The modern behavioral genetic model is a blend of different methodological approaches, based primarily on path analysis and analysis of variance techniques that were part of "biometrical genetics" (Neale and Cardon 1992). It is characterized by (1) the use of multiple kinship groups, often more than just MZ and dizygotic (DZ) twin groups; (2) the statement of a path analytic model using pairs of relatives' trait variances and covariances; and (3) the fitting of this model to observations (usually covariance matrices) with structural equation programs (e.g., Mx, LISREL) that estimate goodness-of-fit and various genetic and environmental parameters (e.g., heritability).
Behavioral genetic research designs help to identify models that would otherwise be unavailable for use in social science. The concept of model identification is a complex one. To make a simple analogy, given two separate equations with x and y variables where each variable retains the same value in both equations, one can solve for x and y. However, given two equations, one with x, y, and z variables and the other with x and y variables, one cannot solve for x, y, and z because two equations do not give enough information to solve for three unknowns. The first pair of equations is an "identified" model; the sec-
ond is not. In other words, an identified model makes it possible to assign values to variables. The ability of behavioral genetic designs to "identify" complex models allows for the exploration of many interesting questions about influences on behavior, including how to use data from multiple informants or raters, how to explain behavioral development, and what creates correlation between phenotypes. Although these topics do not exhaust research activity in behavioral genetics, they serve as good illustrations and they are considered in turn in the next sections.
Consider an interesting problem in explaining trait variation: how to "distill" a trait from various ratings of behavior (Rowe and Kandel, 1997). The multiple raters may have a shared view of a target child, based on their observations of behavior. However, each rater may also have an individual view of the child, resulting from many influences (e.g., experiences one rater but not the other has had with the child, perceptual biases, and so on). The shared view may be interpreted as the child's generalized or global behavioral trait. Multiple raters are often used to estimate this general trait more accurately. With only a single child and two ratings of behavior, however, there is no way to separate the influence of shared and individual views on raters' judgments.
Figure 1 shows an example of two measures of a phenotype. One is the mother's rating; the other is the child's self-rating. Both measures help to define the child's true trait (path coefficients a and b). However, the mother's rating is also influenced by her individual view (path coefficient c). This model nicely states the sources of variation, but it is not an identified model because there is not enough information to solve for all three path coefficients simultaneously. It can be rescued, however, by putting it into a behavioral genetic research design (i.e., a sibling pair design).
Figure 2 draws the comparable model for sibling pairs. As each mother rates two of her children, her individual view is allowed to influence both ratings (path coefficient q). The child's trait influences both the mother's rating (a) and the child's self-rating (b). The path coefficients c and h reflect shared environment and genetic influences, respectively. This model permits two sources of sibling resemblance—one through correlated genes (genetic effects, designated by the path hxrgxh) and the other through shared environments (shared environmental effects, designated by the path cxl.Oxc). Nonshared environmental influences (e) also have an impact on the traits. Although this model is more complex than the model in which each mother rates only one child (as in Figure 1), it is also mathematically identified when covariance matrices are fit for two or more groups (e.g., MZ and DZ twins, or full and half-siblings). Thus, in addition to answering the behavioral genetic question of the estimation of genetic and environmental effects, this model also addresses a question of broad interest to personality researchers: how maternal and child ratings contribute to the estimate of the "true" trait. Neither of these questions could be answered by using the model shown in Figure 1.
As an example, Simonoff et al. (1995) explored a variety of behavioral genetic models of rater effects for rating disruptive behavior in childhood (i.e., 8-16 years) using twin data. Ratings on the twin children's disruptive behavior came from child self-reports, mothers' reports, and fathers' reports. Although their article should be consulted for details, these authors found that genetic effects increased for the shared view, that is, for a global trait inferred from multiple ratings, and shared environmental influences decreased (Simonoff et al. 1995, 318).1 Shared environmental influences decreased because some of the within-parent correlation was explained by the individual view of each
Figure 1. Single child model (MR = mother rating, CSR = child self-rating).
rater. The individual (parental) views were also correlated in the best-fitting model. Parents may agree separately about their children because of shared expectations or other influences that do not relate to a child's global traits. Regardless of the correct explanation of the individual view, it is clear that behavioral genetics offers a method for examining the etiology of a global trait as defined by multiple ratings. In general, in the models of Siminoff et al., a global trait inferred from shared view ratings showed greater genetic influence than behavior associated with individual views.
