Modeling interinformant agreement in the absence of a "gold standard"

Epidemiological surveys of child and adolescent mental disorders often rely on multiple informants to get a complete diagnostic picture. A consistent finding in the literature is that different informants often do not identif y the same children as being disordered. However, because current strategie s for estimating interinformant agreement often involve categorizing childr en using less than perfectly sensitive and/or specific symptoms, biased est imates of interinformant agreement are likely. The aim of this report was t o illustrate how latent class analysis (LCA) can be used to model interinfo rmant agreement in the absence of a "gold standard". The proposed model con sists of informant-specific latent variables each made up of two or more la tent classes corresponding to different levels of symptomatology. Unlike mo st previous applications of LCA this model allows us to model the extent to which the prevalence of the disorder is the same across informants: and, i n addition, the association between informants. The data set comes from a p rospective longitudinal study of 2264 children from Quebec (1155 boys and 1 109 girls). In grade 2, teachers and mothers independently rated each child on three physical aggression behavior symptoms. We satisfactorily accounte d for the cross-classification of the behavior symptoms by postulating the existence of two latent variables - one for each informant - each made up o f three latent classes of children: low-, medium-, and high-aggressive. The results showed that the prevalence of low- and medium-aggressive children in the population differed from teacher to mother, but that the prevalence of high-aggressive children did not. We found that the association between teacher and mother was large and positive and did not vary according to the child's physical aggression state or gender; in contrast, the association between physical aggression and gender was not the same for mother and teac her. Limitations and other potential applications of the proposed model are discussed.