Improving the aggregate performance of psychiatric diagnostic methods when not all subjects receive the standard test.

Family studies of disease incidence often include some subjects who receive a diagnostic evaluation less accurate than that obtained with a standard method. This is particularly true of family studies of mental illness, where the standard is a consensus diagnosis based on both direct interview and corroborating family history from an informant in the family. Family members who are not interviewed have diagnoses based on history alone. Interviewed and uninterviewed relatives differ in several factors that influence the probability of disease, so that interview is not independent of disease. Thus, one cannot use the interview data to estimate the overall rate of illness. Family history-based illness rates may substantially underestimate true rates, so the observed rate in uninterviewed subjects is also a biased estimate. We discuss several models to reduce the bias in estimation of incidence rates. We propose explicit modelling and imputation as alternatives to the implicit assumptions that constitute the basis of the methods in current use. A clinical example involving 4806 relatives of probands with major affective illness illustrates the statistical issues.