6 Discussion of CBC Application

The data we have analyzed are restricted to parental checklist reports of their twin children's behavior problems, without the benefit of self reports, teachers' reports or clinical interviews. As such, they are limited by the ability of parents to provide reliable and valid integrative assessments of their children, using cursorily defined concepts like `Sulks a lot,' `Worrying,' or `Fears going to school.' It is clear from meta-analyses of intercorrelations of ratings of children by different types informants that while the level of agreement between mothers and fathers is often moderate (e.g., yielding correlations around .5 to .6) the level of agreement between parents and other informants (e.g., parent with child or parent with teacher) is modest and generally yields a correlation around 0.2 to 0.3 (Achenbach et al., 1987). Thus, parental consistency in evaluating their children does not guarantee cross situational validity, although it does provide evidence that ratings of behavior observable by parents are not simply reflecting individual rater biases. In assessing the importance of the home environment on children's behavior this becomes a critical issue since studies of children's behavior based on ratings by a single individual in each family, e.g., the mother, confound the rater bias with the influence of the home environment. This may have the dual effect of inflating global estimates of the home environment's influence while at the same time either attenuating the relationship between objective indices of the environment and children's behavior (which is being assessed by a biased observer) or spuriously augmenting apparent relationships which are in fact relationships between environmental indices and maternal or paternal rating biases. An issue distinguishable from that of bias is that of behavior sampling or situational specificity. Thus maternal and paternal ratings of children may differ not because of the tendency of individual parents to rate children in general as more or less problematic (bias), but because they are exposed to different samples of behavior. If this is so, then treating informants' ratings as if they were assessing a common phenotype, albeit in a biased or unreliable way, will be misleading. It is of considerable psychological importance to know whether different observers are being presented with different behaviors. The approach outlined in this chapter first enables us to examine the adequacy of the assumption that different informants are assessing the same behaviors and then, if that assumption is deemed adequate, to separate the contributions of rater bias and unreliability from the genetic and environmental contribution to the common behavioral phenotype. For our particular example, all the models fit our data adequately and the bias model, even in its restricted version, does not fit significantly worse than the psychometric or biometric models. Although not presented here, there is some evidence that for externalizing behavior mothers and fathers cannot be assumed to be simply assessing the same phenotype with bias. In this context it is worth noting, however, that the adequacy of the assumption that parents are assessing the same phenotype in their children does not imply a high parental correlation (which may be lowered by bias and unreliability) and, conversely, even though parents may be shown to be assessing different phenotypes in their children to a significant degree the parental correlation in assessments may predominate over variance specific to a given parent. Our comparison of the bias with the psychometric and biometric models provides important evidence of the equivalence of the internalizing behaviors assessed by mothers and fathers using this instrument. This equivalence does not preclude bias or unreliability and the evidence presented in Table 11.4 provides a striking illustration of the impact of these sources of variation on maternal or paternal assessments. A shared environmental component which might be estimated to account for 34% of variance if mothers' ratings alone were considered, may correspond to only 20% of the variance when maternal biases have been removed. Similarly, a non-shared environmental variance component of 19% of variance may correspond to 9% of variance in individual differences between children that can be consistently rated by both parents. Finally, once allowance has been made for bias and inconsistency or unreliability, the estimated heritability rises from 47% to 70% in this case. We have not been concerned here to seek the most parsimonious submodel within each of the model types. We should be aware that although we have, for the younger children, presented the full models with sex limitation, differences between boys and girls are not necessarily significant (for example, although the biometric model without sex limitation fit our data significantly worse than the corresponding model allowing for sex limitation ( $\chi^2_{9} = 21.31$, $p<.05$), the overall fit without sex limitation is still adequate, $\chi^2_{41} = 42.26$). Furthermore, individual parameter estimates reported for our full models may not depart significantly from zero. Other limitations of the method are that it does not allow for interaction effects between parents and children and, in our application, assumed the independence of maternal and paternal biases. The analysis of parental bias under this model requires that both parents rate each of two children. Distinguishing between correlated parental biases and shared environmental influences would require a third, independent, rater (e.g., a teacher); thus we cannot rule out a contribution of correlated biases to our estimates of the remaining shared family environmental influence. The final caveat against overinterpretation of particular parameter estimates is that we have reported analyses for families in which both parents have returned a questionnaire and we have made no distinction between different biological or social parental statuses. Clearly, we anticipate that the inclusion and exclusion criteria are not neutral with respect to children's behavior problems and their perception by parents. However, we have illustrated that behavior genetic analyses are possible even when we have to rely on ratings by observers, providing that we have at least two degrees of relatedness among those being rated (e.g., MZ and DZ twins). Without an approach of this sort we have no way of establishing whether parents are assessing the same behaviors in their children and whether analyses will spuriously inflate estimates of the shared environment as much as parental biases inflate the correlations for pairs of twins independent of zygosity. Extension of the model to include other raters, for example, teachers, is straightforward.