alleles interact with each other to influence the liability for a trait, either at a single genetic locus (i.e., dominance) or at different loci (i.e., epistasis). Shared environ- mental influences (c 2) represent the amount of liability variance that is due to environmental influences that are experienced in common and make family members similar to one another, whereas nonshared environmental influences (e 2) represent the amount of liability variance that is due to environmental influences that are experienced uniquely and make family members different from one another. It is customary in contemporary behavior genetic analyses to compare alternative models, containing different sets of causal influences, for their fit to the observed data (i.e., twin or familial correlations or covariances). These models posit that antisocial behavior is affected by the types of influences described above: additive genetic influences (A), shared envi- ronmental influences (C), nonadditive genetic influences (D), and non- shared environmental influences (E). In the present meta-analysis, the ACE model, the AE model, the CE model, and the ADE model were compared. It is not possible to estimatec 2andd 2simultaneously or test an ACDE model with data only from twin pairs reared together because the estima- tion ofc 2andd 2both rely on the same information (i.e., the difference between the MZ and DZ twin correlations). If the DZ correlation is greater than half of the MZ correlation, the ACE model is the correct model and the estimate ofd 2in the ADE model is always zero. If the DZ correlation is less than half of the MZ correlation, the ADE model is the correct model and the estimate ofc 2in the ACE model is always zero. If another type of data, such as the correlations between adoptees and their adoptive and biological parents, also is included in the analyses, this provides another source of information for the estimation ofc 2and the ACDE model can be 509 ANTISOCIAL BEHAVIOR tested. Given that the ACDE model can be tested only when both twin and adoption studies are included in the analysis, it was only possible to estimatec 2andd 2simultaneously when analyzing all of the data included in the meta-analysis. For other analyses (i.e., the comparison of including and excluding weight matrices, the comparison between twin and adoption studies, and the tests of moderators), both twin and adoption studies were not always available across different types of studies. Therefore, we were limited to comparing the ACE, AE, CE, and ADE models for analyses other than those that included all data included in the meta-analysis. Two types of adoption studies and two types of twin studies were included in the meta-analysis. The adoption studies provided data on the comparison of the correlation between adoptees and their adoptive parents versus the correlation between adoptees and their biological parents (i.e., parent–offspring adoption studies) and the comparison of the correlation between adoptive siblings and the correlation between biological siblings (i.e., sibling adoption studies). Data from both studies of twin pairs reared together and twin pairs reared apart were included. The effect sizes (i.e., Pearson or intraclass correlations or the tetrachoric correlations plus the weight matrices) from each study were entered in separate groups in the model-fitting program Mx (Neale, 1995). Stem and leaf plots of the effect sizes from the adoption studies and the twin studies are shown in Tables 5 and 6, respectively. In the model-fitting program, the correlations between pairs of relatives are explained in terms of the components of variance that are shared between the relatives. These can include A, or additive genetic influences; C, or shared environmental influences; and D, or nonadditive genetic influences. Nonshared environmental influences, or E, do not explain any part of the correlation between the pairs of relatives because, by definition, nonshared environmental influences are not shared between relatives. The correlation between different types of relatives is explained by different sets of influences and their appropriate weights as shown in Appendix B. These weights reflect the genetic or environmental similarity between pairs of relatives. For example, the correlation between an adoptee and his or her adoptive parent is explained only by shared environmental influences (1*C), whereas the correlation between an adoptee and his or her biological parent is explained only by additive genetic influences (.5*A). The example Mx script in Appendix C shows how an analysis was set up to test an ACDE model, and Figure 1 shows the path diagram for the ACDE model. In Appendix C, Group 1 defines the parameters of the model:a 2(additive genetic influences),c 2(shared environmental influenc- es),d 2(nonadditive genetic influences), ande 2(nonshared environmental influences). Groups 2 to 9 show how the correlation matrix for each type of relative pair (adoptee and biological parent, adoptee and adoptive parent, biological siblings, adoptive siblings, MZ twins reared together, DZ twins reared together, MZ twins reared apart, and DZ twins reared apart) is defined in the Mx script according to the information shown in Appendix B. For each study, the effect size, or the correlation matrix for each type of relative pair (e.g., MZ twin pairs and DZ twin pairs), is listed in a separategroup. If a study listed separate correlation matrices for independent groups (e.g., males and females, younger children and older children), these correlation matrices were listed in separate groups. In analyzing behavior genetic data for two generations, as in the parent– offspring adoption studies, it is important to consider the possibility of estimating separatea 2andc 2values for children and parents becausea 2 andc 2estimates may differ across the generations. Unfortunately, the adoptee–adoptive parent and adoptee–biological parent correlations do not provide enough information for such analyses. In the parent–offspring adoption studies, a problem of nonindependence exists because the same adoptees are in the adoptee–adoptive parent groups and the adoptee–biological parent groups. Therefore, the adoptee–adoptive parent data were included only in comparisons between the twin studies and the two types of adoption studies. Analyses of All Data The analyses were first conducted for all data, including the two types of twin studies and the two types of adoption studies. The ACDE model, the ACE model, the AE model, the CE model, and the ADE model were compared. The fit of each model, as well as of competing models, was assessed using both the chi-square statistic and the Akaike information criterion (AIC), a fit index that reflects both the fit of the model and its parsimony (Loehlin, 1992b). The AIC has been used extensively in both the structural equation modeling and behavior genetics literatures. Among competing models, that with the lowest AIC and the lowest chi-square value relative to its degrees of freedom is considered to be the best fitting model. Assessment of Possible Outliers and High-Influence Studies We examined the possibility that certain studies may be outliers or exert undue influence on the results by analyzing the data both including and excluding these studies. Specifically, we reanalyzed the data both including and excluding studies with construct validity concerns—that is, studies examining psychopathy (eight samples) or using measures with question- able validity (two samples)—to examine the sensitivity of the results to the effects from these studies. The data also were analyzed both including and excluding the Centerwall and Robinette (1989) study for three reasons. First, there was a much larger difference between the MZ and DZ corre- lations as compared with other studies included in the meta-analysis, thus raising the possibility that the study represented an outlier. Second, with almost 10,000 participants, this study was far larger than any other study, which meant that it could exert undue influence on the results. Third, this study used an unusual operationalization for criminality (i.e., dishonorable discharge from the military). Table 5 Stem and Leaf Plot of the Effect Sizes (Correlations) in Adoption Studies Adoptee–biological parentAdoptee–adoptive parent Biological siblings Adoptive siblings Stem Leaf Stem Leaf Stem Leaf Stem Leaf .5 3 .5 .5 2 .5 .4 0 9 .4 .4 2 6 .4 .3 .3 .3 1 .3 7 .2 .2 .2 .2 .1 02467 .1 .1 .1 119 .0 0 .0 1 9 .0 .0 0 .0 .0 2 .0 .0 510 RHEE AND WALDMAN Assessment of Potential Moderators We examined whether operationalization (i.e., diagnoses, criminality, aggression, and antisocial behavior), assessment method (i.e., self-report, report by others, objective test, reaction to aggressive material, and records), zygosity determination method (i.e., blood typing, questionnaire, and a combination of the two), sex (i.e., male, female, and both or opposite sex), and age (i.e., children, adolescents, and adults) were significant moderators by contrasting the fit of a model in which the parameter estimates are constrained to be equal across levels of the relevant variables to the fit of a model in which the parameter estimates are free to vary across levels of the relevant variables on the same dataset. If the fit of the two models is significantly different, this indicates the significance of the moderator. It is possible that a nonsignificant result may be due to lack of power, especially if there is little variability in the levels of a moderator. Assessment of Confounding Among Moderators When testing a moderator for significance, one tests whether estimating separate parameter estimates (e.g.,a 2,c2, ande 2) for studies at each level of the moderator leads to a better fit than when the parameter estimates are constrained to be equal across the different levels of the moderator. When testing for one moderator’s significance after another moderator has been statistically controlled for, one tests whether estimating separate parameter estimates for studies at each level of both moderators leads to a better fit than estimating separate parameter estimates for studies at each level of only one of the moderators. For example, when examining whether assess- ment method is a significant moderator after the effects of operationaliza- tion have been statistically controlled for, one compares two models. In the first model, parameter estimates are allowed to vary only across the four operationalizations. This model is compared with a second, less restrictive model, in which parameter estimates are allowed to vary across both the four operationalizations and the five assessment methods conjointly. If the fit of the second model (i.e., with both operationalization and assessment method as moderators) is significantly better than the fit of the first model (i.e., with only operationalization as a moderator), this indicates that assessment method is a significant moderator even after the effects of operationalization as a moderator are statistically controlled. Comparisons Between Twin and Adoption Studies Three comparisons between the twin studies and the two types of adoption studies were conducted. Twin studies were not divided into thetwo types (i.e., twin pairs reared together and twin pairs reared apart), given that there were only two samples of twin pairs reared apart. First, twin studies were compared with all adoption studies. Second, twin studies were compared with adoption studies examining adoptees and their adop- tive or biological parents (i.e., parent–offspring adoption studies). Third, twin studies were compared with adoption studies examining adoptive and biological siblings (i.e., sibling adoption studies). These comparisons were made by contrasting the fit of two models: (a) a model in which the parameter estimates are constrained to be equal across the twin and adoption studies and (b) a model in which the parameter estimates are free to vary across the twin and adoption studies. If the fit of the two models is significantly different, this would indicate that the estimates of genetic and environmental influences from twin and adoption studies are significantly different. Note that this is the same procedure used for the assessment of potential moderators that was described above. Effect of Excluding Weight Matrices As discussed above, we did not have access to the raw data for most of the studies and were limited to analyzing the data published in the studies (i.e., Pearson product–moment correlations or intraclass correlations), which meant that ML estimation had to be used rather than the preferred WLS estimation. Although there was no other option, this is a limitation of the meta-analysis, given that WLS estimation is more appropriate than ML estimation when the normal distribution assumption is violated and when correlations rather than covariances are analyzed (Neale & Cardon, 1992). In order to examine the potential effects of using ML estimation rather than WLS estimation on the results, we examined the effects of excluding the weight matrices in two ways. First, in the studies with estimated weight matrices, the data were analyzed both including and excluding the weight matrices. Second, we contrasted the results from studies with the estimated weight matrices with those from studies for which the estimation of weight matrices was not possible. This contrast was tested by comparing the fit of the model in which the parameter estimates were constrained across the two types of studies with the fit of the model in which the parameter estimates were free to vary across the two types of studies. Results Analyses of All Data In this section, the number ofsamplesrefers to the number of independent studies in the analyses. The number ofgroupsrefers Table 6 Stem and Leaf Plot of the Effect Sizes (Correlations) in Twin Studies MZ twin pairs DZ twin pairsMZ twin pairs reared apartDZ twin pairs reared apart Stem Leaf Stem Leaf Stem Leaf Stem Leaf .9 0 .9 .9 .9 .8 011435 .8 0 .8 .8 .7 000244446689 .7 .7 .7 .6 1223666778 .6 002224 .6 2 .6 .5 222457899 .5 0012236679 .5 .5 .4 23334568 .4 022244566777888999 .4 .4 .3 1235679 .3 001444678 .3 .3 .2 2499 .2 012235579 .2 .2 .1 .1 01222344566789 .1 0 .1 4 .0 9 .0 0 .0 .0 .0 .0 8 .0 .0 .1 .1 .1 .1 .2 .2 4 .2 .2 Note.MZ monozygotic; DZ dizygotic. 511 ANTISOCIAL BEHAVIOR to the total number of independently analyzed units in the samples. For example, Slutske, Heath, et al. (1997) and Torgersen, Skre, Onstad, Edvardsen, and Kringlen (1993) examined two indepen- dent samples, the Australian adult twins and the Norwegian twins. There are five groups (male–male MZ twin pairs, male–male DZ twin pairs, female–female MZ twin pairs, female–female DZ twin pairs, and male–female DZ twin pairs) in Slutske, Heath, et al. and two groups (MZ twin pairs and DZ twin pairs) in Torgersen et al. Therefore, if an analysis is conducted using data from Slutske, Heath, et al. and Torgersen et al., there would be two samples and seven groups in the analysis. The results of analyses of the data from all of the samples meeting the inclusion criteria (N 52 samples, 149 groups, 55,525 pairs of participants) are presented in Table 7. The full ACDE model fit best as compared with the other, more restrictive models. Excluding possible outliers—that is, studies that examined psy- chopathy (8 samples), and studies using measures with question- able validity (2 samples), and the Centerwall and Robinette (1989) study—did not alter the results of the meta-analysis, as parameter estimates did not differ after excluding these studies. (The specific results can be obtained from the authors.) Assessment of Potential Moderators Table 8 shows the results of analyses examining operationaliza- tion, assessment method, zygosity determination method, sex, and age as moderators of the magnitude of genetic and environmental influences on antisocial behavior. The chi-square difference be- tween a model in which the parameter estimates are constrained to be equal and a model in which the parameter estimates are free to vary across the different levels of the moderator is shown for each moderator. Operationalization The chi-square difference test is significant for operationaliza- tion, indicating significant differences in the magnitude of genetic and environmental influences on diagnosis (14 samples, 40 groups, 11,681 pairs of participants), criminality (5 samples, 13 groups, 34,122 pairs of participants), aggression (14 samples, 40 groups, 4,408 pairs of participants), and antisocial behavior (15 samples, 48 groups, 4,365 pairs of participants), 2(9, N 54,576) 339.87,p .01. The ACE model was the best fitting model for diagnosis (a 2 .44,c 2 .11,e 2 .45), Figure 1.ACDE model. A additive genetic influences; C shared environmental influences; D nonadditive genetic influences; E nonshared environmental influences; a-ap adoptee–adoptive parent pairs; a-bp adoptee–biological parent pairs; b-bp biological child–biological parent pairs; a sibs adoptive sibling pairs; b sibs biological sibling pairs; MZ monozygotic twin pairs reared together; DZ dizygotic twin pairs reared together; MZ ra monozygotic twin pairs reared apart; DZ ra dizygotic twin pairs reared apart. 512 RHEE AND WALDMAN aggression (a 2 .44,c 2 .06,e 2 .50), and antisocial behavior (a 2 .47,c 2 .22,e 2 .31), whereas the ADE model was the best fitting model for criminality (a 2 .33,d 2 .42,e 2 .25). Within the operationalization of diagnosis, significant differences were found between studies examining ASPD (8 samples, 17 groups, 5,019 pairs of participants) and CD (5 samples, 22 groups, 6,560 pairs of participants). Although the magnitude of shared environmental influences was similar, thea 2estimate washigher in studies examining CD (a 2 .50,c 2 .11,e 2 .39), whereas thee 2estimate was higher in studies examining ASPD (a 2 .36,c 2 .10,e 2 .54). The possible effects of confounding between operationalization and assessment method and between operationalization and age should be considered when interpreting these results (see Tables 1 and 2). Parent report was more frequently used in studies exam- ining antisocial behavior than in studies examining diagnosis or Table 7 Standardized Parameter Estimates and Fit Statistics—Inclusion of All Data ModelParameter estimate Fit statistic a 2 c2 e2 d2 2 df pAIC ACE .38 .18 .44—1,420.38 147 .001 1,126.38 AE .55—.45—1,707.89 148 .001 1,411.89 CE—.45 .55—2,364.90 148 .001 2,068.90 ADE .41—.42 .17 1,590.58 147 .001 1,296.58 ACDE .32 .16 .43 .09 1,394.46 146 .001 1,102.46 Note.Dashes indicate that data are not applicable.a 2 the magnitude of additive genetic influences (A);c 2 the magnitude of shared environmental influences (C);e 2 the magnitude of nonshared environmental influences (E);d 2 the magnitude of nonadditive genetic influences (D); AIC Akaike information criterion. Table 8 Standardized Parameter Estimates and Fit Statistics for the Best Fitting Models— Test of Moderators ModeratorFit statistic 2 df pAIC Operationalization Parameters constrained to be equal 1,406.50 139 .001 1,128.50 Parameters free to vary 1,066.63 130 .001 806.63 Chi-square difference test 339.87 9 .001 321.87 Assessment method Parameters constrained to be equal 1,361.73 139 .001 1,083.73 Parameters free to vary 530.47 128 .001 274.47 Chi-square difference test 831.26 11 .001 809.26 Zygosity determination method Parameters constrained to be equal 1,305.79 110 .001 1,085.79 Parameters free to vary 945.65 104 .001 737.65 Chi-square difference test 360.14 6 .001 348.14 Age Parameters constrained to be equal 1,351.30 133 .001 1,085.30 Parameters free to vary 1,107.35 127 .001 853.35 Chi-square difference test 243.95 6 .001 231.95 Sex (studies examining one sex or both sexes: males, females, and both) Parameters constrained to be equal 1,420.38 147 .001 1,126.38 Parameters free to vary 1,383.43 141 .001 1,101.43 Chi-square difference test 36.95 6 .001 24.95 Sex (studies examining one sex or both sexes: males and females) Parameters constrained to be equal 1,057.03 76 .001 905.03 Parameters free to vary 1,037.67 73 .001 891.67 Chi-square difference test 19.36 3 .001 13.36 Sex (studies examining both sexes: males and females) Parameters constrained to be equal 870.61 66 .001 738.61 Parameters free to vary 869.07 63 .001 743.07 Chi-square difference test 1.53 3 .68 4.47 Note.AIC Akaike information criterion. 513 ANTISOCIAL BEHAVIOR aggression, and there were more studies examining antisocial behavior in children and adolescents than studies examining anti- social behavior in adults. Also, all of the behavior genetic studies of criminality were those examining adults using the assessment method of official records. The specific comparison between stud- ies examining the diagnoses of ASPD and CD showed that the magnitude of genetic influences was higher for CD, whereas the magnitude of nonshared environmental influences was higher for ASPD. These results may be explained by age differences (ASPD being assessed in adulthood and CD being assessed in childhood) or differences in assessment method (self-report being used more often to assess ASPD and parent report being used more often to assess CD). Assessment Method The chi-square difference test indicates that assessment method is a moderator of the magnitude of genetic and environmental influences on antisocial behavior, 2(11,N 54,533) 831.26, p .01. Self-report (23 samples, 69 groups, 13,329 pairs of participants), report by others (14 samples, 51 groups, 6,851 pairs of participants), records (5 samples, 13 groups, 34,122 pairs of participants), reaction to stimuli (2 samples, 6 groups, 146 pairs of participants), and objective assessment (1 sample, 2 groups, 85 pairs of participants) were compared. The ACE model was the best fitting model for self-report (a 2 .39,c 2 .06,e 2 .55) and report by others (a 2 .53,c 2 .22,e 2 .25), whereas the AE model was the best fitting model for reaction to aggressive stimuli (a 2 .52,e 2 .48). All of the studies using the assessment method of records were also studies examining criminality, and the ADE model was the best fitting model (a 2 .33,d 2 .42,e 2 .25). Model fitting could not be conducted for the assessment method of objective test because of lack of information (i.e., only one study used an objective test). Caution is recommended in interpreting these results, given that only one study (Plomin et al., 1981) used an objective test, and only two studies (Owen & Sines, 1970; G. D. Wilson, Rust, & Kasriel, 1977) used reaction to aggressive material. Also, all of the studies using the assessment method of records were studies ex- amining the operationalization of criminality. When the assess- ment methods of self-report and report by others were compared, the magnitude of familial influences (a 2andc 2) was higher for report by others than for self-report. These results differ slightly from the conclusions of Miles and Carey (1997), who found lower a 2and higherc 2estimates for parent reports than for self-reports of aggression. Again, the possibility of confounding between mod- erators should be considered. Studies using the assessment method of self-report were more likely to be those examining the opera- tionalization of diagnosis in adults or adolescents, whereas studies using the assessment method of parent report were more likely to be those examining the operationalization of antisocial behavior in children. Zygosity Determination Method The chi-square difference test indicates that zygosity determi- nation method is a significant moderator, as the magnitude of genetic and environmental influences differed significantly for studies using blood grouping (8 samples, 18 groups, 1,020 pairs ofparticipants), a combination of blood grouping and the question- naire method (15 samples, 55 groups, 27,631 pairs of participants), and the questionnaire method (11 samples, 39 groups, 8,249 pairs of participants), 2(6,N 36,900) 360.14,p .01. The ADE model was the best fitting model for studies using blood grouping (a 2 .14,d 2 .33,e 2 .53), whereas the ACE model was the best fitting model for studies using the questionnaire method (a 2 .43,c 2 .27,e 2 .30) and a combination of the two methods (a 2 .39,c 2 .11,e 2 .50). These parameters estimates are difficult to interpret, given that studies using the most stringent method of zygosity determination (i.e., blood grouping) and the least stringent method of zygosity determination (i.e., questionnaire) yielded higher estimates of ge- netic influences (broadh 2 .43 to .47) than studies using a combination of the two methods (broadh 2 .39). Age The chi-square difference test indicates that age is a significant moderator and that the magnitude of genetic and environmental influences on antisocial behavior in children (15 samples, 54 groups, 7,807 pairs of participants), adolescents (11 samples, 31 groups, 2,868 pairs of participants), and adults (17 samples, 50 groups, 27,671 pairs of participants) is significantly different, 2(6,N 38,346) 243.95,p .01. The ACE model was the best fitting model for children (a 2 .46,c 2 .20,e 2 .34), adolescents (a 2 .43,c 2 .16,e 2 .41), and adults (a 2 .41, c 2 .09,e 2 .50). The magnitude of familial influences (a 2and c 2) decreased with age, whereas the magnitude of nonfamilial influences (e 2) increased with age. These results should be interpreted with caution for two reasons. First, although many studies examined a wide age range, either the mean or the midpoint age had to represent this age range, given that access to the raw data for each study was not possible. Second, age was simplified into a categorical variable (i.e., children, ado- lescents, and adults) in our meta-analysis, given the limitations of including continuous moderators in model-fitting analyses. As age increased, the magnitude of familial influences (i.e., botha 2and c 2) decreased. These findings for behavior genetic studies of antisocial behavior differ somewhat from the general finding in the behavior genetics literature (Loehlin, 1992a; Plomin, 1986) thata 2 ande 2estimates increase andc 2estimates decrease with increasing age. These findings also differ from Miles and Carey’s (1997) conclusion thata 2estimates increase andc 2estimates decrease with age. The confounding among moderators should again be considered in interpreting our results. The same pattern of results found for age was found for assessment method, with studies using report by others (viz., used more with children) yielding higher estimates of familial influences than those using self-report (viz., used more with adolescents and adults), and for operationalization, with studies examining antisocial behavior (viz., assessed more in children) yielding higher estimates of familial influences than those examining diagnosis (viz., assessed more in adults and adolescents). The results were not consistent with L. F. DiLalla and Gottes- man’s (1989) hypothesis given that the magnitude of genetic influences was lower in both adolescence and adulthood than in childhood, but again, the presence of confounding among the moderators should be considered. Unfortunately, it is difficult to 514 RHEE AND WALDMAN interpret the results of analyses examining age as a moderator after statistically controlling for assessment method because only one study examining children used self-report and only two studies examining adolescents used parent report. Sex The chi-square difference test examining the differences among studies examining males (21 samples, 42 groups, 22,521 pairs of participants), females (19 samples, 38 groups, 7,375 pairs of participants), and both sexes or opposite-sex pairs (41 samples, 69 groups, 25,629 pairs of participants) was significant, 2(6, N 55,525) 36.95,p .01. The ACE model was the best fitting model for males (a 2 .38,c 2 .17,e 2 .45), females (a 2 .41,c 2 .19,e 2 .40), and both sexes/opposite-sex pairs (a 2 .35,c 2 .17,e 2 .48). The magnitude of familial influences (a 2andc 2) was higher in same-sex twin pairs (a 2 .39, c 2 .18,e 2 .43) than in data including both sexes or opposite- sex twin pairs (a 2 .35,c 2 .17,e 2 .48). These results support Cloninger, Christiansen, Reich, and Gottesman’s (1978) conclu- sion that although many of the etiologic factors that influence antisocial behavior in males and females are shared in common, they are not fully identical. The difference between males and females also was significant, 2(3,N 29,896) 19.36,p .01, indicating that thea 2and thec 2estimates are higher in females. These results are not consistent with those of Miles and Carey (1997), who found higher heritability estimates for aggres- sion in males. Given the fact that several studies examined only one sex and the fact that these studies varied a great deal in the operational- ization examined (e.g., dishonorable discharge for males and ag- gression for females) and the assessment method used (e.g., offi- cial records for males and parent report for females), the comparison of results for males and females was repeated after excluding these studies (see Tables 3 and 4 for studies including only one sex). When the analyses were limited to studies that examined antisocial behavior in both males (17 samples, 34 groups, 5,610 pairs of participants) and females (17 samples, 34 groups, 7,225 pairs of participants)—that is, when studies exam- ining antisocial behavior in only one sex were excluded—the difference between males (a 2 .43,c 2 .19,e 2 .38) and females (a 2 .41,c 2 .20,e 2 .39) was no longer significant, 2(3,N 12,835) 1.53,p .68. This result is consistent with those of traditional literature reviews (e.g., Widom & Ames, 1988) in which the authors have concluded that the magnitude of genetic and environmental influences on antisocial behavior in males and females is similar. Assessment of Confounding Among Moderators The possibility of confounding was assessed between the fol- lowing pairs of moderators: operationalization and assessment method, age and operationalization, and age and assessment method. All analyses showed that each moderator is significant even after the effects of the possible confounding moderator are controlled for statistically. For example, the model estimating separate parameter estimates for each level of operationalization and each level of assessment method fit significantly better than the model estimating separate estimates for each level of opera-tionalization only, 2(13,N 54,122) 633.67,p .001, and the model estimating separate estimates for each level of assess- ment method only, 2(12,N 54,122) 112.56,p .01. This result indicates that assessment method is a significant moderator after controlling statistically for the effects of operationalization as a moderator, and that operationalization is a significant moderator after controlling statistically for the effects of assessment method as a moderator. Similarly, assessment method was a significant moderator after controlling for age, 2(7,N 38,071) 676.28, p .01; operationalization was a significant moderator after controlling for age, 2(18,N 37,935) 410.52,p .01; and age was a significant moderator after controlling for operational- ization, 2(15,N 37,935) 335.44,p .01, and after controlling for assessment method, 2(7,N 38,071) 102.73, p .01. Comparisons Between Twin and Adoption Studies Comparisons of the results from twin (42 samples, 131 groups, 37,700 pairs of participants) and adoption studies (10 samples, 21 groups, 31,272 pairs of participants) are presented in Table 9. Twin (a 2 .45,c 2 .12,e 2 .43) and adoption (a 2 .32,c 2 .05,e 2 .63) studies yielded different parameter estimates, as there was a significant chi-square difference between the model in which the parameter estimates were constrained to be equal across twin and adoption studies and the model in which the parameter estimates were free to vary for each type of study, 2(3,N 68,972) 119.68,p .01. Results from twin studies were next compared with results from adoption studies after di- viding the adoption studies into two types: (a) studies comparing the correlations between adoptees and their adoptive parents with the correlations between adoptees and their biological parents (i.e., parent–offspring adoption studies; 7 samples, 12 groups, 30,504 pairs of participants) and (b) studies comparing the correlations between adoptive siblings with the correlations between biological siblings (i.e., sibling adoption studies; 3 samples, 9 groups, 768 pairs of participants). There was a significant difference between the results from twin studies (a 2 .45,c 2 .12,e 2 .43) and parent–offspring adoption studies (a 2 .31,c 2 .05,e 2 .64), 2(3,N 68,204) 130.81,p .001, but not between the results from twin studies and sibling adoption studies (a 2 .48, c 2 .13,e 2 .39), 2(3,N 38,468) 0.75,p .86. Given the similar results of twin and sibling adoption studies, results from the parent–offspring adoption studies were compared with those from the twin and sibling adoption studies combined (45 samples, 140 groups, 38,468 pairs of participants). The results were found to differ, such that the twin and sibling adoption studies (a 2 .44, c 2 .13,e 2 .43) yielded highera 2andc 2estimates and lower e 2estimates than the parent–offspring adoption studies (a 2 .31, c 2 .05,e 2 .64), 2(3,N 68,972) 131.65,p .01. Effect of Excluding Weight Matrices Table 10 shows the results of two analyses assessing the effect of excluding the weight matrices. First, it shows the effect of excluding the weight matrices in the samples where the estimation of weight matrices was possible. When the weight matrices were included, the best fitting model was the ACE model (a 2 .54, c 2 .28,e 2 .18), but when the weight matrices were omitted, 515 ANTISOCIAL BEHAVIOR the best fitting model was the ADE model (a 2 .43,d 2 .26, e 2 .31). This analysis shows that excluding the weight matrices results in an overestimation of the magnitude of genetic influences and an underestimation of shared environmental influences, al- though a significance test is not possible (i.e., given that the same data were analyzed in this comparison). Second, Table 10 shows the comparison between studies with and without estimable weight matrices. There was a significant chi-square difference between a model in which all estimates were constrained to be equal and a model in which estimates were free to vary between studies with estimated weight matrices (10 samples, 27 groups, 22,584 pairs of participants) and studies without estimated weight matrices (42 samples, 122 groups, 32,941 pairs of participants), 2(3, N 55,525) 303.68,p .01. Studies with estimated weight matrices (a 2 .54,c 2 .28,e 2 .18) had highera 2andc 2 estimates than studies without estimated weight matrices (a 2 .35,c 2 .17,e 2 .48). Discussion Overview of the Results When all available data from both twin and adoption studies were analyzed together and the magnitude of nonadditive geneticinfluences was estimated in addition to the magnitude of shared environmental influences, the best fitting model was the ACDE model. On the basis of this analysis, there were moderate additive genetic (a 2 .32), nonadditive genetic (d 2 .09), shared envi- ronmental (c 2 .16), and nonshared environmental (e 2 .43) influences on antisocial behavior. Operationalization, assessment method, zygosity determination method, and age accounted for significant differences in the ge- netic and environmental influences on antisocial behavior. Al- though sex was a significant moderator when data from all studies were examined, there were no statistically significant sex differ- ences in studies that examined both sexes. In the three pairs of moderators that are confounded in the literature (i.e., age and operationalization, age and assessment method, and operational- ization and assessment method), each moderator was found to be significant even after the other potentially confounding moderator was controlled for statistically. Parent–offspring adoption studies showed a lower magnitude of familial influences on antisocial behavior (i.e., lowera 2andc 2and highere 2) than the twin and sibling adoption studies. There are several possible reasons for this result. First, the age difference between the children and their parents may lead to lower correla- Table 9 Standardized Parameter Estimates and Fit Statistics for the Best Fitting Models—Comparison Between Twin and Adoption Studies ModelsParameter estimate Fit statistic a 2 c2 e2 d2 2 df pAIC Comparison between all twin studies and all adoption studies Parameters constrained to be equal .46 .10 .44—1,541.06 150 .001 1,241.06 Parameters free to vary 1,421.38 147 .001 1,127.38 Twin studies .45 .12 .43—1,355.28 129 .001 1,097.28 Adoption studies .32 .05 .63—66.10 19 .001 28.10 Chi-square difference test 119.68 3 .001 113.68 Comparison between all twin studies and parent–offspring adoption studies Parameters constrained to be equal .46 .10 .44—1,531.82 141 .001 1,249.82 Parameters free to vary 1,401.01 138 .001 1,125.01 Twin studies .45 .12 .43—1,355.28 129 .001 1,097.28 Parent–offspring studies .31 .05 .64—45.73 10 .001 25.72 Chi-square difference test 130.81 3 .001 124.81 Comparison between all twin studies and sibling adoption studies Parameters constrained to be equal .44 .13 .43—1,363.69 138 .001 1,087.69 Parameters free to vary 1,362.94 135 .001 1,092.94 Twin studies .45 .12 .43—1,355.28 129 .001 1,097.28 Sibling adoption studies .48 .13 .39—7.66 7 .36 6.34 Chi-square difference test 0.75 3 .86 5.25 Comparison between twin–sibling adoption studies and parent–offspring adoption studies Parameters constrained to be equal .46 .10 .44—1,541.06 150 .001 1,241.06 Parameters free to vary 1,409.41 147 .001 1,115.41 Twin–sibling adoption studies .44 .13 .43—1,363.69 138 .001 1,087.69 Parent–offspring studies .31 .05 .64—45.72 10 .001 25.72 Chi-square difference test 131.65 3 .001 125.65 Note.Dashes indicate that none of the best fitting models included nonadditive genetic influences.a 2 the magnitude of additive genetic influences;c 2 the magnitude of shared environmental influences;e 2 the magnitude of nonshared environmental influences;d 2 the magnitude of nonadditive genetic influences; AIC Akaike information criterion. 516 RHEE AND WALDMAN tions, given that there may be age- or cohort-specific genetic and/or environmental influences. This age difference is absent in the twin studies and smaller in the sibling adoption studies. Sec- ond, because of the practical obstacles involved in conducting an adoption study, in several studies, different operationalizations and methods of assessment were used for the adoptees and their parents (e.g., criminality via official records for the parents and aggression via self-report for the adoptees). There was not a statistically significant difference between the results of twin studies and sibling adoption studies. This result should be interpreted while considering the fact that 42 indepen- dent twin samples were compared with only 3 independent sibling adoption samples. Although the power to detect a statistically significant difference between the two types of studies may have been limited by the small number of sibling adoption studies, the parameter estimates for the twin studies (a 2 .45,c 2 .12,e 2 .43) and the sibling adoption studies (a 2 .48,c 2 .13,e 2 .39) were very similar. When data from studies with estimated weight matrices were analyzed both including and excluding the weight matrices, we found that excluding the weight matrices led to an overestimation of the magnitude of genetic influences and an underestimation of the magnitude of shared environmental influences. This suggests that simply using ML estimation without a weight matrix to analyze covariances, as is typical of contemporary twin studies of antisocial behavior, may bias parameter estimates when analyzing data that do not meet the assumption of multivariate normality. Limitations of the Present Meta-Analysis Analyses of Correlations Without Weight Matrices Most of the studies included in the meta-analysis simply re- ported Pearson or intraclass correlations in their publications, and we were limited to using this information in the meta-analysis. This leads to two major methodological limitations in the meta-analysis.One assumption of model fitting is that the variances on the outcome measures are equal for the different groups of relatives examined. Given that only correlations are analyzed, there was no way to compare the variances of different types of relatives (e.g., MZ twins vs. DZ twins; twin studies vs. adoption studies) or across other variables such as gender or age. This is an important con- sideration because there may be genuine differences in the vari- ances of outcome measures across the different groups of relatives. For example, the variance in the antisocial behavior of adoptees may be restricted because antisocial behavior is more common in adoptees than nonadoptees (e.g., Sharma et al., 1998) or because most adoptees are placed in middle-class homes (e.g., Fergusson et al., 1995). Also, given that we were not able to test for differences in variances between MZ and DZ twins, we were not able to examine sibling influences (i.e., cooperation or contrast effects), which have been found to be important in antisocial behavior (e.g., Carey, 1992). Another significant limitation in analyzing only the correlations reported in the individual studies was the limitation of having to use ML estimation rather than WLS estimation. WLS estimation is preferable to ML estimation for obtaining asymptotically correct standard errors of parameter estimates and chi-square goodness- of-fit tests when the normal distribution assumption is violated or when correlations rather than covariances are analyzed. As stated above, in the present meta-analysis, we found that excluding the weight matrices and using ML estimation led to an overestimation of the magnitude of genetic influences and an underestimation of the magnitude of shared environmental influences. Again, this result suggests that using ML estimation without weight matrices may bias parameter estimates when analyzing data that do not meet the normality assumption. Effects of Censored Variables It is possible that many of the scales measuring antisocial behavior fail to distinguish differences among the majority of the population who do not show significant problems with antisocial Table 10 Standardized Parameter Estimates and Fit Statistics for the Best Fitting Models—Effect of Excluding Weight Matrices ModelParameter estimate Fit statistic a 2 c2 e2 d2 2 df pAIC Studies with estimable weight matrices: Weight matrices included and weight matrices omitted Weight matrices included .54 .28 .18—66.03 25 .001 16.03 Weight matrices omitted .43—.26 .31 685.26 25 .001 635.26 Direct comparison between studies with and without estimable weight matrices Parameters constrained to be equal .38 .18 .44—1,420.38 147 .001 1,126.38 Parameters free to vary 1,116.70 144 .001 828.70 With weight matrices .54 .28 .18—66.03 25 .001 16.03 Without weight matrices .35 .17 .48—1,050.67 120 .001 810.67 Chi-square difference test 303.68 3 .001 297.68 Note.Dashes indicate that data are not applicable.a 2 the magnitude of additive genetic influences;c 2 the magnitude of shared environmental influences;e 2 the magnitude of nonshared environmental influences;d 2 the magnitude of nonadditive genetic influences; AIC Akaike information criterion. 