Behavioral development is marked by both stability and change. For example, the stability of self-reports of delinquency fall into the r = 0.50-0.70 range
across three years of adolescence (Rowe and Britt 1991). Nonetheless, many adolescents who start out as delinquents desist from crime by the end of their teenaged years. This mixed picture of stability and change can be found for many traits. The longitudinal design allows one to examine individuals as they mature, as they both change and stay the same. The strengths of the design for establishing causal order and for examining change are widely understood. However, the contribution of longitudinal studies to advancing theory about environmental and genetic influences on development is less appreciated. Furthermore, although several scholars have enthusiastically promoted developmental behavioral genetics (Plomin 1986; Rutter 1994), there has been limited use of these designs because data that are both longitudinal and kinship based are rare.
Developmental behavioral genetics deals with the determinants of stability and change in behavioral traits. The field combines a longitudinal design with a kinship design; that is, more than one age of measurement for pairs of relatives who differ in their genetic relatedness. Three periods of observation are necessary to mathematically identify (i.e., make solvable) the more complex developmental designs.
Developmental behavioral genetic designs distinguish between two broad classes of developmental mechanisms—transmission and liability or common factors. In the transmission model, earlier experiences send their influence forward in time so that successive periods of functioning are causally linked. The idea of transmission is present in many developmental theories. In criminology, labeling theory assumes that arrest or other contact with the justice system changes the self-concept so that the individual is more likely to commit future crimes. Similarly, gang involvement might temporarily increase the likelihood of later criminal offenses. In attachment theory, the security of infant attachment is assumed to create a "mental model" of relationships that may extend to romance in adolescence. Developmental concepts that highlight critical periods or developmental tasks claim that outcomes of certain phases affect future functioning.
The transmission model resembles the concept of state dependence in demography and sociology. In state dependence, past behavior affects future behavior. For instance, the mere commission of a crime may loosen social commitments toward conventionality so that another crime becomes more probable. State dependence also comes into play as life events occur that may alter an individual's state and thus redirect behavior. In the example of crimi nal acts, marriage is thought to cause some criminals to desist. As new influences enter and affect behavioral development, change is instigated.
The second class of developmental models can be referred to as liability or common factor models. In these models, no causal relation between subsequent time points is assumed. In the liability model, the stability of development arises from underlying individual differences, which may be only partly understood. The continuum of these individual differences is called a "liability." This liability may act as a "third variable" that makes the stability of behavior from one age to the next noncausal; temporal order alone does not establish that a prior variable is a cause of one that follows it in time.
The trait of blood pressure can be used to illustrate the liability model. Blood pressure is highly stable over the life course. This stability, however, does not mean that high blood pressure at one age directly causes high blood pressure at a later age. Instead, it is more likely that individual differences in physiology—some of which are targeted by blood pressure medications— make some people prone to life-long high blood pressure. Liability mechanisms can also account for change, because at each point in time there may be new effects of other variables, such as the introduction of medicines, which might cause changes in blood pressure. The essential assumption, however, is that continuity is completely explained by the underlying liability, not by previous scores on a trait or behavior.
In demography and criminology, the liability model sometimes goes under the name of heterogeneity. Heterogeneity refers to all sources of individual differences that are not directly measured in a research design. The known co-variates, of course, can be used in regression analyses as statistical controls, and their effects on other variables partialled away, but unknown covariates cannot be controlled for and remain as validity threats. For example, divorced and nondivorced groups may be matched for social class, but such matching might leave personality differences uncontrolled. One advantage of a developmental genetic design is that the liability can be treated as a latent variable that includes the influences of all sources of heterogeneity, whether measured or not.
