Using diagnostic stability data from independent diagnostic interviews conducted 6 years apart, we determine which diagnoses are predictive of diagnoses 6 years later. Logistic analysis using categorical predictors is used to establish ordinal relationships and to suggest diagnostic hierarchies. The multiple-threshold multifactorial model is used to estimate the within-person correlation over time. Rather than use a simple dichotomy of "affected" or "unaffected," we provide odds ratios for mania, hypomania, and major depressive disorder in terms of diagnostic hierarchies, allowing a ranking of these diagnoses. This division increases the information for genetic studies or studies of a phenotype with correlated biological or environmental continuous covariates. The diagnosis of schizophrenia shows remarkable specificity across occasions. We find significant error in a cross-sectional assessment in this nonclinical sample. Assuming a multifactorial model, the proportion of variance in liability due to assessment error is approximately 30 percent under all schemes considered. The use of repeated measures in family studies is thus strongly recommended.