517 ANTISOCIAL BEHAVIOR behavior. This failure can lead to a“floor effect”—that is, most of the sample having scores close to the lower end of the scale. This type of censoring may be the primary reason that the normality assumption is not met in many of the studies included in the meta-analysis. When variables are censored, correlations in the middle range (i.e., .50 to .60) are decreased more than correlations in the lower range. This means that if the uncensored MZ corre- lation is in the middle range, the magnitude of genetic influences is underestimated, and if the uncensored DZ correlation is in the middle range, the magnitude of genetic influences is overestimated (van den Oord & Rowe, 1997). Unfortunately, the possible effects of censoring on the results could not be assessed in the present meta-analysis. Simultaneous Estimation of Shared Environmental Influences and Nonadditive Genetic Influences The findings of this meta-analysis demonstrate the importance of comparing the results of twin and adoption studies, given the finding of significant differences between twin and parent– offspring adoption studies. Another reason for examining twin and adoption study results simultaneously is the ability to estimate the magnitude of shared environmental influences in the presence of nonadditive genetic influences, and vice versa. We found that the ACDE model, a model that includes both shared environmental influences and nonadditive genetic influences, was the best fitting model when analyzing all of the data included in the meta-analysis. Unfortunately, we were limited to comparing the more restrictive ACE, AE, CE, and ADE models when testing the significance of moderators because both twin and adoption study data were not available for each level of the moderators examined. Given that the ACE model was the best fitting model for most of these analyses, the results may give the false impression that nonadditive genetic influences are unimportant for antisocial behavior. The inability to estimate the magnitude of shared environmental influences and nonadditive genetic influences simultaneously is a limitation of both the twin study design and the adoption study design consid- ered separately. The fact that the ACE model was the best fitting model for most of the analyses examining moderators does not mean that nonadditive genetic influences are unimportant for an- tisocial behavior. Future Directions Examination of Other Operationalizations Although we were able to contrast the results from a number of different operationalizations of antisocial behavior, further mean- ingful distinctions in the operationalizations of antisocial behavior should be examined. The results of behavior genetic studies of violent versus nonviolent crime illustrate the importance of this issue. Two adoption studies and one twin study have contrasted violent and nonviolent crimes. Mednick et al. (1984) found that in Danish adopted males, the frequency of property crime was related to the number of convictions of the biological father, whereas the frequency of violent crime was not. Bohman, Cloninger, Sigvards- son, and von Knorring (1982) also found evidence that property crime and violent crime may differ in their etiology. Genetic influences were found to be significant for property crimes, but notfor cases of violent crime associated with alcoholism. Cloninger and Gottesman (1987) analyzed the data from the Danish twin sample and found that the heritability for property crimes was .78, whereas the heritability for violent crime was .50. When cross- correlations were examined, they found that there was no genetic overlap between property crime and violent crime, suggesting a distinct and specific etiology for property crime and violent crime. In this meta-analysis, the data on violent and nonviolent crimes could not be analyzed separately because most studies reported results on crime in general. In the past, researchers have disagreed about the role of genetic influences on delinquency, with some arguing that there are ge- netic influences on criminality but not on delinquency (e.g., L. F. DiLalla & Gottesman, 1991) and others arguing that there are genetic influences on delinquency as well (e.g., Rowe, 1983). This debate could not be resolved in the present meta-analysis. Many studies with child or adolescent samples did find genetic influ- ences of substantial magnitude for antisocial behavior in general, but no study examined criminality or delinquency in children or adolescents without the inclusion of aggression items. In order to resolve the past debate, new studies on juvenile delinquency (i.e., studies without the inclusion of aggression items or the method- ological problems of the early twin studies) are needed. We were unable to examine another meaningful distinction between two different kinds of aggression, that of relational and overt aggression (Crick, Casa, & Mosher, 1997; Crick & Grotpe- ter, 1995), because there are no published twin or adoption studies of relational aggression.Overt aggressionharms others through physical damage or threat of physical damage, whereasrelational aggressionharms others by damaging their peer relationships or reputation (e.g., spreading rumors, excluding them from the peer group). Although relational aggression does not lead to physical harm to the victims, it has serious consequences for both the aggressors (e.g., higher levels of loneliness, depression, and neg- ative self-perceptions, as well as concurrent and future peer rejec- tion; Crick & Grotpeter, 1995) and the victims (e.g., depression, anxiety; Crick & Grotpeter, 1996). The distinction between rela- tional and overt aggression is an especially important consider- ation when examining sex differences in aggression and its causes, given that females are significantly more relationally aggressive and less overtly aggressive than males (Crick et al., 1997; Crick & Grotpeter, 1995). Given the evidence that overt and relational aggression are correlated but distinct (Crick et al., 1997), future behavior genetic studies of overt and relational aggression should examine the degree of genetic and environmental influences that are common to both types of aggression and specific to each type of aggression. Validity of the Assessment Method It was often difficult to make conclusive statements about the moderators examined in the present meta-analysis given concerns regarding the validity of the assessment method. Confounding between assessment method and other moderators was a serious problem, and in some cases, there is convincing evidence that the results reflect the assessment method rather than other moderators that have more conceptual importance, such as operationalization or age. 518 RHEE AND WALDMAN Given the current evidence, it is not possible to distinguish whether the behavior genetic results on criminality refer to the operationalization of criminality per se or the assessment method of official records, as all of the studies examining criminality used the assessment method of official records. Beyond the problem with confounding, official records also have a validity problem, given that many criminal activities escape detection and therefore do not appear in official records (J. Q. Wilson & Herrnstein, 1985). The additional use of self-reports may lessen this problem, given self-reports’potential to assess criminality in people who are able to escape arrest or incarceration because of intelligence or high social status (Raine & Venables, 1992). Use of self-report alone (e.g., Rowe, 1983), however, also has led to debate regarding the validity of the assessment method (e.g., L. F. DiLalla & Gottes- man, 1991). If the results of studies that examine the same operationalization but use different assessment methods do not agree, questions of validity of the assessment method are raised. In this meta-analysis, studies assessing aggression with parent and self-report found that genetic influences are important, but the one study (Plomin et al., 1981) that examined aggression using an objective test found no evidence for genetic influences. The objective test used by Plomin et al. (1981) has been validated against peer ratings and teacher ratings of aggression (Johnston, DeLuca, Murtaugh, & Diener, 1977), but the sample size in Plomin et al.’s (1981) study is small. Larger behavior genetic studies using different types of validated, objective tests of aggression are necessary to resolve this question. Given these conflicting findings, there is reason to suspect that one of the assessment methods does not validly assess the construct of aggression. Thus, the finding of genetic influences on antisocial behavior or the lack thereof may be influenced by the method used to assess antisocial behavior. No matter which operationalization was being examined (i.e., diagnosis, aggression, or antisocial behavior), the magnitude of familial influences (a 2andc 2) was lower in studies using the assessment method of self-report than in studies using the assess- ment method of report by others. The only exception occurred in studies examining antisocial behavior, where thea 2estimate was .47 for both report by others and self-report. These results suggest the possibility that the lowerh 2andc 2estimates may be more a function of the assessment method of self-report than a function of any of the operationalizations that were examined. Two separate raters are involved in the assessment method of self-report, whereas only one rater rates both twins or siblings when parent report is used. It is possible that this difference between the assessment methods led to lower familial correlations and a lower estimate of the magnitude of familial influences in studies using self-report. The confounding between age and assessment method pre- cluded our ability to test L. F. DiLalla and Gottesman’s (1989) hypothesis regarding genetic influences on continuous versus tran- sitory antisocial behavior. The assessment method of report by others was used only in children and adolescents, whereas the assessment method of self-report was used only in adolescents and adults. Given the fact that the pattern of results for age (i.e., familial influences decreasing and nonfamilial influences increas- ing as age increases) was identical to the pattern of results for assessment method (i.e., familial influences smaller and nonfamil- ial influences larger for self-report than for report by others) andthat age and assessment method are confounded, it is impossible to conclude whether age moderates the magnitude of genetic and environmental influences on antisocial behavior. The assessment methods used in future behavior genetic studies of antisocial behavior should be diversified given the common concerns regarding the validity of the assessment method. For example, a combination of official records and self-report should be used to assess criminality given the shortcomings of each assessment method. Larger behavior genetic studies using different types of validated, objective tests of aggression are needed. Most important, the limitations of the assessment method chosen for a behavior genetic study of antisocial behavior should be acknowl- edged and considered given the evidence that the assessment method can influence the results. If multiple assessment methods are used to assess antisocial behavior in a single twin study, the common pathways model (see Figure 2) can be used to estimate the magnitude of the genetic and environmental influences that are common to the latent construct being examined (i.e., antisocial behavior) and the genetic and environmental influences that are specific to each assessment method (e.g., Riemann, 1999). Genotype–Environment Interaction The adoption study is the ideal method for testing genotype– environment interactions because the genetic and environmental influences on a trait are disentangled and can be measured dis- tinctly. In contrast, genotype–environment interactions may be more difficult to test in twin studies because the genetic and environmental influences on a trait are likely to be correlated. Data from several adoption studies (Cadoret et al., 1983; Clon- inger, Sigvardsson, Bohman, & von Knorring, 1982; Mednick et al., 1983) show evidence of genotype–environment interaction for antisocial behavior, although there were not enough relevant stud- ies in the meta-analysis to conduct a quantitative review of this issue. Mednick et al. (1983) conducted a cross-fostering analysis of Danish adoptees. Among adoptees who had a criminal back- ground in both their biological and adoptive parents, 24.5% be- came criminal themselves. This is in comparison to 20% of adopt- ees who have a criminal background only in their biological parents, 14.7% of adoptees who have a criminal background only in their adoptive parents, and 13.5% of adoptees with no criminal background. Cloninger et al. (1982) found similar results for petty criminality in Swedish adoptees when they considered both bio- logical variables (i.e., criminality in biological parents) and envi- ronmental variables (i.e., negative rearing experiences and adop- tive placement). Among adoptees with both biological and environmental risks, 40% were criminal. This is in comparison to 12.1% of those with only biological risk factors, 6.7% of those with only environmental risk factors, and 2.9% of those with neither biological nor environmental risk factors. Also, in a sample of adoptees from Iowa, Cadoret et al. (1983) found that when both genetic and environmental risk factors were present, they ac- counted for a greater number of antisocial behaviors than an additive combination of the two kinds of risk factors acting independently. The genotype–environment interactions were not statistically significant in Cloninger et al. (1982) or Mednick et al. (1983). Unfortunately, the power to test the genotype–environment inter- action term may be reduced in adoption studies of antisocial 519 ANTISOCIAL BEHAVIOR behavior because of range restriction in the variables used to indicate the environmental and/or genetic influences on antisocial behavior. McClelland and Judd (1993) demonstrated that restrict- ing the range of the predictor variables reduces the residual vari- ance of the product of the two predictors and, in turn, the statistical power to detect an interaction. The problem with range restriction is especially a concern in adoption studies of antisocial behavior because the chance of adoptees being placed in adoptive homes with criminal or antisocial adoptive parents is very low. For example, one reviewer of this article noted that in Cloninger et al. (1982), none of the 862 adoptees came from an adoptive family in which a parent had an arrest record. Therefore, the statistical difficulties of detecting interactions should be considered in inter- preting adoption studies examining genotype–environment inter- actions. Also, future behavior genetic studies should consider alternative research design strategies, such as oversampling ex- treme observations (McClelland & Judd, 1993). For example, such studies may oversample children with a low genetic predisposition to antisocial behavior who are reared in environments that predis- pose them to antisocial behavior. Multivariate Analyses In the present meta-analysis, four operationalizations of antiso- cial behavior were studied: diagnosis, criminality, aggression, andantisocial behavior. Operationalization was a significant modera- tor, suggesting that the magnitude of genetic and environmental influences is different for the different operationalizations. In order to determine the extent to which these operationaliza- tions have common or specific genetic and environmental influ- ences, multivariate behavior genetic analyses of two or more operationalizations should be conducted. One example of such an analysis is Cloninger and Gottesman’s (1987) finding that there is little genetic overlap between violent and nonviolent crime. According to several reviewers (e.g., Carey, 1994; L. F. DiLalla & Gottesman, 1991; Nigg & Goldsmith, 1994), the next important step in clarifying the role of genes and environment on antisocial behavior is multivariate behavior genetic research on personality and psychopathology. These researchers have suggested a number of personality variables that may share common genetic influences with antisocial behavior, including low harm avoidance, high novelty seeking, low reward dependence, overattribution of hos- tility, and many others. Gottesman and Goldsmith (1994) sug- gested that the statistical line of evidence that must be established is the documentation of the heritability of the personality variables, demonstration that the personality variables predict antisocial be- havior, documentation that the patterns of antisocial behavior are heritable, and demonstration that the genetic influences underlying both the personality variables and antisocial behavior overlap. Figure 2.Common pathway model. A additive genetic influences; C shared environmental influences; E nonshared environmental influences; ASB antisocial behavior; AM assessment method. 520 RHEE AND WALDMAN Age of Onset and Developmentally Different Subtypes of Antisocial Behavior L. F. DiLalla and Gottesman (1989) and Moffitt (1993) have suggested that in order to show conclusive evidence regarding their hypotheses, future studies of antisocial behavior should in- clude longitudinal data of the same individuals. Five studies in- cluded in the current meta-analysis examined the same individuals at two time points, but none of these studies provide the kind of evidence needed to examine L. F. DiLalla and Gottesman’s (1989) hypothesis. Two of these studies (Loehlin, Willerman, & Horn, 1987; Lytton, Watts, & Dunn, 1988) only assessed antisocial behavior at the second assessment. Dworkin, Burke, Maher, and Gottesman (1976) found that heritability for psychopathy (i.e., as measured by the MMPI Psychopathic Deviate scale and the CPI Socialization scale) was significant during adolescence (mean age 15.9) but not during adulthood (mean age 27.9 years); however, the sample size was very small (i.e., 27 MZ pairs and 17 DZ pairs). McGue, Bacon, and Lykken (1993) found that there are nonadditive genetic influences on aggression at both late adoles- cence (mean age 20 years) and adulthood (mean age 30 years), but the same sample size was small (79 MZ pairs and 48 DZ pairs). Deater-Deckard, Reiss, Hetherington, and Plomin (1997) also reported results on longitudinal data for the same individuals, but the two waves of assessments were only 3 years apart and both assessments occurred when the twins were adolescents. Environmental Influences on Antisocial Behavior The most frequently cited candidate for a specific environmental influence on antisocial behavior is parenting style. Patterson and his colleagues (e.g., Patterson et al., 1992) contended that inade- quate parental supervision can lead to antisocial behavior in chil- dren. They have describedcoercive cyclesduring which a child responds to a mother’s command with aggression or a temper tantrum, the mother responds in turn by backing off, and the aggression or temper tantrum is thus reinforced. Several experi- mental studies using random assignment show that parent man- agement training, which attempts to alter these coercive cycles by training parents to reinforce prosocial behavior rather than aggres- sive behavior, is effective in improving parenting skills and reduc- ing aggressive behavior in children (Brestan & Eyberg, 1998; Kazdin, 1987). Further evidence for coercion theory is provided by studies that show that the intervention’s effect on the child’s aggressive behavior is mediated by the improvement in parenting practices. For example, Forgatch and DeGarmo (1999) showed that parent training reduced coercive parenting, prevented decay in positive parenting, and improved effective parenting practices, and that these improvements in turn led to improvements in child adjustment, including reduced externalizing behavior. Similarly, Eddy and Chamberlain (2000) showed that the positive effects of multidimensional treatment foster care on severe antisocial behav- ior were mediated by improved family management skills. Also, parenting style may influence children’s antisocial behavior indi- rectly through sibling influences (Bank, Patterson, & Reid, 1996) and peer influences (Forgatch & Stoolmiller, 1994). The results of these studies support the view that parenting styles and behavior represent important environmental influences on antisocial behav- ior and that they should be included as specific environmental indices in future behavior genetic studies.In contrast to previous theories that emphasize the influence of parenting, Harris’s (1995, 1998) group socialization theory of development emphasizes the importance of peer group influences on personality development. Harris’s (1995) main criticism of the previous research emphasizing the influence of parenting styles is the failure to consider the possibility of genetic influences on children’s behavior and the possibility that parents could be react- ing to their children’s behavior rather than causing it. Harris (1995) cited examples of significant peer group influences on several variables including smoking (Rowe, Chassin, Presson, Edwards, & Sherman, 1992, as cited in Harris, 1995) and motivation to do well in school (Kindermann, 1993; as cited in Harris, 1995) and sug- gested that neighborhood and peer group influences are also im- portant environmental influences on antisocial behavior. Accord- ing to the group socialization theory of development, delinquency is pervasive during adolescence (Moffitt, 1993) not because ado- lescents are aspiring to adult status but because adolescents are contrasting themselves from adults as a group by exhibiting de- linquent behaviors that set them apart from adults. Harris’s (1995, 1998) theory is consistent with previous studies that have reported a significant relationship between exposure to deviant peers and antisocial behavior (e.g., Keenan, Loeber, Zhang, Stouthamer- Loeber, & van Kammen, 1995). On the other hand, Rowe, Woulbroun, and Gulley (1994) raised the possibility that the relationship between exposure to deviant peers and antisocial behavior may be due to peer selection (i.e., deviant children being more likely to select deviant peers than nondeviant children) rather than peer influence. For example, Rowe and Osgood (1984) found that children’s antisocial behavior was significantly related to association with deviant peers, but the cross-correlation between the children’s own antisocial behavior and association with deviant peers was higher in MZ twins than in DZ twins. This result suggests that there are genetic influences on the relation between antisocial behavior and association with de- viant peers and supports peer selection as an explanation for this relationship. Deater-Deckard and Dodge (1997) attempted to integrate the results of studies examining the influence of parenting and those of behavior genetic studies of childhood antisocial behavior. They concluded that there is a significant relation between harsh phys- ical discipline and childhood antisocial behavior but that the magnitude of the effect depends on several variables. First, the association between harsh physical discipline and childhood ag- gression includes a nonlinear component, in that the degree of association should be larger for the upper end of the continuum of physical discipline (i.e., harsh discipline or abusive discipline). Stoolmiller, Patterson, and Snyder (1997) also found evidence for a nonlinear relation between harsh, abusive discipline and antiso- cial behavior but suggested that the causal effect may be limited to families with out-of-control children and unskilled parents. Sec- ond, the association varies across cultural groups, in that there is a positive correlation between physical discipline and childhood aggression for European American children, but not for African American children. Third, parental discipline effects vary accord- ing to the context of the broader parent–child relationship, such as parent–child warmth. Fourth, the relation between harsh physical discipline and childhood aggression is stronger for same-gender parent–child dyads. Turkheimer and Gottesman (1996) discussed the lack of evi- dence for shared environmental influences from behavior genetic 521 ANTISOCIAL BEHAVIOR studies and offered a possible explanation for this finding. They conducted a study simulating the dynamics of genes, environment, and development and concluded that environmental variation is only detectable when the genotype is held constant. Turkheimer and Gottesman explained that two siblings with different geno- types can both be affected by the same shared environmental influences but that the effect of the shared environmental influ- ences may make them dissimilar rather than similar given the differences in their genotype. They also found that small changes in environment can result in large and sudden changes in pheno- typic outcomes that would be difficult to capture with traditional linear models. In contrast, linear models fit the phenotypic varia- tion associated with genotype well. In future studies examining shared environmental influences on antisocial behavior, research- ers should consider the possibility of nonlinear relations. Given the strong evidence of both shared and nonshared envi- ronmental influences on antisocial behavior, more studies exam- ining specific shared and nonshared environmental influences within behavior genetic designs are needed. Behavior genetic studies are uniquely equipped to examine these issues, given their ability to estimate the true magnitude of parental and peer envi- ronmental influences on antisocial behavior while controlling for genetic influences, including those on peer selection. Although difficult to implement, the examination of specific environmental influences in a combined twin–adoption design is especially rec- ommended given the ability to examine measured environmental variables, shared and nonshared environmental influences, and additive and nonadditive genetic influences simultaneously. There also are a number of genetically noninformative designs that can be used to evaluate the effects of the environment while control- ling for genetic effects (see Rutter, Pickles, Murray, & Eaves, 2001, for a detailed review). These include migration designs (i.e., comparing the incidence of a disorder in a migrant population to that in the country of origin and the host country), time trends analyses (e.g., changes in marriage rates, secular trends in suicide), and intervention designs (e.g., the parent training studies discussed above). Conclusion In the current meta-analysis, we found that there were moderate additive genetic (a 2 .32), nonadditive genetic (d 2 .09), shared environmental (c 2 .16), and nonshared environmental influences (e 2 .43) on antisocial behavior. When twin and adoption studies were compared, there was a significant difference between twin studies and parent–offspring adoption studies, but not between twin studies and sibling adoption studies. There was a lower magnitude of familial influences (i.e., botha 2andc 2) in the parent–offspring adoption studies as compared with the twin or sibling adoption studies. All of the potential moderators examined except for sex (i.e., operationalization, assessment method, zygos- ity determination method, and age) were found to account for significant differences in the genetic and environmental influences on antisocial behavior. Although there was a significant difference between studies examining both sexes simultaneously and studies examining the two sexes separately, there was no statistically significant difference in the results for males and females in studies that included both sexes. Several future directions were recommended for overcoming the limitations of the present meta- analysis and for improving our understanding of genetic andenvironmental influences on antisocial behavior. These include examining further distinctions in the operationalization of antiso- cial behavior, diversifying the assessment methods used to mea- sure antisocial behavior, examining genotype–environment inter- actions, conducting multivariate behavior genetic analyses, conducting longitudinal studies to more effectively examine the effects of age of onset and developmentally different subtypes on the genetic and environmental influences underlying antisocial behavior, examining specific environmental influences, and con- trolling for the effects of assortative mating. References References marked with an asterisk indicate studies considered for the meta-analysis. Abram, K. M. (1989). The effect of co-occurring disorders on criminal careers: Interaction of antisocial personality, alcoholism, and drug dis- orders.International Journal of Law and Psychiatry, 12,133–148. Achenbach, T. M., & Edelbrock, C. S. (1983).Manual for the Child Behavior Checklist and revised behavior profile.Burlington: University of Vermont. American Psychiatric Association. (1994).Diagnostic and statistical man- ual of mental disorders(4th ed.). Washington, DC: Author. *Baker, L. A. (1986). Estimating genetic correlations among discontinuous phenotypes: An analysis of criminal convictions and psychiatric-hospital diagnoses in Danish adoptees.Behavior Genetics, 16,127–142. *Baker, L. A., Mack, W., Moffitt, T., & Mednick, S. (1989). Sex differ- ences in property crime in a Danish adoption cohort.Behavior Genet- ics, 19,355–370. Bank, L., Patterson, G. R., & Reid, J. B. (1996). Negative sibling interac- tion patterns as predictors of later adjustment problems in adolescent and young adult males. In G. H. Brody (Ed.),Sibling relationships: Their causes and consequences(pp. 197–229). Norwood, NJ: Ablex. *Blanchard, J. M., Vernon, P. A., & Harris, J. A. (1995). A behavior genetic investigation of multiple dimensions of aggression.Behavior Genetics, 25,256. *Bohman, M. (1978). Some genetic aspects of alcoholism and criminality: A population of adoptees.Archives of General Psychiatry, 35,269–276. *Bohman, M., Cloninger, C. R., Sigvardsson, S., & von Knorring, A.-L. (1982). Predisposition to petty criminality in Swedish adoptees: I. Ge- netic and environment heterogeneity.Archives of General Psychia- try, 39,1233–1241. *Bouchard, T. J., & McGue, M. (1990). Genetic and rearing environmental influences on adult personality: An analysis of adopted twins reared apart.Journal of Personality, 58,263–292. *Brandon, K., & Rose, R. J. (1995). A multivariate twin family study of the genetic and environmental structure of personality, beliefs, and alcohol use [Abstract].Behavior Genetics, 25,257. Brestan, E. V., & Eyberg, S. M. (1998). Effective psychosocial treatments of conduct-disordered children and adolescents: 29 years, 82 studies, and 5,272 kids.Journal of Clinical Child Psychology, 27,180–189. *Cadoret, R. J. (1978). Psychopathology in adopted-away offspring of biologic parents with antisocial behavior.Archives of General Psychi- atry, 35,176–184. Cadoret, R. J., Cain, C. A., & Crowe, R. R. (1983). Evidence for gene– environment interaction in the development of adolescent antisocial behavior.Behavior Genetics, 13,301–310. *Cadoret, R. J., Cunningham, L., Loftus, R., & Edwards, J. (1975). Studies of adoptees from psychiatrically disturbed biological parents: II. Tem- perament, hyperactive, antisocial and developmental variables.Journal of Pediatrics, 87,301–306. *Cadoret, R. J., O’Gorman, T. W., Troughton, E., & Heywood, E. (1985). Alcoholism and antisocial personality: Interrelationships, genetic and environmental factors.Archives of General Psychiatry, 42,161–167. 522 RHEE AND WALDMAN *Cadoret, R. J., & Stewart, M. A. (1991). An adoption study of attention deficit/hyperactivity/aggression and their relationship to adult antisocial personality.Comprehensive Psychiatry, 32,73–82. *Cadoret, R. J., Troughton, E., Bagford, J., & Woodworth, G. (1990). Genetic and environmental factors in adoptee antisocial personality.Eu ropean Archives of Psychiatry and Neurological Sciences, 239,231– 240. *Cadoret, R. J., Troughton, E., & O’Gorman, T. W. (1987). Genetic and environmental factors in alcohol abuse and antisocial personality.Jour- nal of Studies on Alcohol, 48,1–8. *Cadoret, R. J., Troughton, E., O’Gorman, T. W., & Heywood, E. (1986). An adoption study of genetic and environmental factors in drug abuse. Archives of General Psychiatry, 43,1131–1136. *Cadoret, R. J., Yates, W. R., Troughton, E., Woodworth, G., & Stewart, M. A. (1995). Adoption study demonstrating two genetic pathways to drug abuse.Archives of General Psychiatry, 52,42–52. *Cadoret, R. J., Yates, W. R., Troughton, E., Woodworth, G., & Stewart, M. A. (1996). An adoption study of drug abuse/dependency in females. Comprehensive Psychiatry, 37,88–94. *Carey, G. (1992). Twin imitation for antisocial behavior: Implications for genetic and family environment research.Journal of Abnormal Psychol- ogy, 101,18–25. Carey, G. (1994). Genetics and violence. In A. J. Reiss, K. A. Miczek, & J. A. Roth (Eds.),Understanding and preventing violence(Vol. 2, pp. 21–58). Washington, DC: National Academy Press. *Cates, D. S., Houston, B. K., Vavak, C. R., Crawford, M. H., & Uttley, M. (1993). Heritability of hostility-related emotions, attitudes, and be- haviors.Journal of Behavioral Medicine, 16,237–256. *Centerwall, B. S., & Robinette, C. D. (1989). Twin concordance for dishonorable discharge from the military: With a review of genetics of antisocial behavior.Comprehensive Psychiatry, 30,442–446. *Christiansen, K. O. (1973). Mobility and crime among twins.Interna- tional Journal of Criminology and Penology, 1,31–45. *Christiansen, K. O. (1974). Seriousness of criminality and concordance among Danish twins. In R. Hood (Ed.),Crime, criminology, and public policy(pp. 63–77). London: Heinemann. *Christiansen, K. O. (1977a). A preliminary study of criminality among twins. In S. A. Mednick & K. O. Christiansen (Eds.),Biological bases of criminal behavior(pp. 89–108). New York: Gardner Press. Christiansen, K. O. (1977b). A review of studies of criminality among twins. In S. A. Mednick & K. O. Christiansen (Eds.),Biosocial bases of criminal behavior(pp. 45–88). New York: Gardner Press. Clerget-Darpoux, F., Goldin, L. R., & Gershon, E. S. (1986). Clinical methods in psychiatric genetics: III. Environmental stratification may simulate a genetic effect in adoption studies.Acta Psychiatrica Scan- danavica, 74,305–311. *Cloninger, C. R., Christiansen, K. O., Reich, T., & Gottesman, I. I. (1978). Implications of sex differences in the prevalences of antisocial personality, alcoholism, and criminality for familial transmission.Ar- chives of General Psychiatry, 35,941–951. Cloninger, C. R., & Gottesman, I. I. (1987). Genetic and environmental factors in antisocial behavior disorders. In S. A. Mednick, T. E. Moffitt, & S. A. Stack (Eds.),The causes of crime: New biological approaches (pp. 92–109). New York: Cambridge University Press. Cloninger, C. R., & Reich, T. (1983). Genetic heterogeneity in alcoholism and sociopathy. In S. S. Kety, L. P. Rowland, R. L. Sidman, & S. W. Matthysse (Eds.),Genetics of neurological and psychiatric disorders (pp. 145–166). New York: Raven Press. Cloninger, C. R., Sigvardsson, S., Bohman, M., & von Knorring, A.-L. (1982). Predisposition to petty criminality in Swedish adoptees: II. Cross-fostering analysis of gene–environment interaction.Archives of General Psychiatry, 39,1242–1247. *Coccaro, E. F., Bergeman, C. S., Kavoussi, R. J., & Seroczynski, A. D. (1997). Heritability of aggression and irritability: A twin study of theBuss–Durkee aggression scales in adult male subjects.Biological Psy- chiatry, 41,264–272. *Coid, B., Lewis, S. W., & Reveley, A. M. (1993). A twin study of psychosis and criminality.British Journal of Psychiatry, 162,87–92. Cooney, N. L., Kadden, R. M., & Litt, M. D. (1990). A comparison of methods for assessing sociopathy in male and female alcoholics.Journal of Studies on Alcohol, 51,42–48. Corey, L. A., Nance, W. E., Kang, K. W., & Christian, J. C. (1979). Effects of type of placentation on birthweight and its variability in monozygotic and dizygotic twins.Acta Geneticae Medicae et Gemellologiae (Roma), 28,41–50. Crick, N. R., Casa, J. F., & Mosher, M. (1997). Relational and overt aggression in preschool.Developmental Psychology, 33,579–588. Crick, N. R., & Grotpeter, J. K. (1995). Relational aggression, gender, and social-psychological adjustment.Child Development, 66,710–722. Crick, N. R., & Grotpeter, J. K. (1996). Children’s treatment by peers: Victims of relational and overt aggression.Development and Psychopa- thology, 8,367–380. Crowe, R. R. (1972). The adopted offspring of women criminal offenders: A study of their arrest records.Archives of General Psychiatry, 27, 600–603. Crowe, R. R. (1974). An adoption study of antisocial personality.Archives of General Psychiatry, 31,785–791. Crowe, R. R. (1975). An adoptive study of psychopathy: Preliminary results from arrest records and psychiatric hospital records. In R. Fieve, D. Rosenthal, & H. Brill (Eds.),Genetic research in psychiatry(pp. 95–103). Baltimore: Johns Hopkins University Press. *Cunningham, L., Cadoret, R. J., Loftus, R., & Edwards, J. E. (1975). Studies of adoptees from psychiatrically disturbed biological parents: Psychiatric conditions in childhood and adolescence.British Journal of Psychiatry, 126,534–549. Dalgard, O. S., & Kringlen, E. (1976). A Norwegian twin study of criminality.British Journal of Criminology, 16,213–232. Deater-Deckard, K., & Dodge, K. A. (1997). Externalizing behavior prob- lems and discipline revisited: Nonlinear effects and variation by culture, context, and gender.Psychological Inquiry, 8,161–175. *Deater-Deckard, K., & Plomin, R. (1999). An adoption study of the etiology of teacher and parent reports of externalizing behavior problems in middle childhood.Child Development, 70,144–154. *Deater-Deckard, K., Reiss, D., Hetherington, E. M., & Plomin, R. (1997). Dimensions and disorders of adolescent adjustment: A quantitative genetic analysis of unselected samples and selected extremes.Journal of Child Psychology and Psychiatry, 38,515–525. Defilippis, N. A. (1979). Concurrent validity of the Missouri Children’s Picture Series.Journal of Clinical Psychology, 35,433–435. Devlin, B., Daniels, M., & Roeder, K. (1997). The heritability of IQ. Nature, 388,468–471. *DiLalla, D. L., Carey, G., Gottesman, I. I., & Bouchard, T. J. (1996). Heritability of MMPI personality indicators of psychopathology in twins reared apart.Journal of Abnormal Psychology, 105,491–499. DiLalla, L. F., & Gottesman, I. I. (1989). Heterogeneity of causes for delinquency and criminality: Lifespan perspectives.Development and Psychopathology, 1,339–349. DiLalla, L. F., & Gottesman, I. I. (1991). Biological and genetic contrib- utors to violence: Widom’s untold tale.Psychological Bulletin, 109, 125–129. Dodge, K. A., Bates, J., & Pettit, G. S. (1990, December 21). Mechanisms in the cycle of violence.Science, 250,1678–1683. *Dworkin, R. H., Burke, B. W., Maher, B. A., & Gottesman, I. I. (1976). A longitudinal study of the genetics of personality.Journal of Person- ality and Social Psychology, 34,510–518. Eaves, L. J. (1988). Dominance alone is not enough.Behavior Genet- ics, 18,27–33. *Eaves, L. J., Silberg, J. L., Meyer, J. M., Maes, H. H., Simonoff, E., Pickles, A., et al. (1997). Genetics and developmental psychopathology: 523 ANTISOCIAL BEHAVIOR II. The main effects of genes and environment on behavioral problems in the Virginia Twin Study of Adolescent Behavioral Development. Journal of Child Psychology and Psychiatry and Allied Disciplines, 38, 965–980. Eddy, J. M., & Chamberlain, P. (2000). Family management and deviant peer association as mediators of the impact of treatment condition on youth antisocial behavior.Journal of Consulting and Clinical Psychol- ogy, 68,857–863. *Edelbrock, C., Rende, R., Plomin, R., & Thompson, L. A. (1995). A twin study of competence and problem behavior in childhood and early adolescence.Journal of Child Psychology and Psychiatry and Allied Disciplines, 36,775–785. *Eley, T. C., Lichtenstein, P., & Stevenson, J. (1999). Sex differences in the etiology of aggressive and nonaggressive antisocial behavior: Re- sults from two twin studies.Child Development, 70,155–168. Emde, R. N., Plomin, R., Robinson, J. A., Corley, R., DeFries, J., Fulker, D. W., et al. (1992). Temperament, emotion, and cognition at fourteen months: The MacArthur longitudinal twin study.Child Develop- ment, 63,1437–1455. Falconer, D. S., & Mackay, T. F. C. (1996).Introduction to quantitative genetics(4th ed.). Harlow, England: Longman. Fergusson, D. M., Lynskey, M., & Horwood, L. J. (1995). The adolescent outcomes of adoption: A 16-year longitudinal study.Journal of Child Psychology and Psychiatry, 36,597–615. *Finkel, D., & McGue, M. (1997). Sex differences and nonadditivity in heritability of the multidimensional personality questionnaire scales. Journal of Personality and Social Psychology, 72,929–938. Forgatch, M. S., & DeGarmo, D. S. (1999). Parenting through change: An effective prevention program for single mothers.Journal of Consulting and Clinical Psychology, 67,711–724. Forgatch, M. S., & Stoolmiller, M. (1994). Emotions as contexts for adolescent delinquency.Journal of Research on Adolescence, 4,601– 614. *Gabrielli, W. F., & Mednick, S. A. (1984). Urban environment, genetics, and crime.Criminology, 22,645–652. Gau, J. S., Silberg, J. L., Erickson, M. T., & Hewitt, J. K. (1992). Childhood behavior problems: A comparison of twin and non-twin samples.Acta Geneticae Medicae et Gemellologiae, 41,53–63. *Ghodsian-Carpey, J., & Baker, L. A. (1987). Genetic and environmental influences on aggression in 4- to 7-year-old twins.Aggressive Behav- ior, 13,173–186. Gjone, H., & Nøvik, T. S. (1995). Parental ratings of behaviour problems: A twin and general population comparison.Journal of Child Psychology and Psychiatry, 36,1213–1224. *Gottesman, I. I. (1963). Heritability of personality: A demonstration. Psychological Monographs, 77(9, Whole No. 572). *Gottesman, I. I. (1965). Personality and natural selection. In S. G. Van- denberg (Ed.),Methods and goals in human behavior genetics(pp. 63–80). New York: Academic Press. *Gottesman, I. I. (1966). Genetic variance in adaptive personality traits. Journal of Child Psychology and Psychiatry, 7,199–208. *Gottesman, I. I., Carey, G., & Bouchard, T. J. (1984, May).The Minne- sota Multiphasic Personality Inventory of identical twins raised apart. Paper presented at the meeting of the Behavior Genetics Association, Bloomington, IN. Gottesman, I. I., & Goldsmith, H. H. (1994). Developmental psychopa- thology of antisocial behavior: Inserting genes into its ontogenesis and epigenesis. In C. A. Nelson (Ed.),Threats to optimal development: Integrating biological, psychological, and social risk factors(pp. 69– 104). Hillsdale, NJ: Erlbaum. Gough, H. G. (1969).Manual for the California Psychological Inventory (Rev. ed.). Palo Alto, CA: Consulting Psychologists Press. Gough, H. G., & Heilbrun, A. B. (1972).The Adjective Checklist manual. Palo Alto, CA: Consulting Psychologists Press.Grayson, D. A. (1989). Twins reared together: Minimizing shared envi- ronmental influences.Behavior Genetics, 19,593–604. *Grove, W. M., Eckert, E. D., Heston, L., Bouchard, T. J., Segal, N., & Lykken, D. T. (1990). Heritability of substance abuse and antisocial behavior: A study of monozygotic twins reared apart.Biological Psy- chiatry, 27,1293–1304. *Gustavsson, J. P., Pedersen, N. L.,Åsberg, M., & Schalling, D. (1996). Exploration into the sources of individual differences in aggression-, hostility-, and anger-related (AHA) personality traits.Personality and Individual Differences, 21,1067–1071. Hare, R. D., Hart, S. D., & Harpur, T. A. (1991). Psychopathy and the DSM–IVcriteria for antisocial personality disorder.Journal of Abnormal Psychology, 100,391–398. Harris, J. R. (1995). Where is the child’s environment? A group socializa- tion theory of development.Psychological Review, 102,458–489. Harris, J. R. (1998).The nurture assumption: Why children turn out the way they do.New York: Free Press. Hathaway, S. R., & McKinley, J. C. (1942).Minnesota Multiphasic Per- sonality Inventory.Minneapolis: University of Minnesota Press. Hayashi, S. (1967). A study of juvenile delinquency in twins.Bulletin of Osaka Medical School, 12,373–378. *Hershberger, S. L., Billig, J. P., Iacono, W. G., & McGue, M. (1995). Life events, personality, and psychopathology in late adolescence: Genetic and environmental factors [Abstract].Behavior Genetics, 25,270. *Horn, J. M., Plomin, R., & Rosenman, R. (1976). Heritability of person- ality traits in adult male twins.Behavior Genetics, 6,17–30. *Hutchings, B., & Mednick, S. A. (1971). Criminality in adoptees and their adoptive and biological parents: A pilot study. In S. A. Mednick & K. O. Christiansen (Eds.),Biosocial bases of criminal behavior(pp. 127–141). New York: Gardner Press. Hyde, J. S. (1984). How large are gender differences in aggression? A developmental meta-analysis.Developmental Psychology, 20,722–736. Jary, M. L., & Stewart, M. A. (1985). Psychiatric disorder in the parents of adopted children with aggressive conduct disorder.Neuropsychobiol- ogy, 1,7–11. Johnston, A., DeLuca, D., Murtaugh, K., & Diener, E. (1977). Validation of a laboratory play measure of child aggression.Child Development, 48, 324–327. Kasriel, J., & Eaves, L. (1976). The zygosity of twins: Further evidence on the agreement between diagnosis by blood groups and written question- naires.Journal of Biosocial Science, 8,263–266. Kazdin, A. E. (1987). Treatment of antisocial behavior in children: Current status and future directions.Psychological Bulletin, 102,187–203. Keenan, K., Loeber, R., Zhang, Q., Stouthamer-Loeber, M., & van Kam- men, W. B. (1995). The influence of deviant peers on the development of boys’disruptive and delinquent behavior: A temporal analysis.De- velopment and Psychopathology, 7,715–726. Krueger, R. F., Moffitt, T. E., Caspi, A., Bleske, A., & Silva, P. A. (1998). Assortative mating for antisocial behavior: Developmental and method- ological limitations.Behavior Genetics, 28,173–186. Langbehn, D. R., Cadoret, R. J., Yates, W. R., Troughton, E. P., & Stewart, M. A. (1998). Distinct contributions of conduct and oppositional defiant symptoms to adult antisocial behavior: Evidence from an adoption study. Archives of General Psychiatry, 55,821–829. *Livesley, W. J., Jang, K. L., Jackson, D. N., & Vernon, P. A. (1993). Genetic and environmental contributions to dimensions of personality disorder.American Journal of Psychiatry, 150,1826–1831. Loehlin, J. C. (1992a).Genes and environment in personality development. Newbury Park, CA: Sage. Loehlin, J. C. (1992b).Latent variable models: An introduction to factor, path, and structural analysis(2nd ed.). Hillsdale, NJ: Erlbaum. *Loehlin, J. C., & Nichols, R. C. (1976).Heredity, environment, and personality.Austin: University of Texas Press. *Loehlin, J. C., Willerman, L., & Horn, J. M. (1985). Personality resem- 524 RHEE AND WALDMAN blance in adoptive families when the children are late-adolescent or adult.Journal of Personality and Social Psychology, 48,376–392. *Loehlin, J. C., Willerman, L., & Horn, J. M. (1987). Personality resem- blance in adoptive families: A 10-year follow-up.Journal of Personality and Social Psychology, 53,961–969. Lykken, D. T. (1997). Incompetent parenting: Its causes and cures.Child Psychiatry and Human Development, 27,129–137. *Lykken, D. T., Tellegen, A., & DeRubeis, R. (1978). Volunteer bias in twin research: The rule of two-thirds.Social Biology, 25,1–9. *Lyons, M. J., True, W. R., Eisen, S. A., Goldberg, J., Meyer, J. M., Faraone, S. V., et al. (1995). Differential heritability of adult and juvenile antisocial traits.Archives of General Psychiatry, 52,906–915. *Lytton, H., Watts, D., & Dunn, B. E. (1988). Stability of genetic deter- mination from age 2 to age 9: A longitudinal twin study.Social Biol- ogy, 35,62–73. Mason, D. A., & Frick, P. J. (1994). The heritability of antisocial behavior: A meta-analysis of twin and adoption studies.Journal of Psychopathol- ogy and Behavioral Assessment, 16,301–323. Matheny, A. P. (1989). Children’s behavioral inhibition over age and across situations: Genetic similarity for a trait during change.Journal of Personality, 57,215–235. McCartney, K., Harris, M. J., & Bernieri, F. (1990). Growing up and growing apart: A developmental meta-analysis of twin studies.Psycho- logical Bulletin, 107,226–237. McClelland, G. H., & Judd, C. M. (1993). Statistical difficulties of detect- ing interactions and moderator effects.Psychological Bulletin, 114, 376–390. *McGue, M., Bacon, S., & Lykken, D. T. (1993). Personality stability and change in early adulthood: A behavioral genetic analysis.Developmental Psychology, 29,96–109. *McGue, M., Sharma, A., & Benson, P. (1996). The effect of common rearing on adolescent adjustment: Evidence from a U.S. adoption cohort. Developmental Psychology, 32,604–613. *Mednick, S. A., Gabrielli, W. F., & Hutchings, B. (1983). Genetic influences in criminal behavior: Evidence from an adoption cohort. In K. T. Van Dusen & S. A. Mednick (Eds.),Prospective studies of crime and delinquency(pp. 39–56). Boston: Kluwer-Nijhoff. Mednick, S. A., Gabrielli, W. F., & Hutchings, B. (1984, May 25). Genetic influences in criminal convictions: Evidence from an adoption cohort. Science, 224,891–894. *Meininger, J. C., Hayman, L. L., Coates, P. M., & Gallagher, P. (1988). Genetics or environment? Type A behavior and cardiovascular risk factors in twin children.Nursing Research, 37,341–346. Melnick, M., Myrianthopoulos, M. N., & Christian, J. C. (1978). The effects of chorion type on variation in IQ in the NCPP twin population. American Journal of Human Genetics, 30,425–433. Miles, D. R., & Carey, G. (1997). Genetic and environmental architecture of human aggression.Journal of Personality and Social Psychology, 72, 207–217. Moffitt, T. E. (1993). Adolescence-limited and life-course-persistent anti- social behavior: A developmental taxonomy.Psychological Review, 100,674–701. Mullen, B. (1989).Advanced BASIC meta-analysis.Hillsdale, NJ: Erlbaum. Mutaner, C., Walter, D., Nagoshi, C., Fishbein, D., Haertzen, C. A., & Jaffe, J. H. (1990). Self-report vs. laboratory measures of aggression as predictors of substance abuse.Drug and Alcohol Dependence, 25,1–11. *Nathawat, S. S., & Puri, P. (1995). A comparative study of MZ and DZ twins on Level I and Level II mental abilities and personality.Journal of the Indian Academy of Applied Psychology, 21,87–92. Neale, M. C. (1995).Mx: Statistical modeling.Richmond: Virginia Com- monwealth University, Medical College of Virginia, Department of Psychiatry. Neale, M. C., & Cardon, L. R. (1992).Methodology for genetic studies of twins and families.Dordrecht, The Netherlands: Kluwer Academic.