The possibility that separation of a categorical depressive disease ('melancholia') from remaining depressive disorders can be improved by assessment of mental state signs was examined in patients treated by representative Sydney psychiatrists and patients referred to a specialised mood disorders unit. A set of signs, principally assessing retardation, was derived within the two samples by principal-components and latent-class analyses. Scores were significantly correlated with clinical, DSM-III, and RDC diagnoses, and appeared independent of severity, suggesting that melancholia can be defined phenomenologically. Scores were also associated with several 'validating' factors. Comparative analyses of a refined list of melancholia symptoms suggested that ratings of defined signs are likely to have greater capacity than symptom ratings to differentiate melancholia from residual depressive disorders.