Both the transmission and liability models assume a certain amount of developmental stability. These models, however, imply rather different patterns of stability among assessments in a longitudinal study. In the transmission model, assessments closer in time should be more highly correlated than more distant ones. This "autoregressive" pattern occurs because, while the effects of prior experiences are felt for some time, new experiences and events also enter the developmental process and can create change. For example, suppose we have data on self-reported delinquency from assessments at three ages that yielded these correlations: r = 0.55, r23 = 0.55, and r = 0.30. The most separated assessments (r13) show the lowest stability correlation; furthermore, as expected under an autoregressive process, the stability from the first to the last assessment exactly equals the product of the correlations between adjacent assessments (i.e., 0.30 = 0.55 x 0.55). This hypothetical example also demonstrates the need for assessments from three or more time periods, because two assessments provide no test of this autoregressive property. In contrast, the liability model assumes that the same stable mechanisms exert their effects at each assessment and does not imply a decrease in stability with an increase of the time lag between assessments. In a pure liability model, stability from the first to the third assessment of delinquency would be 0.55, the same as the correlations between adjacent time points.
The goodness-of-fit of liability and transmission models can be compared using behavioral genetic designs. For example, Van den Oord and Rowe (1997) looked at the stability of problem behaviors in children assessed at three ages: 4-6 years, 6-8 years, and 8-10 years. The measure of problem behavior was a maternal checklist of problem behaviors, where items were rated as "often true," "sometimes true," and "not true" of a child. The children were either siblings or cousins, and the children who were siblings were either full or maternal half-siblings. These three groups were used to fit both liability and transmission models that estimated behavioral genetic parameters. On the whole, the liability model performed substantially better than the transmission model. The liability model did not require time-specific effects, and it showed that the combination of a genetic liability and a shared environmental liability together accounted for the stability of problem behaviors from 4 to 10 years. Thus, results from this study suggest that previous problem behavior does not cause later problem behavior; rather, stable genetic and environmental factors influence problem behavior at each of the three time points. In addition, the study also revealed that the largest contribution to age-related change in problem behavior was due to nonshared environmental influences. In summary, developmental behavioral genetic methods examine genetic, shared environmental, and nonshared environmental mechanisms underlying age-to-age continuity and change (see Eaves et al. 1986; Hewitt et al. 1988; Philips and Fulker 1989).
Genetic and Environmental Mediators of
Perhaps the most common use of behavioral genetic designs is apportioning covariance between phenotypes to genetic and environmental sources of variation. Many questions about causality depend on the source of variation between two phenotypes. Analysis of the covariance or correlation between phenotypes may be bivariate when just two variables are involved, or may involve complex multivariate models (e.g., psychometric and biometric models, see Neale and Cardon 1992). For example, traits X and Y may be correlated within individuals. Now, if both traits are heritable, then their correlation might arise because of a genetic correlation between them. That is, genes that affect more than one phenotype can produce a genetic correlation (technically, pleiotropy). As an illustration, consider that the gene for albino skin coloration in mice also produces behavioral inhibition in an open field test (Plomin et al. 1997). Thus, coat color and behavioral inhibition are correlated traits in mice because they are partly caused by the same gene.
Sources of covariance can also be environmental. Shared environmental influences common to two or more phenotypes may induce correlation between them. For example, in adolescence, shared environments (e.g., parental beliefs) influence social attitudes (Eaves et al. 1997). The covariance typically found among different attitudes tends to occur because some parental beliefs influence more than one attitude. As an adolescent models parental beliefs, he accepts the same "suite" of attitudes as the parent; thus a shared environmental effect on the covariance among attitudes occurs. It is interesting that the effects of shared family environments on social attitudes sharply diminish once the adolescent leaves home and enters the working world; at the same time, the heritability (i.e., the genetic influence) of social attitudes increases.