*Neiderhiser, J. M., Pike, A., Hetherington, E. M., & Reiss, D. (1998). Adolescent perceptions as mediators of parenting: Genetic and environ- mental contributions.Developmental Psychology, 34,1459–1469. Nigg, J. T., & Goldsmith, H. H. (1994). Genetics of personality disorders: Perspectives from personality and psychopathology research.Psycho- logical Bulletin, 115,346–380. Norland, S., Shover, N., Thornton, W., & James, J. (1979). Intrafamily conflict and delinquency.Pacific Sociological Review, 22,233–237. *O’Connor, M., Foch, T., Sherry, T., & Plomin, R. (1980). A twin study of specific behavior problems of socialization as viewed by parents.Jour- nal of Abnormal Child Psychology, 8,189–199. *O’Connor, T. G., McGuire, S., Reiss, D., Hetherington, E. M., & Plomin, R. (1998). Co-occurrence of depressive symptoms and antisocial behav- ior in adolescence: A common genetic liability.Journal of Abnormal Psychology, 107,27–37. *O’Connor, T. G., Neiderhiser, J. M., Reiss, D., Hetherington, E. M., & Plomin, R. (1998). Genetic contributions to continuity, change, and co-occurrence of antisocial and depressive symptoms in adolescence. Journal of Child Psychology and Psychiatry, 39,323–336. Ollendick, D. G., & Woodward, G. L. (1982). Use of the Missouri Chil- dren’s Picture Series with school-referred children.Psychology in the Schools, 19,290–292. *Owen, D., & Sines, J. O. (1970). Heritability of personality in children. Behavior Genetics, 1,235–248. *Parker, T. (1989).Television viewing and aggression in four and seven year old children.Paper presented at Summer Minority Access to Research Training meeting, University of Colorado, Boulder. Partanen, J., Bruun, K., & Markkanen, T. (1966).Inheritance of drinking behavior: A study on intelligence, personality, and use of alcohol of adult twins.Helsinki, Finland: Finnish Foundation for Alcohol Studies. Patterson, G. R., Reid, J. B., & Dishion, T. J. (1992).Antisocial boys. Eugene, OR: Castalia. *Pike, A., McGuire, S., Hetherington, E. M., Reiss, D., & Plomin, R. (1996). Family environment and adolescent depressive symptoms and antisocial behavior: A multivariate genetic analysis.Developmental Psy- chology, 32,590–603. Plomin, R. (1981). Heredity and temperament: A comparison of twin data for self-report questionnaires, parental ratings, and objectively assessed behavior. In L. Gedda, P. Parisi, & W. E. Nance (Eds.),Twin research: Vol. 3. Intelligence, personality, and development(pp. 269–278). New York: Alan R. Liss. Plomin, R. (1986).Development, genetics, and psychology.Hillsdale, NJ: Erlbaum. Plomin, R., DeFries, J. C., McClearn, G. E., & Rutter, M. (1997).Behav- ioral genetics(3rd ed.). New York: Freeman. *Plomin, R., & Foch, T. T. (1980). A twin study of objectively assessed personality in childhood.Journal of Personality and Social Psychol- ogy, 39,680–688. *Plomin, R., Foch, T. T., & Rowe, D. C. (1981). Bobo clown aggression in childhood: Environment, not genes.Journal of Research in Person- ality, 15,331–342. Plomin, R., Nitz, K., & Rowe, D. C. (1990). Behavioral genetics and aggressive behavior in childhood. In M. Lewis & S. M. Miller (Eds.), Handbook of developmental psychopathology(pp. 119–133). New York: Plenum. *Pogue-Geile, M. F., & Rose, R. J. (1985). Developmental genetic studies of adult personality.Developmental Psychology, 21,547–557. Prescott, C. A., Johnson, R. C., & McArdle, J. J. (1999). Chorion type as a possible influence on the results and interpretation of twin study data. Twin Research, 2,244–249. Pulkkinen, L., & Pitka¨nen, T. (1993). Continuities in aggressive behavior from childhood to adulthood.Aggressive Behavior, 19,249–263. *Rahe, R. H., Hervig, L., & Rosenman, R. H. (1978). Heritability of Type A behavior.Psychosomatic Medicine, 40,478–486. Raine, A., & Venables, P. H. (1992). Antisocial behaviour: Evolution, 525 ANTISOCIAL BEHAVIOR genetics, neuropsychology, and psychophysiology. In A. Gale & M. W. Eysenck (Eds.),Handbook of individual differences: Biological perspec- tives(pp. 287–321). Chichester, England: Wiley. Reed, T., Carmelli, D., & Rosenman, R. H. (1991). Effects of placentation on selected Type A behaviors in adult males in the National Heart, Lung, and Blood Institute (NHLBI) twin study.Behavior Genetics, 21,9–19. Reiss, D., Hetherington, E. M., Plomin, R., Howe, G. W., Simmens, S. J., Henderson, S., et al. (1995). Genetic questions for environmental stud- ies: Differential parenting and psychopathology in adolescence.Archives of General Psychiatry, 52,925–936. Ridenour, T. A., & Heath, A. C. (1997). Meta-analysis of conduct disorder heritability and tests for parameter heterogeneity across potentially me- diating factors [Abstract].Behavior Genetics, 27,603. Riemann, R. (1999, July).Multi-method measurement of personality: First results from the German observational study of adult twins.Paper presented at the meeting of the Behavior Genetics Association, Vancou- ver, British Columbia, Canada. Riese, M. L. (1999). Effects of chorion type on neonatal temperament differences in monozygotic twin pairs.Behavior Genetics, 29,87–94. Robins, L. N. (1966).Deviant children grown up.Baltimore: Williams & Wilkins. Robins, L. N., & Regier, D. A. (1991).Psychiatric disorders in America. New York: Free Press. Rosanoff, A. J., Handy, L. M., & Rosanoff, I. A. (1934). Criminality and delinquency in twins.Journal of Criminal Law & Criminology, 24, 923–934. *Rose, R. J. (1988). Genetic and environmental variance in content dimen- sions of the MMPI.Journal of Personality and Social Psychology, 55, 302–311. Rose, R. J., Uchida, I. A., & Christian, J. C. (1981). Placentation effects on cognitive resemblance of adult monozygotes. In L. Gedda, P. Parisi, & W. E. Nance (Eds.),Twin research: Vol. 3. Epidemiological and clinical studies(pp. 35–41). New York: Alan R. Liss. Rosenthal, R. (1991).Meta-analytic procedures for social research.New- bury Park, CA: Sage. *Rowe, D. C. (1983). Biometrical genetic models of self-reported delin- quent behavior: A twin study.Behavior Genetics, 13,473–489. Rowe, D. C., & Osgood, D. W. (1984). Heredity and sociological theories of delinquency: A reconsideration.American Sociological Review, 49, 526–540. Rowe, D. C., & Rodgers, J. L. (1989). Behavioral genetics, adolescent deviance, and“d”: Contributions and issues. In G. R. Adams, R. Mon- temayor, & T. P. Gullotta (Eds.),Biology of adolescent behavior and development(pp. 38–67). Newbury Park, CA: Sage. Rowe, D. C., Woulbroun, E. J., & Gulley, B. L. (1994). Peers and friends as nonshared environmental influences. In E. M. Hetherington, D. Reiss, & R. Plomin (Eds.),Separate social worlds of siblings: The impact of nonshared environment on development(pp. 159–173). Hillsdale, NJ: Erlbaum. *Rushton, J. P. (1996). Self-report delinquency and violence in adult twins. Psychiatric Genetics, 6,87–89. *Rushton, J. P., Fulker, D. W., Neale, M. C., Nias, D. K. B., & Eysenck, H. J. (1986). Altruism and aggression: The heritability of individual differences.Journal of Personality and Social Psychology, 50, 1192–1198. Rutter, M., Pickles, A., Murray, R., & Eaves, L. (2001). Testing hypotheses on specific environmental causal effects on behavior.Psychological Bulletin, 127,291–324. *Scarr, S. (1966). Genetic factors in activity motivation.Child Develop- ment, 37,663–673. Scarr, S., & Weinberg, R. A. (1978). The influence of“family background” on intellectual attainment.American Sociological Review, 43,674–692. *Schmitz, S., Cherny, S. S., Fulker, D. W., & Mrazek, D. A. (1994).Genetic and environmental influences on early childhood behavior. Behavior Genetics, 24,25–34. *Schmitz, S., Fulker, D. W., & Mrazek, D. A. (1995). Problem behavior in early and middle childhood: An initial behavior genetic analysis.Journal of Child Psychology and Psychiatry, 36,1443–1458. Schulsinger, F. (1972). Psychopathy: Heredity and environment.Interna- tional Journal of Mental Health, 1,190–206. *Seelig, K. J., & Brandon, K. O. (1997, July).Rater differences in gene– environment contributions to adolescent problem behavior.Paper pre- sented at the meeting of the Behavior Genetics Association, Toronto, Ontario, Canada. Sharma, A. R., McGue, M. K., & Benson, P. L. (1998). The psychological adjustment of United States adopted adolescents and their nonadopted siblings.Child Development, 69,791–802. *Sigvardsson, S., Cloninger, C. R., Bohman, M., & von Knorring, A. L. (1982). Predisposition to petty criminality in Swedish adoptees: III. Sex differences and validation of the male typology.Archives of General Psychiatry, 39,1248–1253. *Silberg, J. L., Erickson, M. T., Meyer, J. M., Eaves, L. M., Rutter, M. L., & Hewitt, J. K. (1994). The application of structural equation modeling to maternal ratings of twins’behavioral and emotional problems.Jour- nal of Consulting and Clinical Psychology, 62,510–521. *Silberg, J., Rutter, M., Meyer, J., Maes, H., Hewitt, J., Simonoff, E., et al. (1996). Genetic and environmental influences on the covariation be- tween hyperactivity and conduct disturbance in juvenile twins.Journal of Child Psychology and Psychiatry, 37,803–816. *Simonoff, E., Pickles, A., Hewitt, J., Silberg, J., Rutter, M., Loeber, L., et al. (1995). Multiple raters of disruptive child behavior: Using a genetic strategy to examine shared views and bias.Behavior Genetics, 25, 311–326. *Slutske, W. S., Heath, A. C., Dinwiddie, S. H., Madden, P. A. F., Bucholz, K. K., Dunne, M. P., et al. (1997). Modeling genetic and environmental influences in the etiology of conduct disorder: A study of 2,682 adult twin pairs.Journal of Abnormal Psychology, 106, 266–279. Slutske, W., Lyons, M., True, W., Eisen, S., Goldberg, J., & Tsuang, M. (1997, July).Testing a developmental taxonomy of antisocial behavior. Paper presented at the meeting of the Behavior Genetics Association, Toronto, Ontario, Canada. Smith, S. M., & Penrose, L. S. (1955). Monozygotic and dizygotic twin diagnosis.Annals of Human Genetics, 19,273–289. Sokol, D. K., Moore, C. A., Rose, R. J., Williams, C. J., Reed, T., & Christian, J. C. (1995). Intrapair differences in personality and cognitive ability among young monozygotic twins distinguished by chorion type. Behavior Genetics, 25,457–466. *Stevenson, J., & Graham, P. (1988). Behavioral deviance in 13-year-old twins: An item analysis.Journal of the American Academy of Child and Adolescent Psychiatry, 27,791–797. Stoolmiller, M. (1999). Implications of the restricted range of family environments for estimates of heritability and nonshared environment in behavior-genetic adoption studies.Psychological Bulletin, 125, 392–409. Stoolmiller, M., Patterson, G. R., & Snyder, J. (1997). Parental discipline and child antisocial behavior: A contingency-based theory and some methodological refinements.Psychological Inquiry, 8,223–229. *Taylor, J., McGue, M., Iacono, W. G., & Lykken, D. T. (2000). A behavioral genetic analysis of the relationship between the Socialization scale and self-reported delinquency.Journal of Personality, 68,29–50. *Tellegen, A., Lykken, D. T., Bouchard, T. J., Wilcox, K., Segal, N., & Rich, S. (1988). Personality similarity in twins reared apart and together. Journal of Personality and Social Psychology, 54,1031–1039. *Thapar, A., & McGuffin, P. (1996). A twin study of antisocial and neurotic symptoms in childhood.Psychological Medicine, 26, 1111–1118. 526 RHEE AND WALDMAN *Torgersen, S., Skre, I., Onstad, S., Edvardsen, J., & Kringlen, E. (1993). The psychometric–genetic structure ofDSM–III–Rpersonality disorder criteria.Journal of Personality Disorders, 7,196–213. Turkheimer, E., & Gottesman, I. I. (1996). Simulating the dynamics of genes and environment in development.Development and Psychopa- thology, 8,667–677. *van den Oord, E. J. C. G., Boomsma, D. I., & Verhulst, F. C. (1994). A study of problem behaviors in 10- to 15-year-old biologically related and unrelated international adoptees.Behavior Genetics, 24,193–205. van den Oord, E. J. C. G., Koot, H. M., Boomsma, D. I., Verhulst, F. C., & Orlebeke, J. F. (1995). A twin–singleton comparison of problem behaviour in 2–3-year-olds.Journal of Child Psychology and Psychia- try, 36,449–458. van den Oord, E. J. C. G., & Rowe, D. C. (1997). Effects of censored variables on family studies.Behavior Genetics, 27,99–112. *van den Oord, E. J. C. G., Verhulst, F. C., & Boomsma, D. I. (1996). A genetic study of maternal and paternal ratings of problem behaviors in 3-year-old twins.Journal of Abnormal Psychology, 105,349–357. Verhulst, F. C., Versluis-den Bieman, H., van der Ende, J., Berden, G. F. M. G., & Sanders-Woudstra, J. A. R. (1990). Problem behavior in international adoptees: III. Diagnosis of child psychiatric disorders. Journal of the American Academy of Child and Adolescent Psychia- try, 29,420–428. Vlietinck, R., Derom, R., Neale, M. C., Maes, H., van Loon, H., Derom, C., & Thiery, M. (1989). Genetic and environmental variation in the birth weight of twins.Behavior Genetics, 19,151–161. *Waldman, I. D., Levy, F., & Hay, D. A. (1995). Multivariate genetic analyses of the overlap amongDSM–III–Rdisruptive behavior disorder symptoms.Behavior Genetics, 25,293–294. Waldman, I. D., Levy, F., & Hay, D. A. (1997, June).Etiological valida- tion of a developmental taxonomy of antisocial behavior.Paper pre- sented at the meeting of the International Society for Research in Child and Adolescent Psychopathology, Paris, France.*Waldman, I. D., McGue, M. K., Pickens, R. W., & Svikis, D. S. (in press). Sex and cohort differences in genetic and environmental influences underlying childhood and adolescent antisocial behavior.Behavior Genetics. Walters, G. D. (1992). A meta-analysis of the gene–crime relationship. Criminology, 30,595–613. Widom, C. S., & Ames, A. (1988). Biology and female crime. In T. E. Moffitt & S. A. Mednick (Eds.),Biological contributions to crime causation(pp. 308–331). Dordrecht, the Netherlands: Martinus Nijhoff. *Willcutt, E. G., Shyu, V., Green, P., & Pennington, B. F. (1995, April).A twin study of the comorbidity of the disruptive behavior disorders of childhood.Paper presented at the annual meeting of the Society for Research in Child Development, Indianapolis, IN. *Wilson, G. D., Rust, J., & Kasriel, J. (1977). Genetic and family origins of humor preferences: A twin study.Psychological Reports, 41, 659–660. Wilson, J. Q., & Herrnstein, R. J. (1985).Crime and human nature.New York: Simon & Schuster. *Young, S. E., Stallings, M. C., Corley, R. P., Hewitt, J. K., & Fulker, D. W. (1996, June).Parent–offspring transmission of substance use, antisocial behavior, and cognitive factors in selected, adoptive, and control families.Paper presented at the meeting of the Behavior Genetics Association, Pittsburgh, PA. *Young, S. E., Stallings, M. C., Corley, R. P., Hewitt, J. K., & Fulker, D. W. (1997, July).Sibling resemblance for conduct disorder and attention deficit-hyperactivity disorder in selected, adoptive, and control families.Paper presented at the meeting of the Behavior Genetics As- sociation, Toronto, Ontario, Canada. *Zahn-Waxler, C., Schmitz, S., Fulker, D., Robinson, J., & Emde, R. (1996). Behavior problems in 5-year-old monozygotic and dizygotic twins: Genetic and environmental influences, patterns of regulation, and internalization of control.Development and Psychopathology, 8,103– 122. Appendix A Terms Used in PsycINFO andMedlineSearches We searched for each of the words in the left column in combination with any of the words in the right column: aggressive twin(s) aggression adoptee(s) antisocial adoptive conduct genetic psychopathy genetics sociopathy genes crime environmental criminal environment criminality delinquent delinquency behavior problem(s) problem behavior(s) Appendix B Correlations for Adoption and Twin Relationships Relationship Correlation Adoption studies Adoptee–adoptive parent 1*C Adoptee–biological parent .5*A Biological child–biological parent .5*A 1*C Adoptive siblings 1*C Biological siblings .5*A 1*C .25*D Twin studies MZ twin pairs reared together 1*A 1*C 1*D DZ twin pairs reared together .5*A 1*C .25*D MZ twin pairs reared apart 1*A 1*D DZ twin pairs reared apart .5*A .25*D Note.C shared environmental influences; A additive genetic influ- ences; D nonshared environmental influences; MZ monozygotic; DZ dizygotic. 527 ANTISOCIAL BEHAVIOR (Appendixes continue) Appendix C Example of an Mx Script for a Model Testing an ACDE Model G1: model parameters Calculation Ngroups 9 Matrics X Lower 1 1 Free ! a: additive genetic parameter Y Lower 1 1 Free ! c: shared environmental parameter Z Lower 1 1 Free ! e: unique environmental parameter W Lower 1 1 Free ! d: non-additive genetic influence parameter I Iden 22 H Full 1 1 ! scalar, .5 Q Full 1 1 ! scalar, .25 End Matrics; Matrix H .5 Matrix Q .25 Begin Algebra; A X*X ;!a∧ 2: additive genetic variance C Y*Y ;!c∧ 2: shared environmental variance E Z*Z ;!e∧ 2: unique environmental variance D W*W ;!d∧ 2: non-additive genetic variance V A C E D; ! total variance P A C E D; ! put parameter estimates in one matrix S [email protected] ; ! standardized parameter estimates End Algebra; Labels Row X parest_a Labels Row Y parest_c Labels Row Z parest_e Labels Row W parest_d Labels Row A a∧ 2 Labels Row C c∧ 2 Labels Row E e∧ 2 Labels Row D d∧ 2 Labels Row V variance Labels Row P estimate Labels ColPaced Labels Row S standest Labels Col S a∧ 2c∧ 2e∧ 2d∧ 2 End Title G2: adoptee-biological parents – Loehlin 1987 Data NInput_vars 2 NObservations 81 KMatrix Symm 1 .095 1 Matrices Group 1 Covariances A C E D [email protected] _ [email protected] A C E D/ Option Rsiduals End Title G3: adoptee-adoptive mother – Loehlin 1985 Data NInput_vars 2 NObservations 253 Kmatrix Symm 1 .02 1 Matrices Group 1 Covariances A C E D C_ C A C E D/ Option Rsiduals End Title G4: biological siblings – van den Oord 1994 Data NInput_vars NObservations 35 KMatrix Symm 1 .42 1 Matrices Group 1 Covariances A C E D [email protected] C [email protected] _ [email protected] C [email protected] A C E D/ Option Rsiduals End 528 RHEE AND WALDMAN Received August 31, 1998 Revision received November 26, 2001 Accepted November 26, 2001 Appendix C (continued) Title G5: adoptive siblings – van den Oord 1994 Data NInput_vars 2 NObservations 48 KMatrix Symm 1 .37 1 Matrices Group 1 Covariances A C E D C_ C A C E D/ Option Rsiduals End Title G6: MZ twin pairs reared together – Cates 1993 Data NInput_vars 2 NObservations 77 KMatrix Symm 1 .29 1 Matrices Group 1 Covariances A C E D A C D_ A C D A C E D/ Option Rsiduals End Title G7: DZ twin pairs reared together – Cates 1993 Data NInput_vars 2 NObservations 21 KMatrix Symm 1 .16 1 Matrices Group 1 Covariances A C E D [email protected] C [email protected] _ [email protected] C [email protected] A C E D/ Option Rsiduals End Title G8: MZ twin pairs reared apart – DiLalla 1996 Data NInput_vars 2 NObservations 66 KMatrix Symm 1 .62 1 Matrices Group 1 Covariances A C E D A D_ A D A C E D/ Option Rsiduals End Title G9: DZ twin pairs reared apart – DiLalla 1996 Data NInput_vars 2 NObservations 54 KMatrix Symm 1 .14 1 Matrices Group 1 Covariances A C E D [email protected] [email protected] _ [email protected] [email protected] A C E D/ Option Rsiduals Option NDecimals 4 Option DF 15 Option Iterations 200 Option Check End 529 ANTISOCIAL BEHAVIOR
Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m
ARTICLE The VNTR 2 repeat inMAOAand delinquent behavior in adolescence and young adulthood: associations andMAOApromoter activity Guang Guo* ,1, Xiao-Ming Ou 2, Michael Roettger 1and Jean C Shih 2 1Department of Sociology, Carolina Center for Genome Sciences, Carolina Population Center, University of North Carolina – Chapel Hill, Chapel Hill, NC, USA; 2Department of Molecular Pharmacology and Toxicology, School of Pharmacy, University of Southern California, Los Angeles, CA, USA Genetic studies of delinquent and criminal behavior are rare in spite of the wide recognition that individuals may differ in their propensity for delinquency and criminality. Using 2524 participants in Add Health in the United States, the present study demonstrates a link between the rare 2 repeat of the 30-bp VNTR in theMAOAgene and much higher levels of self-reported serious and violent delinquency. The evidence is based on a statistical association analysis and a functional analysis ofMAOApromoter activity using two human brain-derived cell lines: neuroblastoma SH-SY5Y and human glioblastoma 1242-MG. The association analysis shows that men with a 2R report a level of serious delinquency and violent delinquency in adolescence and young adulthood that were about twice (CI: (0.21, 3.24),P¼0.025; and CI: (0.37, 2.5), P¼0.008 for serious and violent delinquency, respectively) as high as those for participants with the other variants. The results for women are similar, but weaker. In the functional analysis, the 2 repeat exhibits much lower levels of promoter activity than the 3 or 4 repeat. European Journal of Human Genetics(2008)16,626 – 634; doi:10.1038/sj.ejhg.5201999; published online 23 January 2008 Keywords:delinquency; crime; violence; MAOA; genotype; antisocial behavior Introduction Studies that investigate the connections between genetic variants and delinquent and criminal behavior in humans have been rare in spite of the wide recognition that individuals may differ in the propensity to commit serious delinquent and criminal acts. 1–3 TheMAOAgene has been a focus in the investigation of aggression in animals and violent behavior in humans. Monoamine oxidase A (MAOA) is one major enzyme thatcatalyzes the oxidative deamination of a number of biogenic amines in the brain, including dopamine. Because of its ability to catabolize neurotransmitters,MAOAis frequently a candidate gene in the study of psychiatric diseases and behavioral traits. Evidence that implicates the MAOAgene in aggressive behavior has come from knock- out mouse models and human data. Caseset al 4and Shih and Thompson 5developed a line of mice with a targeted disruption of theMAOAgene. They observed an increase in the brain levels of dopamine, serotonin, and norepinephr- ine, and an increase in manifested aggression among men. Brunneret al 6reported mental retardation and impulsive aggression among eight men in an extended Dutch family with an uncommon sex-specific point mutation in the MAOAgene. TheMAOAgenomic sequences and promoters 7–9 were studied extensively in search of polymorphisms that might Received 9 June 2007; revised 10 December 2007; accepted 11 December 2007; published online 23 January 2008 *Correspondence: Professor G Guo, Department of Sociology, Carolina Center for Genome Sciences, Carolina Population Center, University of North Carolina – Chapel Hill, CB no. 3210, Chapel Hill, NC 27599-3210, USA. Tel:þ919 962 1246; Fax: 919 962 7568; E-mail: [email protected] European Journal of Human Genetics (2008) 16,626 – 634 &2008 Nature Publishing Group All rights reserved 1018-4813/08 $30.00 www.nature.com/ejhg be potentially associated with psychiatric disorders and behavioral traits. These basic studies led to a discovery of a 30-bp promoter region VNTR inMAOA, affecting level of transcriptional activity by Sabolet al. 10 The PCR product usually consists of five possible fragment sizes that include 2, 3, 3.5, 4, and 5 copies of the repeat sequence; the 3 and 4 repeats are much more common than the 2, 3.5, and 5 repeats in human populations. Sabolet al 10 showed that alleles with 3.5 (3.5R) or 4 (4R) copies of the repeat sequence are transcripted more efficiently than alleles with 3 (3R) or 5 (5R) copies of the repeat. Sabol and colleagues did not examine theMAOA2R. Caspiet al 11did not find a main effect ofMAOAvariants, but reported that maltreated male children in New Zealand with the 3R or 5R of the VNTR inMAOAwere more likely to engage in violent behavior than maltreated children with the 3.5R or 4R of the VNTR. Widom and Brzustowicz 12 reported a replication of these results. Habersticket al 13 failed to replicate the gene – environment interaction findings in this same Add Health data set used in this study. The present study focuses on the main effects of MAOAvariants. Meyer-Lindenberget al 14 studied the impact of theMAOAVNTR on brain structure and function with MRI in a large sample of healthy human volunteers. The study showed that the low expression variants predict differences in the size of limbic structures such as the amygdala and that men with the low expression variants exhibited increased reactivity of the left amygdale and hippocampus during the recall of aversive information. The objective of this study is twofold. The First, to investigate the association between the self-reported serious and violent delinquency and the 30-bp VNTR in theMAOAgene in a cohort of 2524 adolescents and young adults in the United States in the National Longitudinal Study of Adolescent Health (Add Health). The specific hypothesis is that theMAOA2R is associated with higher levels of delinquency. Second, to perform a functional analysis that evaluates the promoter activity in anMAOA 1.3-kb promoter-luciferase construct containing 2-, 3-, and 4-repeat sequences of the 30-bp VNTR, using two human brain-derived cell lines: neuroblastoma SH-SY5Y and hu- man glioblastoma 1242-MG. Materials and methodsSubject The data source for our analysis is the sibling subsample of 2524 participants in the National Longitudinal Study of Adolescent Health (Add Health), which started as a nationally representative sample of more than 20 000 adolescents in grades 7 – 12 in 1994 – 1995 (Wave I) in the United States. 15 Add Health is longitudinal; the respon- dents have been followed by two additional in-home interviews in 1995 – 1996 (Wave II) and 2001 – 2002 (Wave III). Add Health was stratified by region, ethnic mix, size,urbanicity (urban/suburban/rural), and school type (pub- lic/private/parochial). Our analysis uses the sibling sample of Add Health because DNA measures collected at Wave III in 2002 are available only for this subset of the Add Health respondents. Table 1 provides the mean (SD), proportion, and frequency of the sample characteristics. Measures We constructed a serious delinquency scale and a violent delinquency scale using the 12 questions asked to all the Add Health respondents at Waves I – III. The questions and scaling weights used to create the scales are given in Appendix 1. These two scales are variations of a widely used type of scales in contemporary research on delin- quency and criminal behavior. 16 Our scales are closely related to the scales used by, for example, Hagan and Foster 17and Haynie 18,19 in the analysis of Add Health data and by Hannon 20in the analysis of data from the National Longitudinal Study of Youth. Our serious delinquency scale overlaps with the delinquency scale of Hagan and Foster 17 to a substantial extent. As the descriptor suggests, our violent delinquency scale focuses on an array of violent delinquent behavior that could potentially be classified as violent offenses by the criminal justice system. Table 1 Mean (SD), proportion, or frequency of sample characteristics Males Females Serious delinquency Wave I 2.42 (4.31) 1.14 (2.60) Wave II 1.64 (3.42) 0.75 (1.85) Wave III 1.20 (2.39) 0.32 (1.15) Violent delinquency Wave I 1.67 (3.11) 0.70 (1.75) Wave II 1.05 (2.35) 0.43 (1.75) Wave III 0.71 (1.66) 0.15 (0.64) Genotype Any2R 0.92% 2.39% Only 3R or 5R 42.4% 17.3% 3.5R or 4R 56.7% 80.35% Ethnicity Caucasian 56.7% 57.6% Non-Caucasian 43.2% 42.3% Age (mean; in years) Wave 1 15.6 15.5 Wave II 16.57 16.42 Wave III 21.34 21.29 Sibling type MZ pairs 92 94 DZ pairs (same sex) 184 164 Full sib pairs (same sex) 97 102 Singletons 554 504 Sample size 1200 1324 VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 627 European Journal of Human Genetics Following the delinquency literature, 17 – 19 we divided the 12 items into the nonviolent and violent types. The nonviolent delinquency includes stealing amounts larger or smaller than $50, breaking and entering, and drug selling. Violent delinquency includes serious physical fighting that resulted in injuries needing medical treat- ment, use of weapons to get something from someone, involvement of physical fighting between groups, shooting or stabbing someone, deliberately damaging property, and pulling a knife or gun on someone. The serious delin- quency scale (nonviolent and violent) is based on the entire 12 items and the violent scale is based on a subset (8) of the 12 items. The Cronbach’sa-values for the serious delinquency scale for Waves I, II, and III are 0.81, 0.79, and 0.73, respectively. For Waves I, II, and III, the Cronbach’sa- values for the violent delinquency scale are 0.75, 0.74, and 0.66, respectively. Measuring delinquency and crime is challenging. Offi- cial measures based on the police reports and the justice system have been long known to substantially under- estimate delinquency and crime, 16,21 – 23 because official measures reflect not only the behavior of offenders, but also the decisions made by the justice system. For these reasons, many criminologists have turned to self-reports in recent decades. 24,25 Self-reports are now a fundamental method of measuring criminality and seem capable of yielding reliable and valid data. 16 As with any survey of sensitive private information, reporting accuracy is a concern. To protect confidentiality, reduce nonresponses, and increase reporting accuracy, this section of the interview in Add Health was self-adminis- tered by audio-CASI (computer-assisted self-interview). The sensitive question was read to respondents by means of audio headphones. Respondents were given instructions by the computer on how to complete their answers. Self- reported rates of illegal and embarrassing behavior are higher when computer-assisted techniques, particularly self-administered techniques, are used. 26,27 DNA preparation and genotyping At Wave III, in collaboration with the Institute for Behavioral Genetics in Boulder (CO, USA), Add Health collected, extracted, and quantified DNA samples from the sibling subsample. Genomic DNA was isolated from buccal cells using a modification of published methods. 28 – 31 All of the methods employed Applied Biosystems instruments and reagents. Microsatellite and VNTR polymorphisms were performed using fluorescent primers that were analyzed on an ABI capillary electrophoresis instrument. To reduce errors, two individuals independently scored all genotyping. The additional details on DNA collection and genotyping can be found at Add Health website (Smolen and Hewitt, http://www.cpc.unc.edu/projects/addhealth/). The MAOA-uVNTR polymorphism was assayed by a modified method. 10,13 The primer sequences for the30-bp VNTR in the promoter region of theMAOAopen reading frame were: forward, 5 0ACAGCCTGACCGTGG AGAAG-3 0(fluorescently labeled); and reverse, 5 0-GAACGTG ACGCTCCATTCGGA-3 0.10 The reaction yielded five frag- ment sizes that included 291, 321, 336, 351, and 381 bp (2, 3, 3.5, 4, and 5 repeats, respectively). This analysis focuses on the effect of 2R (291)vsall the other alleles. Only 11 men possess a 2R. These 11 individuals are from 11 families with six from six pairs of DZ twins and five from another five families of full siblings. Of 31 women who possess one or two 2R alleles, six are from three pairs of MZ twins, three from three pairs of DZ twins, and 22 from 19 families of full siblings. We performed a Hardy – Weinberg equilibrium test for the 2R allele among women and obtained aw 2-value of 0.012 for one degree of freedom indicating that the equilibrium is not violated. Analytical strategies To test the associations of theMAOA30-bp VNTR polymorphisms with serious and violent delinquency, we followed a three-step analytical strategy. The first step is a contingency table analysis in which the mean scores of serious delinquency and violent delinquency across geno- types were compared within each Add Health Wave and gender (Add Health Waves refer to the initial Add Health study in 1994 and two follow-up studies in 1995 – 1996 and 2002). The second step is regression analysis. Our sample consists of twins and siblings as well as the repeated observations of the same individual over different Add Health Waves; these observations are not independent. The mixed model has long been established in the statistical literature for the analysis of data that are not indepen- dent. 32,33 The following equation describes the basic structure of the mixed models used in our analysis Delinquency jitðsÞ ¼b 0þb 1genotype jiþb 2age jitþb 3age 2 jitþb 4gender ji þb 5ethnicity jiþu j0ðsÞ þv jiþe jitðsÞ ð1Þ wherej,i, andtindex sibling pair or cluster, individual, and Add Health Waves, respectively;s¼m,d,orfindicates whether the sibling cluster or pair are MZ twins, DZ twins, or full biological siblings. The basic trajectory of serious and violent delinquency is described by age and age 2, and their parameters. The model allows the random effect at the sibling cluster level and the level of observations to vary by type of sibling cluster because the strength of the correlation in these types of sibling clusters varies con- siderably. Conditional on the three random intercepts at the level of sibling clusters and one random intercept at the individual level, the siblings and repeated measures are assumed to be independent. The models are estimated by SAS. VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 628 European Journal of Human Genetics In our third step, we used two strategies to address the potential impact of population structure. First, we adjusted for self-reported race/ethnicity in all regression analysis. Tanget al 34showed a near-perfect correspondence between the four self-reported ethnic categories (European Amer- icans, African Americans, East Asians, and Hispanics) and the categories determined by 326 microsatellite markers. As a second strategy, we applied the procedure by Allison et al 35 to test for possible population stratification. Following the idea used in the development of sibship tests of linkage and association, 36 – 38 Allisonet alreasoned that the probabilities of genotypes of siblings depended entirely on parental genotypes and that controlling for the effects of sibship would be equivalent to controlling for parental genotypes. Indexing sibships byj, individuals by k, and genotypes byi, they proposed a procedure that can be written as a mixed model Y ijk¼mþa iþb jþðabÞ ijþe ijk ð2Þ wherea i, or the effect of genotypei, is assumed to be fixed; b j, or the effect of sibshipj, is assumed to be random; and (ab) ijis an interaction term specifying the dependence of the random effect of sibship on genotype. This model is a special case of the mixed model. 32,33 Functional analysis of promoter activity The analysis of promoter activity was based upon two human brain-derived cell lines: neuroblastoma SH-SY5Y and human glioblastoma 1242-MG. TheMAOA1.3-kb promoter-luciferase construct was generated by PCR using anMAOA2-kb luciferase reporter gene construct (contain- ing 4.5 repeat) as a template. The PCR product (MluI/ HindIII)ofMAOApromoter fragment ( 1336/ 64 bp) was cloned into the polylinker site (MluI/HindIII) upstream of the luciferase gene in the pGL2-Basic vector. Site-directed mutagenesis was utilized to generate 4, 3, and 2 repeats, respectively, using theMAOA1.3-kb promoter-luciferase construct as a template (Figure 1). The three primers used for mutagenesis of VNTR sites were the following (deletednucleotides are underlined and in lowercase): 5 0- GCACCA GTACCCGCACCAGT accggcaccggcaccGAGCGCAAGGCGG AGGGCCCGCC-3 0( 1117 to 1113; for generating 4- repeat sequence; deleted 15-bp nucleotides usingMAOA 1.3-kb promoter-luciferase construct containing 4.5-repeat sequence as a template); 5 0-GCACCAGTACCCGCACCAG T accggcaccggcaccagtacccgcaccagtGAGCGCAAGGCGGAGG GCCCGCC-3 0( 1199 to 1113 bp; for generating 3-repeat sequence; deleted a 30-bp nucleotides usingMAOA1.3-kb promoter-luciferase construct containing 4-repeat se- quence as a template); and 5 0-GCACCAGTACCCGCACCA GT accggcaccggcaccagtacccgcaccagtGAGCGCAAGGCGGAG GGCCCGCC-3 0( 1231 to 1113 bp; for generating 2- repeat sequence; deleted another 30-bp nucleotides using MAOA1.3-kb promoter-luciferase construct containing 3- repeat sequence as a template). The mutated nucleotide sequences of all mutant con- structs were confirmed by DNA sequencing. Transfections in SH-SY5Y and 1242-MG cells were performed using Lipofectamine 2000 (Invitrogen). Cells were plated at a density of 5 10 5 cells/well in 6-well plates. In the following day, 0.5mgMAOApromoter-luciferase construct (for one well) was co-transfected with 20 ng of plasmid pRL-TK (the herpes simplex virus thymidine kinase promoter fused upstream to theRenillaluciferase gene, which is used as an internal control; Promega) into the cells as described previously. Controls were the pGL2-basic luciferase reporter gene vector instead ofMAOApromoter- luciferase construct. After 24 h, cells were harvested with luciferase assay lysis buffer (Promega). The cell lysates were assayed for luciferase activity using the Promega Dual Luciferase Assay system. ResultsContingency table analysis Table 2 compares the mean score of serious and violent delinquency across genotypes within each gender and each of the three Add Health Waves. The number of observa- tions and the standard deviation are also given for each mean score in parentheses. The declining trend of both serious and violent delinquency over the Waves is a reflection of the well-known age pattern of delinquent and criminal behavior. For men, the genetic variants on the single X chromosome are grouped into three categories based on the finding by Sabolet al(1998): 2R; 3R or 5R; and 3.5R or 4R. The most striking result for men is the much higher scores of those with a 2R allele than those in the other two categories. The much higher score for 2R holds for both serious and violent delinquency and is present at all Waves of Add Health. In contrast, the delinquency scores of those with 3R or 5R do not seem to differ systematically from individuals with the 3.5R or 4R. The delinquency score for 2R tends to be twice as high as those in the other two genotype categories. The basic -2072 bp -1336 bpVNTR VNTR 4.5-repeats PCR & site-directed mutagenesis 4-repeats 3-repeats 2-re peatsSpl sites Spl sites Luciferase Luciferase Figure 1 The human MAOA promoter-luciferase construct. The MAOA 1.3-kb promoter containing 2, 3, and 4 repeats were generated by PCR and site-directed mutagenesis using MAOA 2-kb promoter luciferase reporter gene vector as a template. VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 629 European Journal of Human Genetics findings remain the same whether 3.5R and 5R are included or not, because 3.5R and 5R each account for only about 1% of the samples in the data of Sabolet alas well as ours. Grouping women into genotype categories is less straightforward because each woman possesses two X chromosomes and it is unknown which of the two alleles is inactivated. The female participants were classified into three genotype categories: Any2R, those with only 3R or 5R, and those with 3.5R or 4R. For women in Table 2, the 2R stands out being associated with much higher scores for both serious and violent delinquency at Waves II and III. The serious and violent delinquency scores for the other two genotype categories do not seem to differ from one another.Because of the sibling clustering in the data, standard significance tests are not valid for these comparisons. The next section presents significance tests for the genotype effects obtained from the mixed regression models that take the correlations into consideration. These exploratory results suggest that the genotype effects may be relatively constant over ages in adolescence and young adulthood, or the trajectories of delinquency across genotypes appear to be parallel over the age range. Regression analysis The regression analysis compares serious and violent delinquency scores between the 2R genotype and all the other genotypes within each gender after adjusting for the effects of age and self-reported race/ethnicity (Table 3). The Table 2 Mean score of serious and violent delinquency scales, number of observations, SD byMAOAgenotype, Add Health Wave (age), and gender Genotype Serious delinquency (n; SD) Violent delinquency (n; SD) Wave I Wave II Wave III Wave I Wave II Wave III Age (in years) 12 – 18 13 – 19 19 – 23 12 – 18 13 – 19 19 – 23 Males 2R 5.63 (11; 9.05) 3.18 (10; 6.08) 1.59 (10; 2.65) 3.95 (11; 7.04) 2.45 (10; 5.03) 1.36 (10; 2.26) 3R or 5R 2.33 (508; 4.27) 1.61 (478; 3.24) 1.55 (379; 2.45) 1.13 (508; 3.15) 1.18 (478; 2.25) 0.74 (397; 1.84) 3.5R or 4R 2.41 (681; 4.21) 1.66 (625; 3.49) 1.19 (580; 2.33) 1.64 (681; 2.95) 1.04 (625; 2.36) 0.66 (520; 1.51) Females Any2R 1.16 (31; 2.18) 1.96 (28; 3.56) 1.03 (27; 3.76) 0.77 (31; 1.75) 1.46 (28; 3.12) 0.68 (27; 2.05) Only 3R or 5R 1.17 (227; 3.10) 0.75 (212; 1.62) 0.43 (189; 1.10) 0.78 (227; 2.16) 0.50 (212; 1.17) 0.16 (189; 0.55) Any3.5R and Any4R 1.17 (1066; 2.50) 0.75 (988; 1.82) 0.33 (853; 0.96) 0.68 (1066; 1.65) 0.42 (988; 1.13) 0.15 (853; 0.56) Table 3 Estimated association of the genetic variants in the VNTR ofMAOAwith serious and violent delinquency among adolescents and young adults a Serious delinquency Violent delinquency Males Females Males Females Model 1 Model 2 Model 3 Model 4 b a 95% CIb95% CIb95% CIb95% CI Intercept 3.65 ( 7.81, 0.51) 5.88 (3.59, 8.19) 1.35 ( 4.30, 1.59) 4.65 (3.14, 6.18) Age (in years) 0.73 (0.26, 1.20) 0.45 ( 0.714, 0.201) 0.37 (0.048, 0.70) 0.40 ( 0.570, 0.229) Age squared (in years) 0.024 ( 0.035, 0.012) 0.009 (0.002, 0.016) 0.013 ( 0.022, 0.003) 0.008 (0.004, 0.014) European AmericanFFFFFFFF African American 0.38 ( 0.054, 0.82) 0.072 ( 0.15, 0.29) 0.34 (0.038, 0.65) 0.24 (0.099, 0.383) Hispanic 0.49 (0.031, 0.96) 0.16 ( 0.080, 0.41) 0.38 (0.055, 0.71) 0.17 (0.013, 0.322) Asian 0.26 ( 0.33, 0.85) 0.12 ( 0.46, 0.21) 0.12 ( 0.52, 0.29) 0.050 ( 0.259, 0.160) Others b FFFFFFFF 2R 1.72 (0.21, 3.24)FF1.46 (0.37, 2.5)FF Others c FFFFFFFF Any2RFF0.48 ( 0.087, 1.04)FF0.29 ( 0.065, 0.65) 2 logL16 844.0 14 893.5 14 626.6 11 896.4 No. of persons 1198 1314 1198 1314 No. of measures 3221 3592 3221 3592 aEstimated regression coefficient.bReference category for 2R for men: 3.5R, 4R, 3R, or 5R.cReference category for Any2R for women: Any3.5R, Any4R, Any3R, or Any5R. Random parameters are not presented in Table 3. VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 630 European Journal of Human Genetics regression analysis has yielded findings that are consistent with those from the contingency table analysis (Table 2). The regression coefficients in mixed models can be interpreted exactly as those in the ordinary least square linear regression. For men, the 2R genotype scored 1.72 (P¼0.025) points higher than the other genotypes on the serious delinquency scale. For the violent delinquency scale, men with the 2R genotype scored 1.45 (P¼0.008) points higher than the other genotypes. Figure 2 plots the model-predicted serious delinquency over age for those with the 2R and those with 3R, 3.5R, 4R, or 5R. The 2Rs on average scored about twice as high as the non-2Rs. The effects of 2R for violent delinquency are similar to those in Figure 1 and the data not presented. Using 3.5R or 4R as the reference category in a reestimated regression model shows that the 3R or 5R genotype does not differ from the 3.5R or 4R genotype.The female effect of Any2R genotype is similar to the male effect. Women with the Any2R genotype scored 0.47 (P¼0.097) and 0.29 (P¼0.108) points higher than the other genotypes on the serious and violent delinquency scales, respectively. These regression coefficients represent a large increase in the serious and violent delinquency scores among women. To address potential bias from population stratification, we controlled for self-reported race/ethnicity in all regres- sion models presented in Table 3. In addition, we performed the procedure of Allisonet al 4(Equation (2)) The model includes one random effect (b j) at the sibling- cluster level, a second random effect at the individual level (e ijk), and the key interaction term [(ab) ij] between genotype and the random effect at the sibling-cluster level. In all Allison’s models we have estimated, the random interaction term is not significant indicating that within- family effects may not be sufficiently influential to generate population stratification. Promoter activity To test whether the statistical link between theMAOA VNTR 2R and delinquency has a biochemical basis, we carried out an analysis ofMAOApromoter activity by genotype. As shown in Figure 3, the transient transfection and luciferase assay reveals three levels of promoter activity for the three 30-bp nucleotide repeat sequences inMAOA in SH-SY5Y and 1242-MG cells. The 4R sequence exhibited a higher level of promoter activity than the 3R sequence. Alleles with the 3.5R or 4R sequence had been previously shown to transcribe more efficiently than those with the 3R or 5R sequence. 