A difficulty in the interpretation of the reliability/stability of a lifetime diagnosis of mental disorders is the lack of a theoretical perspective. A model expressed in terms of the three unknowns-sensitivity, specificity and true base rate-is problematic due to the lack of a "gold standard", so that only two of these unknowns can be estimated. We extend this model to allow for clinical covariates that increase the likelihood that a positive case at time 1 will be positive at time 2. Under the assumption that all observed cases are true cases at the highest covariate values, we obtain a direct estimate of the sensitivity, so that all unknowns can be estimated. Moreover, we then calculate the likelihood that an observed case with given covariate levels is in fact a true case. The implications of diagnostic error for the estimation of incidence and population rates and the fitting of genetic models are given. These methods are applied to stability data collected as part of the NIMH Psychobiology of Depression Program. A total of 982 relatives have been assessed with interviews separated by a 6 year interval. A logistic function was used to model the stability in 264 relatives with an initial lifetime diagnosis of Major Depressive Disorder. The significant covariate predictors in a step-wise analysis were the number of depressive symptoms in the worst episode, the number of episodes and receiving medication for the worst episode. The sensitivity was estimated to be 0.918 and the specificity was estimated to be 0.940. However, even with these high values, less than half of the reported first episode cases with 3 symptoms who did not receive medication are true cases. These data illustrate that the modelling of diagnostic error, as well as including clinical covariates, will be useful in analyzing epidemiologic or family data for psychiatric illnesses.