Nonshared environmental influences are also a potential source of covari-ance among phenotypes within an individual. However, nonshared environmental effects, as noted earlier in this chapter, are unique to each individual and thus always reduce the behavioral resemblance between family members. For example, some part of the correlation of height and weight within an individual is explained by common genes that influence body size. However, another part of this within-person correlation may arise from nonshared environmental influences on body size, including almost accidental aspects of embryological development, or factors such as nutrition, which might vary across individuals.
The analysis of covariance has many uses. Two examples illustrate two extremely different applications, one analyzing the physiological substrates of traits, the other analyzing their environmental correlates. Although heritabil-ity implies the existence of a physiological substrate for a trait, the phenotypic association between a trait and a biological marker for it is not sufficient proof of a common genetic effect. Better early nutrition, for example, may increase both body and brain size. If so, a correlation between brain size and IQ could be mediated by environmental influences (both shared and nonshared) related to nutrition, rather than by genetic influences.
Another physical variable that has a phenotypic association with IQ— nerve conduction velocity—has been analyzed in a twin study (Rijsdik and Boomsma 1997). Nerve conduction velocity is measured in peripheral nerves in the arm and reflects the speed with which nerve impulses move through axons and across synapses. In this twin sample, conduction velocity and IQ had a correlation of 0.20. The bivariate analysis of this within-person correlation revealed that the association between IQ and nerve conduction velocity was entirely genetic. Thus, the same genes that influence conduction velocity also influence IQ, and the former is a good physiological marker for individual differences in IQ that are due to genes.
The second type of bivariate analysis deals with the association of children's traits and environmental measures. A typical interpretation of the association is that it is causal; that is to say, the environmental influences causally affect the trait. This interpretation suggests that a decomposition of the association between shared environmental influences (such as parental treatment) and children's traits involves the environmental influences only and not genetic influences. However, another possibility is that the decomposition involves genetic effects alone. This situation could arise if genes in a parent affect parental behaviors related to childrearing and if copies of those same genes, expressed in the children, produce the behavioral trait that is the dependent variable of interest. A causal association between parental treatment and children's behavior, then, would be spurious (i.e., noncausal) because genes, not environmental effects, create it.
An example of a study examining the relationship between children's traits and environmental measures is the bivariate genetic analysis of the relationship between parental negativity and adolescent antisocial behavior carried out by Pike et al. (Pike et al. 1996). The genetically informative groups consisted of MZ twins, DZ twins, full siblings, half-siblings, and biologically unre lated siblings. Parental negativity was measured by parental self-reports and observations of parent-child interactions. Antisocial behavior was measured by the children's (10-18 years of age) self-report of antisocial acts and by the extent to which the adolescents were disruptive or disrespectful during home observations of them.
It is possible to anticipate the findings of Pike et al.'s more complex model fitting by examining the cross-correlations between siblings. A sibling correlation is simply trait X on sibling 1 correlated with trait X on sibling 2, whereas a sibling cross-correlation is trait X on sibling 1 correlated with trait Y on sibling 2 (or vice versa). In this study, the sibling cross-correlation is the parental negativity directed toward sibling 1 correlated with sibling 2's antisocial behavior (and vice versa). To the extent that genetic influences are important, cross-correlations should increase in step with the genetic relatedness of each sibling group. In Pike et al.'s study, the correlations of the mother's negativity toward sibling 1 and the antisocial behavior of sibling 2 were MZ twins, 0.54; DZ twins, 0.34; full siblings, 0.30; half-siblings, 0.34; and unrelated siblings, 0.18. Although the half-siblings are more alike than the genetic expectation, as are the unrelated siblings reared together, the pattern of cross-correlations clearly suggests a genetic effect.