10Our finding is that the 2R sequence of theMAOApromoter displayed the lowest level of promoter + pGL2 Basic + MAO A promoter 1242-MG SH-SY5Y MAO A Promoter 1.3 kb 2 repeats 3 repeats 4 repeats Luc Luc Luc 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Figure 3 Relative luciferase activity. The effect of the 30-bp nucleotide repeat sequence on the human MAOA promoter activity in SH-SY5Y and 1242-MG cells. The MAOA promoter 1.3-kb luciferase constructs were transfected into either SH-SY5Y or 1242-MG cells for 24 h. Then cells were harvested and luciferase activity was determined. Controls were pGL2-Basic vector as indicated. Please note that 2-repeat sequence of MAOA promoter shows the lowest activity and 4-repeat sequence of MAOA promoter shows the highest activity. Data were the mean± SD from three independent experiments with triplicates for each experiment. The sixt-tests (three for each cell) between 2 and 3 repeats, 2 and 4 repeats, and 3 and 4 repeats are all highly significant (t46). 4 3.5 3 2.5 2 1.5 1 0.5 0 Serious Deliquency 12 13 14 15 16 17 18 19 20 21 22 23 Age No 2R 2R Figure 2 ProjectedMAOA2R and non-2R serious delinquency by age. VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 631 European Journal of Human Genetics activity. In both SH-SY5Y and 1242-MG cells, the level of promoter activity for the 2R sequence is substantially lower than that of the 3R and only a fraction of that for 4R. The extremely low level of promoter activity in the 2R sequence corresponds to the much heightened level of serious and violent delinquency for the 2R genotype. We performed a series oft-tests between the 2R, 3R, 4R, and the controls (pGL2-Basic) and among 2R, 3R, and 4R (3Rvs2R, 4Rvs2R, and 3Rvs4R). All thet-tests are highly significant, witht-ratios reaching 10 or above in most cases. The smallestt-ratio is 6. Discussion Genetic studies of serious delinquent and criminal beha- vior in non-patient or general human populations are rare. Caspiet al 11reported an interaction between maltreatment in childhood and level ofMAOAactivity for violent behavioral problems; they did not find a main effect of theMAOAgene. Chenet al 39 hypothesized the association of aggressive behavior in adolescents with both the dopamine D2 receptor gene and the dopamine transporter gene. They provided suggestive evidence from a small study of 11 adolescents diagnosed to have impulsive – aggressive violent behavior. Guoet al 40 reported the main effects of theTa qI polymorphism in theDRD2gene and the 40-bp VNTR in theDAT1gene on serious and violent delinquency among men in the same Add Health data. The present study demonstrated a link between the 2 repeat of the 30-bp VNTR in theMAOAgene and much higher levels of self-reported serious and violent delin- quency. The finding is supported by a statistical association analysis and a functional analysis ofMAOApromoter activity. The association analysis showed that men with a 2R reported a level of serious delinquency in adolescence and young adulthood that were at least twice as high as that for those with the other variants in the VNTR. A very similar finding was obtained for violent delinquency for men. Women with Any2R also reported much higher levels of serious and violent delinquency than those with other repeats, although theP-values associated with these estimates are 0.097 and 0.108, respectively. In the functional analysis, the 2 repeat exhibited the lowest level of promoter activity, that is, 25 – 30% of the promoter activity exhibited by the 4 repeat. The level of promoter activity for the 3 repeat is located between the 2 and 4 repeats, which is consistent with the previous report. 10 The excessively low promoter activity of the 2R suggests a biochemical basis for the excessively high levels of serious and violent delinquency. Although the promoter activity differs significantly between 3R and 4R, we did not find a significant difference in serious and violent delinquency between those posses- sing 3R and those possessing 4R. This result is notinconsistent with the findings reported by Caspiet al, 11 who found a higher level of violent behavior for 3R (and 5R) than 4R (and 3.5R) only among men who were maltreated in childhood. Kim-Cohenet al 41 reported a main effect of 3R against 4R as well as an interaction effect in a sample of 7-year old Caucasian boys born in England and Wales. Although our sample is large, the 2 repeat is rare. Though the findings concerning 2R for men have passed the standard tests of significance in spite of the small category of 2R, it is possible that some of our findings could be attributable to chance. For this reason, it is important that these findings are replicated in a much larger population-based study. Future replications may prove the importance of the 2R allele, but the allele cannot possibly be involved in most delinquent behavior because of its rarity, just like the rare mutation in the MAOAgene the Dutch family 6cannot explain most of the delinquency. The weaker results for women could be due to the ambiguity in defining the X-linkedMAOAVNTR genotypes for women, which is equivalent to measurement errors. The Any2R category used in the present study may include those whose 2 repeat is not fully active. A test of this hypothesis requires a much larger sample that contains a sufficient number of participants homozygous for the 2 repeat. Our measures of serious and violent delinquency are constructed in the tradition of research on delinquency and crime. Serious delinquency measures the overall delinquency including violent delinquency, but it does not include acts more typically viewed as common adolescent deviance such as lying to parents/guardians about where they had been, minor vandalism, being loud in a public place, and driving a car without its owner’s permission. Violent delinquency measures violent behavior that is typically treated as violent offense by the criminal justice system. Our findings suggest that the MAOA*2R may be more predictive of violent delinquency than nonviolent delinquency. Future work should test this hypothesis explicitly using nonviolent and violent delinquency measures. Acknowledgements This research uses data from Add Health, a program project designed by J Richard Udry, Peter S Bearman, and Kathleen Mullan Harris, and funded by the Grant P01-HD31921 from the National Institute of Child Health and Human Development, with cooperative funding from 17 other agencies (http://www.cpc.unc.edu/addhealth/con- tract.html). Special acknowledgment is due to Andrew Smolen and John K Hewitt of the Institute for Behavior Genetics, University of Colorado for DNA isolation and genotyping. We gratefully acknowl- edge grant supports from NIH, P01-HD31921 to Add Health; R03 HD042490-02 and R03 HD053385-01 to Guang Guo; and from NSF, SES-0210389 to Guang Guo. VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 632 European Journal of Human Genetics References1 Gottfredson MR, Hirschi T:A General Theory of Crime. Stanford, CA: Stanford University Press, 1990. 2 Moffitt TE: Adolescence-limited and life-course-persistent anti- social behavior – a developmental taxonomy.Psychol Rev1993; 100: 674 – 701. 3 Paternoster R, Brame R: Multiple routes to delinquency? A test of developmental and general theories of crime.Criminology1997; 35: 49 – 84. 4 Cases O, Seif I, Grimsby Jet al: Aggressive behavior and altered amounts of brain serotonin and norepinephrine in mice lacking MAOA.Science1995;268: 1763 – 1766. 5 Shih JC, Thompson RF: Monoamine oxidase in neuropsychiatry and behavior.Am J Hum Genet1999;65: 593 – 598. 6 Brunner HG, Nelen M, Breakefield XO, Ropers HH, van Oost BA: Abnormal behavior associated with a point mutation in the structural gene for monoamine oxidase A.Science1993;262: 578 – 580. 7 Zhu QS, Grimsby J, Chen K, Shih JC: Promoter organization and activity of human monoamine-oxidase (MAO) A and B genes.J Neurosci1992;12: 4437 – 4446. 8 Zhu QS, Shih JC: An extensive repeat structure down-regulates human monoamine oxidase A promoter activity independent of an initiator-like sequence.J Neurochem1997;69: 1368 – 1373. 9 Zhu QS, Chen K, Shih JC: Bidirectional promoter of human monoamine oxidase A (MAO A) controlled by transcription factor SP1.J Neurosci1994;14: 7393 – 7403. 10 Sabol SZ, Hu S, Hamer D: A functional polymorphism in the monoamine oxidase A gene promoter.Hum Genet1998;103: 273 – 279. 11 Caspi A, McClay J, Moffitt TEet al: Role of genotype in the cycle of violence in maltreated children.Science2002;297: 851 – 854. 12 Widom CS, Brzustowicz LM: MAOA and the ‘cycle of violence’: childhood abuse and neglect, MAOA genotype, and risk for violent and antisocial behavior.Biol Psychiatry2006;60: 684 – 689. 13 Haberstick BC, Lessem JM, Hopfer CJet al: Monoamine oxidase A (MAOA) and antisocial behaviors in the presence of childhood and adolescent maltreatment.Am J Med Genet B2005;135B:59– 64. 14 Meyer-Lindenberg A, Buckholtz JW, Kolachana Bet al: Neural mechanisms of genetic risk for impulsivity and violence in humans.Proc Natl Acad Sci USA2006;103: 6269 – 6274. 15 Harris KM, Florey F, Tabor J, Bearman PS, Jones J, Udry JR:The National Longitudinal Study of Adolescent Health: Research Design, Available online at http://www.cpc.unc.edu/projects/addhealth/ design, 2003, Vol 2005. 16 Thornberry TP, Krohn MD:The Self-Report Method for Measuring Delinquency and Crime: Criminal Justice 2000. Washington, DC: National Institute of Justice, 2000, Vol 4, pp 33 – 83. 17 Hagan J, Foster H: S/He’s a rebel: toward a sequential stress theory of delinquency and gendered pathways to disadvantage in emerging adulthood.Soc Forces2003;82: 53 – 86. 18 Haynie DL: Delinquent peers revisited: does network structure matter?Am J Sociol2001;106: 1013 – 1057. 19 Haynie DL: Contexts of risk? Explaining the link between girls’ pubertal development and their delinquency involvement.Soc Forces2003;82: 355 – 397. 20 Hannon L: Poverty, delinquency, and educational attainment: cumulative disadvantage or disadvantage saturation?Sociological Inquiry2003;73: 575 – 594. 21 Hood R, Sparks R:Key Issues In Criminology. Wallop, New Hampshire: BAS, 1970. 22 Murphy F, Shirly M, Witmer F: The incidence of hidden delinquency.Am J Orthopsychiatry1946;16: 686 – 696. 23 Robison SM:Can Delinquency Be Measured?. New York: Columbia University Press, 1936. 24 Hindelang MJ: Variations in sex-race-age-specific incidence rates of offending.Am Sociol Rev1981;46: 461 – 474.25 Hindelang MJ, Hirschi T, Weis JG: Correlates of delinquency – illusion of discrepancy between self-report and official measures. Am Sociol Rev1979;44: 995 – 1014. 26 Tourangeau R, Smith TW: Asking sensitive questions – the impact of data collection mode, question format, and question context. Public Opin Q1996;60: 275 – 304. 27 Wright DL, Aquilino WS, Supple AJ: A comparison of computer- assisted and paper-and-pencil self-administered questionnaires in a survey on smoking, alcohol, and drug use.Public Opin Q1998; 62: 331 – 353. 28 Freeman B, Powell J, Ball D, Hill L, Craig I, Plomin R: DNA by mail: an inexpensive and noninvasive method for collecting DNA samples from widely dispersed populations.Behav Genet1997;27: 251 – 257. 29 Lench N, Stanier P, Williamson R: Simple non-invasive method to obtain DNA for gene analysis.Lancet1988;1: 1356 – 1358. 30 Meulenbelt I, Droog S, Trommelen GJ, Boomsma DI, Slagboom PE: High-yield noninvasive human genomic DNA isolation method for genetic-studies in geographically dispersed families and populations.Am J Hum Genet1995;57: 1252 – 1254. 31 Spitz E, Moutier R, Reed Tet al: Comparative diagnoses of twin zygosity by SSLP variant analysis, questionnaire, and dermato- glyphic analysis.Behav Genet1996;26: 55 – 63. 32 Searle SR:Linear Models. New York: Wiley & Sons, 1971. 33 Searle SR, Casella G, McCulloch C:Variance Components.New York: Wiley & Sons, 1992. 34 Tang H, Quertermous T, Rodriguez Bet al: Genetic structure, self- identified race/ethnicity, and confounding in case – control association studies.Am J Hum Genet2005;76: 268 – 275. 35 Allison DB, Heo M, Kaplan N, Martin ER: Sibling-based tests of linkage and association for quantitative traits.Am J Hum Genet 1999;64: 1754 – 1764. 36 Boehnke M, Langefeld CD: Genetic association mapping based on discordant sib pairs: the discordant-alleles test.Am J Hum Genet1998;62: 950 – 961. 37 Curtis D: Use of siblings as controls in case – control association studies.Ann Hum Genet1997;61: 319 – 333. 38 Spielman RS, Ewens WJ: A sibship test for linkage in the presence of association: the sib transmission/disequilibrium test.Am J Hum Genet1998;62: 450 – 458. 39 Chen TJ, Blum K, Mathews Det al: Are dopaminergic genes involved in a predisposition to pathological aggression? Hy- pothesizing the importance of ‘super normal controls’ in psychiatricgenetic research of complex behavioral disorders. Med Hypotheses2005;65: 703 – 707. 40 Guo G, Roettgei ME, Shih JC: Contributions of the DAT1 and DRD2 genes to serious and violent delinquency among adoles- cents and young adults.Hum Genet2007;121: 125 – 136. 41 Kim-Cohen J, Caspi A, Taylor Aet al: MAOA, maltreatment, and gene – environment interaction predicting children’s mental health: new evidence and a meta-analysis.Mol Psychiatry2006; 11: 903 – 913. Appendix 1 The serious and the violent delinquency scales 1. In the past 12 months, how often did you hurt someone badly enough to need bandages or care from a doctor or nurse? a 2. In the past 12 months, how often did someone hurt you badly enough to need bandages or care from a doctor or nurse? a 3. In the past 12 months, how often did you use or threaten to use a weapon to get something from someone? a VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 633 European Journal of Human Genetics 4. In the past 12 months, how often did you take part in a fight where a group of your friends was against another group? a 5. In the last 12 months, how often did you deliberately damage property that did not belong to you? a 6. In the past 12 months, how often did you carry a handgun to school or work? a 7. In the past 12 months, how often did you steal something worth more than $50? a 8. In the past 12 months, how often did you steal something worth less than $50? a 9. In the past 12 months, how often did you go into a house or building to steal something? a 10. In the past 12 months, how often did you sell marijuana or other drugs? a 11. In the past 12 months, have you shot or stabbed someone? b 12. In the past 12 months, have you pulled a knife or gun on someone? b aFor this question, the score value on the scale is determined in the following manner: the score is coded as zero if the event did not occur in the past 12 months; the score is coded as one if the event occurred once or twice in the past 12 months; the score is coded as two if the event occurred three or four times in the past 12 months; the score is coded as three if the event occurred five or more times in the past 12 months. bFor this question, the score value on the scale is determined in the following manner: the score is coded as zero if the event did not occur in the past 12 months; the score is coded as three if the event did occur once or more during the past 12 months. In the construction of the serious delinquency scale, individuals with more than two missing responses were excluded from analysis. In the construction of the violent delinquency scale, individuals with more than one missing response were excluded from analysis. The violent scale is based upon 8 of the 12 items and they are items 1 – 6, 11, and 12. VNTR 2 repeat inMAOAand delinquent behavior G Guoet al 634 European Journal of Human Genetics
Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m
Geneticall y Informativ e Design s fo r th e Stud y of Behavioura l Developmen t Kathry n S. Lemer y an d H. Hil l Goldsmit h Universit y of W isconsin – Madison , US A Geneti c an d environmen ta l in uence s on behaviou r an d developm en t ca n be examine d by studyin g mor e tha n on e individua l withi n a family , usin g quantitativ e geneti c theor y an d behaviour al geneti c (BG ) methodolog y. Speci Ž c environmen ta l an d geneti c in uence s ca n be measure d an d effec t size s estimated , an d man y assumption s of th e methodolog y ca n be explicitl y tested . BG design s ca n identif y speci Ž c aspect s of th e environmen t tha t hav e th e greates t in uenc e on behaviour al variation , an d the y ca n pinpoin t critica l period s in whic h environmen ta l in uence s ar e mos t malleable , bot h of whic h ar e usefu l whe n designin g interventio ns . Trait s tha t ar e show n to be th e mos t heritabl e throug h traditiona l famil y resemblanc e method s ca n no w be explore d furthe r an d actua l gene s ma y be identi Ž ed , usin g ne w molecula r methods . By identifyin g speci Ž c geneti c an d environmen ta l in uence s on behaviour , an d modellin g th e structur e of thes e in uence s ove r time , we ca n rapidl y advanc e ou r understand in g of huma n developme nt . Importanc e of B ehavioura l Geneti c M ethodolog y fo r Developmenta l P sycholog y Mos t psychologic al theor y consider s th e observable , or phenotypi c level , bu t phenotyp e is a functio n of bot h gene s an d environment . Eve n longitudina l design s ar e confounde d by geneti c change s withi n an d geneti c difference s amon g individuals . On e wa y to examin e geneti c an d environ – menta l in uence s withou t activel y manipulatin g individual ’ s biolog y or experienc e is to stud y mor e tha n on e individua l withi n a family , usin g quantitativ e geneti c theor y an d behavioura l geneti c (BG ) methodology . In a geneticall y informe d design , tru e environment al effect s ma y be identi Ž ed an d studied . Environment al in uence s ar e furthe r parse d int o thos e tha t mak e individual s simila r (shared , or commo n environment ) an d Request s fo r reprint s shoul d be sen t to Kathry n S. Lemery , Psycholog y Department , 120 2 W. Johnso n Street , Universit y of Wisconsin – Madison , WI 53706 , USA ; e-mail : [email protected] . c 199 9 Th e Internationa l Societ y fo r th e Stud y of Behavioura l Developmen t INTERNATIONA L JOURNA L OF BEHAVIORA L DEVELOPMENT , 1999 , 23 (2) , 293 – 31 7 294 LEM E R Y AN D GOLDSM IT H thos e tha t mak e individual s differen t fro m on e anothe r (nonshare d environmen t) . Interestingl y, man y kinshi p studie s in th e personalit y domai n hav e demonstrat ed tha t it is share d gene s tha t accoun t fo r th e similarit y betwee n individuals , rathe r tha n th e share d environmen t (Rowe , 1994) . Finding s suc h as thes e ar e usefu l in identifyin g th e speci Ž c aspect s of th e environmen t tha t hav e th e greates t in uenc e on behaviour , in recognisin g critica l period s durin g whic h thes e phenotype s ar e mos t malleable , an d fo r designin g interventions . Clearly , a geneticall y informa – tiv e desig n is a usefu l wa y to stud y developmen t. Thi s pape r is an introductio n to methodologic al an d conceptua l frontier s in BG designs . Thi s Ž el d ha s move d beyon d simple , univariat e biometri c estimate s of proportion s of behaviour al varianc e du e to geneti c or environmen ta l in uences . Speci Ž c environmen ta l an d geneti c in uence s ca n be measure d an d effec t size s estimated . Man y assumption s of th e methodolog y ca n be teste d explicitly . Developmen ta l function s ca n be Ž t to explai n th e structur e of continuit y an d chang e acros s age . Th e differentia l heritabilit y of behaviou r ca n be examine d betwee n categorica l groups . Thi s pape r illustrate s th e powe r of BG design s to enric h th e inference s possibl e fro m studie s of developmen t. We begi n wit h a brie f explanatio n of BG logi c an d design , procee d wit h an explicatio n of multipl e regressio n an d structura l equatio n mode l Ž ttin g wit h famil y data , an d continu e by introducin g commo n longitudina l model s an d BG methodolog y fo r categorica l data . B G Logi c an d Methodolog y We provid e an outlin e of BG logi c an d methodology . Thes e issue s ar e treate d in mor e detai l in an y behavioura l geneti c textboo k (e.g . Plomin , DeFries , McClearn , & Rutter , 1997) , an d quantitativ e genetic s backgroun d is provide d in text s suc h as Falcone r (1989) . Quantitativ e genetic s theor y is a genera l accoun t of th e aetiolog y of individua l differences . It is a progressiv e theor y in tha t it lead s to ne w prediction s tha t ca n be teste d empirically , an d th e assumption s of th e theor y ca n be teste d as well . Th e basi c linea r mode l of quantitativ e genetic s parse s th e phenotypi c varianc e int o tha t du e to geneti c an d environmen ta l in uences . Symbolically , V P = V G + V E + 2Cov ( G ) ( E ) + V G E , where V P is th e phenotypi c varianc e (th e su m of th e individuals ’ square d deviation s fro m th e mean , divide d by th e numbe r of individuals) , V G is th e geneti c variance , V E is th e environment al variance , 2Cov(G)(E ) is th e covarianc e betwee n G an d E, an d V G E represent s an y nonadditiv e effec t of G an d E. Man y model s do no t estimat e th e covarianc e an d interactio n terms , assumin g tha t the y ar e insigni Ž cant . However , by considerin g th e formula , we ca n se e tha t eve n if th e covarianc e betwee n G GE NETI C ALL Y IN F OR M ATIV E DESIGN S 295 an d E is ignored , th e estimate s of th e varianc e of G an d E remai n th e same . Also , th e correlatio n betwee n G an d E add s to V P substantiall y onl y if bot h G an d E ar e large . Becaus e we canno t estimat e al l of thes e component s directl y by measurin g gene s an d environment s, we indirectl y estimat e the m fro m th e resemblanc e of relatives . If relative s ar e mor e simila r to eac h othe r tha n individual s picke d randoml y fro m th e populatio n on a particula r trait , the n ther e is covariatio n fo r tha t trait . Thi s phenotypi c covariatio n is the n parse d int o geneti c an d environmen ta l in uences , dependin g on ho w simila r individual s ar e geneticall y an d environmen tally . Geneti c effect s ca n be partitione d int o additive , dominant , an d epistati c components . Th e additiv e componen t is th e aspec t tha t is mos t typicall y estimate d an d is th e su m of th e averag e effect s of individua l allele s acros s th e genotype . Sibling s shar e 1 2 of th e additiv e geneti c varianc e becaus e the y receiv e th e sam e allele s fro m thei r parent s 1 2 of th e time . Parent-offspri ng pair s als o shar e 1 2 , half-sib s shar e 1 4 , and , of course , identica l twin s shar e al l of thei r additiv e geneti c variance . If a dominanc e componen t is present , the n th e averag e effect s of allele s do no t ad d up in a linea r fashion . Dominanc e is th e interactio n betwee n allele s at th e sam e locus . Assumin g rando m mating , childre n do no t shar e dominanc e deviatio n wit h thei r parent s becaus e the y canno t receiv e bot h allele s at on e locu s fro m th e sam e parent . Sibling s shar e 1 4 of the dominanc e varianc e becaus e the y shoul d receiv e th e sam e allele s fro m bot h parent s 1 4 of th e time . Half-sib s do no t shar e an y of thei r dominanc e deviation , becaus e the y do no t hav e bot h parent s in common . Identica l twin s shar e al l of thei r dominanc e deviation . Thus , if dominanc e is present , the n monozygoti c (MZ ) correlation s woul d be mor e tha n hal f dizygoti c (DZ ) correlations , an d th e twi n desig n woul d yiel d highe r heritabilit y estimate s tha n othe r famil y design s (e.g . parent-offspr in g studies , adoptio n studies) . In th e presenc e of dominance , siblin g correlation s woul d als o be highe r tha n parent-offspri ng correlations , bu t on e mus t als o conside r tha t th e heritabilit y of th e chil d an d adul t form s of th e trai t ma y be different . Dominanc e effect s ar e als o confounde d wit h siblin g environmen ta l effects , whic h bot h lea d to greate r siblin g tha n parent-offspri ng resemblanc e. A secon d typ e of geneti c interactio n is terme d epistasi s an d is th e interactio n of gene s at differen t loci . Ther e ar e severa l differen t kind s of epistati c interaction , tha t is , betwee n additiv e geneti c value s at differen t loci , betwee n dominanc e deviation s at differen t loci , or betwee n additiv e geneti c value s at on e locu s an d dominanc e deviation s at anothe r locus . As wit h dominance , Ž rst-degre e relative s ar e les s tha n hal f as simila r as identica l twin s in th e presenc e of epistasis . Identica l twin s shar e 100 % of thei r epistati c interactions , wherea s fraterna l twin s an d sibling s shar e 25 % of thei r additiv e additiv e interaction s betwee n tw o loci , an d 12.5 % of 296 LEM E R Y AN D GOLDSM IT H thei r additiv e additive additiv e interaction s amon g thre e loci , or additive dominanc e interactio n betwee n tw o loci . We do no t kno w ho w man y loc i or wha t type s of epistati c interaction s ar e involve d in mos t behaviour s an d disorders . Se e Cro w (1986 ) fo r detail s concernin g geneti c interaction s an d famil y designs . In sum , geneti c effect s ca n be partitione d int o additive , dominance , an d epistati c components — al l of whic h ca n mak e famil y member s simila r to on e another . Othe r tha n fo r identica l twins , ther e ar e geneti c in uence s tha t mak e famil y member s differen t fro m on e anothe r as well . In addition , ther e ar e aspect s of th e environmen t tha t mak e individual s similar , an d othe r aspect s tha t mak e the m different . Share d environment al effect s explai n similarit y betwee n relative s in additio n to tha t accounte d fo r by genetics , an d als o explai n similarit y betwee n thos e wh o ar e geneticall y unrelate d bu t reare d together . Th e nonshare d environmen t is th e remainde r of th e varianc e no t explaine d by genetic s or by share d environmen t an d include s environ – menta l in uence s tha t ar e uniqu e to eac h individua l in a famil y an d thu s creat e differences . Th e estimat e of nonshare d environmen t is ofte n confounde d wit h measureme nt error . Studie s sho w tha t th e majorit y of environmen ta l in uence s on behaviou r ac t to mak e sibling s differen t fro m eac h othe r (Plomi n & Daniels , 1987 ; Rowe , 1994) . Wha t is th e implicatio n of thi s Ž nding ? Doe s it mea n tha t globa l famil y variables , suc h as parentin g style , ar e no t importan t fo r th e developmen t of children ’ s behaviour ? No , becaus e commo n factor s suc h as parentin g styl e ca n affec t individual s differently , an d thu s ma y contribut e to individua l differences . Th e impac t of environmen ta l factor s is speci Ž c to eac h child , dependin g on his / he r individua l phenotype , suc h as temperamen t. Dun n an d Plomi n (1986 ) foun d tha t mother s als o diffe r in ho w similarl y the y trea t thei r children . Thi s effec t is in uence d by materna l ag e an d personality , education , an d siblings ’ temperament . Parent s hav e differen t relationship s wit h eac h child , an d differentia l treatmen t is associate d wit h difference s in th e qualit y of th e siblin g relationship . Heritability. Thi s mean s les s tha n th e layperso n commonl y assume s bu t th e concep t nevertheles s is importan t fo r BG . Broad-sens e heritabilit y (h B 2 ) is th e proportio n of th e phenotypi c varianc e du e to geneti c varianc e amon g individual s in a population . Narrow-sens e heritabilit y (h 2 ) is the proportio n of phenotypi c varianc e du e to additiv e geneti c variance . Becaus e heritabilit y is a proportion , estimate s of heritabilit y wil l be differen t in differen t environmen ts . Som e environmen ts ar e mor e conduciv e to th e expressio n of geneticall y in uence d behaviour s tha n others . Heritabilit y decrease s whe n th e relevan t environmen t varie s a lo t fro m individua l to individual , an d heritabilit y increase s if th e environmen t GE NETI C ALL Y IN F OR M ATIV E DESIGN S 297 is nearl y th e sam e fo r al l individuals . Additionally , heritabilit y is clearl y greate r if th e populatio n ha s greate r variabilit y in th e relevan t genes , an d conversel y bu t les s intuitively , heritabilit y decrease s if th e populatio n share s nearl y al l gene s tha t affec t a particula r phenotype . Thus , th e heritabilit y of a huma n bein g bor n wit h tw o leg s is abou t zero , becaus e nearl y al l human s shar e th e gene s tha t contro l thi s phenotype . Fro m thes e quali Ž cations , it become s eviden t tha t estimate s of heritabilit y ar e speci Ž c to th e populatio n studie d becaus e population s diffe r in thei r distributio n of relevan t gene s an d environment s. A behaviou r tha t is heritabl e is no t necessaril y presen t at birt h or unmodi Ž able . Gen e actio n is dynami c an d th e in uenc e of gene s ma y chang e throughou t th e lifespan . Thus , a behaviou r tha t is foun d to be heritabl e at a particula r ag e ma y no t be heritabl e at another . Aspect s of th e environmen t (e.g . therapeuti c intervention ) ma y modif y heritabl e behaviou r as well . Fo r behaviour al trait s mos t commonl y studied , heritabilit y is typicall y in th e rang e of 30 – 70 % (Goldsmith , 1989) . Behavioura l Geneti c D esigns . Ther e ar e thre e basi c behavioura l geneti c designs : th e twi n study , adoptio n study , an d famil y study , an d thes e design s ca n be combined . In th e twins-reared- togethe r design , heritabilit y is implie d if monozygoti c (MZ ) twi n correlation s ar e highe r tha n dizygoti c (DZ ) twi n correlation s fo r th e trai t unde r study . Ther e ar e tw o mai n assumption s of th e twi n design : (1 ) tha t twin s ar e representat iv e of th e norma l population ; an d (2 ) tha t environment al similarit y doe s no t diffe r fo r MZ pair s an d DZ pair s fo r th e trai t unde r study . Dat a on singleton s wit h th e sam e assessmen t procedur e ca n be compare d to th e twi n dat a to tes t th e Ž rs t assumption . Th e secon d assumptio n ma y be examine d by considerin g th e parents ’ belief s abou t thei r twins ’ zygosity . As man y as 30 % of parent s ca n be mistake n in thei r belief s abou t thei r twins ’ zygosity , an d informatio n give n at birt h ca n be wron g in man y case s (Goldsmith , 1991) . Dat a fro m parent s wh o believ e tha t the y hav e fraterna l twin s wh o reall y hav e identica l twin s (o r vic e versa ) ca n be compare d to dat a fro m parent s wh o correctl y believ e thei r twin s ar e identical . Ther e ar e severa l limitation s of th e twi n desig n as well . First , it doe s no t allo w fo r an estimatio n of assortativ e mating . Assortativ e matin g is th e tendenc y to mat e wit h thos e simila r (positiv e assortativ e mating ) or dissimila r (negativ e assortativ e mating ) on som e characteristic . In th e presenc e of assortativ e mating , th e geneti c similarit y of sibling s an d parent-offspri ng is greate r tha n th e 50 % expecte d by chanc e du e to th e transmissio n of th e sam e gene s fro m bot h mothe r an d father . Assortativ e matin g increase s th e similarit y of DZ twin s compare d to MZ twin s (wh o ar e alread y at th e maximu m geneti c similarity) , thu s de atin g estimate s of heritability . In th e twi n design , assortativ e matin g effect s ar e confounde d 298 LEM E R Y AN D GOLDSM IT H wit h share d environmen t effects . Second , nonadditiv e geneti c varianc e an d gene-enviro nmen t correlatio n canno t be convincingl y isolated . Again , mor e informatio n is neede d (i.e . othe r kinshi p data ) to explor e thes e ideas . Th e secon d desig n is th e adoptio n study . Ther e ar e thre e mai n correlation s to conside r in an adoptio n design . Th e Ž rs t is th e correlatio n betwee n th e birt h paren t an d th e adoptee , wh o shar e 1 2 thei r gene s an d non e of thei r environmen t. (O f course , nongeneti c effect s ca n als o be transmitte d fro m biologica l mother s vi a prenata l effects. ) Th e secon d is th e correlatio n betwee n tw o biologicall y unrelate d relatives , suc h as childre n adopte d int o th e sam e family , whic h represent s effect s du e to th e commo n environmen t. Third , is th e correlatio n betwee n geneticall y relate d individual s wh o ar e reare d together , suc h as biologica l siblings , wh o shar e 1 2 thei r gene s an d al l of thei r commo n environmen t. Th e adoptio n desig n is strengthene d by havin g a contro l grou p of nonadoptiv e familie s studie d in th e sam e way . Heritabilit y is estimate d as twic e th e differenc e betwee n th e nonadoptiv e ( 1 2 h 2 + c 2 ) an d adoptiv e (c 2 ) correlations . Th e share d environmen t is usuall y estimate d directl y fro m adoptiv e siblin g correla – tions . Th e nonshare d environmen t is indirectl y estimate d by subtraction . Ther e ar e thre e mai n assumption s of th e adoptio n design . Th e Ž rs t is th e absenc e of selectiv e placement . Selectiv e placemen t is th e tendenc y of adoptio n agencie s to matc h childre n an d familie s on som e trait , usuall y physica l characteristics . If selectiv e placemen t is present , the n geneti c an d share d environmen t estimate s ma y be in ated . However , if bot h biologica l (typicall y onl y mothers ) an d adoptiv e parent s ar e measured , selectiv e placemen t ca n be controlle d fo r in mode l Ž tting . Th e secon d mai n assumptio n is tha t adoptiv e famil y environmen ts ar e represent ativ e of genera l famil y environmen ts , an d th e thir d assumptio n is th e absenc e of assortativ e mating . Bot h of thes e assumption s als o onl y appl y if the y affec t th e behaviou r unde r study , an d ca n be controlle d fo r in mode l Ž tting . Th e adoptio n study , unlik e th e twi n study , allow s fo r estimate s of assortativ e mating . However , development al maturit y an d generationa l difference s ar e confounde d in th e adoptio n desig n whe n comparin g individual s of differen t ages . Th e las t desig n is th e famil y study . Th e famil y desig n is ver y usefu l in testin g th e generalisati on of twi n an d adoptio n designs . On e ca n asses s sibling s (wh o shar e 1 2 thei r genes) , parent-offspri ng (wh o shar e 1 2 their genes) , half-sib s (wh o shar e 1 4 thei r genes) , uncle-nephew , aunt-niece , grandparen t-grandchil d (wh o shar e 1 4 thei r genes) , an d Ž rst-cousi n (wh o share 1 8 thei r genes ) pairs . If th e trai t unde r stud y is geneticall y in uenced , the n ther e shoul d be a linea r decreas e in similarit y acros s thes e groups . Today , ther e ar e ne w opportunitie s fo r studyin g reconstitute d familie s tha t migh t includ e full- , half- , an d step-siblings . Th e Nonshare d Environmen t an d Adolescen t Developmen t (NEAD ) project , fo r example , include d MZ GE NETI C ALL Y IN F OR M ATIV E DESIGN S 299 an d DZ twins , full-sibs , full-sib s in step-families , half-sib s in step-families , an d geneticall y unrelate d sib s in step-familie s (Pike , McGuire , Hether – ington , Reiss , & Plomin , 1996) . Th e larges t drawbac k wit h th e famil y desig n is tha t it confound s share d environmen t an d geneti c similarity . Ideally , al l of thes e method s shoul d be used . Thre e practica l way s to approac h thi s idea l ar e to includ e biologica l non-twi n sibling s in twi n designs , parent s in twi n designs , an d biologica l sibling s in adoptiv e designs . Ne w Methodologi ca l an d C onceptua l P er s pective s on Geneticall y Informe d D esign s Geneti c Analysi s of Fa m il y D ata . A quic k estimat e of heritabilit y ca n be obtaine d usin g Falconer ’ s (1989 ) formul a (2 time s th e differenc e betwee n MZ an d DZ twi n correlations) . Thi s formula , writte n in term s of varianc e components , is : 2{[(100 % geneti c variance ) + (100 % share d environmen ta l variance) ] ± [(50 % geneti c variance ) + (100 % share d environmen ta l variance)] } = 2{50 % geneti c variance } = heritability. Heritability , share d environmen tal , an d nonshare d environmen ta l esti – mate s ar e obtaine d wit h thi s method . However , mode l Ž ttin g give s a mor e accurat e estimat e becaus e it take s int o accoun t informatio n abou t phenotypi c variance s an d sampl e size s in additio n to th e covarianc e informatio n use d in th e Falcone r formula . Simultaneou s equation s mode l Ž ttin g solve s som e of th e problem s wit h constrainin g parameters , obtainin g statistica l tests , an d unequa l sampl e size s tha t th e intraclas s correlationa l approac h doe s no t address . Mode l parameter s explai n th e basi s of behavioura l covariation . Model s mak e assumption s explici t an d ca n incorporat e severa l differen t familia l relationship s simultaneous ly (e.g . family , twin , an d adoption) . In mode l Ž tting , th e expecte d covariance s base d on a theoretica l mode l ar e compare d wit h th e observe d covariances . Matrice s ar e forme d tha t carr y th e informatio n abou t covariance s of measure s an d occasions , an d covariance s of twins . Geneti c an d share d environment al effect s contribut e to betwee n pai r correlations , an d nonshare d environmen t an d erro r contribut e to residua l variance , as wel l as geneti c difference s betwee n individual s wh o do no t shar e al l of thei r genes . Multipl e regressio n geneti c analysi s (‘ ‘ DF regression ’ ’ ; DeFrie s & Fulker , 1985 , 1988 ) an d structura l equatio n modellin g (Neal e & Cardon , 1992 ) ar e use d to for m a serie s of equation s tha t tes t geneti c an d environmen ta l contribution s to behaviour . Thes e approache s permi t simultaneou s testin g an d estimatio n of bot h geneti c an d share d environ – menta l in uences . Structura l equatio n model- Ž ttin g programs , suc h as LISRE L (Joresko g & Sorbom , 1989 ) or Mx (Neale , 1994) , ar e use d to obtai n paramete r estimates , whic h ar e the n teste d fo r Ž t. 300 LEM E R Y AN D GOLDSM IT H Test s of Fit . If th e mode l hold s an d is identi Ž ed , the n th e Ž t is represente d in larg e sample s as chi-squar e wit h degree s of freedo m equa l to th e numbe r of independen t value s in th e covarianc e matri x minu s th e numbe r of unknown s bein g estimated . Th e chi-squar e tes t is use d to conclud e tha t a mode l doe s no t Ž t th e data , so a smal l chi-squar e correspond s to goo d Ž t an d a larg e chi-squar e correspond s to ba d Ž t. Th e chi-squar e goodnes s of Ž t tes t is in uence d by sampl e size . Wit h too-smal l samples , a poo r Ž t ma y no t be rejected , an d wit h too-larg e samples , a goo d Ž t ma y stil l resul t in a signi Ž can t chi-square . Th e change s in goodnes s of Ž t ca n be assesse d by th e chang e in th e chi-squar e valu e fro m on e neste d mode l to th e next . (Eac h successiv e mode l constitute s a mor e restricte d for m of th e prio r model. ) A nonsigni Ž can t differenc e in th e chi-squar e value s betwee n tw o model s implie s tha t th e additiona l speci Ž catio n di d no t signi Ž cantl y reduc e th e Ž t; thus , th e new , mor e restricte d mode l is tentativel y accepte d as th e mor e parsimoniou s model . Ther e is a larg e literatur e on alternativ e test s of goodnes s of Ž t tha t ar e typicall y use d in additio n to chi-squar e (se e Tanaka , 1993 , fo r a compariso n of man y of thes e indices) . Th e DF Regressio n Approac h to G eneti c M ode l Fitting . With DF regression , on e twin ’ s measur e is predicte d fro m th e othe r twin ’ s measur e an d th e degre e of geneti c relationshi p (DeFrie s & Fulker , 1985 , 1988) . To th e exten t tha t th e behaviou r is heritable , th e co-twi n regressio n to th e populatio n mea n wil l be les s fo r individual s wit h a close r geneti c relationshi p (e.g . MZ vs . DZ twins) . Th e regressio n equatio n simulta – neousl y estimate s bot h geneti c an d share d environmen ta l effects . Th e equatio n fo r th e augmente d mode l (fo r us e wit h ful l rang e variatio n rathe r tha n extrem e groups ) follows : T 2 = B 3 T 1 + B 4 R + B 5 T 1 R + A, where T 2 is th e co-twin ’ s predicte d score , T 1 is th e othe r twin ’ s score , R is th e coef Ž cien t of relationshi p (e.g . 