In studying the risk of affective disorder in children, the investigator must deal with the problem that there are two possible units of analysis: the child and the family. An analysis based on children must take account of the intercorrelation within a sibship to produce correct results, while a family-based analysis makes it difficult to investigate individual characteristics of children that help determine the net risk. A two-stage iterative approach to this problem is proposed, yielding estimates of the effect of family-based factors (parental illness, family social class, marital status of parents) and individual factors (age and sex of child, previous non-affective illness). This technique is applied to a sample of 275 children from 143 families representing a wide range of familial risk for affective disorder. The final family-based model (predicting at least one child with affective disorder in the sibship) indicates a six-fold increase in risk to the child associated with maternal affective disorder (P less than 0.001), a three-fold increase in risk associated with paternal affective disorder (P less than 0.05) and divorce or separation of the biological parents, and a suggestion of increased risk in the highest social class (P = 0.06). The excess sibship risk, due to child factors age, prior anxiety disorder, and prior childhood diagnosis, contributed significantly to the family prediction (P less than 0.001).