This genetic effect was also borne out by the results of Pike et al.'s model fitting. The phenotypic correlation between negativity and antisocial behavior (r = 0.60) was apportioned to genetic influence (0.40), shared environment (0.16), and nonshared environment (0.05). In other words, 66 percent of the total association between parent negativity and child antisocial behavior was mediated genetically (i.e., 0.40/0.60). Two explanations can be offered for this genetic mediation. One is that negativity in a parent is simply another expression of antisocial behavior, so that the same genetic trait surfaces differently, depending on age and social role (i.e., antisocial behavior in adolescents vs. negative parenting in adults). Adolescents, who usually have no children of their own, lack the opportunity to display their antisocial behavior in terms of poor parenting behaviors. On the other hand, parental treatment also differs from sibling to sibling. This favors a different explanation: namely, that a child's heritable antisocial behavior partly elicits negativity from a parent (see also Ge et al. 1996; McGue et al. 1996). In both explanations, however, the commonplace view that an association between parental behavior and child outcome can be interpreted as solely environmental in origin must be abandoned.
As mentioned earlier, surprisingly few g x e interactions are found in behavioral genetic research. In most cases, additive models, which do not allow for g x e interactions, fit data extremely well (Rowe 1994). There are a few cases of g x e interactions (Eaves et al. 1997). In one case, the relative importance of genetic and shared environmental influences on social attitudes was found to shift with age. In adolescence, shared environmental influences dominate the total variation. For example, in the mid-teens, both MZ and DZ twin correlations of social attitudes fell into the range 0.40-0.60. A genetic effect would show up in a greater MZ twin correlation; thus it would appear as though genetics has little influence on the social attitudes of adolescents. In contrast, the presence of a significant correlation between twins suggests some shared environmental influence. After age 20, however, the twin correlations strikingly diverged. The MZ twin correlation settled to a value of about 0.70 and the DZ twin correlation to about 0.40. In other words, in one environment (i.e., adolescents living together in their home), the genes relevant to social attitudes were not expressed. However, in the environment of adult life, when twins usually live apart from one another, enter the workforce, and make individual political choices that might increase the relevance of their social attitudes, social attitudes become genetically influenced.
In another case of g x e interaction, Rowe et al. (in press) explored the magnitude of genetic and environmental influences on adolescent verbal IQ for different levels of parental education. The overall heritability of verbal IQ was about 0.55. The heritability was significantly greater among more well-educated familes. In contrast, however, at lower levels of parental education (i.e., less than high school), shared environmental effects increased. Thus, shared environmental influences varied from about zero in well-educated families to 28 percent of the total variation in poorly educated families. One explanation for this pattern is that intellectual stimulation may be widely available in the non-family environments of well-educated children, such as through their schools and peer groups. Thus, a child from a well-educated family that provides low intellectual stimulation can compensate by finding intellectual stimulation outside the home. In contrast, a child from a poorly educated family may not have this opportunity, or it may be less available. Thus, in a neighborhood of better-educated families, the particular level of stimulation within families becomes less important, reducing the shared environmental effect to nil.
Overall, the general absence of many g x e interactions is noteworthy, yet an absence of replicable interaction effects is not new in social science. For example, Cronbach and Snow (1975) concluded that aptitude x treatment interactions were rare. The absence of interactions in behavioral genetics may partly reflect the great technical demands needed to uncover them, such as the need for large samples and the relative rarity of the genotypes or environments most productive of interaction effects (McClelland and Judd 1993). However, Mc-Call (1991) suggests a more profound reason for the relative absence of interactions: Nature may avoid interactions because the dependence of development on particular environmental circumstances is maladaptive. If the successful development of a trait were to become too tightly constrained by environmental influences, it would be exposed to the "struggle for existence" of natural selection, and would tend to result in death, and thereby the loss of the particular gene (or genes) from the population. Whatever the reason, the general absence of interactions has been one of the more unexpected findings of behavioral genetic research.
Was this article helpful?