1 fo r MZ twins ; .5 fo r DZ twin s fo r additiv e geneti c effects) , an d A is th e regressio n constant . Th e B’ s ar e least-square s regressio n coef Ž cients , wit h B 3 providin g an estimat e of th e share d environmen ta l in uenc e an d B 5 providin g an estimat e of heritability . (B 4 allow s th e simultaneou s analysi s of dat a fro m tw o or mor e groups , suc h as MZ an d DZ twins , an d is use d to tes t th e differenc e betwee n grou p heritabilit y an d ful l rang e heritabilit y whe n analysin g dat a fro m extrem e groups , se e later. ) Alternativ e model s ca n be tested . Fo r example , if th e estimat e of th e share d environmen t is nonsigni Ž cant , the n on e ca n dro p tha t ter m an d tes t th e reduced , neste d model . To conside r nonadditiv e effect s as wel l as additiv e geneti c effects , on e ca n ad d an additiona l coef Ž cien t of relationshi p (R 2 ) in an interactio n ter m wit h T 1 , coded 1 for MZ twins , .2 5 fo r DZ twin s an d full-sibs . Th e mode l ca n als o be extende d by includin g measure s of age , sex , ethni c group , socioeconomi c status , etc . GE NETI C ALL Y IN F OR M ATIV E DESIGN S 301 Fo r example , to conside r se x effects , on e ca n ad d a ter m fo r se x (code d 1, ± 1) and sex genotyp e (se x * T 2 R) to th e equation . DF regressio n is a exibl e approac h tha t is usefu l fo r considerin g speci Ž c measure s of gene s an d environmen t in th e equatio n an d fo r considerin g whethe r or no t heritabilit y differ s fo r extrem e scorer s compare d to th e norma l range , se e later . Univariat e Structura l E quatio n Models . Univariat e model s estimat e th e heritability , commo n environmen t, an d uniqu e environmen t fo r a particula r trai t fro m pattern s of observe d covariances , obtaine d fro m individual s of variou s degree s of geneti c relationshi p (e.g . adoption , twin , an d parent-offspri ng studies) . Figur e 1 depict s a pat h diagra m of a univariat e geneti c model . Fo r a twi n study , covariance s amon g th e phenotypi c measure s woul d be compute d separatel y fo r MZ an d DZ pairs . Thi s mode l evaluate s th e effect s of additiv e geneti c in uence s (A) , commo n environmen ta l in uence s (C) , an d nonshare d environmen ta l in uence s (E) . Thi s is calle d th e univariat e AC E model . Usin g pat h analysis , laten t geneti c an d environment al in uence s on th e phenotyp e ar e represente d by unidirectiona l arrows . Correlation s ar e double-heade d arrows . C is completel y share d betwee n individual s in thi s model , an d E is completel y independen t an d noncorrelate d betwee n individuals . Becaus e F IG. 1. Univariat e pat h diagra m tha t depict s th e correlatio n betwee n twin s or siblings . Additiv e geneti c (A) , share d environmenta l (C) , an d nonshare d environmenta l (E ) in uence s on th e phenotyp e ar e illustrated . R G is th e additiv e geneti c correlatio n betwee n th e twin s or sibling s fo r th e sam e trait . Lowe r cas e h, c, an d e ar e pat h coef Ž cient s fo r additiv e genetic , share d environmental , an d nonshare d environmenta l in uences , respectively . 302 LEM E R Y AN D GOLDSM IT H thes e individual s shar e som e genes , th e laten t A in uence s ar e correlated . Fo r MZ twins , th e geneti c correlatio n equal s 1 becaus e the y shar e 100 % of thei r genes , an d fo r DZ twin s an d ful l siblings , th e geneti c correlatio n equal s .5 becaus e the y shar e on averag e 50 % of thei r genes . Thus , in Fig . 1, th e phenotyp e is du e to additiv e geneti c effects , share d environmen ta l effects , an d nonshare d environment al effects . Th e percen t varianc e explaine d by eac h pat h is it s standardise d pat h coef Ž cien t squared . To comput e th e expecte d phenotypi c correlatio n betwee n sibling s usin g pat h analysis , on e sum s th e product s of th e path s tha t connec t thei r phenotype s (Wright , 1934) . Fo r MZ twins , th e expecte d phenotypi c correlatio n equal s h 2 + c 2 . Wit h DZ twin s or siblings , it is .5 h 2 + c 2 . If geneti c source s ar e importan t fo r th e phenotypi c covarianc e betwee n tw o traits , the n cross-siblin g correlation s shoul d yiel d th e followin g pattern : MZ twins . DZ twin s an d ful l sibling s . hal f sibling s . unrelated siblings , as illustrate d in dat a by Pik e et al . (1996) , fo r example . Simila r correlation s acros s al l siblin g type s indicat e an effec t of th e share d environmen t. If share d environmen t an d geneti c source s of varianc e canno t explai n th e phenotypi c correlations , thi s indicate s an effec t of th e nonshare d environment . A Multivariat e Extensio n of th e Univariat e Model . Th e purpos e of a multivariat e desig n is to uncove r th e aetiolog y of th e phenotypi c covariance s amon g th e variables . Th e mode l illustrate s th e exten t to whic h geneti c (o r environmen tal ) effect s mediat e th e phenotypi c correla – tio n amon g th e trait s (o r amon g differen t rater s of th e sam e trait) . Th e sam e gene s ma y in uenc e bot h phenotypes , a conditio n calle d ‘ ‘ pleio – tropy ’ ’ . Th e phenotypi c correlatio n coul d als o be du e to assortativ e matin g or commo n environment . Analogou s to th e intraclas s correlatio n use d in th e univariat e case , a cros s siblin g correlatio n is th e correlatio n betwee n on e sibling ’ s scor e at tim e 1 an d th e othe r sibling ’ s scor e at tim e 2 wit h a longitudina l design , or th e correlatio n betwee n on e sibling ’ s teacher-rate d scor e an d th e othe r sibling ’ s tester-rate d score , if considerin g differen t measure s of th e sam e trai t on on e occasion . If th e correlatio n amon g tim e point s or rater s is geneticall y in uenced , the n MZ cros s siblin g correlation s wil l be highe r tha n DZ cros s siblin g correlations . Cros s correlation s ca n als o be use d in parent-offspri ng designs . Th e geneti c correlatio n betwee n trait s indicate s ho w muc h tw o or mor e trait s ar e in uence d by th e sam e genes , an d th e environment al correlatio n betwee n trait s is ho w muc h tw o trait s ar e in uence d by th e sam e environment . Multivariat e mode l Ž ttin g typicall y begin s wit h th e Cholesk y decom – positio n (Gorsuch , 1983) , whic h sometime s is use d as a nul l mode l fo r mode l comparisons . Th e Cholesk y is simpl y a mathematical ly tractabl e startin g point , no t necessaril y a mode l of grea t substantiv e interest . In th e GE NETI C ALL Y IN F OR M ATIV E DESIGN S 303 Cholesky , al l variable s loa d on th e Ž rs t laten t facto r (muc h lik e a principa l component s analysis) , al l variable s excep t th e Ž rs t on e loa d on th e secon d factor , etc . Figur e 2 portray s th e Cholesk y decompositio n in th e bivariat e case , illustratin g bot h sibling s an d th e correlation s betwee n siblin g geneti c an d commo n environmen ta l laten t factors . Th e laten t variable s in th e to p portio n of th e Ž gur e represen t th e AC E in uence s commo n to bot h phenotypes . Th e laten t variable s belo w re ec t in uence s uniqu e to phenotyp e 2. Th e path s fro m th e commo n laten t factor s to phenotyp e 1 indicat e geneti c an d environment al in uence s on phenotyp e 1. Th e commo n A, C, an d E path s to phenotyp e 2 includ e varianc e speci Ž c to th e Ž rs t phenotyp e as well . Thes e path s indicat e th e exten t to whic h geneti c an d environment al in uence s on phenotyp e 1 als o in uenc e phenotyp e 2. Th e path s fro m AC E uniqu e factor s (below ) to phenotyp e 2 F IG. 2. Ful l bivariat e geneti c mode l illustratin g bot h siblings . Th e laten t variable s abov e represen t in uence s commo n to bot h phenotypes . Th e laten t variable s belo w represen t in uence s uniqu e to phenotyp e 2. 304 LEM E R Y AN D GOLDSM IT H represen t geneti c an d environmen ta l in uence s independen t of th e commo n geneti c an d environmen ta l in uences . To estimat e th e proportio n of varianc e accounte d fo r by phenotyp e 2, on e sum s th e square d pat h coef Ž cien t fro m th e pat h leadin g fro m th e commo n facto r to th e phenotyp e 2 an d th e square d pat h coef Ž cien t leadin g fro m th e uniqu e facto r to phenotyp e 2. Th e produc t of th e geneti c pat h fro m commo n A to phenotyp e 1 an d commo n A to phenotyp e 2 estimate s th e geneti c contributio n to th e covarianc e of phenotyp e 1 an d phenotyp e 2. Th e orde r of inpu t variable s int o thi s mode l in uence s th e speci Ž c pat h coef Ž cients , bu t th e varianc e accounte d fo r an d Ž t statistic s wil l no t differ . Severa l othe r multivariat e model s ar e commonl y use d in geneticall y informe d design s an d ar e introduce d next . Figur e 3A illustrate s Marti n an d Eaves ’ (1977 ) biometri c common – factor s model , whic h is a standar d mode l in th e Ž eld . In thi s model , AC E is speci Ž ed as bot h commo n an d uniqu e to th e thre e phenotypi c measures , an d eac h ha s a direc t effec t on th e phenotype . On th e othe r hand , Fig . 3B depict s McArdl e an d Goldsmith ’ s (1990 ) psychometri c common-facto r model . Thi s mode l assume s relate d phenotype s ar e in uence d by a singl e commo n factor , F, tha t in uence s al l phenotypes , an d uniqu e factor s (ACE ) speci Ž c to eac h phenotype . First , th e phenotypi c facto r structur e of th e observe d variable s is considered . Then , th e correlate d geneti c an d correlate d environment al effect s ar e speci Ž ed throug h th e singl e genera l phenotypi c factor , F. Last , commo n an d uniqu e facto r score s ar e teste d fo r geneti c an d environmen ta l source s of in uence . Comparin g Fig . 3A wit h Fig . 3B , we se e tha t Fig . 3B is mor e restrictiv e tha n Fig . 3A becaus e th e thre e matrice s of covariance s ar e constraine d to be proportiona l to on e anothe r (whe n formin g F) wherea s in Fig . 3A the y ar e allowe d to vary . Figur e 3B is a mor e restricte d for m of Fig . 3A , an d thu s th e tw o neste d model s ar e testabl e alternatives . Mor e comple x model s includ e model s tha t estimat e GE correlation , interactio n an d nonadditiv e source s of varianc e (Neal e & Cardon , 1994) . Longitudina l M odels . Thes e allo w fo r th e examinatio n of th e consistenc y of geneti c and / or environmen ta l in uence s acros s age . Ther e ar e tw o fundamenta l approache s to analysin g longitudina l data : th e autoregress iv e an d differenc e scor e models . In th e behaviour al geneti c literature , th e autoregress iv e approac h dominates . Fro m th e autoregress iv e perspective , investigator s typicall y begi n wit h th e Cholesk y (describe d earlie r an d illustrate d in Fig . 2) , an d dro p unnecessar y path s (wit h zer o or nonsigni Ž can t pat h coef Ž cients ) unti l the y obtai n th e simplest , mos t parsimoniou s model . Anothe r approac h is to begi n wit h severa l a-prior i models , som e of whic h ar e describe d later , an d tes t thei r Ž t to th e data . FIG.3. Biometriccommon-factorsmodel(PartA)andpsychometriccommon-factormodel(PartB)illustratingoneindividualforsimplicity.A c ,C c ,andE c arecommontoallthreephenotypicmeasures,whereastheA u ,C u ,andE u latentfactorsbelowareuniquetoeachphenotype.Fisasinglecommonfactorthatinuencesallthreephenotypes. 305 306 LEM E R Y AN D GOLDSM IT H Th e autoregress iv e simple x mode l speci Ž es tha t al l stabilit y is mediate d throug h intermediat e form s of th e trait , tha t is , ther e is no effec t of relativ e standin g at tim e 1 on tim e 3 othe r tha n tha t mediate d throug h tim e 2 (Guttman , 1954 ; Jo ¨ reskog , 1970) . Thus , thi s model , depicte d in Fig . 4, woul d Ž t dat a tha t is represente d by geneti c an d environment al correlatio n matrice s wit h th e highes t value s clos e to th e principa l diagonal , an d progressive ly lowe r value s farthe r fro m th e diagonal . Th e common-facto r model , whic h Ž ts if th e correlatio n matri x indicate s a time-independ en t relationship , is anothe r a-prior i longitudina l mode l fro m th e autoregress iv e perspective . Th e common-facto r mode l implie s a singl e underlyin g sourc e of variatio n tha t explain s th e patter n of acros s tim e continuit y (James , Mulaik , & Brett , 1982) . Thi s mode l represent s bot h pleiotropi c an d longitudina l effects . Th e commo n facto r ha s equa l effect s at al l ages . Th e persistenc e of th e geneti c deviatio n acros s tim e is F IG. 4. Autoregressiv e simple x mode l illustratin g additiv e geneti c (A ) an d nonshare d environmenta l (E ) in uence s on a singl e phenotyp e acros s time . Partia l regressio n coef Ž cient s b e and b g re ec t environmenta l an d geneti c stability , respectively ; l e and l g represen t uniqu e environmenta l an d geneti c effect s at a give n age . GE NETI C ALL Y IN F OR M ATIV E DESIGN S 307 represente d by th e path s fro m th e commo n geneti c facto r to th e measureme nt occasions . Thus , ther e ar e tw o source s of geneti c variance : tha t commo n acros s ages , an d tha t uniqu e to eac h age . Common-facto r model s ar e Ž t fo r eac h of th e effect s on th e phenotype : genetic , share d environmen t, an d nonshare d environmen t. Whe n thes e thre e source s of varianc e ar e considere d at once , th e covarianc e betwee n th e geneti c an d share d environment al factor s ar e als o considered . Wit h dat a on 24 7 adoptiv e familie s an d 24 6 nonadoptiv e familie s fro m th e Colorad o Adoptio n Projec t (CAP) , Phillip s an d Fulke r (1989 ) employe d th e longitudina l common-facto r mode l to represen t cognitiv e abilit y in th e CA P childre n fro m 1 to 7 year s of ag e an d thei r parents . Th e onl y signi Ž can t sourc e of longitudina l stabilit y fo r IQ in thi s dat a wa s genetic . Estimate s of newl y introduce d geneti c varianc e tapere d of f by ag e 3, an d th e increas e in parent-chil d correlation s ove r tim e wa s foun d to be du e to an increas e in th e geneti c correlation s betwee n chil d an d adul t measure s of IQ . Ther e wa s an absenc e of selectiv e placemen t an d cultura l transmissio n effect s fro m paren t IQ to chil d IQ . Fro m th e differenc e scor e perspective , individua l chang e ca n be depicte d wit h laten t growt h curv e analysis , whic h combine s an ANOV A approac h wit h a longitudina l facto r analysi s approac h (McArdl e & Epstein , 1987) . Laten t growt h curv e model s us e correlations , variances , an d mean s to structur e th e developmen t acros s time . Repeate d measure s on th e sam e subjects , wit h th e sam e variables , an d in th e sam e unit s of measureme nt ar e needed . Laten t growt h factor s summaris e th e long – itudina l means , correlations , an d variances , an d hypothese s abou t thes e laten t variable s (e.g . DZ twin s shar e 1 2 of thei r geneti c material ) ar e speci Ž ed . Parameter s ar e the n estimated , th e Ž t of th e mode l is tested , an d th e Ž t of alternativ e model s is compared . Parameter s ar e estimate d fo r bot h th e grou p an d th e individual . Facto r scores , th e l path s depicte d in Fig . 5, ca n be use d to describ e th e similarit y betwee n an individua l curv e an d th e grou p curve . Wherea s traditiona l structura l equatio n modellin g de Ž ne s change s as independen t of prio r changes , growt h curv e model s de Ž ne chang e as dependen t on prio r change . Conside r Fig . 5, a phenotypi c pat h mode l of a laten t growt h curv e model , wit h thre e measuremen t occasions . Thi s mode l ha s on e laten t commo n variable , F, wit h path s to al l observe d variables . Th e triangl e is a uni t constan t whos e pat h permit s F to hav e a mea n of M C (nonzero). Conside r thi s pat h mode l togethe r wit h Fig . 3B , th e psychometri c commo n facto r mode l to se e ho w thi s mode l coul d be adapte d fo r a geneticall y informe d design . Thi s mode l test s whethe r or no t th e mean s an d covariance s ca n be structure d wit h a commo n laten t factor . Th e facto r loadings , th e l’ s in Fig . 5, on thi s commo n facto r represen t an individual ’ s curv e scores . (Th e resultin g parameter s ar e average s of individua l curves. ) 308 LEM E R Y AN D GOLDSM IT H An individual ’ s laten t score , F, portray s an individual ’ s similarit y to th e grou p curve , wit h a hig h scor e portrayin g hig h similarity , an d a lo w scor e portrayin g larg e differences . Figur e 5 ca n be expande d wit h multipl e laten t variables , an d intercep t term s ca n be adde d as a varianc e component . Goldsmith , McArdle , an d Thompso n (1989 ) compare d an d contraste d th e autoregres siv e approac h wit h th e differenc e scor e approac h usin g longitudina l twi n temperamen t dat a at 4 an d 7 year s of age . Fro m psychologist s’ ratings , the y identi Ž ed tw o factors , Reactivit y an d Persis – tence . Fro m th e autoregres siv e perspective , geneti c difference s in uence d variabilit y at ag e 4 an d accounte d fo r essentiall y al l of th e stability . Geneti c difference s affecte d residua l varianc e at 7 year s in Reactivit y bu t no t Persistence . Fro m th e differenc e scor e perspective , geneti c difference s wer e relativel y les s in uentia l on chang e fro m 4 to 7 years . Thus , conclusion s abou t th e geneti c architectur e of developmen t ma y diffe r by th e metho d used . Measurin g Speci Ž c Geneti c an d Environme nta l C ontribution s to Vari – abilit y ove r Time . Th e classi c BG desig n parse s th e varianc e int o tha t F IG. 5. Phenotypi c laten t growt h curv e mode l (adapte d fro m McArdl e & Epstein , 1987) . F is a commo n laten t variabl e tha t in uence s al l thre e measuremen t occasions . Mc is th e mea n of F. Th e l’ s ar e facto r loading s whic h represen t averag e curv e scores . Th e U’ s ar e laten t factor s uniqu e to eac h measuremen t occasion . GE NETI C ALL Y IN F OR M ATIV E DESIGN S 309 whic h is du e to genetic , an d tha t whic h is du e to share d an d nonshare d environmen ta l in uences . However , thi s desig n doe s no t identif y th e speci Ž c aspect s of th e environmen t (o r speci Ž c genes ) tha t produc e th e effect . Contemporar y geneticall y informe d design s includ e speci Ž c measure s of th e environmen t and / or gene s in an attemp t to identif y th e causa l aspects . Thes e measure s ca n be adde d to curren t mode l Ž tting . DF regressio n is an exampl e of a metho d in whic h speci Ž c measure s ca n be adde d to th e mode l quit e easily . Row e an d Waldma n (1993 ) extende d th e DF regressio n approac h (describe d previously ) to includ e speci Ž c measure s of th e environment . Speci Ž c environment al measure s ar e adde d to th e regressio n equatio n as interaction s betwee n th e siblings ’ phenotyp e an d th e measure , an d th e coef Ž cien t of relationshi p (e.g . 1 fo r MZ twins , .5 fo r DZ twins) , an d th e measure . On e limitatio n of thi s metho d of accountin g fo r environmen ta l contex t is th e lo w statistica l powe r associate d wit h interactio n term s in th e regressio n equation . Row e an d Waldma n (1993 ) demonstrat e tha t to achiev e hig h powe r to detec t thes e effect s wit h thi s method , on e mus t oversampl e th e extremes . Identifyin g Example s of Gene-Env ironmen t (G E ) Covarianc e an d Interaction. Behavioura l geneticist s ofte n theoris e abou t GE covariance s an d interactions , bu t empirically , thes e concept s ca n be dif Ž cul t to measur e an d ofte n ar e lef t unmeasured . GE covarianc e embrace s th e concep t of differentia l exposur e to environment s contributin g to th e developmen t of heritabl e traits . Ther e ar e thre e type s of GE covariance : passive , reactive , an d activ e (Plomin , DeFries , & Loehlin , 1977) . If a child ’ s genotyp e is correlate d wit h th e environmen t of his / he r parent s an d sibling s (wh o hav e simila r genotypes ) thi s is calle d passiv e gene-enviro nmen t covariance . A reactiv e GE covarianc e arise s fro m th e situatio n in whic h other s reac t to a particula r individua l on th e basi s of som e of th e individual ’ s inherite d characteristics . Fo r example , an inattentiv e chil d ma y be taugh t les s materia l in a les s effectiv e manne r in school . Thus , th e environmen t become s correlate d wit h genotypi c differences . Last , an activ e gene – environmen t covarianc e ca n aris e whe n an individua l seek s an environ – men t conduciv e to furthe r developin g som e of his / he r geneti c tendencies . Thus , aggressiv e youth s ma y activel y choos e to associat e wit h peer s wh o ar e als o easil y frustrate d an d pron e to attribut e hostil e inten t to benig n action s of others . Thes e friendship s woul d contribut e to furthe r develop – men t of th e aggressiv e phenotype . Clearly , man y development al pathway s migh t entai l th e presenc e of al l thre e type s of covariance . GE covariance s ma y be positiv e or negative . Caregiver s ma y provid e emotionall y labil e childre n wit h unsettle d an d unpredictabl e environment s (positiv e covariance) , or the y ma y provid e especiall y stabl e an d 310 LEM E R Y AN D GOLDSM IT H predictabl e environmen ts (negativ e covariance) . If thes e in uence s ar e no t take n int o accoun t in th e dat a analyti c model , positiv e GE covarianc e wil l increas e estimate s of geneti c an d share d environmen ta l in uences , an d negativ e GE covarianc e wil l decreas e thes e estimate s (Goldsmith , Gottesman , & Lemery , 1997) . Th e larges t portio n of th e covarianc e betwee n parenta l practice s an d chil d adjustmen t is accounte d fo r by GE correlation s (Pik e et al. , 1996 ; Reiss , 1995) . Geneticall y in uence d characteristic s of th e chil d (e.g . dif Ž cul t temperamen t) elici t speci Ž c response s fro m parents , an d thes e parenta l response s furthe r affec t th e developmen t of th e child . Patterson ’ s (1986 ) investigation s of coerciv e cycle s involvin g aggressiv e youth s an d thei r parent s provid e an exampl e of GE correlation . Initially , th e childre n ar e noncompliant , an d th e parent s ar e no t goo d disciplinarian s (i.e . th e parent s us e hars h discipline , lac k involvement , avoi d positiv e reinforce – ment , an d ar e remis s in monitoring) . Thes e initia l circumstance s se t up a situatio n in whic h th e chil d is reinforce d fo r coerciv e behaviour s suc h as whining , teasing , an d yelling . Th e proces s escalates , resultin g in increase d chil d aggression , whic h lead s to harshe r parenta l discipline . As explaine d by Goldsmit h et al . (1997) , som e behavioura l geneticist s questio n th e utilit y of th e concep t of GE correlation , claimin g tha t it canno t be distinguishe d fro m direc t geneti c effects . Th e argumen t is tha t direc t geneti c effect s als o requir e a relevan t environmen t in orde r to be expressed . However , wit h case s in whic h th e environmen t ma y be measured , th e concep t of GE covarianc e ca n be useful . GE interaction s refe r to th e sam e environmen t havin g differen t effects , dependin g on th e genotyp e of th e individual . GE interaction s ar e statistica l interactions . An exampl e woul d be a sociabl e chil d (sociabilit y bein g a geneticall y in uence d personalit y trait ) reactin g differentl y to a nove l environmen t tha n an inhibite d child . Thes e interaction s ar e dif Ž cul t to recognis e withou t identifyin g speci Ž c genotype s an d environmen ts fo r th e phenotype . On e dif Ž cult y is tha t scalin g irregularitie s ofte n mimi c interactions . Th e fe w attempt s to isolat e thes e interaction s in huma n data , usin g quantitativ e methods , hav e bee n largel y unsuccessfu l (Plomi n & Daniels , 1984) . Adoptio n studie s ar e th e mos t powerfu l quantitativ e metho d fo r detectin g GE interaction s if goo d measure s of th e birt h parent s ar e used . Fo r a heritabl e trait , th e parent ’ s (o r midparent ’ s) phenotyp e ma y be take n as an estimat e of th e child ’ s genotype , dependin g on th e heritabilit y an d th e associatio n betwee n th e childhoo d an d matur e form s of th e phenotyp e in question . Th e environmen ta l measur e ca n be , fo r example , a characteristi c of th e adoptiv e parents . A 2 2 ANOV A tabl e is se t up wit h on e variabl e bein g genotyp e an d th e othe r environmen t, wit h th e dependen t variabl e bein g som e characteristi c of th e chil d (Plomi n et al. , GE NETI C ALL Y IN F OR M ATIV E DESIGN S 311 1977) . Eac h mai n effec t indicate s an independen t effec t of genotyp e or environmen t, an d th e interactio n of th e tw o yield s th e GE interaction . Likewise , multipl e regressio n wit h famil y dat a ca n be use d to analys e continuou s data , an d logisti c regressio n ca n be use d whe n onl y th e outcom e is discontinuous . Th e varianc e explaine d previously , an d beyon d th e mai n effect s of genotyp e an d environmen t ar e measure s of GE interaction . Th e analysi s of GE interactio n in design s suc h as thos e describe d earlie r require s larg e sampl e sizes . GE interactio n effect s wil l be muc h easie r to detec t onc e gene s associate d wit h th e phenotyp e unde r stud y ar e identi Ž ed , usin g molecula r geneti c analysis . Onc e gene s associate d wit h a particula r phenotyp e ar e identi Ž ed , the n we ca n tes t whethe r thes e sam e gene s affec t th e phenotyp e in differen t environmen ts . Detectin g Heritabilit y D ifference s in E xtrem e Group s co m pare d wit h th e Norma l Range . On e metho d of testin g th e aetiolog y of extrem e score s on a continuou s variabl e is to us e DF regression , introduce d earlie r (DeFrie s & Fulker , 1985 , 1988) . Thi s metho d ha s bee n use d whe n on e membe r of eac h twi n pai r is selecte d du e to a devian t scor e (e.g . psychopathol ogy) . First , th e basi c mode l is Ž t to determin e th e exten t to whic h th e differenc e betwee n th e proban d mea n an d th e mea n of th e populatio n is heritable . Th e equatio n fo r th e basi c mode l follows : T 2 = B 1 T 1 + B 2 R + A, where T 2 is th e co-twin ’ s predicte d score , T 1 is th e proband ’ s score , R is th e coef Ž cien t of relationshi p (e.g . 1 fo r MZ twins ; .5 fo r DZ twin s fo r additiv e geneti c effects) , an d A is th e regressio n constant . Th e B’ s ar e least-square s regressio n coef Ž cients , wit h B 1 providin g an estimat e of resemblanc e independen t of zygosit y an d B 2 providin g an estimat e of grou p heritability , afte r dat a ar e transforme d to adjus t fo r covarianc e difference s betwee n th e mean s fo r MZ an d DZ probands . Second , th e augmente d mode l (describe d earlier ) is Ž t to obtai n an estimat e of ordinar y heritability . Ordinar y heritabilit y an d grou p heritabilit y ca n the n be compare d to tes t th e hypothesi s tha t th e aetiolog y of devian t or extrem e score s is differen t fro m tha t of norma l rang e variation . Wit h th e transforme d data , B 4 from th e augmente d mode l is a tes t of th e differenc e betwee n grou p heritabilit y an d ordinar y heritability . Regressio n is a goo d approac h fo r considerin g devian t group s becaus e it is les s affecte d by restrictio n of rang e tha n correlatio n (Cohe n & Cohen , 1983) . However , larg e sample s ar e neede d to hav e enoug h powe r to compar e grou p heritabilit y to norma l rang e heritability , an d difference s in heritabilit y ca n be du e to difference s in measureme nt erro r alon g th e continuu m of measuremen t. Deater-Decka rd , Reiss , Hetherington , an d Plomi n (1997 ) obtaine d mothe r an d fathe r rating s on a shortene d Chil d Behavio r Checklis t in a 312 LEM E R Y AN D GOLDSM IT H twi n an d step-famil y stud y includin g ove r 70 0 earl y adolescen t siblin g pairs . Fo r internalisin g an d externalisin g behavioura l symptoms , the y uncovere d moderat e geneti c (heritabilit y estimate s aroun d 50% ) an d modes t share d environmen ta l in uence s (estimate s aroun d 10% ) fo r th e ful l range . Grou p heritabilit y estimate s wer e lowe r tha n ordinar y heritabilit y fo r bot h externalisin g (estimate s aroun d 40% ) an d internalis – in g (estimate s aroun d 25%) , an d grou p share d environmen t estimate s wer e highe r fo r bot h internalisin g an d externalisin g (aroun d 20%) . Althoug h thes e difference s wer e no t statisticall y signi Ž cant , th e patter n of result s support s environmen t bein g mor e importan t in th e extremes . Parent-Of fsprin g (P- O ) M odels . Thes e ar e use d to stud y behaviour s tha t sho w heritabilit y durin g bot h childhoo d an d adulthood . It is dif Ž cul t to detec t geneti c effects , bu t interestin g to se e geneti c relationship s acros s man y years . In orde r to detec t a geneti c relationship , th e childhoo d an d adulthoo d behaviour s mus t bot h be heritable , an d th e childhoo d an d adulthoo d behaviour s mus t be geneticall y correlated . A geneti c correlatio n is th e exten t to whic h th e sam e gene s affec t bot h phenotypes . P- O model s typicall y includ e bot h geneti c an d cultura l transmission ; however , in biologica l families , chil d behaviou r an d paren t behaviou r coul d be geneticall y or sociall y linked . Cultura l transmissio n is th e in uenc e of paren t behaviou r on th e share d environmen t of thei r offspring . Rando m environmen ta l effect s ar e assume d no t to transmi t acros s generations . A, C, an d E in bot h th e paren t an d offsprin g generation s determin e th e phenotypes . It is importan t to us e additiona l information , suc h as establishe d heritabilitie s of eac h of th e trait s at eac h age , whe n usin g thes e model s (DeFries , Plomin , & LaBuda , 1987) . Also , th e magnitud e of th e P- O correlatio n predicte d fro m a geneti c hypothesi s is low . Heritabilitie s rangin g fro m .4 0 to .50 — wit h no share d environmen ta l component — predic t lo w parent-offspri ng correlation s rangin g fro m .2 0 to .2 5 (Gold – smith , Losoya , Bradshaw , & Campos , 1994) , an d th e predicte d P- O correlation s ar e eve n lowe r if ther e ar e differen t geneti c in uence s on th e earl y an d late r form s of th e trait . Speci Ž cally , ‘ ‘ th e regressio n of offsprin g on midparen t value s (b po ) is equa l to th e produc t of th e squar e root s of th e infant (h i ) an d adul t (h a ) heritabilitie s an d th e geneti c correlatio n (r g ) from infanc y to adulthood , or b po = h i h a r g ’ ’ (Goldsmit h et al. , 1994 , p. 258) . Wit h thi s formul a th e rang e of reasonabl e correlation s is typicall y in th e .20s , give n fe w heritabilitie s of behaviou r in childhoo d or adulthoo d ove r .50. P- O design s ar e mos t powerfu l whe n use d in th e contex t of an adoptio n or twi n design . P- O model s withi n adoptio n design s allo w separatio n of effect s of gene s an d th e share d environmen t by comparin g childre n to bot h GE NETI C ALL Y IN F OR M ATIV E DESIGN S 313 thei r biologica l an d adoptiv e parents . Th e parent-twi n desig n allow s fo r th e separatio n of geneti c an d environmen ta l in uence s as well , bot h betwee n co-twin s an d betwee n paren t an d offspring . Thus , th e share d environmen ta l effec t on th e behaviou r of th e twin s is mad e up of tw o components : environmen ta l in uence s share d wit h th e co-twin , an d cultura l transmissio n fro m th e parents . Thes e model s ca n als o allo w fo r differen t geneti c factor s operatin g in th e paren t generatio n an d chil d generation , an d the y ca n includ e path s accountin g fo r assortativ e mating , selectiv e placemen t in adoptio n designs , an d passiv e GE correlation . Koopman s an d Boomsm a (1996 ) use d a parent-twi n desig n to decompos e familia l resemblanc es in alcoho l use . Questionnair e dat a on health-relate d behaviou r wa s collecte d fro m adolescen t twin s an d thei r parents . Fo r 15 – to 16-year-old s (40 3 families) , share d environmen t accounte d fo r mos t of th e varianc e in alcoho l us e (M Z correlatio n = .80, DZ correlatio n = .8 9 fo r males , averag e P- O correlatio n = .25) . Onl y 10 % of thi s effec t ma y be du e to parenta l alcoho l us e throug h cultura l transmission . Th e resemblanc e betwee n parent s an d offsprin g coul d be geneti c or cultura l transmission ; th e model s Ž t equall y wel l fo r thi s ag e range . On th e othe r hand , fo r 17 – to 21-year-old s (80 5 families) , geneti c in uence s bes t explaine d th e varianc e in alcoho l us e (M Z correlatio n = .74 , DZ correlatio n = .6 0 fo r males , averag e P- O correlatio n = .31), both betwee n co-twin s an d betwee n parent s an d thei r children . Th e geneti c correlatio n betwee n parent s an d offsprin g wa s high , wit h estimate s rangin g fro m .6 4 to 1.00 , dependin g on th e model . Laten t Clas s Analysis . Base d on pattern s of response s to intervie w an d questionnair e items , laten t clas s analysi s identi Ž es underlyin g categorie s of peopl e (Lazarsfeld , 1950 , applie d to BG design s by Eave s et al. , 1993) . Laten t clas s analysi s is an excellen t metho d fo r considerin g whethe r a dimensiona l or categorica l approac h explain s th e dat a better . It is typicall y use d wit h categorica l dat a suc h as psychopathol og y diagnoses . Dimen – siona l variation s in personalit y coul d affec t liabilit y to psychopathol ogy , an d at th e sam e time , categorica l distinction s ca n be impose d by th e actio n of majo r gene s or majo r environmen ta l events . Ther e ar e tw o part s to executin g a geneti c laten t clas s analysis . First , laten t clas s analysi s is performe d at th e phenotypi c leve l to determin e ho w man y classe s ar e neede d to explai n th e phenotype s (Eave s et al. , 1993) . Thi s proces s is simila r to facto r analysis , onl y rathe r tha n searchin g fo r linea r associations , on e seek s underlyin g classe s of peopl e tha t woul d predic t al l of th e differen t pattern s of responses . Then , fo r eac h putativ e class , th e probabilitie s tha t an individua l in a give n clas s wil l respon d in a particula r wa y to a questio n ar e examined . Fo r psychopathol – ogy , conside r th e probabilit y tha t a perso n in a particula r clas s woul d sho w 314 LEM E R Y AN D GOLDSM IT H a symptom , an d if he / sh e does , the n comput e th e probabilit y tha t it wa s mil d versu s severe . Next , usin g maximu m likelihood , th e Ž t of differen t model s is compare d (e.g . 3 classe s vs . 4 classes) , an d th e mos t parsimoniou s mode l is retained . Thi s metho d ha s prove n usefu l fo r heterogeneo us disorder s lik e attentio n de Ž ci t hyperactivit y disorde r (ADHD ) (Eave s et al. , 1993) . Fo r ADH D data , ther e is evidenc e of dimensiona l orderin g of severit y at th e phenotypi c level , an d evidenc e fo r a majo r gen e effec t contributin g to risk . Second , on e ca n conside r whethe r or no t ther e is an associatio n betwee n twi n (o r sib ) pair s fo r membershi p in a laten t clas s (Eave s et al. , 1993) . If MZ twin s ar e mor e likel y to be in th e sam e laten t class , the n a geneti c effec t of clas s membershi p is indicated . Laten t classe s ca n be geneti c categorie s or environment al ris k factors . Variou s model s ar e explored , tha t is , no famil y resemblanc e (th e phenotypi c mode l above) , a geneti c effec t (M Z associatio n greate r tha n DZ) , an d a share d environmen ta l effec t (M Z associatio n equal s DZ) . In addition , mor e speci Ž c geneti c hypothesi s testing , suc h as singl e gen e model s wit h complet e or incomplet e penetranc e ca n be examine d wit h thi s method . Again , maximu m likelihoo d estimate s th e probabilitie s of eac h of th e models . Laten t clas s analysi s ca n be extende d to th e analysi s of ver y larg e pedigree s an d linkag e analysis . Molecula r G eneti c M ethods . On e of th e mos t excitin g frontier s in behavioura l genetic s is th e ne w molecula r method s available . Trait s tha t ar e show n to be th e mos t heritabl e throug h traditiona l famil y resemblanc e method s ca n no w be explore d furthe r an d actua l gene s ma y be identi Ž ed . Onc e a gen e is identi Ž ed , we ca n trac k th e correspondi ng protei n an d stud y ho w thi s protei n affect s behaviour , elucidatin g geneti c link s betwee n behaviours , geneti c interaction s an d correlations , an d tracin g th e aetiolog y of developmen ta l courses . Wit h th e continuin g succes s of th e Huma n Genom e Project , gene s fo r single-gen e disorder s ar e bein g foun d routinely . A gen e tha t in uence s a complex , quantitativ e trai t is calle d a quantitativ e trai t locu s (QTL ) (Gelderman , 1975) . Behavioura l phenotypes , suc h as aggression , ar e mos t likel y in uence d by multipl e genes . Curren t method s of considerin g th e geneti c aetiolog y of comple x trait s include : (1 ) examinin g phenotypi c result s of larg e crosse s of know n anima l strains ; (2 ) allele-sharin g (i.e . seekin g an associatio n betwee n particula r allele s an d a particula r phenotyp e in pair s of affecte d relative s in man y differen t families) ; (3 ) linkag e analysi s (i.e . mappin g gene s to chromosome s by examinin g whethe r or no t a DN A marke r an d a particula r allel e ar e inherite d together) ; an d (4 ) associatio n studie s (i.e . comparin g unrelate d affecte d an d unaffecte d individual s on presenc e or absenc e of a particula r allele) . Detaile d methodolog y is reviewe d by Lande r an d Schor k (1994 ) an d text s GE NETI C ALL Y IN F OR M ATIV E DESIGN S 315 suc h as Ot t (1991 ) an d McGuf Ž n, Owen , O’ Donovan , Thapar , an d Gottesma n (1994) . C onc lusion Contemporar y behavioura l geneti c design s ar e usefu l an d powerfu l fo r answerin g a variet y of researc h questions . Thi s pape r encourage s investigator s to mov e beyon d th e univariat e biometri c analysi s of th e proportio n of varianc e accounte d fo r by geneti c an d environment al effects . BG design s ar e a powerfu l wa y to stud y environment al in uence s on behaviour . Additionally , developmen ta l BG design s ar e usefu l whe n designin g interventions . The y ca n identif y speci Ž c aspect s of th e environmen t tha t hav e th e greates t in uenc e on behaviour , an d the y ca n pinpoin t critica l period s in whic h environment al in uence s ar e mos t malleable . By identifyin g speci Ž c geneti c an d environmen ta l in uence s on behaviour , an d modellin g th e structur e of thes e in uence s ove r time , we ca n rapidl y advanc e ou r understandi ng of huma n development . We no w hav e th e conceptua l an d methodologica l sophisticatio n necessary ; remain – in g barrier s primaril y involv e resource s neede d to obtai n larg e enoug h sample s an d theoreticall y interestin g measures . Manuscrip t receive d Jun e 199 8 Revise d manuscrip t receive d Septembe r 199 8 R E F E R E NC E S Cohen , J. , & Cohen , P. (1983) . Applie d multipl e regression /correlatio n analysi s fo r th e behaviora l science s (2n d ed.) . Hillsdale , NJ : Erlbaum . Crow , J.F . (1986) . Basi c concept s in population , quantitativ e, an d evolutionar y genetic s . New York : Freeman . Deater-Deckard , K. , Reiss , D. , Hetherington , E.M. , & Plomin , R. (1997) . Dimension s an d disorder s of adolescen t adjustment : A quantitativ e geneti c analysi s of unselecte d sample s an d selecte d extremes . Journa l of Chil d Psycholog y an d Psychiatr y , 38 , 515– 525. DeFries , J.C. , & Fulker , D.W . (1985) . Multipl e regressio n analysi s of twi n data . Behavior Genetics , 15 , 467– 473. DeFries , J.C. , & Fulker , D.W . (1988) . Multipl e regressio n analysi s of twi n data : Etiolog y of devian t score s versu s individua l differences . Act a Genetica e Medica e Gemellolog ia e , 37 , 205– 216. DeFries , J.C. , Plomin , R. , & LaBuda , M.C . (1987) . Geneti c stabilit y of cognitiv e develop – men t fro m childhoo d to adulthood . Development al Psycholog y , 23 , 4– 12. Dunn , J. , & Plomin , R. (1986) . Determinant s of materna l behavio r towar d three-year-ol d siblings. Britis h Journa l of Developmen ta l Psycholog y , 4 , 127– 137. Eaves , L. , Silberg , J. , Hewitt , J.K. , Meyer , J. , Rutter , M. , Simonoff , E. , Neale , M. , & Pickles , A. (1993) . Genes , personality , an d psychopathology : A laten t clas s analysi s of liabilit y to symptom s of attention-de Ž ci t hyperactivit y disorde r in twins . In R. Plomi n & G.E . McClear n (Eds) , Natur e an d nurtur e an d psycholog y . Washington , DC : America n Psychologica l Association . 316 LEM E R Y AN D GOLDSM IT H Falconer , D.S . (1989) . Introductio n to quantitativ e genetic s (3r d ed.) , Ne w York : Longman . Gelderman , H. (1975) . Investigation s on inheritanc e of quantitativ e character s in animal s by gen e markers : I. Methods . Theoretica l an d Applie d Genetic s , 46 , 319– 330. Goldsmith , H.H . (1989) . Behavior-geneti c approache s to temperament . In G.A . Kohn – stamm , J.E . Bates , & M.K . Rothbar t (Eds.) , Temperamen t in childhoo d (pp.111 – 132) . Ne w York : Wiley . Goldsmith , H.H . (1991) . A zygosit y questionnair e fo r youn g twins : A researc h note . Behavio r Genetic s , 21 , 257– 269. Goldsmith , H.H. , Gottesman , I.I. , & Lemery , K.S . (1997) . Epigeneti c approache s to developmenta l psychopathology . Developmen t an d Psychopathol og y , 9 , 365– 387. Goldsmith , H.H. , Losoya , S.H. , Bradshaw , D.L. , & Campos , J.J . (1994) . Genetic s of personality : A twi n stud y of th e Ž ve facto r mode l an d parent-offsprin g analyses . In C. Halverson , R. Martin , & G. Kohnstam m (Eds.) , Th e developin g structur e of temperamen t an d personalit y fro m infanc y to adulthoo d (pp.241 – 265) . Hillsdale , NJ : Erlbaum . Goldsmith , H.H. , McArdle , J.J. , & Thompson , B. (1989) . Longitudina l twi n analyse s of childhoo d temperament . Behavio r Genetic s , 19 , 759 – 760 . (abstract ) Gorsuch , R.L . (1983) . Facto r analysi s (2n d ed.) . Hillsdale , NJ : Erlbaum . Guttman , L. (1954) . A ne w approac h to facto r analysis : Th e radex . In P.F . Lazarsfel d (Ed.) , Mathematic al thinkin g in th e socia l science s (pp.258 – 349) . Glencoe , IL : Fre e Press . James , L.R. , Mulaik , S.A. , & Brett , J.M . (1982) . Causa l analysis : Assumptions , models , an d data . Beverle y Hills : Sage . Joreskog , K.G . (1970) . Estimatio n an d testin g of simple x models . Britis h Journa l of Mathematic al an d Statistica l Psycholog y , 23 , 121– 145. Joreskog , K.G. , & Sorbom , D. (1989) . LISRE L 7: A guid e to th e progra m an d application s (2n d ed.) . Chicago , IL : SPSS . Koopmans , J.R. , & Boomsma , D.I . (1996) . Familia l resemblance s in alcoho l use : Geneti c or cultura l transmission ? Journa l of Studie s on Alcoho l , 57 , 19– 28. Lander , E.S. , & Schork , N.J . (1994) . Geneti c dissectio n of comple x traits . Science , 265 , 2037– 2048. Lazarsfeld , P.F . (1950) . Th e logica l an d mathematica l foundation s of laten t structur e analysis . In S.A . Stouffe r et al . (Eds) , Measurem en t an d predictio n . Princeton , NJ : Princeto n Universit y Press . Martin , N.G. , & Eaves , L.J . (1977) . Th e genetica l analysi s of covarianc e structure . Heredity , 38 , 79– 95. McArdle , J.J. , & Epstein , D. (1987) . Laten t growt h curve s withi n developmenta l structura l equatio n models . Chil d Developmen t , 58 , 110– 133. McArdle , J.J. , & Goldsmith , H.H . (1990) . Alternativ e common-facto r model s fo r multi – variat e biometri c analysis . Behavio r Genetic s , 20 , 569– 608. McGuf Ž n, P. , Owen , M.J. , O’ Donovan , M.C. , Thapar , A. , & Gottesman , I.I . (1994). Seminar s in psychiatri c genetic s . London : Roya l Colleg e of Psychiatrists . Neale , M.C . (1994) . Mx : Statistica l Modelin g (2n d ed.) . Bo x 71 0 MCV , Richmond , VA 23298 : Departmen t of Psychiatry . Neale , M.C. , & Cardon , L.R . (1992) . Methodolog y fo r geneti c studie s of twin s an d familie s . Dordrecht : Kluwer . Ott , J. (1991) . Analysi s of huma n geneti c linkag e . Baltimore , MD : John s Hopkin s Universit y Press. Patterson , G.R . (1986) . Performanc e model s fo r antisocia l boys . America n Psychologis t , 41 , 432– 444. Phillips , K. , & Fulker , D.W . (1989) . Quantitativ e geneti c analysi s of longitudina l trend s in adoptio n design s wit h applicatio n to IQ in th e Colorad o Adoptio n Project . Behavior Genetics , 19 , 621– 658. GE NETI C ALL Y IN F OR M ATIV E DESIGN S 317 Pike , A. , McGuire , S. , Hetherington , E.M. , Reiss , D. , & Plomin , R. (1996) . Famil y environmen t an d adolescen t depressiv e symptom s an d antisocia l behavior : A multivariat e geneti c analysis . Development al Psycholog y , 32 , 590– 603. Plomin , R. , & Daniels , D. (1984) . Th e interactio n betwee n temperamen t an d environment : Methodologica l considerations . Merrill-Pa lme r Quarterl y , 30 , 149– 162. Plomin , R. , & Daniels , D. (1987) . Wh y ar e childre n in th e sam e famil y so differen t fro m eac h other? Behaviora l an d Brai n Science s , 10 , 1– 16. Plomin , R. , DeFries , J.C. , McClearn , G.E. , & Rutter , M. (1997) . Behaviora l Genetic s (3rd ed.) . Ne w York : Freeman . Plomin , R. , DeFries , J.C. , & Loehlin , J.L . (1977) . Genotype-environmen t interactio n an d correlatio n in th e analysi s of huma n variation . Psychologic al Bulleti n , 84 , 309– 322. Reiss , D. (1995) . Geneti c in uenc e on famil y systems : Implication s fo r development . Journa l of Marriag e an d th e Famil y , 57 , 1– 18. Rowe , D.C . (1994) . Th e limit s of famil y in uence : Genes , experience , an d behavio r . New York : Guilfor d Press . Rowe , D.C. , & Waldman , I.D . (1993) . Th e questio n ‘‘how ’’? reconsidered . In R. Plomi n & G.E . McClear n (Eds.) , Nature , nurture , an d psycholog y (pp.355 – 373) . Washington , DC : America n Psychologica l Association . Tanaka , J.S . (1993) . Multifacete d conception s of Ž t in structura l equatio n models . In K.A . Bolle n & J.S . Lon g (Eds.) , Testin g structura l equatio n model s (pp.10 – 39) . Newbur y Park , CA: Sage Wright , S. (1934) . Th e metho d of pat h coef Ž cients . Annal s of Mathemati ca l Statistic s , 5 , 161– 215.
Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m
This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright School-level genetic variation predicts school-level verbal IQ scores: Results from a sample of American middle and high schools☆ Kevin M. Beaver a,⁎, John Paul Wright b aFlorida State University, College of Criminology and Criminal Justice, 634 W. Call Street, Tallahassee, FL 32306-1127, United StatesbUniversity of Cincinnati, United States article info abstract Article history: Received 3 February 2011 Received in revised form 28 March 2011 Accepted 28 March 2011 Available online 20 April 2011 Research has consistently revealed that average IQ scores vary significantly across macro-level units, such as states and nations. The reason for this variation in IQ, however, has remained at the center of much controversy. One of the more provocative explanations is that IQ across macro-level units is the result of genetic differences, but empirical studies have yet to examine this possibility directly. The current study partially addresses this gap in the literature by examining whether average IQ scores across thirty-six schools are associated with differences in the allelic distributions of dopaminergic polymorphisms across schools. Analysis of data drawn from subjects (ages 12–19 years) participating in the National Longitudinal Study of Adolescent Health provides support in favor of this perspective, where variation in school-level IQ scores was predicted by school-level genetic variation. This association remained statistically significant even after controlling for the effects of race. © 2011 Elsevier Inc. All rights reserved. Keywords: Aggregation Dopaminergic system Polymorphisms Schools 1. Introduction Substantial variation exists across macro-level units for virtually every measurable characteristic. For example, research has revealed that indicators of wealth, measures of health, and crime rates vary significantly across neighborhoods, states, and nations (Beaver & Wright, 2011; Kanazawa, 2008; McDaniel, 2006; Pesta, McDaniel, & Bertsch, 2010). This same line of research has also documented that variation and inequality tend to be the most pronounced at the level of the nation. Statedsimply, some nations are rich and others are poor; some nations are healthy and others are not; some nations have high rates of crime and others have low rates of crime (Braithwaite, 1989). The question that has plagued researchers, however, is what accounts for such disparities. Most of the explanations that have been advanced to explain nation-level differences have focused on culture, socialization, access to resources, and other socio- environmental factors (e.g.,Diamond, 1997; Messner & Rosenfeld, 1994). Perhaps the most controversial explanation for inequality across nations was advanced inLynn and Vanhanen’s (2002) book,IQ and a Wealth of Nations. In this book Lynn and Vanhanen empirically examined the association between the average IQ of the nation and measures of wealth. The result of their analyses revealed a statistically significant association, where nations with higher average IQ scores tended to have more wealth than nations with lower IQ scores. More recently, they expanded their analyses and examined whether nation- level IQ scores were related to other measures of inequalities, such as educational level, life expectancy, and literacy rates (Lynn & Vanhanen, 2006). Their results once again indicated a statistically significant association between IQ and an assort- ment of measures of inequality. Intelligence 39 (2011) 193–197 ☆ This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Information on how to obtain the Add Health datafiles is available on the Add Health website (http://www.cpc.unc.edu/addhealth). No direct support was received from grant P01-HD31921 for this analysis. ⁎Corresponding author. Tel.: + 1 850 644 9180; fax: + 1 850 644 9614. E-mail address:[email protected](K.M. Beaver). 0160-2896/$–see front matter © 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.intell.2011.03.005 Contents lists available atScienceDirect Intelligence With evidence mounting in favor of the position that nation- level IQ scores are related to inequality across nations, the next logical question to ask is what accounts for variation in IQ across nations?Lynn and Vanhanen (2002, 2006)(see alsoHart, 2007; Rushton, 1997) advanced a very provocative and controversial claim that variation in nation-level IQ scores is produced by genetic variation across nations. Much of the evidence that they cite and discuss in relation to this claim, however, centers on heritability estimates that were generated using data at the individual level. For example,Lynn and Vanhanen (2006) describe the results of twin studies showing that IQ is about .75, meaning that about 75% of the variance in IQ is due to genetic factors. Although the evidence indicating that variation in individual-level IQ scores is due largely to genetic factors is overwhelming, the connection between individual-level IQ scores and nation-level IQ scores is not entirely clear. Heri- tability estimates are point estimates that are designed to explain variance generated fromindividualscores and thus whether these results can be extrapolated to higher levels of aggregation remains to be determined. The goal of the current study is to provide a partial test of Lynn and Vanhanen’s (2002, 2006)thesis that variation in IQ scores at the nation level is the result of genetic differences. Our analysis focuses on examining whether polymorphisms in dopaminergic genes are related to IQ scores. We employed dopaminergic genes because prior research has provided some theoretical and empirical evidence linking the dopami- nergic system, including dopaminergic polymorphisms, to cognitive abilities and IQ (Beaver, DeLisi, Vaughn, & Wright, 2010; Berman & Noble, 1995; Previc, 1999). Due to data limitations we were unable to obtain data that included IQ scores at the nation level and genetic data at the nation level. We were, however, able to locate data that included IQ scores and DNA markers that could be aggregated to the school level. In this way, we were able to test whether variation in IQ scores at the school level was associated with variation in DNA markers that were aggregated to the school level. While using data aggregated to the school level cannot be considered a definitive test of Lynn and Vanhanen’s hypotheses at the nation level, the results based on schools can be considered an initial test of their statements for two main reasons. First, schools, like nations, show tremendous variation in terms of health, wealth, crime, and even IQ (Herrnstein & Murray, 1994; Saab & Klinger, 2010; Weissberg, 2010). Second,Lynn and Vanhanen’s (2002) arguments linking IQ to various outcomes have been shown to exist at levels of aggregation other than the nation, including the state level and the county level (Beaver & Wright, 2011; Kanazawa, 2008; McDaniel, 2006; Pesta et al., 2010). It is quite likely, then, that Lynn and Vanahanen’s explanation may apply to all types of aggregate units of analysis, not just nations. We use this possibility as a springboard to provide thefirst partial test of Lynn and Vanhanen’s provocative thesis that IQ varies across nations because of variation in genetic factors. 2. Method 2.1. Sample Data for this study come from waves 1 and 3 of the National Longitudinal Study of Adolescent Health (Add Health). The Add Health is a four-wave study of a nationally representativesample of American youths who were enrolled in seventh through twelfth grades during the 1994–1995 school year (Udry, 2003). Multi-stage stratified sampling techniques were employed to select 132 middle and high schools included in the study. Students attending these schools were adminis- tered a self-report survey during a specified school day. More than 90,000 youths were included in the wave 1 in-school component of the Add Health study. A subsample of youths was then selected to be reinterviewed at their homes to gain more detailed information. Altogether, 20,745 adolescents participated in the wave 1 in-home component of the study (Harris, Florey, Tabor, Bearman, Jones, & Udry, 2003). One of the distinguishing features of the Add Health data is that at wave 3 a subsample of respondents was genotyped. To be eligible for participation in the DNA subsample, respondents had to have a sibling who was also included in the study. If they were eligible, and if they agreed to participate, then they submitted samples of their buccal cells to be genotyped. Genotyping was conducted in coordination with the Institute of Behavioral Genetics in Boulder, Colorado and Add Health. In total, more than 2500 participants were included in the DNA subsample of the study (Harris, Halpern, Smolen, & Haberstick, 2006). Thefinal analytic sample consisted of only schools where there were at least 19 students who were included in the sample. In that way, the school-level estimates, which were based on aggregated individual-level scores, were less subject to variability associated with small sample sizes. After removing schools with less than 19 students, we were left with afinal analytical sample of 1265 youths nested within 36 schools. Given that our analysis is based on a small sample size (N = 36 schools), the power to detect small-to-moderate effect sizes is compromised. Any statistically significant effects that are detected will thus be moderate-to-large in magnitude. 2.2. Measures 2.2.1. School-level IQ scores At wave 1, Add Health participants completed the Picture Vocabulary Test (PVT). The PVT is an abbreviated version of the full-length Peabody Picture Vocabulary Test (PPVT), a test used to assess verbal abilities and receptive vocabulary. The PVT measure has been used previously as a measure of IQ among researchers analyzing the Add Health data (Rowe, Jacobson, & Van den Oord, 1999). School-level IQ was esti- mated by aggregating and averaging individual PVT scores at the school level. A similar technique has been used previously to estimate county-level IQ (Beaver & Wright, 2011). Thefinal score represents the average IQ score for respondents at- tending that school. The average school-level IQ was 99.08 with a standard deviation of 7.54. 2.2.2. School-level dopamine scores To estimate school-level dopamine scores, we aggregated and averaged (at the school level) genotypic scores for three dopaminergic polymorphisms: one in the dopamine trans- porter gene (DAT1), one in the dopamine D2 receptor gene (DRD2/ANKK1), and one in the dopamine D4 receptor gene (DRD4). Detailed information about the genotyping of these polymorphisms is available elsewhere (Beaver, Vaughn, Wright, DeLisi, & Howard, 2010; Hopfer, Timberlake, Haberstick, 194K.M. Beaver, J.P. Wright / Intelligence 39 (2011) 193–197 Lessem, Ehringer, Smolen et al., 2005). We used prior research examining the link between dopaminergic genes and cognitive abilities to determine which alleles should be coded as the risk alleles (Beaver,Vaughnetal.,2010). Briefly, respondents were genotyped for a 40 base pair variable number of tandem repeats (VNTR) in the 3′untranslated region of DAT1 (SLC6A3). For this polymorphism, the 10R allele was coded as the risk allele, while the 9R allele was coded as the non-risk allele. Following the lead of prior researchers (Hopfer et al., 2005), alleles other than the 9R and 10R were removed from the analysis. The second dopaminergic polymorphism included in the current study was the DRD2/ANKK1 TaqIA polymorphism. For DRD2/ANKK1, the A1 allele was scored as the risk allele and the A2 allele was scored as the non-risk allele. Last, DRD4 has a 48 base pair VNTR located at 11p15.5 on exon III. Two groups of alleles were created: one that included the 2R, 3R, 4R, 5R, and 6R alleles and one that included the 7R, 8R, 9R, and 10R alleles. The group of alleles that contained alleles of 7R or greater were coded as the risk alleles while the other group of alleles were coded as the non-risk alleles. All of the polymorphisms were coded co- dominantly, where the value indexed the number of risk alleles that each respondent possessed (0 risk alleles, 1 risk allele, or 2 risk alleles). Following prior research indicating that the combination of genes (as opposed to each gene in isolation) tends to have the strongest and most consistent effects on human pheno- types (Belsky & Beaver, 2011; Li et al., 2010), we created an additive dopamine index (Beaver, Vaughn et al., 2010). To create this index, the scores for each of the three poly- morphisms were summed at the individual level with values ranging between zero (0) and six. We then aggregated and averaged the individual-level dopamine scores for each school. The average school-level dopamine score was 2.54, with a standard deviation of 0.28. 2.2.3. Percentage African American We included a measure of percentage African American in the analyses as a control variable. To create this measure, we aggregated and averaged scores on an individual-level self- reported race question, where 0 = white and 1 = African American. The resulting value indexed the percentage of African Americans who were attending the school. 2.3. Analytical strategy The analysis for this study was conducted in two main steps. First, our analysis was focused on the interrelationships between IQ and dopaminergic polymorphisms at the indi- vidual-level. Specifically, we estimated the means and standard deviations for IQ by each genotype and we alsoexamined the bivariate correlations between the dopaminer- gic genes and IQ scores. In addition, we examined whether IQ scores and dopaminergic scores varied across schools by calculating F-tests. The second step in the analysis was to estimate whether school-level dopamine scores predicted school-level IQ scores before and after controlling for per- centage African American. To do so, we conducted ordinary least squares (OLS) regression models. 3. Results The analysis begins byfirst estimating the association between each of the dopaminergic polymorphisms and individual-level IQ scores.Table 1presents the means, stan- dard deviations, sample sizes, and correlations for each of the genotypes. The results indicate that DAT1 and DRD2 maintain statistically significant and negative associations with IQ scores, while the effect of DRD4 on IQ is non-significant. To further explore the association between dopaminergic poly- morphisms and IQ, we employed the additive dopamine scale as a predictor of IQ scores. The results of this analysis in- dicated a statistically significant and negative association between IQ and dopamine scores, where higher scores on the dopamine index correspond to lower IQ scores (r =−.15, pb.05, two-tailed test). We continue our analysis of the individual-level data by examining whether IQ scores and dopamine scores vary signifi cantly across the 36 schools. Our aggregate-level analyses hinge on significant variation across schools in both IQ and dopamine scores, otherwise it would be akin to trying to explain a constant with a constant, a variable with a constant, or a constant with a variable. The results of the F-tests revealed that IQ scores varied significantly across schools (F = 11.227, pb.05) as do dopamine scores (F = 2.239, pb.05).Fig. 1reveals additional support that IQ scores and dopamine scores vary significantly across schools. The distributions in thisfigure reveal the scores for IQ and dopamine, respectively, across schools and clearly indicate a significant amount of dispersion for both variables. The next set of analyses examines the association between school-level dopamine scores and school-level IQ. Model 1 in Table 2shows the results of the bivariate analyses revealing a strong and statistically significant negative association be- tween dopamine scores and IQ scores (as measured with a standardized regression coefficient [i.e., Beta]). Given that allelic distributions for certain genes and IQ scores both vary across race/ethnicity, it is possible that the results would be rendered spurious by the confounding effects of race. As a result, in Model 2 we introduce the percentage of African American variable. As can be seen, even after including race in Table 1 Means, standard deviations, and correlations for IQ by dopaminergic genotypes (N = 1265). DAT1 DRD2 DRD4 9R/9R 9R/10R 10R/10R A2/A2 A1/A2 A1/A1b7R/b7Rb7R/≥7R≥7R/≥7R Mean 102.29 99.49 97.42 99.84 97.47 92.97 98.67 97.82 97.07 SD 13.05 14.22 14.56 13.88 14.53 15.79 14.54 14.10 15.21 N 56 428 781 659 502 104 814 397 54 r−.09⁎ −.13⁎ −.03 ⁎ Significant at the .05-level, two-tailed tests.195 K.M. Beaver, J.P. Wright / Intelligence 39 (2011) 193–197 the analysis, the partial correlation between school-level dopamine scores and school-level IQ scores remained large and statistically significant. Last, to examine convergence in the results generated at the individual level with those generated at the school level, we plotted predicted IQ scores across scores on the dopamine scale index. The dopamine scale indexes were z-transformed so that the individual-level analysis could be compared with the school-level analysis.Fig. 2portrays these plots and shows a high degree of convergence in the slopes and by implication the predicted values, where IQ scores decrease as the total number of risk alleles increases. 4. Discussion Research has consistently revealed that IQ and other mea- sures of cognitive abilities vary significantly across macro- level units of analysis, such as states, nations, and even schools. Although various explanations have been set forth to explain variation in IQ at the macro-level, the most contro- versial explanation is that genetic variation across macro-level units explains variation in IQ. To this point, however, empirical research had not directly examined this potential link. The current study partially addressed this gap in the literature by examining whether variation in IQ at the school level was associated with dopaminergic scores aggregated to the school-level. Analysis of data drawn from the Add Health revealed support in favor of this position, where schools that had higher dopamine scores were the same schools that had, on average, lower IQ scores. Our results also examined the association between dopa- minergic polymorphisms and IQ at the individual level.Consistent with prior research (e.g.,Beaver, DeLisi et al., 2010; Berman & Noble, 1995), the associations between dopaminergic genes and individual-level IQ scores were either small and statistically significant or non-significant. Recall, however, that the association between school-level dopamine scores and school-level IQ scores was relatively large in magnitude, which necessarily begs the question of why the effects differed so markedly. While not exhaustive we offer two potential explanations. First, given the small sample size that was employed in the school-level analysis, our statistical power to detect small-to-moderate effect sizes was severely compro- mised and detecting large effect sizes could be due, in part, to methodological and statistical artifacts. We addressed this possibility by comparing the predicted values of IQ scores at the individual- and school-levels of analysis. The results of these models converged suggesting that the significant effects at the school-level are not solely due to a methodological or statistical artifact. Second, it is well known thatfindings detected at one level of analysis cannot be extrapolated to other levels of aggregation (Piantadosi, Byar, & Green, 1988; Samuelson, 1955). This 10 8 6 4 2 0 80.00 90.00 100.00 110.00 120.00 Frequency Frequency School-Level IQ School-Level Dopamine Score 12 10 8 6 4 2 0 1.80 2.002.20 2.40 2.60 2.80 3.00 3.20 Mean = 99.08 Std. Dev. = 7.537 N = 36Mean = 2.54 Std. Dev. = 0.279 N = 36 Fig. 1.Distributions of IQ scores and dopamine scores across schools. Table 2 OLS regression models examining the association between school-level dopamine scores and school-level IQ scores. Model 1 Model 2 b SE Beta b SE Beta School-level dopamine score−12.57 4.1−.47⁎ −9.56 3.8−.35⁎ Percentage African American−10.58 3.6−.42⁎ ⁎ Significant at the .05-level, two-tailed tests. 85 90 95 100 105 110 -3 SD -2 SD -1 SD Mean +1 SD +2 SD +3 SD Dopamine Scale Scores (Z-Scores) Predicted IQ Scores Individual-Level IQ Scores School-Level IQ Scores Fig. 2.Predicted IQ scores for individuals and schools across scores on the dopamine scale. 196K.M. Beaver, J.P. Wright / Intelligence 39 (2011) 193–197 phenomenon is particularly salient in the social sciences where research often spans multiple units of analysis, but the effects can differ considerably among units of analysis (Kramer, 1983). Criminological research, for instance, consistently reveals a strong and robust association between poverty and crime rates among macrosocial units (e.g., states or neighborhoods), while the association between poverty and criminal involvement at the individual-level is weak and oftentimes non-significant. It is quite possible that this pattern also applies to genetic research, where the usual small effects of single genes detected at the individual level become much larger at higher levels of aggregation. Future research will need to explore this possibil- ity in much greater detail. To our knowledge, this is thefirst study to aggregate DNA markers to a unit of analysis higher than the individual. Moreover, this is thefirst study to our knowledge that has revealed that variation in aggregate IQ scores is associated with variation in aggregate DNA markers. These results are in line withLynn and Vanhanen’s (2002, 2006)(see alsoHart, 2007; Rushton, 1997) thesis that the average IQ of nations is the result of genetic differences across those nations. Of course, the current study used schools, not nations, as the unit of analysis, meaning that the results reported here may not generalize to other levels of aggregation, including the nation level. There is good reason to believe, however, that the association between DNA and IQ would be even stronger at the nation level in comparison with the school level. There is much more variation in both genetic markers and IQ scores cross-nationally than there is across schools. Schools in the current study were all drawn from the same country (i.e., the United States) creating more genetic homogeneity among schools than there is among nations. Given that nations can vary quite drastically in terms of the allelic distributions of certain genes (Cavalli-Sforza, Menozzi, & Piazza, 1994), it stands to reason that this increased genetic variation would be able to explain more of the variance in IQ scores. Future research is needed to address this issue more fully and examine whether the link between DNA markers and IQ scores would be detected at other levels of aggregation. The results of the current study provide some of thefirst evidence indicating that IQ scores across macro-level units are the result of genetic factors. As with all research, though, the current study is host to at least three limitations that need to be rectified in follow-up studies. First, only three dopaminergic genes were used to create the dopamine scale. Although the dopaminergic system has previously been linked to IQ (Beaver, DeLisi et al., 2010; Berman & Noble, 1995; Previc, 1999), future research would benefit by examining a broader range of genes from the dopaminergic system and other systems of genes that may be linked to IQ. Second, the data that were available only allowed for IQ and DNA to be aggregated to the school level. It would be interesting to examine what types of associations are visible at other levels of aggregation, including the neighbor- hood level, the state level, and the nation level. Third, the measure of IQ was based on scores garnered from the PVT, a test designed to assess verbal skills. Whether the results would be observed using different measures of IQ is an empirical question awaiting future research. Until these limitations are addressed, the results of the current study should be interpreted with caution. If future researchers are able to replicate thesefindings, then the results would begin to provide additional support thatcross-national inequalities may be produced, in part, by genetic variation. References Beaver, K. M., DeLisi, M., Vaughn, M. G., & Wright, J. P. (2010). Association between the A1 allele of the DRD2 gene and reduced verbal abilities in adolescence and early adulthood.Journal of Neural Transmission,117, 827–830. Beaver, K. M., Vaughn, M. G., Wright, J. P., DeLisi, M., & Howard, M. O. (2010). Three dopaminergic polymorphisms are associated with academic achievement in middle and high school.Intelligence,38, 596–604. Beaver, K. M., & Wright, J. P. (2011). The association between county-level IQ and county-level crime rates.Intelligence,39,22–26. Belsky, J., & Beaver, K. M. (2011). Cumulative-genetic plasticity, parenting, and adolescent self-regulation.Journal of Child Psychology and Psychiatry, 52, 619–626. Berman, S. M., & Noble, E. P. (1995). Reduced visuospatial performance in children with the D2 dopamine receptor A1 allele.Behavior Genetics,25, 45–58. Braithwaite, J. (1989).Crime, shame, and reintegration.New York NY: Cambridge University Press. Cavalli-Sforza, L. L., Menozzi, P., & Piazza, A. (1994).History and geography of human genes.Princeton NJ: Princeton University Press. Diamond, J. (1997).Guns, germs, and steel: The fates of human societies.New York, NY: W. W. Norton. Harris, K. M., Florey, F., Tabor, J., Bearman, P. S., Jones, J., & Udry, J. R. (2003). The national longitudinal study of adolescent health: Research design [www document].http://www.cpc.unc.edu/projects/addhealth/designURL. Harris, K. M., Halpern, C. T., Smolen, A., & Haberstick, B. C. (2006). The national longitudinal study of adolescent health (Add Health) twin data. Twin Research and Human Genetics,9, 988–997. Hart, M. H. (2007).Understanding human history: An analysis including the effects of geography and differential evolution.Augusta, GA: Washington Summit. Herrnstein, R. J., & Murray, C. (1994).The bell curve: Intelligence and class structure in American life.New York, NY: Free Press. Hopfer, C. J., Timberlake, D., Haberstick, B. C., Lessem, J. M., Ehringer, M. A., Smolen, A., et al. (2005). Genetic influences on quantity of alcohol consumed by adolescents and young adults.Drug and Alcohol Dependence,78,187–193. Kanazawa, S. (2008). IQ and the health of states.Biodemography and Social Biology,54, 200–213. Kramer, G. H. (1983). The ecological fallacy revisited: Aggregate- versus individual-levelfindings on economics and elections, and sociotropic voting.The American Political Science Review ,77,92–111. Li, S., Zhao, J. H., Luan, J., Ekelund, U., Luben, R. N., Khaw, K. -T., et al. (2010). Physical activity attenuates the genetic predisposition to obesity in 20,000 men and women from EPIC-Norfolk prospective population study.PLoS Medicine,7,1–9. Lynn, R., & Vanhanen, T. (2002).IQ and the wealth of nations.Westport: Praeger. Lynn, R., & Vanhanen, T. (2006).IQ & global inequality.Augusta, GA: Washington Summit Publishers. McDaniel, M. A. (2006). Estimating state IQ: Measurement challenges and preliminary correlates.Intelligence,34, 607–619. Messner, S. F., & Rosenfeld, R. (1994).Crime and the American dream. Belmont, CA: Wadsworth. Pesta, B. J., McDaniel, M. A., & Bertsch, S. (2010). Toward an index of well- being for thefifty U.S. states.Intelligence,38, 160–168. Piantadosi, S., Byar, D. P., & Green, S. B. (1988). The ecological fallacy. American Journal of Epidemiology,127, 893–904. Previc, F. H. (1999). Dopamine and the origins of human intelligence.Brain and Cognition,41, 299–350. Rowe, D. C., Jacobson, K. C., & Van den Oord, E. J. C. G. (1999). Genetic and environmental influences on vocabulary IQ: Parental education level as moderator.Child Development,70, 1151–1162. Rushton, J. P. (1997).Race, evolution, and behavior: A life history perspective. New Brunswick, NJ: Transaction. Saab, H., & Klinger, D. (2010). School differences in adolescent health and wellbeing: Findings from the Canadian health behaviour and school- aged children study.Social Science & Medicine,70, 850–858. Samuelson, P. A. (1955).Economics: An introductory analysis.New York, NY: McGraw-Hill. Udry, J. R. (2003).The national longitudinal study of adolescent health (Add Health), waves I and II, 1994–1996; wave III, 2001–2002 [machine- readable datafile and documentation].Chapel Hill, NC: Carolina Population Center, University of North Carolina. Weissberg, R. (2010).Bad students, not bad schools.New Brunswick, NJ: Transaction. 197 K.M. Beaver, J.P. Wright / Intelligence 39 (2011) 193–197
Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m
Journal of Personality and Social Psychology 1997. Vol. 72, No. I, 207-217 Copyright 1997 by die American Psychological Association, Inc. 0022-351*97/$3.00 Genetic and Environmental Architecture of Human Aggression Donna R. Miles and Gregory Carey University of Colorado A meta-analysis was performed on data from 24 genetically informative studies by using various personality measures of aggression. There was a strong overall genetic effect that may account for up to 50% of the variance in aggression. This effect was not attributed to methodological inadequacies in the twin or adoption designs. Age differences were important. Self-report and parental ratings showed genes and the family environment to be important in youth; the influence of genes increased but that of family environment decreased at later ages, Observational ratings of laboratory behavior found no evidence for heritability and a very strong family environment effect. Given that almost all substantive conclusions about the genetics of personality have been drawn from self or parental reports, this last rinding has obvious and important implications for both aggression research in particular and personality research. It has been fairly well established that aggression and antiso- cial behavior run in families. Researchers have either seen them as delinquency (Rowe, Rodgers, & Meseck-Bushey, 1992), criminality (Mednick, Moffit, Gabrielli, & Hutchings, 1986), conduct disorder (Jary & Stewart, 1985), or antisocial personal- ity (Cadoret, 1978). However, although similarity among family members for aggressive or antisocial behavior has been evident, the study of intact nuclear families has not been able to trace this similarity to shared genetic influences, shared familial envi- ronmental factors, or some combination of both genes and environment. The purpose of this article is to provide an overview of twin and adoption studies on the personality construct of aggression so that genetic and environmental influences may be uncon- founded and thus provide more insight into why relatives are similar. Most behavioral genetic studies have used the twin de- sign, estimating heritability and environmentality through the comparison of monozygotic (MZ) and dizygotic (DZ) twin intraclass correlations. However, results have varied from one study to another, even when the same instrument of measure has been used. We illustrate this in the present article with the Psychopathic-deviate (Pd) scale of the Minnesota Multiphasic Personality Inventory (MMPI). Gottesman (1963), reporting on 68 pairs of adolescent twins, gave correlations of .57 for MZ twins and .18 for DZ twins, which suggests a substantial genetic effect. Correlations of similar magnitude were also found by Rose on a larger study of 410 adolescent twin pairs (Rose, personal communication, July 18, 1986). However, al- Donna R. Miles and Gregory Carey, Institute for Behavioral Genetics, University of Colorado. This research was supported in part by Grant DA-05131 from Na- tional Institute of Drug Abuse and by a grant from the National Research Council. We thank two anonymous reviewers for their valuable comments. Correspondence concerning this article should be addressed to Donna R. Miles, Institute for Behavioral Genetics, Campus Box 447, University of Colorado, Boulder, Colorado 80309-0447. Electronic mail may be sent via the Internet to [email protected] though Pogue-Geile and Rose (1985) found significant genetic effects at Age 20, no significant genetic variance was detected at Age 25. Reznikoff and Honeyman (1967) also failed to find significant heritability for the Pd scale in their sample of 34 adult twin pairs. Studies of children also have had variable findings. Some studies have reported significant heritability (Lytton, Watts, & Dunn, 1988; O’Connor, Foch, Sherry, & Plomin, 1980; Scarr, 1966). Others have reported little genetic influence (Owen & Sines, 1970), whereas yet another suggested heritability was important for males but not for females (Stevenson & Graham, 1988). Adoption studies on aggression again have shown variable results. The Texas Adoption Project found modest correlations between biological mother and adoptee, but correlations of nearly zero for adopted relationships (Loehlin, Willerman, & Horn, 1985, 1987). However, adoptive siblings in the Colorado Adoption Project correlated .85 (Rende, Slomkowski, Stocker, Bilker, & Plomin, 1992). Recent reviews of this literature suggest an overall consensus that there is some genetic influence on aggression and antisocial behavior (Carey, 1994; Gottesman & Goldsmith, 1994). How- ever, as discussed above, there is striking variability among these studies. Hence, we think that it is justified to examine in a rigorous quantitative fashion the degree to which this variability can be attributed to factors such as age, sex, measuring instru- ment, and, of course, statistical sampling error. Meta-analysis allows one to use differential scaling for variable methods of measurement and to account for sample size differences by weighing results from samples accordingly. Through this proce- dure, researchers can explore how factors such as age and sex may moderate the genetic and environmental architecture of aggressive and antisocial behavior. Studies The samples selected for this review include the twin and adoption studies tabulated by Carey (1994) in the National Research Council report on violence, plus additional studies 207 208 MILES AND CAREY published since that report was originally written in 1989. We included a study if it used any measure of aggression, hostility, or antisocial behavior or if the scale was specifically constructed to predict juvenile delinquency. Results for male participants and female participants were treated separately unless the origi- nal study reported only results pooled for gender. Table 1 and Table 2 present the samples and data used in this review. We coded the following variables for the meta-analysis: zy- gosity (for twins), biological versus adoptive relationship (for adoption studies), sex, age, and type of measurement. Age was coded as youth (mean sample age of 18 or younger) versus adult (over age 18). (Other methods of coding age, such as children, adolescent, and adult, were also tried, but results did not substantively differ and hence are not presented.) Type of measurement was coded into three categories: self report (SR), parental report (PR), and observational (OB). Table 3 presents the number of studies that fit into each of the categories. We would have liked to code the actual measurement instru- ment according to its construct or its predictive validity with respect to aggression per se as opposed to general antisocial behavior; which would include aggression among a much wider range of behavior. However, we could not think of an objective way of coding these data along this line for two reasons. First, we examined items from those studies that included a scale with the word aggression in the title. Despite the same title, there was a significant amount of heterogeneity of item content, ranging fromprojective techniques (Owen & Sines, 1970), to an intrapunitive sense of guilt and self-blame (Partanen, Bruun, & Markkanen, 1966), to an excellent psychometric scale with item content ranging from relatively minor feelings of anger and retri- bution to overt acts of assault (Tellegen et al., 1988). Second, it is well established that the extreme aggression that would be called violence (sexual assault, robbery, murder, etc.) is strongly correlated with other aspects of antisocial behavior (e.g., van- dalism, theft). As a consequence, some well-validated scales that predict general antisocial behavior (e.g., the MMPI Pd scale) may actually be more predictive of interpersonal, physical aggression than other scales that are specifically called aggres- sion scales. Lacking intercorrelation matrices among these scales and lacking correlations between the scales and a com- mon criterion variable of interpersonal aggression, we decided on a different tactic, outlined below in the Method section, to arrive at an index of measuring a common construct. Study 1 Method The purpose of this study was to examine the validity of two crucial assumptions for twin and adoption data: the absence of selective place- ment for the trait and the equal environments assumption for MZ and DZ twins. Selective placement occurs when there is some correlation in the environments of biological relatives who are raised in separate households. Strong selective placement in environments relevant to ag- gression and antisocial behavior will compromise interpretation of the studies of twins raised apart and adoptees that are presented in Tables 1 and 2. The second assumption is that identical twins raised together do not experience more similar environments for antisocial behavior and aggression than do fraternal twins raised together. Given evidence of twin imitation with regard to registered criminality (Carey, 1992) and the presence of selective placement in some adoption studies (Plomin, DeFries, &Rilker, 1988), it is crucial to examine both of these assump- tions in the data set before proceeding with substantive analysis. We performed three analyses for the three instruments of measure presented in Tables 1 and 2 that have been gathered on both twins and adoptees—the California Psychological Inventory Socialization (CPI So) scale, the MMPI Pd scale, and the Multidimensional Personality Questionnaire (MPQ) Aggression scale. The model of analysis assumes that the phenotype (P) is a linear function of genotype (G) and environment (E), giving the structural equation P = AG + eE, where h and e are regression weights. We assumed the simple model of additive gene action with random mating. We also assumed that the phenotype of parents (Pin and Pr for mother and father, respectively) impinge on the environments of their offspring (Eo), giving the structural equation En = rPm + rPf + U, where U is a residual and the is are regression weights. We also assumed that the environments of siblings are correlated by the amount rfc in addition to the correlation in their environments that results from the influence of the parental phenotypes. We used two parameters to measure selective placement. We used the quantity s to denote the correlation between a biological parent and an adoptive parent of an adopted child. We used the quantity w to measure the extent to which the environments of twins raised apart are correlated with respect to siblings raised in the same household; w = 0 implies that the environments for twins raised apart are random, and w = 1 implies that the environments of twins raised apart are as similar as those of siblings and twins raised together. Finally, we used the parameter a to denote imitative effects with separate as for MZ and DZ twins (see Carey, 1992). The expected correlations are presented in Table 4. This full model is not identified with the data at hand. However, it is possible to fix certain parameters to examine the effect of the violations of assumptions on other parameter estimates. Typically, models used in behavioral genetics have assumed no selective placement of either single- ton adoptees or twins raised apart. This is one extreme model in which s = w = 0. At the other extreme, one can assume perfect selective placement so mat the correlation between biological parent and adoptive parent is at its upper mathematical limit of t and that the correlation for the environments of twins raised apart is equal to that of siblings raised together (i.e., w = 1). We fit five models to each of the three scales. The first allowed perfect selective placement with the possibility of genetic influence. The second also had selective placement but no heritability. The third and fourth models assumed no selective placement with heritability (Model 3) or no heritability (Model 4). The final model fitted only heritability. Results The results of fitting selective placement models to the MMPI Pd scale are presented in Table 5. Model 1, in which there is perfect selective placement, gives a satisfactory fit. However, Model 2 demonstrates that heritability cannot be set to 0; even under such extreme circumstances, this model must be rejected on the basis of both goodness of fit and likelihood ratio (LR), X2(l) ~ 14.83, p < .001. A model that takes the extreme genetic position of no selective placement also fits well (Model 3), but again, heritability cannot be set to 0 (Model 4). In terms of information, the last model fits best. It uses only one parameter, h2. and satisfactorily accounts for all the data. Table 6 gives the results of fitting these data to the CPI So scale. These results parallel those for the MMPI. The two models that assume no heritability, Models 2 and 4, must be rejected. There is no evidence for selective placement, and the simplest ARCHITECTURE OF HUMAN AGGRESSION 209 Table 1 Twin Studies Included in Meta-Analysis Study Gottesman (1966) Partanen, Bruun, & Markkaren (1966) Scarr (1966) Owen &. Sines (1970) Loehlin & Nichols (1976) Rowe (1983) Rushton, Fulker, Neale, Nias, & Eysenck (1986) Lytton, Watts, & Dunn (1988) Stevenson & Graham (1988) Gottesman (1963, 1966) Reznikoff & Honeyman (1967) Canter (1973) O’Connor, Foch, Sherry, & Plomin (1980) Plomin, Foch, & Rowe (1981) Pogue-Geile & Rose (1985) Rose (personal communication July 18, 1986) Ghodsian-Carpey & Baker (1987) Tellegen et al. (1988) Gottesman, Carey, & Bouchard (1984) Tellegen et al. (1988) Bouchard & McGue (1990) Sample Measure used Group Male N Twins raised together, genders analyzed separately 1 2 3 4 5 6 7 8 9 1 10 11 11 12 12 13 14 15 15 15 CFI Socialization Aggression items ACL n Aggression MCPS Aggression CPI Socialization ACL n Aggression No. of delinquent acts 23 aggression items from IBS Rutter Antisocial subscale Rutter Antisocial subscale MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ DZ-OS MZ DZ MZ DZ Twins raised together, genders pooled MMPI Psychopathy Acting Out Hostility of Foulds Hostility Scale Conners’s bullying Median (three objective aggression ratings) MMPI Psychopathy MMPI Psychopathy CBC Aggression MOCL Aggression MPQ Aggression MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ MZ DZ Twins raised apart, genders pooled MMPI Psychopathy MPQ Aggression CPI Socialization MZ DZ MZ DZ MZ DZ 34 32 157 189 — — 10 11 – 202 124 216 135 – 61 38 90 46 98 13 20 46 48 120 132 39 44 52 32 53 32 71 62 228 182 21 17 21 17 217 114 51 25 44 27 45 26 r .32 .06 .25 .16 .09 .24 .52 .15 .20 .05 .62 .52 .33 .16 .12 .89 .67 .61 .40 .48 .27 .14 .30 .72 .42 .39 .42 .35 .18 .47 .23 .78 .31 .65 .35 .43 .14 .64 .34 .46 .06 .53 .39 Female N i 45 36 — 24 28 -. 8 13 288 193 293 195 107 59 106 133 — — — 53 58 — — — — — — — — — — — — — — — — — — — — — — 52 26 35 08 58 22 55 48 24 06 66 46 43 00 29 49 Age (years) 14-25 14-25 28-37 28-37 6-10 6-10 6-14 6-14 18 18 18 18 13-18 13-18 19-60 19-60 19-60 9 9 13 13 14-18 14-18 16-55 16-55 5-11 5-11 5-11 5-11 20-25 20-25 14-34 14-34 4-7 4-7 4-7 4-7 19-41 19-41 19-68 19-68 19-68 19-68 19-68 19-68 Method SR SR SR SR PR PR SR SR SR SR SR SR SR SR SR SR SR PR PR PR PR SR SR SR SR PR PR OB OB SR SR SR SR PR PR PR PR SR SR SR SR SR SR SR SR Note. Dashes indicate data were not obtained. SR = self-report; PR = parental report; OB = observational data; CPI = California Psychological Inventory; ACL = Gough’s Adjective Checklist; n = number of adjectives checked; MCPS = Missouri Children’s Picture Series; IBS = Interpersonal Behavior Survey; MMPI = Minnesota Multiphasic Personality Inventory; CBC = Child Behavior Checklist; MOCL = Mothers’ Observations Checklist; MPQ = Multidimensional Personality Questionnaire; DZ — dizygotic; MZ = monozygotic; DZ-OS = dizygotic opposite sex. model of only heritability (Model 5) gives a satisfactory fit by using only one parameter. The fits for the MPQ Aggression scale are given in Table 7. The results are very similar but are not identical to those for the MMPI and CPI. Once again, the best fitting model is Model 5, which explains the observed correlations by using only h2, and heritability cannot be set to 0 when there is no selective placement. The difference between this MPQ Aggression and 210 MILES AND CAREY Table 2 Adoption Studies Included in Mela-Analysis Study Loehlin, Willerman, & Horn (1985) Loehlin, Willerman, & Horn (1987) Parker (1989) Rende, Slomkawski, Stacker, Fulker, & Plomin (1992) Sample 16 16 17 17 Measure used CPI Socialization MMPI Psychopathy CBC Aggression Conflict scale Relationship Adoptive father-child Adoptive mother-child Biological father-child Biological mother-child Adoptive-adoptive siblings Adoptive-biological siblings Biological-biological siblings Adoptive father-child Adoptive mother-child Biological father-child Biological mother-child Birth mother-adopted child Adoptive-adoptive siblings Adoptive-biological siblings Biological-biological siblings Adoptive siblings Biological siblings Adoptive siblings Biological siblings N 241 253 52 53 76 47 15 180 177 81 81 133 44 69 20 45 66 57 67 r .00 -.02 .16 .06 .03 .10 -.01 .07 .01 .12 .07 .27 .02 .06 -.06 .47 .44 .85 .91 Age (years) parents 39-76 39-76 39-76 14-36 14-36 14-45 parents 39-76 39-76 39-76 39-76 14-36 14-36 14-45 4-6 4-6 3-11 3-11 Method SR SR SR SR SR SR SR SR SR SR SR SR SR SR SR PR PR OB OB Note. SR = self-report; PR = parental report; OB = observational data; CPI = California Psychological Inventory; MMPI = Minnesota Multiphasic Personality Inventory; CBC = Child Behavior Checklist. the MMPI and CPI occurs in comparing Model 2 with Model 1. With the MPQ scale, heritability can be set to 0 under perfect selective placement; it cannot be set to 0 under perfect selective placement with the MMPI or CPI scales. The importance of this analysis lies less in its providing a strong argument for the presence of heritability than in its dem- onstrating that even if selective placement were occurring in the extreme, genetic effects would still be present. The appropriate data points here are the estimates of heritability under Model 5 in contrast with the estimates under Model 1 in Tables 5, 6, and 7. Despite the unreasonable assumption of perfect selective placement, heritability estimates change from .48 to .44 (MMPI Pd), from .53 to .40 (CPI So), and from .43 to .42 (MPQ Aggression). Selective placement and correlations in the envi- ronments of twins raised apart reduce heritability, as they should. They do not make the influence of heritability go away, and they do not dramatically alter the magnitude of the genetic influence. Perfect selective placement implies that the personnel at an Table 3 Studies Separated by Method of Report Twins raised Sex analyzed Method (age in years) Together Apart Adoption N separately” Self-report Young (6-25) 5 0 0 5 4 Old (14-76) 6 3 2 11 2 Parent report (4-11) 5 0 16 3 Observational (3-11) 1 0 1 2 0 a Only twins raised together were analyzed separately by sex. adoption agency would be able to ascertain perfectly all aspects of psychopathy in a biological mother and in an adoptive mother and father and would then be able to match them. Even if MMPI Pd scores were available on all three parents, the limited number of adoptees available at any one time would prevent perfect matching on observed scores. Study 2 Method In this section, we outline the methods that we used for the actual meta-analysis. The first and most general model expresses heritability Table 4 Expected Correlations for Different Kinship Correlations Kinship correlation Expected correlation Adoptive parent-offspring te Adoptive siblings {It2 + r^e2 Genetic parent-adoptive child ‘f2h2 + 2st2e2 Siblings %h2 + {It1 + rs)e2 DZ twins together [(1 + £Dz)yDZ + 2aDZ]/(l + aDZ MZ twins together [(! + OMZJVMZ + 2aMz]/(l + aa. DZ twins apart h2 + w{2t2 + r*)e2 MZ twins apart h2 + w(2t2 + rs)e2 Note, yDZ = h2 + {2t2 + r.y2; >w = ^2 + (2/2 + rB)e2. DZ – dizygotic; MZ = monozygotic; aDZ = imitative effect for DZ twins; «3MZ = imitative effect for MZ twins; e = environmental effect; h2 = heritabil- ity; rs = environmental correlation for siblings; s = selective placement coefficient; t = regression weight; w = environmental correlation for twins raised apart. ARCHITECTURE OF HUMAN AGGRESSION 211 Table 5 Testing Assumptions on the MMPI Pd Scale Model 1: Perfect selective placement” 2: Perfect selective placement,’ no heritability 3: No selective placement11 4: No selective placement,*1 no heritability 5: Only heritabilityc df 8 9 8 9 13 Goodness of fit x2 8.16 22.99 9.12 41.41 9.55 P .42 <.0l .33 <.OO01 .73 h2 .44 — .51 — .48 Parameter estimates r .00 .10 -.04 .06 — r, .16 .22 -.02 .02 — -.03 .00 .00 .11 — aMZ -.04 .13 -.02 .23 — Note. Dashes indicate parameter was set to zero. MMPI = Minnesota Multiphasic Personality Inventory; Pd = Psychopathic-deviate; DZ = dizygotic; MZ = monozygotic; aoz = imitative effect for DZ twins; aya. = imitative effect for MZ twins; h2 = heritability; rs ~ environmental correlation for siblings; s = selective placement coefficient; t = regression weight; w = environmental correlation for twins raised apart. a s = 0.5, w = 1. b 5 = w = 0. e 5 = w ~ t = rs = (JDZ = OMZ = 0. and common environment by using contrast codes for sex, age category, rating type (self vs. parental), and measurement mode (observational vs. psychometric). The specific formula for heritability was heritabiiity = h2 + /3,sex + /32age +• 03report + /^measure, (1) where A2 is a constant and 0t, 02, 03, and /34 are parameters for sex, age, rating type, and measurement mode, respectively, in accordance with the observed correlation in Tables 1 and 2. Sex was coded numerically as — 1 = female sample, 0 = mixed female-male sample, and 1 = male sample. Age was coded as 0 – adult and 1 = youth. Rating type was coded as 0 = self-report and 1 = parental rating. Measurement mode was coded as 1 = observational and 0 = psychometric. A similar equa- tion was written for common environment: common environment = c2 + In family data, the observed correlation between relatives is the prod- uct of the reliability of the measure and the true correlation between the relatives. Thus, differences in reliability between measures will contrib- + /Sjreport, + <53report, + 64nieasure;-)]. (3) ute to differences in observed correlations among samples. To control for this, we used the parameter a,, where or is the reliability coefficient for the measure and i denotes the ith measure. The predicted correlation for a pair of relatives from the jth sample using the /th measure becomes = a,[yj(h2 + Here y, denotes the coefficient of genetic relatedness for the kinship in the jth sample. If the sample involves adoptive relatives, yi = 0; for biological parent and offspring, biological siblings, and DZ twins; y} = .5; and for MZ twins, y, = 1.0. The quantity rfj denotes die coefficient of environmental relationship of the /th sample; TJ, = 1.0 when the relatives are raised together, and rjj = 0 when the relatives are raised apart. This method permits all the predicted correlations for the samples in Tables 1 and 2 to be made functions of 25 parameters: h2, the 4 betas, c2, the 4 deltas, and the 15 alphas. The quantities y and rj will, of course, be fixed for the individual sample. For example, the predicted correlation for the CPI from an adult, mixed-sex sample of identical twins raised apart is Table 6 Testing Assumptions on the CPI So Scale Model 1: Perfect selective placement” 2: Perfect selective placement,3 no heritability 3: No selective placement6 4: No selective placement,6 no heritability 5: Only heritability0 df 12 13 12 13 16 Goodness of fit x2 18.11 27.43 19.01 35.66 22.25 P .11 <.O1 .07 <.OO07 .14 h2 .40 — .46 — .53 Parameter estimates * -.03 .01 -.04 .01 — rs .13 .20 .05 .05 — flDZ .03 .06 .04 .15 — aMz .03 .18 .02 .26 — Note. Dashes indicate parameter was set to zero. CPI = California Personality Inventory; So = Socializa- tion; DZ – dizygotic; MZ = monozygotic; aDZ – imitative effect for DZ twins; OMZ = imitative effect for MZ twins; h2 = heritability; rB = environmental correlation for siblings; s = selective placement coefficient; t = regression weight; w = environmental correlation for twins raised apart. * s = 0.5, w = 1. b s = w = 0. cs = w = t = rs = aDZ = a^z = 0. 212 MILES AND CAREY Table 7 Model-Fitting Results on the MPQ Aggression Scale Goodness of fit Parameter estimates Model df 1: Perfect selective placement” 2: Perfect selective placement,3 no heritability 3: No selective placement 4: No selective placement,b no heritability 5: Only heritability’ 1.01 .60 .42 .04 .07 .02 -.06 3 2 3 6 3.01 0.67 10.24 2.30 .39 .71 <.O2 .89 — .43 .43 .18 -.05 .18 .23 .11 .22 . .08 -.04 .09 -.08 -.07 -.07 Note. Dashes indicate parameter was set to zero. MPQ = Multidimensional Personality Questionnaire; DZ = dizygotic; MZ = monozygotic; aDZ = imitative effect for DZ twins; aMZ = imitative effect for MZ twins; h2 = heritability; rs = environmental correlation for siblings; t — regression weight; w = environmental correlation for twins raised apart. aw=l. b w = 0. c w = t = rs = aDZ = aMZ = 0. «cp](*3). (4) Similarly, the correlation for the young, mixed-sex, DZ twins on the Child Behavior Checklist (CBC) in Ghodsian-Carpey and Baker (1987) is aCBC[l/2(h2 6,)]. (5) An exact method for fitting the model to the data and for assessing the fit of the model is not possible with the information available. The obstacle is the dependence among several observed correlations that are presented in Tables 1 and 2. For example, the correlations for the MMPI, CPf, and MPQ on the Minnesota series of twins raised apart (Bou- chard & McGue, 1990; Gottesman, Carey, & Bouchard, 1984; Tellegen et al., 1988) came from three different reports from the same series of twins. To control for this statistical dependency, we would need the raw data from this sample and would need to use quantitative pedigree analysis. We chose to treat observed correlations from such overlapping sam- ples as independent. This approach leads to very minimal bias in the parameter estimates but may lead to conservative hypothesis testing (McGue, Wette, & Rao, 1984). That is, there is more power to reject a false hypothesis, but this is gained at the expense of increased Type I errors. Simulations, however, suggest that this bias is not large (McGue et al., 1984). Hence, the fitting function we used was z,)2 (6) where i denotes the ith data point in Tables 1 and 2, N denotes the sample size for that correlation, Zi denotes the zeta transformation of the predicted correlation, and z-s denotes the zeta transformation of the observed correlation. Also, at least one of the as must be fixed to permit identification. To avoid estimates of reliability considerably greater than 1.0, we fixed a for the sample with the largest correlations (Rende et al., 1992) to be 1.0 and the as for the MMPT, MPQ, and CPI to be .80, close to their test-retest correlations over a short interval. Results Table 8 presents the results of the fitting of models to all studies included in Tables 1 and 2. The general model fits all the parameters. The numbered models set one or more effects to zero for both heritability and common environment. For ex- ample, in Model 1, sex – 0 sets both /?, and 6, to zero. Because the actual parameter estimates depend on arbitrary decisions for fixing the values of the as, we scaled our estimates so that the sum of the absolute values of each row equals zero. In this way, the absolute values of the estimates for each row give the percentage of the familiality for aggression attributable to a variable. For example, in the general model, f3t = .05 and 6i = -.05, so that 10% of familiality may be attributable to gender differences, with heritability being slightly higher in males and common environment being slightly higher in fe- males. Similarly, the largest source of variance in the general model (and in all other models where the parameters are free) is attributable to mode of measurement. Observational studies account for 45% (.16 + .29) of the familial variability, greatly decreasing heritability and increasing common environment. The fits of the models in Table 8 may be judged by the two columns labeled GOF p and LR p. The first is the p value for the goodness-of-fit chi-square. The second is the p value for the likelihood ratio chi-square, comparing one of the numbered models in Table 8 against the general model. Because of the lack of independence among the correlations, the p values should not be interpreted literally as rejecting or not rejecting models. Instead, they should be viewed as nonparametric estimates of fit, with larger p values suggesting satisfactory fits and smaller p values implying poor fits. The most striking conclusion one can draw from Table 8 concerns the magnitude of the mode of measurement effect— an observational rating of videotaped or real-life performance versus a parental or self-report. Tn every case in which mode of measurement was set to zero, models had poor fits on the basis of either the GOF or the LR chi-square. Similarly, for every model where measurement was fitted, this effect was the largest predictor of familiality, accounting for roughly 50% of the gen- eral differences among the correlations in Tables 1 and 2. The effect was also consistent: Observational ratings greatly de- creased heritability and increased common environment. In interpreting all of those models in which the measurement effect was fitted (Models 1, 2, 3, 5, 6, 8, and 11, and the general ARCHITECTURE OF HUMAN AGGRESSION 213 Table 8 Model Fits and Proportions of Familial Resemblance for All Data Model General 1: Sex = 0 2: Age = 0 3: Report = 0 4: Measurement = 0 5: Sex, age = 0 6: Sex, report = 0 7: Sex, measurement = 0 8: Age, report = 0 9: Age, measurement = 0 10: Report, measurement = 0 11: Sex, age, report = 0 12: Sex, age, measurement = 0 13: Age, report, measurement = 0 14: Sex, age, report, measurement = 15: No heritability 16: No common environment x1 66.44 69.54 72.39 69.01 76.37 76.61 71.38 79.94 82.67 84.86 161.50 86.42 9.33 241.41 0 244.95 239.58 17.77 df 57 59 59 59 59 61 61 61 61 61 61 63 63 63 65 61 61 GOF P .18 .16 .11 .18 .06 .09 .17 .05 .03 .02 <.001 .03 <.01 <.00l <.001 <.001 <.001 LR P .21 .05 .28 <.01 .04 .29 <.01 <.01 <.00I <.001 <.01 <.001 <.O01 <.001 <.001 <.001 h1 .22 .25 .21 .27 .22 .24 .30 .22 .26 .39 .32 .30 .49 .54 .78 .00 .24 Sex /?i .05 .00 .05 .05 .00 .00 .00 .00 .07 .10 .02 .00 .00 .15 .00 .00 -.01 Age Pz -.01 -.03 .00 -.07 -.23 .00 -.08 -.24 .00 .00 -.29 .00 .00 .00 .00 .00 .05 Report ft -.04 -.09 -.06 .00 .24 -.13 .00 .17 .00 -.06 .00 .00 -.22 .00 .00 .00 .18 Measure- ment /?4 -.16 -.17 -.17 -.14 .00 -.19 -.15 .00 -.20 .00 .00 -.24 .00 .00 .00 .00 .51 c1 .00 .00 .01 .00 .00 .01 .00 .00 .02 .09 .00 .03 .11 .15 .22 .18 .00 Sex *i -.05 .00 -.06 -.05 -.01 .00 .00 .00 -.07 -.12 -.03 .00 .00 -.16 .00 -.03 .00 Age .05 .07 .00 .11 .26 .00 .13 .27 .00 .00 .35 .00 .00 .00 .00 .28 .00 Report 6, .13 .07 .15 .00 -.03 .10 .00 -.11 .00 .25 .00 .00 .17 .00 .00 .22 .00 Measure- ment 64 .29 .32 .30 .30 .00 .33 .33 .00 .37 .00 .00 .43 .00 .00 .00 .29 .00 Note. Coded such that for sex (/?,, 6t), male = 1, mixed male-female = 0, female – -1; for age (&, 62), adult = 0, child or adolescent = 1; for report (/33, <53), self = 0, parental = 1; for measurement (J34, d4), psychometric = 0, observational = 1. GOF = goodness of fit; LR = likelihood ratio; h2 = heritability; c2 = common environment. model), it appears that the influence of sex, age, and type of report was consistent but not very strong. Heritability was slightly more influential in male participants than in female participants, whereas common environment was more important in female participants than in male participants. The effect of heritability was lesser and that of common environment greater among younger samples compared with adult samples. Finally, parental reports resulted in lower heritability but greater com- mon environment than did psychometric self-reports. The worst model fits occurred when age was set to zero (Models 2, 5, 8, and 11). Observational data were available for only two studies: Plomin, Foch, and Rowe (1981) and Rende et al. (1992). Be- cause of the large observational effect, we redid the analysis eliminating these two studies as a post hoc exploration of the effects of sex, age, and type of report. Results from this analysis are presented in Table 9. The fundamental patterning of results remained unchanged. Models that assume no heritability or no common environment (Models 9 and 10) gave the worst fits. Models with no age effect had less satisfactory fits than did those that permit no sex and report effect. The most parsimoni- ous model was Model 6, which permits only age differences to moderate the genetic and environmental architecture of aggression. General Discussion The meta-analysis gives four major conclusions. The first major conclusion is that heritability and common environment are definitely responsible for individual differences in aggres- sion. It is highly unlikely that heritability is a methodological artifact. Even the unrealisttcally high estimates of an artifact effect in Study 1 do not dramatically alter heritability. Unless there is some as yet unidentified methodological flaw in the twin or adoption strategies or both, the results from Study 2 suggest an important contribution of heritability for aggression, perhaps accounting for as much as 50% of the variance. The second important conclusion concerns the extent to which observational methods of measuring aggression gave different results for either self-report or parental report. The two observa- tional studies suggested a very strong influence of common environment with little evidence of heritability, a result that cannot be explained by the age of the two samples (both were young). The reason for this is unclear. Perhaps this is a chance finding—there were, after all, only two studies. Perhaps differ- ent aspects of aggression are tabulated by observers than are reported by individuals. Or perhaps one or both of the studies capitalized on state-specific, reciprocal influences of twin or adoptive dyads when they are tested at the same time. The very large sibling correlations of Rende et al. (1992) — .85 for adoptive siblings and .91 for biological siblings—are, to our knowledge, much larger than any other ever reported for any behavioral measure. They are also consistent with the possibility of capturing an episode of dyadic interaction where both siblings are either aggressive or nonaggressive. Whatever the cause, this result suggests the need for returning to multitrait-multimethod strategies (Campbell & Fiske, 1959) for measuring aggression in families. Indeed, in the behavioral genetics literature, there have been very few attempts to assess a similar personality construct in multiple modes of measurement. Virtually all results on adult personality have been based on self-reports (for recent reviews, see Eaves, Eysenck, & Martin, 1989; Loehlin, 1992). The very possibility of obtaining different types of results from a different mode of measurement has strong implications for research on the genetics of all personality traits, not simply aggression. The third conclusion, made in a more tentative fashion than 214 MILES AND CAREY Table 9 Model Fits and Proportions of Familial Resemblance for Aggression, Deleting Two Observational Studies Model df GOF P LR P h2 Sex Age Report Sex Age Report 1: General 2: Sex = 0 3: Age = 0 4: Report = 0 5: Sex, age = 0 6: Sex, report = 0 7: Age, report = 0 8: Sex, age, report = 0 9: No heritability 10: No common environment 66.36 69.46 72.31 68.93 76.53 71.30 82.59 86.35 37.40 83.97 54 56 56 56 58 58 58 60 57 57 .12 .11 .07 .11 .05 .11 .02 .01 <.O01 < 01 .21 .05 .28 .04 .29 <.01 <.01 <.001 <.001 .40 .49 .40 .49 .49 .58 .61 .91 .00 .50 .08 .00 .10 .09 .00 .00 .16 .00 .00 -.03 -.03 -.05 .00 -.13 .00 -.16 .00 .00 .00 .10 -.07 -.18 -.11 .00 -.27 .00 .00 .00 .00 .38 .00 -.01 .01 .00 .02 -.01 .06 .09 .25 .00 -.10 .00 -.11 -.10 .00 .00 -.17 .00 -.04 .00 .09 .13 .00 .20 .00 .25 .00 .00 .40 .00 .23 .14 .27 .00 .22 .00 .00 .00 .31 .00 Note. Coded such that for sex (0U 6,), male = 1, mixed male-female = 0, female = -1; for age ((32, 62), adult = 0, youth = 1; for report (/?3, 6i), self = 0, parental = 1. GOF = goodness of fit; LR = likelihood ratio; h2 = heritability; c2 = common environment. the prior two, is that the genetic and family environmental archi- tecture of aggression may change over time. Longitudinal stud- ies of individuals highlight both the consistency of aggression (Farrington, 1986, 1989) and the changes in developmental pat- terns overtime (Moffitt, 1993; Patterson, 1992). However, there are no longitudinal genetic data researchers could use to test how much consistency and change may be due to genes and environment. The results of our cross-sectional meta-analysis suggest that in youth, genes and common environment equally promote similarity among relatives. For adults, however, the influence of common environment is negligible but that of heri- tability increases. These results are completely consistent with the literature on juvenile versus adult criminality in twins (Carey, 1994; Gottesman & Goldsmith, 1994; Rowe, 1990; Rowe & Rodgers, 1989). Common environment has been very important in juvenile delinquency, whereas genes have been relatively more important for adult criminality. This phenome- non suggests that aspects of the family (imitating or reacting to the same parents, living in the same neighborhood, having overlapping groups of friends, etc.) may be very important in the initiation and early maintenance of aggression but may fade over time. We hypothesized that this age effect is due to what has been informally termed the “smorgasbord” model’ of gene-environ- ment interplay. This model analogizes the various tidbits on the smorgasbord to different types of environments and the taste preference of the diner to a genotype. Faced with a large number of choices, the diner samples a little of each but then returns to the most enjoyable dishes. Thus, the adult genotype ends up choosing the environments most compatible with the genotype. With aggression in youth, the initial choices on the table may be limited by the family environment. Departing from the family household as an adult, one experiences a broader range of envi- ronments that either positively or negatively reinforce aggres- sion. The adult then prefers those environments compatible with his or her genotype, creating a gene-environment correlation that is indistinguishable from heritability. The smorgasbord model is consistent with longitudinal data that suggest that fa- milial factors such as parental inconsistency and failure to set limits predict juvenile and adolescent antisocial behavior (Pat- terson, Capaldi, & Bank, 1991; Patterson, DeBarsyshe, & Ram- sey, 1989). However, testing this hypothesis requires genetically informative, longitudinal data that also include appropriate mea- sures of the environment. The fourth, and indeed most tentative, conclusion is that of a relative lack of common environment, at least in adults, for aggressive behavior. Although this finding is consistent with Rowe’s (1994) review and conclusions about a minor role for family environment, we hesitate to endorse his position in the cases of aggression and antisocial behavior. There are two rea- sons for our view. First, we assumed in creating our model that all gene action is additive. This means that there is virtually no genetic domi- nance or gene-gene interaction for every single locus predicting individual differences in aggression. When this assumption is violated, twin data will overestimate heritability and underesti- mate common environment. Second, the adoption data that might resolve the issue of additivity in the twin design have very serious problems of their own when they are applied to severe antisocial behavior, including aggression. Those families with exceptionally high risk environments for violence are seldom allowed to adopt. Examples would include a single-parent household, extreme poverty, and parental alcohol abuse, substance abuse, and crimi- nality. If this tail of the distribution is missing among adoptive families, the influence of family environment may be underesti- mated, particularly when the effect is nonlinear. That is, those families screened out of the adoption pool may have a very strong influence, whereas the regression line between family environment and offspring aggression is flat among those intact and psychologically healthy families permitted to adopt. For these reasons, we feel confident that heritability is important for aggression and that the causes of familial resemblance change over time, but we await the results of future research before quantifying the effects of the family environment. Much of personality psychology has yet to face the implica- tions of important heritability for aggression. One dominant 1 We attribute the “smorgasbord” analogy to Lindon Eaves (personal communication, August 4, 1984). The ideas of gene-environment covar- iance behind it are discussed in Eaves, Last, Martin, and Jinks (1977); Plomin, DeFries, and Loehlin (1977); and Scarr and McCartney (1983). ARCHITECTURE OF HUMAN AGGRESSION 215 perspective that is used to explain the familial resemblance for aggression is social learning theory (Bandura& Waiters, 1959). According to this theory, a child learns to behave aggressively by observing aggressive behavior; encoding these acts to mem- ory, and incorporating aggressiveness into his or her code of conduct (Bandura, 1986). Although peers and the media are important contributors to a child’s behavior, it is the family environment that may have the greatest influence on the similar- ity found among family members. The family environment may also foster the development of aggression through such means as interactions among family members and parental disciplinary techniques. Through detailed observation of aggressive and nonaggressive families, Patterson (1982) has identified a particular coercive pattern that serves to elicit, maintain, and increase aggression among family members. According to this pattern, a child often acts aggressively in reaction to the aversive behavior of another family member. If the other family member withdraws from the interchange and the aversive behavior is stopped, the child’s aggressive behavior is reinforced, encouraging similar behavior in the future. This coercion process tends to be found in families that are having problems with an aggressive child. Other factors found among aggressive families are lack of parental monitoring, parental aggression, permissiveness or inconsistency in discipline, and parental rejection (Perry, Perry, & Boldizai; 1990). These envi- ronmental factors provide models and reinforcement such that a child learns to act aggressively, thereby increasing a child’s level of aggressive behavior. The results of our meta-analysis suggest that genetics must be seriously considered when explaining the similarity found in the levels of aggression among family members. Adult samples and data on self-report measures and parental ratings on aggres- sive behavior have suggested moderate heritability with some small influence from common environment. These results do not support the social learning theory, which emphasizes the role of common environmental effects. The observed pattern of correlations for parental ratings and self-report data cannot be explained by methodological biases. Individual analyses on the MMPI Pd, CPI So, and the MPQ Aggression scales all lead to similar results. Each of these measures suggests moderate levels of heritability, supporting the importance of genetic influences. It has only been fairly recently that researchers have consid- ered biological-genetic factors important in the development of human aggression. The main areas in which researchers have focused include the relationship between aggression and hor- mones, temperament, arousal, and the central nervous system. Hormone research has concentrated mainly on the correlation between elevated levels of testosterone and aggressive behavior. However, the variable results of several studies have led to diffi- culty in making any conclusions. Other studies have found rela- tionships between aggression and neurochemicals related to adrenaline. For example, delinquents have been reported to show lower than average levels of epinephrine (adrenaline) and nor- epinephrine (Zuckerman, 1989). Zuckerman also reported that undersocialized conduct disorder is associated with low levels of dopamine-beta-hydroxylase, the enzyme that breaks down dopamine. Antisocial individuals have also been reported to show lowered levels of monoamine oxidase, an enzyme that breaks down the neurotransmitters serotonin, dopamine, epi- nephrine, and norepinephrine (Ellis, 1991). For temperament, infant difficultness has been found to be related to parental ratings of hostility in preschool-age children (Parke & Slaby, 1983). Parke and Slaby also suggested that the way an infant’s temperament influences parent-child interac- tions may account for the coercive patterns of aggression seen by Patterson (1982) that were mentioned earlier. Personality characteristics, which may play a role in the expression of ag- gression, have also been found to be heritable. A recent twin study found significant hen labilities for empathy, behavioral inhibition, and expressions of negative affect in infants (Emde et al, 1992). Personality research on adults has found antisocial behavior to be highly correlated with high scores on the three personality dimensions of the Eysenck Personality Question- naire: extraversion, neuroticism, and psychoticism, which have all shown significant heritability (Eysenck & Gudjonsson, 1989). A strong connection between physiological arousal and ag- gression has also been demonstrated by several researchers (Berkowitz, 1988; Brancombe, 1985; Zillmann, 1988). How- ever, researchers have emphasized that the relationship is inter- active and that aggression occurs only in very specific circum- stances. On the other hand, Eysenck and Gudjonsson (1989) proposed the general arousal theory of criminality, which sug- gests that the inheritance of a nervous system insensitive to low levels of stimulation may make an individual inclined to participate in high-risk activities associated with antisocial be- havior, such as crime, substance abuse, and sexual promiscuity, in order to increase their arousal. Studies on the central nervous system have suggested that defects in the limbic system and cortical functioning may also play a role in aggressive behavior (Eichelman, 1983; Gorenstein, 1990). Research in each of these areas has just begun to suggest possible biological influences on aggression. Aggression is also a common behavior among many primate species, and agonistic behavior is often adaptive. For example, in many groups, agonistic behavior seems to be about maintaining dominance relationships (Lauer, 1992). In turn, a dominance hierarchy helps to limit aggression through the development of stable, cooperative, and nonaggressive relations among members of the group. Such relationships promote individual fitness by diminishing the likelihood of injury or death that might result from continued aggression. Perhaps similar adaptive mecha- nisms exist for humans. Observational data available on pre- school children suggest that agonistic exchanges ending with submission revealed dominance hierarchies (Strayer, 1992). Strayer speculated that differences in dominance ranking during the preschool years may contribute to individual differences in later acquisition of social skills. However, few ethological stud- ies focus on the development of agonistic behavior, and most studies are performed on young children. An evolutionary theory of sociopathy proposed by Mealey (1995) suggests that sociopathy develops through two path- ways. Primary sociopaths participate in antisocial behavior be- cause of a lack of normal moral development and an inability to feel social responsibility, which is a product of their genotype, physioptype, and personality. On the other hand, secondary soci- opathy is contingent on the environment and develops in re- 216 MILES AND CAREY sponse to disadvantages in social competition. These individuals participate in antisocial behavior because they are unable to succeed through socially acceptable means at a particular time. Secondary sociopathy is also associated with other risk factors, such as low socioeconomic status, urban residency, low intelli- gence, and poor social skills, whereas primary sociopathy is not associated with a particular background. Although secondary sociopathy appears to be more responsive to the environment, both types of sociopathy may be heritable and may account for some of the genetic evidence that has been found. This brief overview suggests only some of the possible mechanisms that may lead to aggression and the development of antisocial behav- ior, and the search for other possible adaptive mechanisms merits further research. References Bandura, A. (1986). Social foundations of thought and action: A social cognitive theory. Englewood Cliffs, NJ: Prentice-Hall. Bandura, A., & Walters, R. H. (1959). Adolescent aggression. New Tiork: Ronald Press. Berkowitz, L. (1988). Frustrations, appraisals, and aversively stimulated aggression. Aggressive Behavior, 14, 3-11. Bouchard, T. J., & McGue, M. (1990). Genetic and rearing environmen- tal influences on adult personality: An analysis of adopted twins reared apart. Journal of Personality, 58, 263-292. Brancombe, N. (1985). Effects of hedonic valence and physiological arousal on emotion: A comparison of two theoretical perspectives. Motivation and Emotion, 9, 153-169. Cadoret, R. J. (1978). Psychopathology in adopted-away offspring of biologic parents with antisocial behavior. Archives of General Psychi- atry, 35, 176-184. Campbell, D. T., & Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulle- tin, 56, 81-105. Canter, S. (1973). Personality traits in twins. In G. Claridge, S. Canto; & W. I. Hume (Eds.). Personality differences and biological variations (pp. 21-51). New “fork: Pergamon Press. Carey, G. (1992). Twin imitation for antisocial behavior: Implications for genetic and family environment research. Journal of Abnormal Psychology, 101, 18-25. Carey, G. (1994). Genetics and violence. In A. J.Reiss, K. A. Miczek, & J. A. Roth (Eds.), Understanding and preventing violence: Vol. 2. Biobehavioral influences (pp. 21-57). Washington, DC: National Academy Press. Eaves, L. J., Eysenck, H. J., & Martin, N. G. (1989). Genes, culture, and personality. London: Academic Press. Eaves, L. J., Last, K. A., Martin, N. G., & Jinks, J. L. (1977). A progres- sive approach to non-additivity and genotype-environment covariance in the analysis of human differences. British Journal of Mathematical and Statistical Psychology, 30, 1-42. Eichelman, B. (1983). The limbic system and aggression in humans. Neuroscience and Biobehavioral Reviews, 7, 391-394. Ellis, L. (1991). Monoamine oxidase and criminality: Identifying an apparent biological marker for antisocial behavior. Journal of Re- search in Crime and Delinquency, 28, 227-251. Emde, R. N., Plomin, R., Robinson, J., Corley, R.( DeFries, J., Fulker, D., Reznick, J. S.( Campos, J., Kagan, J., & Zahn-Waxler, C. (1992). Temperament, emotion, and cognition at fourteen months: The MacAr- thur Longitudinal Twin Study. Child Development, 63, 1437-1455. Eysenck, H. J., & Gudjonsson, G. (1989). The causes and cures of criminality. New York: Plenum Press. Farrington, D. P. (1986). Stepping stones to adult criminal careers. In D. Olweus, J. Block, & M. Radke-Yarrow (Eds.). Development of antisocial andprosocial behavior: Research, theories, and issues (pp. 359-384). New York: Academic Press. Farrington, D. P. (1989). Early predictors of adolescent aggression and adult violence. Violence and Victims, 4> 79-100. Ghodsian-Carpey, J., & Baker, L. A. (1987). Genetic and environmental influences on aggression in 4- to 7-year old twins. Aggressive Behav- ior, 13, 173-186. Gorenstein, E. E. (1990). Neuropsychology of juvenile delinquency. Fo- rensic Reports, 3, 15-48. Gottesman, 1.1. (1963). Heritability of personality: A demonstration. Psychological Monographs: General and Applied, 77(9), 1-21. Gottesman, 1.1. (1966). Genetic variance in adaptive personality traits. Journal of Child Psychology and Psychiatry, 7, 199-208. Gottesman, 1.1., Carey, G., & Bouchard, T. J. (1984, May). The Minne- sota Multiphasic Personality Inventory of identical twins raised apart. Paper presented at the 15th annual meeting of the Behavior Genetics Association, Bloomington, IN. Gottesman, 1.1., & Goldsmith, H. H. (1994). Developmental psycho- pathology of antisocial behavior: Inserting genes into its ontogenesis and epigenesis. In C. A. Nelson (Ed.), The Minnesota Symposia on Child Psychology: Vol. 27. Threats to optimal development: Integrat- ing biological, psychological, and social risk factors (pp. 69-104). Hillsdale, NJ: Erlbaum. Jary, M. L., & Stewart, M. A. (1985). Psychiatric disorder in the parents of adopted children with aggressive conduct disorder. Neuropsychobi- ology, 13, 7-11. Lauer, C. (1992). Variability in the patterns of agonistic behavior of preschool children. In J. Silverberg & J. P. Gray (Eds.), Aggression and peacefulness in humans and other primates (pp. 172-188). New ^(ork: Oxford University Press. Loehlin, J. C. (1992). Genes and environment in personality develop- ment. Newbury Park, CA: Sage. Loehlin, J. C, & Nichols, R. C. (1976). Heredity, environment, and personality: A study of 850 sets of twins. Austin: University of Texas Press. Loehlin, J. C, Willerman, L., & Horn, J. M. (1985). Personality resem- blances in adoptive families when the children are late-adolescent or adult. Journal of Personality and Social Psychology, 48, 376-392. Loehlin, J. C, Willerman, L., &. Horn, J. M. (1987). Personality resem- blances in adoptive families: A 10-year follow-up. Journal of Person- ality and Social Psychology, 53, 961-969. Lytton, H., Watts, D., & Dunn, B. E. (1988). Stability of genetic deter- mination from age 2 to age 9: A longitudinal twin study. Social Biol- ogy, 35, 62-73. McGue, M., Wette, R., & Rao, D. C. (1984). Evaluation of path analysis dirough computer simulation: Effect of incorrectly assuming indepen- dent distribution of familial correlations. Genetic Epidemiology, 1, 255-269. Mealey, L. (1995). The sociobiology of sociopathy: An integrated evolu- tionary model. Behavioral and Brain Sciences, 18, 523-599. Mednick, S. A., Moffitt, T, Gabrielli, J., & Hutchings, B. (1986). Ge- netic factors in criminal behavior: A review. In E Olweus, J. Block, & M. Radke-Yarrow (Eds.), Development of antisocial and prosociai behavior (pp. 33-50). New York; Academic Press. Moffitt, T. E. (1993). Adolescence-limited and life-course-persistent an- tisocial behavior: A developmental taxonomy. Psychological Review, 100, 674-701. O’Connor, M., Foch, T. T., Sherry, T, & Plomin, R. (1980). A twin study of specific behavioral problems of socialization as viewed by parents. Journal of Abnormal Child Psychology, 8, 189-199. Owen, D. R., & Sines, J. O. (1970). Heritability of personality in chil- dren. Behavior Genetics, 1, 235-247. Parke, R. D., & Slaby, R. G. (1983). The development of aggression. ARCHITECTURE OF HUMAN AGGRESSION 217 In P. H. Mussen (Series Ed.) & E. M. Hetherington (Vol. Ed.), Hand- book of child psychology: Vol. 4. Socialization, personality, and social development (4th ed., pp. 547-641). New M>rk: Wiley. Parker, T. (1989, June). Television viewing and aggression in four and seven year old children. Paper presented at Summer Minority Access to Research Training meeting, University of Colorado, Boulder. Partanen, J., Bruun, K., & Markkanen, T. (1966). Inheritance of drinking behavior: A study on intelligence, personality, and use of alcohol of adult twins. Helsinki, Finland: Keskuskirjapaino. Patterson, G. R. (1982). Coercive family process. Eugene, OR: Castalia Publishing. Patterson, G. R. (1992). Developmental changes in antisocial behavior. In R. DeV. Peters, R. J. McMahon, & V. L. Quinsey (Eds.), Aggres- sion and violence throughout the Hfespan (pp. 52-82). Newbury Park, CA: Sage. Patterson, G. R., Capaldi, D., & Bank, L. (1991). An early starter model for predicting delinquency. In D. J. Pepler & K. H. Rubin (Eds.), The development and treatment of childhood aggression (pp. 139-168). Hillsdale, NJ: Erlbaum. Patterson, G. R., DeBarsyshe, B. D., & Ramsey, E. (1989). A develop- mental perspective on antisocial behavior. American Psychologist, 44, 329-335. Perry, D. G-, Perry, L. C, & Boldizar, J. P. (1990). Learning of aggres- sion. In M. Lewis & S. M. Miller (Eds.), Handbook of developmental psychopathology (pp. 135-146). New brk: Plenum. Plomin, R., DeFries, J. C, & Fulker, D. W. (1988). Nature and nurture during infancy and early childhood. New brk: Cambridge University Press. Plomin, R., DeFries, J. C, & Loehlin, J. C. (1977). Genotype-environ- ment interaction and correlation in the analysis of human behavior. Psychological Bulletin, 84, 309-322. Plomin, R., Foch, T. T., & Rowe, D. C. (1981). Bobo clown aggression in childhood: Environment, not genes. Journal of Research in Person- ality, 15, 331-342. Pogue-Geile, M. K, & Rose, R. J. (1985). Developmental genetic studies of adult personality. Developmental Psychology, 21, 547-557. Rende, R. D., Slomkowski, C. L., Stocker, C, fiilka; D. W., & Plomin, R. (1992). Genetic and environmental influences on maternal and sibling interaction in middle childhood: A sibling adoption study. Developmental Psychology, 28, 484-490. Reznikoff, M., & Honeyman, M. S. (1967). MMPI profiles of monozy- gotic and dizygotic twin pairs. Journal of Consulting Psychology, 31, 100. Rowe, D. C. (1983). Biometrical genetic models of self-reported delin- quent behavior: A twin study. Behavior Genetics, 13, 473-489. Rowe, D. C. (1990). Inherited dispositions toward learning delinquent and criminal behavior: New evidence. In L. Ellis & H. Hoffman (Eds.), Crime in biological, social, and moral contexts (pp. 121- 133). New %rk: Praeger. Rowe, D. C. (1994). The limits of family influence: Genes, experience, and behavior. New “fork: Guilfbrd Press. Rowe, D. C, & Rodgers, J. L. (1989). Behavioral genetics, adolescent deviance, and “d”: Contributions and issues. In G. R. Adams, R. Montemayor, & T. P. Gullotta (Eds.), Biology of adolescent behavior and development. Advances in adolescent development: An annual book series (Vol. 1, pp. 38-67). Newbury Park, CA: Sage. Rowe, D. C, Rodgers, J. L., & Meseck-Bushey, S. (1992). Sibling delin- quency and the family environment: Shared and unshared influences. Child Development, 63, 59-67. Rushton, J. P., Fulker, D. W, Neale, M. C, Nias, D. K. B., & Eysenck, H. J. (1986). Altruism and aggression: The heritability of individual differences. Journal of Personality and Social Psychology, 50, 1192- 1198. Scan; S. (1966). The origins of individual differences in adjective check list scores. Journal of Consulting Psychology, 30, 354-357. Scarr, S., & McCartney, K. (1983). How people make their own environ- ments: A theory of genotype -* environment effects. Child Develop- ment, 54, 424-435. Stevenson, J., & Graham, P. (1988). Behavioral deviance in 13-year- old twins: An item analysis. Journal of Child and Adolescent Psychia- try, 791-797. Strayer, F. F. (1992). The development of agonistic and affiliative struc- tures in preschool play groups. In J. Silverberg & J. P. Gray (Eds.), Aggression and peacefulness in humans and other primates (pp. 150- 171). New York: Oxford University Press. Tellegen, A., Lykken, D. T, Bouchard, T. J., Wilcox, K. J., Segal, N. L., & Rich, S. (1988). Personality similarity in twins reared apart and together. Journal of Personality and Social Psychology, 54, 1031 — 1039. ZHlmann, D. (1988). Cognition-excitation interdependencies in aggres- sive behavior. Aggressive Behavior, 14, 51-64. Zuckerman, M. (1989). Personality in the third dimension: A psychobio- logical approach. Personality and Individual Differences, 10, 391- 418. Received August 29, 1995 Revision received March 18, 1996 Accepted March 23, 1996 •
Behavioral genetic research designs have often been attacked because they rely on comparing monozygotic twins (MZ) to dizygotic twins (DZ). Critics of twin-based research maintain that MZ twins look m
7 0 2 VO L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 Nat u r e G eNe t i c s A n Al y s i s Despite a century of research on complex traits in humans, the relative importance and specific nature of the influences of genes and environment on human traits remain controversial. We report a meta-analysis of twin correlations and reported variance components for 17,804 traits from 2,748 publications including 14,558,903 partly dependent twin pairs, virtually all published twin studies of complex traits. Estimates of heritability cluster strongly within functional domains, and across all traits the reported heritability is 49%. For a majority (69%) of traits, the observed twin correlations are consistent with a simple and parsimonious model where twin resemblance is solely due to additive genetic variation. The data are inconsistent with substantial influences from shared environment or non-additive genetic variation. This study provides the most comprehensive analysis of the causes of individual differences in human traits thus far and will guide future gene-mapping efforts. All the results can be visualized using the MaTCH webtool. Specifically, the partitioning of observed variability into underlying genetic and environmental sources and the relative importance of additive and non-additive genetic variation are continually debated 1– 5. Recent results from large-scale genome-wide association studies (GWAS) show that many genetic variants contribute to the variation in complex traits and that effect sizes are typically small 6 ,7. However, the sum of the variance explained by the detected variants is much smaller than the reported heritability of the trait 4 ,6– 10 . This ‘missing heritability’ has led some investigators to conclude that non-additive variation must be important 4 ,1 1 . Although the presence of gene-gene interaction has been demonstrated empirically 5 ,12– 17 , little is known about its relative contribution to observed variation1 8. In this study, our aim is twofold. First, we analyze empirical esti – mates of the relative contributions of genes and environment for virtually all human traits investigated in the past 50 years. Second, we assess empirical evidence for the presence and relative importance of non-additive genetic influences on all human traits studied. We rely on classical twin studies, as the twin design has been used widely to disentangle the relative contributions of genes and environment, across a variety of human traits. The classical twin design is based on contrasting the trait resemblance of monozygotic and dizygotic twin pairs. Monozygotic twins are genetically identical, and dizygotic twins are genetically full siblings. We show that, for a majority of traits (69%), the observed statistics are consistent with a simple and parsi – monious model where the observed variation is solely due to additive genetic variation. The data are inconsistent with a substantial influence from shared environment or non-additive genetic variation. We also show that estimates of heritability cluster strongly within functional domains, and across all traits the reported heritability is 49%. Our results are based on a meta-analysis of twin correlations and reported variance components for 17,804 traits from 2,748 publications includ – ing 14,558,903 partly dependent twin pairs, virtually all twin studies of complex traits published between 1958 and 2012. This study provides the most comprehensive analysis of the causes of individual differences in human traits thus far and will guide future gene-mapping efforts. All results can be visualized with the accompanying MaTCH webtool. RESULTS The distribution of studied traits is nonrandom We systematically retrieved published classical twin studies in which observed variation in human traits was partitioned into genetic and environmental influences. For each study, we collected reported Meta-analysis of the heritability of human traits based on fifty years of twin studies Tinca J C Polderman 1 ,10 , Beben Benyamin 2 ,10 , Christiaan A de Leeuw 1 ,3 , Patrick F Sullivan 4 – 6 , Arjen van Bochoven 7, Peter M Visscher 2 ,8 ,11 & Danielle Posthuma 1 ,9 ,11 1Department of Complex Trait Genetics, VU University, Center for Neurogenomics and Cognitive Research, Amsterdam, the Netherlands. 2Queensland Brain Institute, University of Queensland, Brisbane, Queensland, Australia. 3Institute for Computing and Information Sciences, Radboud University Nijmegen, Nijmegen, the Netherlands. 4Center for Psychiatric Genomics, Department of Genetics, University of North Carolina, Chapel Hill, North Carolina, USA. 5Department of Psychiatry, University of North Carolina, Chapel Hill, North Carolina, USA. 6Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Stockholm, Sweden. 7Faculty of Sciences, VU University, Amsterdam, the Netherlands. 8University of Queensland Diamantina Institute, Translational Research Institute, Brisbane, Queensland, Australia. 9Department of Clinical Genetics, VU University Medical Center, Neuroscience Campus Amsterdam, Amsterdam, the Netherlands. 10These authors contributed equally to this work. 11These authors jointly supervised this work. Correspondence should be addressed to D.P. ( [email protected]). Received 13 February; accepted 1 April; published online 18 May 2015; doi:10.1038/ng.328 5 Insight into the nature of observed variation in human traits is impor – tant in medicine, psychology, social sciences and evolutionary biology. It has gained new relevance with both the ability to map genes for human traits and the availability of large, collaborative data sets to do so on an extensive and comprehensive scale. Individual differences in human traits have been studied for more than a century, yet the causes of variation in human traits remain uncertain and controversial. npg © 2015 Nature America, Inc. All rights reserved. Nat u r e G eNe t i c s VO L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 7 0 3 twin correlations for continuous traits and contingency tables for dichotomous traits, estimates from genetic model-fitting and study characteristics (sample size, population, age cohort and ascertain – ment scheme) ( Supplementary Table 1 and Supplementary Note ). We manually classified the investigated traits using the chapter and subchapter levels of the International Classification of Functioning, Disability and Health (ICF) or the International Statistical Classification of Diseases and Related Health Problems (ICD- 10) (Online Methods). ICD-10 and ICF subchapter levels refer to actual diseases (for example, atopic dermatitis) and traits (for example, temperament and personality functions), respectively. We identified 2,748 relevant twin studies, published between 1958 and 2012. Half of these were published after 2004, with sample sizes per study in 2012 of around 1,000 twin pairs ( Supplementary Table 2). Each study could report on multiple traits measured in one or several samples. These 2,748 studies reported on 17,804 traits. Twin subjects came from 39 different countries, with a large proportion of studies (34%) based on US twin samples. The continents of South America (0.5%), Africa (0.2%) and Asia (5%) were heavily under- represented ( Fig. 1a,b and Supplementary Table 3 ). The total number of studied twins was 14,558,903 partly dependent pairs, or 2,247,128 when correcting for reporting on multiple traits per study. The majority of studies (59%) were based on the adult population (aged 18–64 years), although the sample sizes available for studies of the elderly population (aged 65 years or older) were the largest ( Supplementary Table 4). Authorship network analyses showed that 61 communities of authors wrote the 2,748 published studies. The 11 largest authorship communities contained >65 authors and could be mapped back to the main international twin registries, such as the Vietnam Era Twin Registry, the Finnish Twin Cohort and the Swedish Twin Registry ( Supplementary Fig. 1 ). The investigated traits fell into 28 general trait domains. The dis – tribution of the traits evaluated in twin studies was highly skewed, with 51% of studies focusing on traits classified under the psychi – atric, metabolic and cognitive domains, whereas traits classified under the developmental , connective tissue and infection domains together accounted for less than 1% of all investigated traits ( Fig. 1c and Supplementary Tables 5 –7). The ten most investigated traits were temperament and personality functions, weight maintenance functions, general metabolic functions, depressive episode, higher- level cognitive functions, conduct disorders, mental and behavioral disorders due to use of alcohol, anxiety disorders, height and mental and behavioral disorders due to use of tobacco. Collectively, these traits accounted for 59% of all investigated traits. 0 200 400 600 800 1,000 1,200 Connective tissue Developmental InfectionAging Mortality Neoplasms Cell Hematological Social values Muscular Gastrointestinal Nutritional Dermatological Social interactions Ear, nose, throat Immunological Activities Ophthalmological Respiratory Reproduction Endocrine Environment Cardiovascular Skeletal Neurological Cognitive Metabolic Psychiatric a b c d Studies (n)Continuous 89% Dichotomous 10% Dichotomous, ascertained 1% Other, non- disease75% Disease 8%Symptoms of disease17% Number of investigated traits Average number of twin pairs per study 1 2 3 4 –0.5 0 0.5 1.0 rMZ R2 = 0.009468 1 2 3 4 –0.5 0 0.5 1.0 rDZ R2 = 0.0003032 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 0 0.2 0.4 0.6 0.8 1.0 log 10 (n pairs) h2 R2 = 0.001651 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 0 0.2 0.4 0.6 0.8 log 10 (n pairs) c2 R2 = 0.005001 1 6,74610 2,104 Figure 1 Distribution of the investigated traits in virtually all twin studies published between 1958 and 2012. ( a) The number of investigated traits in classical twin studies across all countries. ( b) The average number of twin pairs included per study across countries. ( c) The number of investigated traits according to functional trait domain and trait characteristic (inset). ( d) Monozygotic and dizygotic twin correlations and reported estimates of h 2 and c 2 as a function of sample size. Contour lines indicate the density of the data in that region. The lines are colored by ‘heat’ from blue to red, indicating increasing data density. A n Al y s i s npg © 2015 Nature America, Inc. All rights reserved. 7 0 4 V O L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 Nat u r e G e Ne t i c s A nAl y s i s Equal contribution of genes and environment We did not find evidence of systematic publication bias as a function of sample size (for example, where studies based on relatively small sam – ples were only published when larger effects were reported) ( Fig. 1d, Supplementary Figs. 2 –6 and Supplementary Tables 8 –11 ). We cal – culated the weighted averages of correlations for monozygotic ( r MZ ) and dizygotic ( r DZ) twins and of the reported estimates of the relative contributions of genetic and environmental influences to the investi – gated traits using a random-effects meta-analytic model to allow for heterogeneity across different studies ( Supplementary Tables 12–15). The meta-analyses of all traits yielded an average r MZ of 0.636 (s.e.m. = 0.002) and an average r DZ of 0.339 (s.e.m. = 0.003). The reported heritability ( h 2) across all traits was 0.488 (s.e.m. = 0.004), and the reported estimate of shared environmental effects ( c 2) was 0.174 (s.e.m. = 0.004) ( Fig. 2a,b , Table 1 and Supplementary Fig. 7 ). Estimates of h2 and c2 cluster in functional domains We found that heritability estimates clustered in functional domains, with the largest heritability estimates for traits classified under the ophthalmological domain ( h 2 = 0.712, s.e.m. = 0.041), followed by the ear, nose and throat ( h 2 = 0.637, s.e.m. = 0.064), dermatological ( h 2 = 0.604, s.e.m. = 0.043) and skeletal ( h 2 = 0.591, s.e.m. = 0.018) domains. The lowest heritability estimates were for traits in the environment, reproduction and social values domains ( Fig. 2d and Supplementar y Table 16 ). All weighted averages of h 2 across >500 distinct traits had a mean greater than zero ( Supplementary Tables 17 –24 ). The lowest reported heritability for a specific trait was for gene expression, with an estimated h 2 = 0.055 (s.e.m. = 0.026) and an estimated c 2 of 0.736 (s.e.m. = 0.033) (but note that these trait aver – ages are based on reported estimates of variance components derived from only 20 data points reporting on the expression levels of 20 genes; Supplementary Table 21 ). We found the largest influence of c 2 for traits in the cell domain ( c 2 = 0.674, s.e.m. = 0.048), followed by traits in the infection ( c 2 = 0.351, s.e.m. = 0.153), hematological ( c 2 = 0.324, s.e.m. = 0.090), endocrine ( c 2 = 0.322, s.e.m. = 0.050), reproduction ( c 2 = 0.320, s.e.m. = 0.061), social values ( c 2 = 0.271, s.e.m. = 0.032), environment ( c 2 = 0.269, s.e.m. = 0.020) and skeletal ( c 2 = 0.265, s.e.m. = 0.019) domains ( Fig. 2d and Supplementary Table 16 ). 0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0 All traits Skeletal Reproduction Ear, nose, throat Metabolic Ophthalmological Dermatological Respiratory Neurological CognitiveActivities Cardiovascular Endocrine Psychiatric Environment Gastrointestinal Social values Nutritional Social interactions d rMZ rMZM rMZF rDZ rDZSS rDZM rDZF rDOS h2 h2 SS h2 M h2 F c2 c2 SS c2 M c2 F c2 F Correlation (r) Frequency –0.5 0 0.5 1.00 50 150 250 MZ DZ a 0 0.2 0.4 0.6 0.8 1.0 rMZ rMZM rMZF rDZ rDZSS rDZM rDZF rDOS 0 0.2 0.4 0.6 0.8 1.0 0–11 12–17 18–64 Age (years) 65+ h2 h2 SS h2 M h2 F c2 c2 SS c2 M c –0.5 00.5 1.0 –0.50 0.5 1.0 rMZ rDZ b R2 = 0.4167 Figure 2 Twin correlations and heritabilities for all human traits studied. ( a) Distribution of r MZ and r DZ estimates across the traits investigated in 2,748 twin studies published between 1958 and 2012. r MZ estimates are based on 9,568 traits and 2,563,628 partly dependent twin pairs; r DZ estimates are based on 5,220 traits and 2,606,252 partly dependent twin pairs ( Table 1). ( b) Relationship between r MZ and r DZ, using all 5,185 traits for which both were reported. ( c) Random-effects meta-analytic estimates of twin correlations (top) and reported variance components (bottom) across all traits separately for four age cohorts. Error bars, standard errors. ( d) Random-effects meta-analytic estimates of twin correlations (top) and reported variance components (bottom) across all traits, and within functional domains for which data on all correlations and variance components were available. Error bars, standard errors. npg © 2015 Nature America, Inc. All rights reserved. Nat u r e G eNe t i c s V O L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 7 0 5 A nAl y s i s Heterogeneity of twin correlations across sex and age Across all traits, the weighted averages of twin correlations and reported h 2 and c 2 values did not show evidence of heterogeneity by sex, although there was some evidence for lower correlation in opposite- sex twin pairs in comparison to same-sex dizygotic twin pairs ( Table 1 and Supplementary Note ). The data showed a decrease in monozygotic and dizygotic twin resemblance after adolescence and an accompanying decrease in the estimates of both h 2 and c 2 (Fig. 2 c and Supplementary Table 15 ). In the top 20 most investigated traits for twin correlations, the weighted estimates did not show consistent evidence for heterogene – ity by sex, with r MZM (monozygotic male twin correlation) and r MZF (monozygotic female twin correlation) as well as r DZM (dizygotic male twin correlation) and r DZF (dizygotic female twin correlation) remarkably similar across the majority of the top 20 specific traits investigated ( Fig. 3), although for several traits the correlations for opposite-sex twin pairs were lower than the estimates for same-sex twin pairs, mostly after age 11 years (for example, for weight mainte – nance functions, functions of brain, mental and behavioral disorders due to the use of alcohol, and mental and behavioral disorders due to the use of tobacco ). Heterogeneity of weighted twin correlations by age was more prominent than heterogeneity by sex ( Fig. 3). For example, when considering r MZ , we note that for most of the top 20 investigated traits the estimate tended to decrease with age, especially after adolescence, a trend that was generally mirrored in the r DZ estimates ( Fig. 3). Meta-analysis results across all traits across different countries are provided in Supplementary Table 25. Model-fitting and selection leads to underestimation of h2 Falconer’s equations can be used to calculate ĥ 2 and ĉ 2 on the basis of twin correlations 1 8. In these equations, ĥ 2 = 2 × (r MZ − r DZ) and ĉ 2 = 2 × r DZ − r MZ . When these are applied using the weighted averages of r MZ and r DZ, we find an ĥ 2 estimate of 2 × (0.636 − 0.339) = 0.593 and a ĉ 2 estimate of 2 × 0.339 − 0.636 = 0.042 ( Table 1 and Supplementary Fig. 8 ). We note that the ĥ 2 estimate based on twin correlations is larger than the weighted average of reported h 2 values (Supplementary Figs. 9 and 10). As a consequence, the ĉ2 estimate based on twin correlations is lower than the weighted average of the reported c 2 component. To test whether this discrepancy was due to a bias in studies reporting only twin correlations or only variance compo – nents, we conducted the meta-analysis only on studies reporting both. This analysis yielded similar estimates with a similar discrepancy ( Supplementary Table 13 ), ruling out the explanation that twin cor – relations may have been reported on traits that happened to be more heritable than traits for which the estimates of variance components were reported. Through theory, we show that such a discrepancy can arise when individual studies represent a mixture of traits that follow a pattern of r MZ > 2r DZ and r MZ < 2r DZ and where the choice of fitting a model that includes shared environment or non-additive genetic influences is based on the observed pattern of twin correlations ( Supplementary Note ). More specifically, because c 2 and non-additive genetic influences cannot be estimated simultaneously from twin correlations, an ACE model (for additive genetic (A), common envi – ronmental (C), and error or non-shared environmental (E) influences) is fitted to the data if 2 r DZ − r MZ > 0. In contrast, if 2 r DZ − r MZ < 0, an ADE model, including non-additive genetic (D) instead of com – mon environmental (C) influences, is selected. This leads to sampling bias in estimating h 2 from the full model. We show ( Supplementary Table 26 ) that such per-study choices cause bias and can lead to a 10% downward bias in the reported estimates of h 2 in comparison to those based on twin correlations, consistent with the observed discrepancy between our meta-analysis of variance component estimates calculated from twin correlations and the reported variance components. Overall twin correlations imply a simple additive model There may be many causes of similarities and differences within monozygotic and dizygotic twin pairs, and these are typically inter – preted in terms of additive or non-additive genetic influences and shared or non-shared environmental influences 1 9. Yet, there are essentially only two estimable and testable variance components of interest in the twin design. Therefore, inference from classical twin studies on all underlying, unobserved sources of variation that lead to the resemblance between relatives is limited. However, there are two simple and parsimonious hypotheses that can be tested across traits from estimated correlation coefficients for monozygotic twin pairs ( r MZ ) and dizygotic twin pairs ( r DZ). The first is that the cor- relations for the monozygotic and dizygotic twin populations ( ρMZ and ρDZ) are the same, consistent with twin resemblance being solely due to non-genetic factors. The second hypothesis involves a two – fold ratio of ρMZ to ρDZ, consistent with twin resemblance being solely due to additive genetic variation. Across-trait consistency with either of these hypotheses is not a proof of these simple models but would provide an extremely parsimonious model against which other experimental designs (for example, DNA sequence–based studies) should be tested. For the vast majority of traits (84%), Table 1 Weighted means of twin correlations and variance components across all human traits investigated in a classical twin study and published between 1958 and 2012 Statistic Estimate (s.e.m.)n traitsn pairs r MZ 0.636 (0.002) 9,5682,563,627 r MZM 0.617 (0.004) 4,5181,070,962 r MZF 0.626 (0.004) 4,3601,171,841 r DZ 0.339 (0.003) 5,2202,606,252 r DZSS 0.345 (0.003) 6,1081,752,952 r DZM 0.321 (0.003) 4,4121,039,238 r DZF 0.342 (0.004) 4,2551,068,562 r DOS 0.302 (0.005) 2,342898,610 h 2 0.488 (0.004) 2,9294,341,721 h 2 (same sex) 0.471 (0.005)1,7951,187,837 h2 (male) 0.465 (0.005)2,0951,732,622 h2 (female) 0.472 (0.005)1,9571,539,582 c 2 0.174 (0.004) 2,7714,272,318 c 2 (same sex) 0.189 (0.005)1,7691,185,116 c 2 (male) 0.157 (0.004)1,9881,519,148 c 2 (female) 0.169 (0.005)1,9251,516,192 2(r MZ − r DZ) 0.593 (0.008) 9,5685,169,879 2(r MZ − r DZ) (same sex) 0.581 (0.008) 9,5684,316,578 2(r MZ − r DZ) (male) 0.593 (0.010) 4,5182,110,200 2(r MZ − r DZ) (female) 0.569 (0.010) 4,3602,240,403 2r DZ − r MZ 0.042 (0.007) 9,5685,169,879 2r DZ − r MZ (same sex) 0.055 (0.006) 9,5684,316,578 2r DZ − r MZ (male) 0.025 (0.008) 4,5182,110,200 2r DZ − r MZ (female) 0.057 (0.008) 4,3602,240,403 r, correlation; MZ, monozygotic twins; DZ, dizygotic twins; MZM, monozygo tic twins, male; MZF, monzygotic twins, female; DZSS, dizygotic twins, same sex; DZM, dizygo tic twins, both male; DZF, dizygotic twins, both female; DOS, dizygotic twins, opposite sex; h 2, heritability; c 2, proportion of variance due to shared environmental variation; estimate, estimate based on random-effects meta-analysis; n traits, number of investigated traits; n pairs, number of dependent twin pairs. The pairs are not independent, as the same or an overlapping sample of twins may have been used for multiple traits and across multiple studies. npg © 2015 Nature America, Inc. All rights reserved. 7 0 6 V O L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 Nat u r e G e Ne t i c s A nAl y s i s we found that monozygotic twin correlations were larger than dizy – gotic twin correlations. Using the weighted estimates of r MZ and r DZ across all traits, we showed that, on average, 2 r DZ − r MZ = 0.042 (s.e.m. = 0.007) ( Table 1), which is very close to a twofold differ – ence in the correlation of monozygotic twins relative to dizygotic twins ( Supplementar y Figs. 11 and 12). The proportion of single Blood pressure funct. Conduct dis. Depr. episode Endocr. gland funct. Food Funct. of brain General metab. funct. Heart funct. Height High-L. cognitive funct. Hyperkin- etic dis. Imm. system funct. Ment. beh. dis. alc. Ment. beh. dis. tob. Other anxiety dis. Spec. personal. dis. Structure of the eyeball Structure of mouth Temp. pers. funct. Weight maint.funct. rMZ 0.47 0.34 0.430.520.85 0.60 0.40 0.50 0.47 0.790.36 0.59 rMZM 0.630.29 0.58 0.42 0.45 0.60 0.88 0.48 0.50 0.34 0.59 rMZF 0.410.440.52 0.46 0.90 0.52 0.29 0.50 0.710.720.35 0.63 rDZ 0.260.23 0.270.17 0.31 0.490.16 0.23 rDZSS 0.140.09 0.290.250.49 0.25 0.19 0.33 0.33 0.14 0.33 rDZM 0.220.08 0.46 0.21 0.13 0.38 0.50 0.28 0.23 0.10 0.24 rDZF 0.100.310.19 0.22 0.49 0.23 0.12 0.32 0.430.380.15 0.35 rDOS 0.210.24 0.330.15 0.25 0.18 0.12 0.14 0 0.2 0.4 0.6 0.8 1.0 Age 65+ years Blood pressure funct. Conduct dis. Depr. episode Endocr. gland funct. Food Funct. of brain General metab. funct. Heart funct. Height High-L. cognitive funct. Hyperkin- etic dis. Imm. system funct. Ment. beh. dis. alc. Ment. beh. dis. tob. Other anxiety dis. Spec. personal. dis. Structure of the eyeball Structure of mouth Temp. pers. funct. Weight maint.funct. rMZ 0.60 0.63 0.50 0.52 0.49 0.71 0.78 0.55 0.94 0.69 0.67 0.70 0.77 0.82 0.55 0.71 0.85 0.75 0.51 0.87 rMZM 0.62 0.56 0.49 0.59 0.69 0.78 0.58 0.94 0.60 0.66 0.75 0.80 0.510.83 0.86 0.52 0.85 rMZF 0.55 0.63 0.520.55 0.72 0.75 0.53 0.91 0.55 0.69 0.75 0.84 0.500.83 0.86 0.50 0.87 rDZ 0.31 0.38 0.31 0.31 0.32 0.45 0.30 0.49 0.38 0.23 0.41 0.56 0.59 0.35 0.29 0.44 0.38 0.21 0.42 rDZSS 0.36 0.43 0.36 0.42 0.17 0.36 0.45 0.36 0.58 0.38 0.21 0.63 0.62 0.320.39 0.42 0.24 0.51 rDZM 0.44 0.36 0.33 0.27 0.29 0.43 0.35 0.50 0.36 0.21 0.62 0.64 0.310.35 0.57 0.26 0.49 rDZF 0.29 0.42 0.360.42 0.41 0.46 0.36 0.54 0.35 0.31 0.58 0.60 0.330.42 0.43 0.29 0.50 rDOS 0.22 0.33 0.28 0.54 0.30 0.41 0.29 0.44 0.27 0.15 0.46 0.49 0.270.39 0.31 0.19 0.30 0 0.2 0.4 0.6 0.8 1.0 Age 12–17 years Blood pressure funct. Conduct dis. Depr. episode Endocr. gland funct. Food Funct. of brain General metab. funct. Heart funct. Height High-L. cognitive funct. Hyperkin- etic dis. Imm. system funct. Ment. beh. dis. alc. Ment. beh. dis. tob. Other anxiety dis. Spec. personal. dis. Structure of the eyeball Structure of mouth Temp. pers. funct. Weight maint.funct. rMZ 0.59 0.67 0.39 0.53 0.42 0.65 0.65 0.52 0.92 0.68 0.58 0.56 0.55 0.69 0.41 0.41 0.68 0.89 0.42 0.76 rMZM 0.54 0.55 0.40 0.52 0.42 0.70 0.63 0.53 0.91 0.57 0.58 0.62 0.53 0.64 0.36 0.39 0.79 0.42 0.70 rMZF 0.53 0.52 0.44 0.58 0.41 0.60 0.66 0.54 0.89 0.44 0.62 0.52 0.49 0.72 0.39 0.40 0.710.42 0.73 rDZ 0.29 0.34 0.18 0.37 0.20 0.19 0.36 0.24 0.47 0.28 0.26 0.30 0.30 0.41 0.17 0.22 0.33 0.52 0.21 0.34 rDZSS 0.30 0.43 0.21 0.34 0.24 0.24 0.33 0.29 0.53 0.27 0.31 0.38 0.36 0.44 0.16 –0.09 0.330.17 0.39 rDZM 0.25 0.36 0.18 0.32 0.22 0.39 0.32 0.21 0.53 0.34 0.35 0.29 0.31 0.37 0.14 0.25 0.330.17 0.36 rDZF 0.33 0.34 0.22 0.37 0.22 0.35 0.36 0.30 0.51 0.25 0.34 0.30 0.31 0.47 0.16 0.20 0.300.19 0.36 rDOS 0.28 0.29 0.14 0.26 0.14 0.10 0.30 0.16 0.450.250.22 0.26 0.15 0.14 0.32 0.13 0.25 –0.20 0.2 0.4 0.6 0.8 1.0 Age 18–64 years Blood pressure funct. Conduct dis. Depr. episode Endocr. gland funct. Food Funct. of brain General metab. funct. Heart funct. Height Higher- level cognitive funct. Hyperkin- etic dis. Imm. system funct. Ment. beh. dis. alc. Ment. beh. dis. tob. Other anxiety dis. Spec. personal. dis. Structure of the eyeball Structure of mouth Temp. pers. funct. Weight maint.funct. rMZ 0.54 0.68 0.60 0.61 0.81 0.69 0.81 0.49 0.88 0.77 0.65 0.64 0.67 0.62 0.68 0.75 0.84 0.61 0.84 rMZM 0.68 0.62 0.73 0.77 0.71 0.78 0.41 0.91 0.71 0.640.580.690.62 0.83 rMZF 0.70 0.66 0.58 0.76 0.72 0.76 0.44 0.88 0.72 0.61 0.620.710.61 0.85 rDZ 0.39 0.43 0.42 0.44 0.52 0.30 0.47 0.26 0.59 0.54 0.27 0.38 0.41 0.32 0.33 0.40 0.34 0.54 rDZSS 0.41 0.44 0.39 0.22 0.55 0.33 0.52 0.33 0.61 0.58 0.24 0.41 0.43 0.410.30 0.51 0.33 0.56 rDZM 0.45 0.39 0.08 0.56 0.32 0.36 0.26 0.56 0.48 0.18 0.420.250.39 0.54 rDZF 0.42 0.40 0.32 0.62 0.32 0.54 0.26 0.59 0.50 0.20 0.360.410.37 0.56 rDOS 0.42 0.43 0.300.26 0.55 0.47 0.23 0.400.30 0.48 0.39 0.51 0 0.2 0.4 0.6 0.8 1.0 Age 0–11 years Figure 3 Twin correlations for the top 20 most investigated specific traits by age and sex. Alc., alcohol; dis., disorders; depr., depressive; endocr., endocrine; imm., immunological; funct., functions; maint., maintenance; metab., metabolic; ment. beh., mental and behavioral; spec. personal., specific personality; temp. pers., temperament and personality; tob., tobacco; r, correlation; MZ, monozygotic twins; DZ, dizygotic twins; M, males; F, females; SS, same-sex pairs only; DOS, dizygotic opposite-sex pairs. Inclusion for the top 20 most investigated traits was conditional on the reporting of r MZ and r DZ. Empty cells denote insufficient information available to calculate weighted estimates; error bars, standard errors. We note that estimates and graphs for all specific traits are available from the online MaTCH webtool. npg © 2015 Nature America, Inc. All rights reserved. Nat u r e G eNe t i c s V O L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 7 0 7 A nAl y s i s Table 2 Weighted means of twin correlations and proportion ( π0) of studies that are consistent with a model where trait resemblance is solely due to additive genetic variation for the main trait domains and the top 20 most investigated traits π0 rMZ rDZ n traits Estimate Estimate (s.e.m.) n traitsn pairsEstimate (s.e.m.) n traitsn pairs All traits 5,1850.690.636 (0.002) 9,5682,563,628 0.339 (0.003) 5,2202,606,252 General trait domains Activities 620.35 0.570 (0.019) 11858,227 0.340 (0.022) 6355,864 Cardiovascular 2670.95 0.564 (0.008) 38041,669 0.295 (0.010) 26825,544 Cell 540.59 0.722 (0.022) 723,1880.523 (0.043) 541,667 Cognitive 4500.57 0.646 (0.007) 931288,867 0.371 (0.010) 454304,720 Dermatological 740.45 0.729 (0.025) 10919,509 0.402 (0.017) 7523,245 Ear, nose, throat 1650.97 0.760 (0.013) 20027,882 0.332 (0.015) 17214,222 Endocrine 1080.69 0.555 (0.017) 16210,112 0.387 (0.022) 1109,140 Environment 1450.50 0.551 (0.014) 295120,606 0.396 (0.017) 14599,137 Gastrointestinal 320.59 0.551 (0.024) 6410,982 0.274 (0.028) 3928,431 Hematological 190.65 0.764 (0.023) 505,5410.560 (0.032) 193,218 Immunological 2300.67 0.608 (0.012) 28018,051 0.357 (0.013) 23136,075 Metabolic 464 0.60 0.746 (0.005) 912210,189 0.405 (0.008) 464197,921 Neurological 7021.00 0.685 (0.005) 1,751129,076 0.289 (0.006) 70589,103 Nutritional 1100.72 0.479 (0.016) 20575,751 0.220 (0.015) 11079,188 Ophthalmological 1060.87 0.730 (0.017) 19926,139 0.385 (0.017) 10616,189 Psychiatric 1,7780.620.552 (0.004) 2,8651,232,382 0.306 (0.005) 1,7811,374,817 Reproduction 160.44 0.767 (0.034) 3412,130 0.333 (0.063) 1627,879 Respiratory 1250.74 0.697 (0.018) 18434,443 0.325 (0.019) 12751,150 Skeletal 1900.51 0.830 (0.008) 395111,282 0.504 (0.012) 191113,080 Social interactions 240.63 0.338 (0.017) 14643,501 0.267 (0.041) 2422,764 Social values 450.69 0.489 (0.030) 12052,492 0.414 (0.062) 4528,071 Top 20 investigated traits for r MZ and r DZ Blood pressure functions 1100.93 0.581 (0.010) 17920,621 0.307 (0.013) 11011,620 Conduct disorder 2160.41 0.663 (0.009) 289147,974 0.408 (0.010) 216192,651 Depressive episode 1150.60 0.454 (0.014) 17398,315 0.253 (0.015) 115121,936 Endocrine gland functions 920.72 0.538 (0.017) 1398,5330.382 (0.025) 927,295 Food 1100.72 0.479 (0.016) 20575,751 0.220 (0.015) 11079,188 Functions of brain 5940.99 0.676 (0.006) 1,01069,7220.287 (0.006) 59458,951 General metabolic functions 2190.69 0.682 (0.007) 46262,108 0.371 (0.010) 21958,338 Heart functions 1401.00 0.529 (0.009) 17415,070 0.268 (0.011) 14011,109 Height 870.29 0.908 (0.005) 12853,076 0.543 (0.008) 8768,358 Higher-level cognitive functions 1880.44 0.710 (0.009) 419152,197 0.441 (0.016) 188158,626 Hyperkinetic disorders 1000.37 0.651 (0.013) 14486,450 0.260 (0.016) 100121,139 Immunological system functions 2230.67 0.606 (0.012) 27616,703 0.357 (0.013) 22332,964 Mental and behavioral disorders due to the use of alcohol 100 0.36 0.630 (0.015) 15894,477 0.409 (0.020) 10194,196 Mental and behavioral disorders due to the use of tobacco 70 0.47 0.719 (0.016) 11051,102 0.468 (0.022) 7234,186 Other anxiety disorders 1450.29 0.548 (0.013) 191105,902 0.327 (0.016) 145153,730 Specific personality disorders 1400.93 0.448 (0.009) 16241,460 0.225 (0.007) 14033,681 Structure of the eyeball 860.91 0.735 (0.022) 12119,276 0.370 (0.019) 8613,580 Structure of mouth 1170.89 0.819 (0.010) 1277,7690.399 (0.012) 1198,493 Temperament and personality functions 5680.84 0.470 (0.008) 1,134334,190 0.234 (0.010) 568296,114 Weight maintenance functions 2150.48 0.810 (0.005) 391141,152 0.437 (0.010) 215134,867 General trait domain categories with <10 entries for π0 were excluded. For definitions of abbreviations, see Table 1. Inclusion of the top 20 investigated traits was conditional on the reporting of r MZ and r DZ. studies in which the pattern of twin correlations was consistent with the null hypothesis that 2 r DZ = r MZ was 69%. This observed pattern of twin correlations is consistent with a simple and parsimoni – ous underlying model of the absence of environmental effects shared by twin pairs and the presence of genetic effects that are entirely due to additive genetic variation ( Table 2). This remarkable fitting of the data with a simple mode of family resemblance is inconsistent with the hypothesis that a substantial part of variation in human traits is due to shared environmental variation or to substantial non-additive genetic variation. Most specific traits follow an additive genetic model Although across all traits 69% of studies showed a pattern of monozy – gotic and dizygotic twin correlations consistent with an r MZ that was exactly twice the r DZ, this finding is not necessarily representative of the majority of studies in functional domains or for every specific trait (that is, at the ICD-10 or ICF subchapter level). We thus calculated the proportion of studies consistent with 2 r DZ = r MZ within functional domains and for each specific trait and found that traits consistent with this hypothesis tended to cluster in specific functional domains ( Supplementar y Tables 27 –29 ). A pattern of twin correlations npg © 2015 Nature America, Inc. All rights reserved. 7 0 8 V O L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 Nat u r e G e Ne t i c s A nAl y s i s consistent with 2 rDZ = r MZ was most prominent for traits included in the neurological, ear, nose and throat, cardiovascular and oph – thalmological domains, with 99.5%, 97%, 95% and 87% of studies, respectively, being consistent with a model where all resemblance was entirely due to additive genetic variance. In only 3 of 28 general trait domains were most studies inconsistent with this model. These domains were activities (35%), reproduction (44%) and dermatologi – cal (45%) ( Table 2 and Supplementary Table 27 ). Of the 59 specific traits (ICD-10 or ICF subchapter classifications) for which we had sufficient information to calculate the proportion of studies consistent with 2 r DZ = r MZ , 21 traits showed a proportion less than 0.50, whereas for the remaining 38 traits the majority of individual studies were con – sistent with 2 r DZ = r MZ (Supplementary Table 29 ). Of the top 20 most investigated specific traits, we found that for 12 traits the majority of individual studies were consistent with a model where variance was solely due to additive genetic variance and non-shared environ – mental variance, whereas the pattern of monozygotic and dizygotic twin correlations was inconsistent with this model for 8 traits, sug – gesting that, apart from additive genetic influences and non-shared environmental influences, either or both non-additive genetic influ- ences and shared environmental influences are needed to explain the observed pattern of twin correlations ( Table 2). These eight traits were conduct disorders, height, higher-level cognitive functions, hyper – kinetic disorders, mental and behavioral disorders due to the use of alcohol, mental and behavioral disorders due to the use of tobacco , other anxiety disorders and weight maintenance functions . For all eight traits, meta-analyses on reported variance components resulted in a weighted estimate of reported shared environmental influences that was statistically different from zero ( Supplementary Table 21). Comparison of weighted twin correlations for these specific traits resulted in positive estimates of 2 r DZ − r MZ , except for hyperkinetic disorders, where 2 r DZ − r MZ was −0.130 (s.e.m. = 0.034) on the basis of 144 individual reports and 207,589 twin pairs, which suggests the influence of non-additive genetic variation for this trait or any other source of variation that leads to a disproportionate similarity among monozygotic twin pairs. DISCUSSION We have conducted a meta-analysis of virtually all twin studies pub – lished in the past 50 years, on a wide range of traits and reporting on more than 14 million twin pairs across 39 different countries. Our results provide compelling evidence that all human traits are heritable: not one trait had a weighted heritability estimate of zero. The relative influences of genes and environment are not randomly distributed across all traits but cluster in functional domains. In general, we showed that reported estimates of variance components from model-fitting can underestimate the true trait heritability, when compared with heritability based on twin correlations. Roughly two-thirds of traits show a pattern of monozygotic and dizygotic twin correlations that is consistent with a simple model whereby trait resemblance is solely due to additive genetic variation. This implies that, for the majority of complex traits, causal genetic variants can be detected using a simple additive genetic model. Approximately one-third of traits did not follow the simple pattern of a twofold ratio of monozygotic to dizygotic correlations. For these traits, a simple additive genetic model does not sufficiently describe the population variance. An incorrect assumption about narrow-sense heritability (the proportion of total phenotypic variation due to addi – tive genetic variation) can lead to a mismatch between the results from gene-finding studies and previous expectations. If the pattern of twin correlations is consistent with a substantial contribution from shared environmental factors, as we find for conduct disorders, religion and spirituality, and education, then gene-mapping studies may yield dis – appointing results. If the cause of departure from a simple additive genetic model is the existence of non-additive genetic variation, as is, for example, suggested by the average twin correlations for recur – rent depressive disorder, hyperkinetic disorders and atopic dermatitis, then it may be tempting to fit non-additive models in gene-mapping studies (for example, GWAS or sequencing studies). However, the statistical power of such scans is extremely low owing to the many non-additive models that can be fitted (for example, within-locus dominance versus between-locus additive-by-additive effects) and the penalty incurred by multiple testing. Our current results signal traits for which an additive model cannot be assumed. For most of these traits, dizygotic twin correlations are higher than half the monozygotic twin correlations, suggesting that shared environmental effects are causing the deviation from a simple additive genetic model. Yet, data from twin pairs only do not provide sufficient information to resolve the actual causes of deviation from a simple additive genetic model. More detailed studies may identify the likely causes of such deviation and may as such uncover epidemiological or biological factors that drive family resemblance. To make stronger inferences about the causes underlying resemblance between relatives for traits that deviate from the additive genetic model, additional data are required, for example, from large population samples with extensive phenotypic and DNA sequence information, detailed measures of environmental exposures and larger pedigrees including non-twin relationships. We note that our inference is based on twin studies published between 1958 and 2012 and that it generally applies to complex traits but does not necessarily generalize to mendelian subtypes of traits. Most mendelian traits are rare in the population and are therefore not studied by researchers of twins because they cannot ascertain enough affected twin pairs to reliably estimate genetic parameters. In the rare case where sufficient numbers of affected twin pairs were available, the mendelian subtypes were analyzed together with the subtypes of the same trait that were due to common causes, as it was unknown whether the studied trait was a mendelian subtype. If the traits we studied were in fact a mix of mendelian and complex subtypes, our inference would be biased away from our main result because mendelian diseases tend to be dominant or recessive, not additive. In addition, there may be heterogeneity in measurement errors between studies for the same trait and between traits. A test-retest correlation would quantify measurement error when contrasted with a correla – tion between monozygotic twins, but few twin studies report such correlations in the same papers that estimate heritability. Our results provide the most comprehensive empirical overview of the relative contributions of genes and environment to all human traits that have been studied in twins thus far, which can guide and serve as a reference for future gene-mapping efforts. URLs. ICF classification, http://apps.who.int/classifications/ ic fbrowser /; ICD-10 classification, http://apps.who.int/classifica tions/ icd10/browse/2 010/en. The data used for this manuscript have been integrated in a web application, where user-specified selections of traits can be made to apply the analyses presented in this work. The web application is called MaTCH (Meta-analysis of Twin Correlations and Heritability) and is accessible via http://match.ctglab.nl/; Gephi, http://ge phi.github.io /. M ETHODS Methods and any associated references are available in the online version of the pap er. npg © 2015 Nature America, Inc. All rights reserved. Nat u r e G eNe t i c s V O L U M E 4 7 | N U M B E R 7 | J U LY 2 0 1 5 7 0 9 A nAl y s i s Note: Any Supplementary Information and Source Data files are available in the on li n e v ersio n o f t h e p ap er. A C know Le D gM en TSWe would like to thank M. Frantsen, M.P. Roeling, R. Lee and D.M. DeCristo for their contribution to collecting the full texts of selected twin studies and data entry. This work was funded by the Netherlands Organization for Scientific Research (NWO VICI 453-14-005, NWO Complexity 645-000-003), by the Australian Research Council (DP130102666) and by the Australian National Health and Medical Research Council (APP613601). AUTHo R ConTRIBUTI onSD.P., B.B., P.F.S. and P.M.V. performed the analyses. D.P. conceived the study. D.P., T.J.C.P. and P.M.V. designed the study. T.J.C.P. and D.P. collected and entered the data. D.P. and P.F.S. categorized traits according to standard classifications. A.v.B. and C.A.d.L. checked data entries, and checked and wrote statistical scripts. A.v.B. designed and programmed the webtool. D.P., T.J.C.P. and P.M.V. wrote the manuscript. All authors discussed the results and commented on the manuscript. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. Reprints and permissions information is available online at http ://w ww.n atu re .c om / r e p rin ts /in dex.h tm l. 1. Moore, J.H. Analysis of gene-gene interactions. Curr. Protoc. Hum. Genet. Chapter 1, Unit 1.14 (2004). 2. Hill, W.G., Goddard, M.E. & Visscher, P.M. Data and theory point to mainly additive genetic variance for complex traits. PLoS Genet. 4, e1000008 (2008). 3. Traynor, B.J. & Singleton, A.B. Nature versus nurture: death of a dogma, and the road ahead. Neuron 68, 196–200 (2010). 4. Zuk, O., Hechter, E., Sunyaev, S.R. & Lander, E.S. The mystery of missing heritability: genetic interactions create phantom heritability. Proc. Natl. Acad. Sci. USA 109, 1193–1198 (2012). 5. Phillips, P.C. Epistasis—the essential role of gene interactions in the structure and evolution of genetic systems. Nat. Rev. Genet. 9, 855–867 (2008). 6. Visscher, P.M., Brown, M.A., McCarthy, M.I. & Yang, J. Five years of GWAS discovery. Am. J. Hum. Genet. 90, 7–24 (2012). 7. Manolio, T.A. et al. Finding the missing heritability of complex diseases. Nature 461, 747–753 (2009). 8. Stranger, B.E., Stahl, E.A. & Raj, T. Progress and promise of genome-wide association studies for human complex trait genetics. Genetics 187, 367–383 (2011). 9. Maher, B. Personal genomes: the case of the missing heritability. Nature 456, 18–21 (2008). 10. Eichler, E.E. et al. Missing heritability and strategies for finding the underlying causes of complex disease. Nat. Rev. Genet. 11, 446–450 (2010). 11. Nelson, R.M., Pettersson, M.E. & Carlborg, Ö. A century after Fisher: time for a new paradigm in quantitative genetics. Trends Genet. 29, 669–676 (2013). 12. Barker, J.S. Inter-locus interactions: a review of experimental evidence. Theor. Popul. Biol. 16, 323–346 (1979). 13. Cockerham, C.C. An extension of the concept of partitioning hereditary variance for analysis of covariances among relatives when epistasis is present. Genetics 39, 859–882 (1954). 14. Cockerham, C.C. in Statistical Genetics and Plant Breeding 53–94 (Nat. Acad. Sci. Nat. Res. Council Publ., 1963). 15. Kempthorne, O. On the covariances between relatives under selfing with general epistacy. Proc. R. Soc. Lond. B Biol. Sci. 145, 100–108 (1956). 16. Crow, J.F. & Kimura, M. An Introduction To Population Genetics Theory (Harper and Row, 1970). 17. Carlborg, O. & Haley, C.S. Epistasis: too often neglected in complex trait studies? Nat. Rev. Genet. 5, 618–625 (2004). 18. Falconer, D.S. & Mackay, T.F.C. Quantitative Genetics (Longman Group, 1996). 19. Lynch, M. & Walsch, B. Genetics and Analysis of Quantitative Traits (Sinauer Associates, 1998). npg © 2015 Nature America, Inc. All rights reserved. Nat u r e G eNe t i c s doi:10.1038/ng.3285 ONLINE METHODS Identifying relevant studies. We searched PubMed for studies published between 1 January 1900 and 31 December 2012 that provided twin correla – tions, concordance rates, and a heritability estimate ( h 2) or an estimate of shared environmental influences ( c 2), using monozygotic and dizygotic twins. The following search term was used (“English”[Language] AND (“1900/01/01”[Date – Publication]: “2012/12/31” [Date – Publication]) AND twin AND “journal article”[Publication Type] AND “humans”[Filter] AND (heritability[Title/Abstract] OR “genetic influence”[Title/Abstract] OR “environmental influence”[Title/Abstract] OR “genetic factors”[Title/Abstract] OR “environmental factors”[Title/Abstract]) AND “journal article”[Publication Type]) NOT review[Title] NOT review [Publication Type]) The search was run on 31 January 2013 and again on 29 April 2013, which yielded an additional 44 publications, with the difference likely due to keywords or tags that had been added to publications in the intermittent period. The last PubMed search yielded 4,388 unique studies. From these, we deleted studies that were not relevant for the current purpose using the fol – lowing exclusion criteria: (i) studies with only monozygotic twins available; (ii) studies with no heritability estimates, twin correlations or concordances available; (iii) review studies; (iv) meta-analyses; and (v) multivariate stud – ies that provided information on completely overlapping traits and samples with previously published univariate studies. Some studies investigating h 2 for the brain (for example, voxel-based brain measures) were not included for practical purposes. These studies typically presented their results in graphs with color-coded point estimates of heritability mapped onto the brain. Such estimates could not be quantified, and these studies were thus not included. From the remaining 2,748 studies, we were able to retrieve the full text from all but 5 papers (99.8%). Of the studies without full-text availability, we included relevant information based on the abstract. An overview of authors and journals and a full reference list of all 2,748 studies are provided in Supplementary Tables 30 –32 . Primar y information obtained from each study. From every study, we retrieved basic information on the PubMed ID, the authors, the trait as named in the study and the year of publication. In addition, the following informa – tion was retrieved: • Country of origin of the study population. We used standard ISO country names, and where possible data entry was done separately for each country investigated in the study. • Age group of the study population. The study population was classified into four age cohorts on the basis of the average age of the included sample: age >0 and <12; age ≥12 and <18; age ≥18 and <65; and age ≥65. • Monozygotic and dizygotic twin correlations. Twin correlations were entered as provided in the study and could be calculated as intraclass, Pearson, polychoric or tetrachoric correlations or on the basis of least- squares or maximum-likelihood estimates. When available, we entered the twin correlations separately for males and females (i.e., monozygotic male (MZM), monozygotic female (MZF), dizygotic male (DZM), dizygotic female (DZF) and dizygotic opposite-sex (DOS) pairs). If correlations were not available for males and females separately, we entered the MZ and DZ correlations, i.e., the correlations based on both sexes. In cases where it was clear that the dizygotic correlation was based on same-sex twins only, we entered the dizygotic same-sex (DZSS) correlation. • Estimates of heritability ( h2) and shared environmental component ( c 2), under the full ACE (or ADE) model. We entered ‘ h 2_FULL’ and ‘c 2_FULL’, on the basis of estimates under the full ACE (including additive genetic and shared and non-shared environmental influences) or ADE (including additive and non-additive genetic and non-shared environmental influ – ences) model. When an ACE model was fitted, the estimate for A was entered in ‘ h 2_FULL’ and the estimate for C was entered in ‘ c 2_FULL’. When an ADE model was fitted, the estimates of A and D were summed and entered for ‘ h 2_FULL’ and zero was entered for ‘ c 2_FULL’. When esti- mates were provided separately for males and females, they were entered separately. In the case of multivariate analyses, univariate estimates were always preferred to allow comparison across studies. • Estimates of heritability ( h2) and shared environmental component ( c 2), under the best-fitting ACE (or ADE) model. We entered ‘ h 2_BEST’ and ‘ c 2_BEST’, on the basis of estimates under the best-fitting ACE or ADE model as provided in the study. When an ACE model was the best-fitting model, the estimate for A was entered in ‘ h 2_BEST’ and the estimate for C was entered in ‘ c 2_ BEST’. When an ADE model was the best-fitting model, the estimates of A and D were summed and entered for ‘h 2_ BEST’ and zero was entered for ‘ c 2_ BEST’. When estimates were provided separately for males and females, they were entered separately. In the case of multivariate analyses, univariate estimates were always preferred to allow comparison across studies. In cases where estimates for the best-fitting model were not directly provided but information available in the paper indicated that the best-fitting model was AE (or CE or E), we entered zero for ‘ c 2_ BEST’ and missing for ‘ h 2_ BEST’ (when the best-fitting model was an AE model), missing for ‘ c2_ BEST’ and zero for ‘ h 2_ BEST’ (when the best-fitting model was a CE model), and zero for ‘ c2_ BEST’ and zero for ‘ h 2_ BEST’ (when the best-fitting model was described to be an E model). • The total number of twin pairs as used for each entered estimate. • Whether the study was a classical twin study. All 2,748 studies in the database included monozygotic and dizygotic twins. However, a classi – cal twin study was defined as a study that involved only reared-together monozygotic and dizygotic twins. From studies that included siblings, extended families, adoptees or reared-apart twins, only estimates based on the reared-together twin sample were used for the meta-analyses. Most of the non-classical twin studies did provide twin correlations for the classical twin design and were thus included in the meta-analysis for twin correla – tions. When A and C estimates were based on extended twin designs, they were excluded from the meta-analyses. • The method used for estimating the variance components. We entered the statistical method used for estimating the variance components, which included, for example, ANOVA, Bayesian, maximum-likelihood (ML), DeFries-Fulker regression, least-squares (LS) or intrapair differences. We also listed a dichotomized version of this indicating whether the method used was ‘ML or LS’ or ‘not ML or LS’, for all other methods. In the meta-analyses for h 2 and c 2 estimates, only those based on maxi – mum likelihood or least squares were included. • Whether the trait was dichotomous or continuous. Traits measured as 0 or 1, as well as traits measured on a quantitative scale but dichot – omized before analysis, were listed as dichotomous. All other traits, including ordinal traits, were listed as ‘not dichotomous’ and treated as continuous. • Whether the study involved ascertainment for the trait. When the trait under investigation was the same trait that was used to select probands, the study was listed as ‘ascertained’. • Number of concordant and discordant pairs. In cases of dichotomous traits, the total numbers of pairs for discordant and concordant affected pairs were entered separately for each zygosity. In cases of dichotomous traits that were not ascertained, the number of concordant unaffected pairs was also entered. • Prevalence. In cases of dichotomous traits, the population prevalence, sep – arately for monozygotic and dizygotic twins when available, was entered. Prevalence was based on what was provided in the study or was calculated using (2 c + d)/2 n, where c is the number of concordant affected pairs, d is the number of discordant pairs and n is the total number of pairs in non-ascertained traits. Thus, provided that there was availability, the statistics in Supplementary Table 1 were obtained for each trait reported on in every study. When the five basic twin correlations were available ( r MZM , rMZF , rDZM , rDZF and r DOS ), we calculated r MZ, rDZSS and r DZ, using the weighted average via Fisher z transformation and using sample size as weights. In situations where r MZM (or r MZF ) was exactly 1 (or −1), we subtracted (or added in the case of −1) 0.00001 to the correlation to ensure non-problematic Fisher z transformation. Sample sizes of MZ, DZSS and DZ were obtained by summing the sample npg © 2015 Nature America, Inc. All rights reserved. Nat u r e G eNe t i c s doi:10.1038/ng.3285 sizes of MZM and MZF, of DZM and DZF, and of DZM, DZF and DOS, respectively. Estimates of h 2 and c 2 were calculated across sex as the n -weighted average across the separate male and female estimates, when available. For the number of concordant and discordant pairs, MZ, DZ and DZSS were calculated on the basis of the numbers available for MZM, MZF, DZM, DZF and DOS. Prevalences for pooled entries were calculated as an n-weighted average. Data entry checks. Studies were entered and cross-checked for obvious typos by T.J.C.P. and D.P. After initial data entry and initial cross-checking, all data points were manually checked (D.P.) by looking up the entered values in the original paper. In addition, automatic checks were run (D.P. and B.B.) to identify typos, strange outliers or obvious mistakes. These checks included: (i) identifying highly unlikely values (clear typos, for example, correlation of 120); (ii) testing whether the sum of h 2 and c 2 was <100; (iii) testing for strange discrepancies between estimates from the full and best-fitting models; and (iv) checking for outliers on the basis of extreme sample size and extreme χ2 values for rejecting the null hypothesis that either 2 × ( r MZ − r DZ) or 2 × r DZ − r MZ is equal to zero. Classification of traits. After data entry, all traits were manually classified using the ICF. The ICF is the framework of the World Health Organization (WHO) for health and disability and provides the conceptual basis for the definition, measurement and policy formulations for health and disability. It is a universal classification of disability and health for use in health and health-related sectors. ICF belongs to the WHO family of international classifications, the best-known member of which is the ICD-10. ICD-10 provides an etiological framework for the classification of diseases, disorders and other health conditions, whereas ICF classifies functioning and disability associated with health conditions. The ICD-10 and ICF are therefore comple – mentary (see URLs). Most traits investigated in twin studies concern healthy functioning (for example, cognitive function, social attitudes, body height and personality) and were classified according to ICF. In cases where the studied traits were diseases or symptoms of disease, ICD-10 was used. Traits were given two hierarchical classifications corresponding to the ICF or ICD-10 hierarchical structure, using the chapter structure (for example, b1) and the level directly under the chapter (for example, b110), which corresponds to the code for the actual disease (ICD-10) or trait (ICF). Six new classes at the chapter level and 17 new classes at the subchapter level were created to accommodate traits that could not be classified under either ICF or ICD-10. For the chapter level, the created classes were cell, func – tion of DNA, functions of the nervous system, medication effects, mortality and structure of DNA. For the subchapter level, the classes created were all-cause mortality, cell cycle, cell growth, diazepam effects, expression, function of brain, gene expression, height, methylation, mortality from heart disease, mtDNA, physical appearance, receptor binding, sister chromatid exchange, structure of DNA, telomeres and X inactivation. In addition to the two standard ICF or ICD-10 classification levels, we added a general classification of functional trait domains. We thus clas – sified all traits using a 3-level scheme that included 28 broad, functional domains, 54 ICF or ICD-10 chapter-level classes and 313 subchapter-level classes. A small proportion of studied traits (<0.1%) could not be classi – fied meaningfully on the chapter level (two traits) or the subchapter level (three traits). There were 326 unique combinations across the 3 levels of trait categorization ( Supplementar y Table 33 ). All analyses were con- ducted on all entries of each of the three levels of classification. In addition, we analyzed all traits together. Although this is unspecific in terms of diseases or traits, it provides a general over view of the relationship between monozygotic and dizygotic twin correlations and shows general patterns of, for example, sex and age differences. The most specific level was the subchapter level, which was the actual ICD-10 diagnosis or a similar ICF classification for normal functioning, reflecting specific traits such as cleft lip, hyperkinetic disorders or higher-level cognitive function. As researchers do not necessarily adhere to the ICD-10 or ICF trait nomen – clature, traits with the same subchapter classification could have different trait names in the original study : for example, for higher-level cognitive function, the original studies included the trait names total IQ score, cognitive ability, intelligence or ‘g’. Tests for publication bias. Publication bias can occur when studies that report relatively large heritability estimates or high twin correlations are more likely to be submitted and/or accepted for publication than studies that report more modest effects. Such a publication bias would lead to an overestimation of the true twin correlations or the true heritability and environmental estimates. We used several standard statistical tools to aid in identifying and quantify – ing possible publication bias, including inspection of funnel plots, Begg and Mazumdar’s test 2 0, Egger’s regression test 21 and Rosenthal’s fail-safe N 22. Meta-analysis methods of twin correlations and variance components. We used the DerSimonian-Laird (DSL) random-effects meta-analytical approach with correlation coefficients as effect sizes, as described by Schulze 2 3 and implemented in the R package metacor. This function transforms a correlation to its Fisher z value with corresponding standard error before the meta-analysis. This method is preferred over conducting a meta-analysis directly on the correlations because the standard error of a twin correlation is a function of not only sample size but also the correlation itself, with larger correlations having a smaller standard error. This can cause problems in a meta-analysis, as it would lead to the larger correlations appearing more precise and being assigned more weight in the analysis, irrespective of sample size. To avoid this problem, the DSL method transforms correlations to the Fisher’s z metric, whose standard error is determined solely by sample size. All n-weighted computations were thus performed using Fisher’s z metric, and the results were converted back to correlations for interpretation. The random-effects approach allows for heterogeneity of the true twin cor – relations across different studies. That is, rather than assuming that there is one true level of twin correlation, the random-effects model allows a distribution of true correlations. The combined effect of the random-effects model represents the mean of the population of true correlations. For computational reasons, correlations of −1 and 1 were converted to −0.99999 and 0.99999 before meta- analysis. We set a threshold of at least five pairs of twins available per estimate and at least two studies available per category. Meta-analyses were conducted for each category of all three levels of classification. We note that twin samples used in different publications were not independ – ent. For example, studies using Australian twins are predominantly based on twins from the Australian Twin Registry. These studies sometimes include dif – ferent subsamples but may also include completely overlapping samples used to investigate different traits. As participants are anonymous, it is not possible to determine the extent of overlap in the studies included in our analyses. We thus assumed independency of samples in the meta-analyses. This assump – tion leads to an underestimation of the variance of weighted estimates and an overestimation of their precision. We expect that the dependency of study samples is lowest at the specific level of the ICD-10 or ICF subchapters and highest for the general functional domains. Meta-analysis for dichotomous, non-ascertained traits. In the DSL random-effects model, the standard error of a correlation is calculated on the basis of the provided n (pairs). The estimated standard errors for continuous traits are correct, but for dichotomous traits the resulting standard error is incorrect. That is because twin correlations for non-ascertained, dichotomous traits are typically based on three categories of pairs: concordant, unaffected (CON −), discordant (DIS) and concordant, affected (CON +) pairs. Whereas the total number of participating pairs is the sum of these, the information that determines the twin correlation and its significance is mostly derived from the latter two categories. For non-ascertained, dichotomous traits, we calculated the study-specific tetrachoric twin correlation on the basis of the contingency table (i.e., CON −, DIS and CON + pairs), under the assumption that the dichotomous traits represent latent variables that follow a bivariate normal distribution 2 4. We used a maximum-likelihood estimator described by Olsson 2 5, implemented in the R function polycor, to calculate the study- specific twin correlation and its standard error. As our meta-analysis required twin correlation and sample size (not standard error) as input and we wanted to be able to pool across continuous and dichotomous traits, we calculated the ‘effective’ number of pairs on the basis of the obtained standard error. npg © 2015 Nature America, Inc. All rights reserved. Nat u r e G eNe t i c s doi:10.1038/ng.3285 The effective number of pairs was defined as the number of pairs that produced the exact same standard error within the DSL meta-analyses as the standard error obtained from the contingency table. Meta-analysis for dichotomous, ascertained traits. For ascertained traits, it was not possible to calculate the twin correlations and standard errors on the basis of the contingency table, as the traits included only pairs with at least one proband. Without information on the number of concordant, unaffected pairs, the prevalence of the affected status would be required to calculate a twin correlation. We used the algorithms derived from Falconer 2 6 and Smith 27. Again, for practical purposes, calculated standard errors were transformed to an effective number of pairs for use in the DSL meta-analysis. Proportion of studies consistent with specific hypotheses. We estimated the proportion of studies that were consistent with H 0: 2 × (r MZ − r DZ) = 0 ( π0( h )) and the proportion of observations consistent with H 0: 2 × r DZ − r MZ = 0 ( π0( c )), using the Jiang and Doerge method 2 8, as well as the q-value method 2 9. Authorship network analysis. We used the approach more fully described previously 3 0. Briefly, we retrieved from PubMed the full Medline listing for all twin studies included in this meta-analysis using NCBI eutils. The output was parsed to capture the names of all authors. The twin study author list was manually reviewed to resolve clear inconsistencies in the spelling of the names of authors between publications. Gephi was used to construct a network to understand twin study publication patterns. For clarity, we removed individuals who had published only one paper (i.e., we required authorship on ≥2 papers). The substructure of the network was investigated by estimating community membership modules using the Louvain method 3 1 implemented in Gephi. 20. Begg, C.B. & Mazumdar, M. Operating characteristics of a rank correlation test for publication bias. Biometrics 50, 1088–1101 (1994). 21. Egger, M., Davey Smith, G., Schneider, M. & Minder, C. Bias in meta-analysis detected by a simple, graphical test. Br. Med. J. 315, 629–634 (1997). 22. Rosenthal, R. The file drawer problem and tolerance for null results. Psychol. Bull. 86, 91–106 (1979). 23. Schulze, R. Meta-Analysis: A Comparison Of Approaches (Hogrefe & Huber, 2004). 24. Drasgow, F. in Encyclopedia of Statistical Sciences (eds. Kotz, S., Read, C.B., Balakrishnan, N. & Vidakovic, B.) Vol. 7, 68–74 (John Wiley & Sons, 2006). 25. Olsson, U. Maximum likelihood estimation of the polychoric correlation coefficient. Psychometrika 44, 443–460 (1979). 26. Falconer, D.S. The inheritance of liability to certain diseases, estimated from the incidence among relatives. Ann. Hum. Genet. 29, 51–76 (1965). 27. Smith, C. Concordance in twins: methods and interpretation. Am. J. Hum. Genet. 26, 454–466 (1974). 28. Jiang, H. & Doerge, R.W. Estimating the proportion of true null hypotheses for multiple comparisons. Cancer Inform. 6, 25–32 (2008). 29. Storey, J.D. & Tibshirani, R. Statistical significance for genomewide studies. Proc. Natl. Acad. Sci. USA 100, 9440–9445 (2003). 30. Bulik-Sullivan, B.K. & Sullivan, P.F. The authorship network of genome-wide association studies. Nat. Genet. 44, 113 (2012). 31. Blondel, V., Guillaume, J., Lambiotte, R. & Lefebvre, E. Fast unfolding of community hierarchies in large networks. J. Stat. Mech. 10, P10008 (2008). npg © 2015 Nature America, Inc. All rights reserved.