Psychiatric diagnoses provide short labels for diseases or discrete symptom clusters. They should designate the same throughout the world, give information about course, outcome and indications for therapy as well as provide an heuristic basis for etiological research. Hence, the core question is how to attain an optimal representation of real morbidity in diagnosis, sets of diagnostic criteria and diagnostic classifications. Clinical observation can be improved considerably by multi-centre field trials, as applied in the preparation of ICD-10 and DSM-IV. But the approach has considerable limitations due to a lack of external measures in many psychiatric disorders and a highly limited representation of many diagnostic groups in clinical populations. Therefore, epidemiological methods are required in validating diagnosis and diagnostic criteria. The simplest way is to supplement clinical multicentre diagnostic studies by general-practice studies, but these, also, cannot fully replace population studies. Operational diagnosis and case criteria can be defined either categorically or dimensionally. Most of the categorical diagnoses in ICD-10 or DSM III also include dimensional characteristics. The impact of various diagnostic criteria, particularly cut-offs of dimensional characteristics, on the assignment of a diagnosis and, thus, on the morbidity figures of a diagnostic category is demonstrated by data from a large representative sample of first-admitted schizophrenics. Attempts at etiological validation by methods of genetic epidemiology provide limited support for Kraepelin's dichotomous model of functional psychoses. Validation by epidemiological course studies has shown that the stability of diagnosis in functional psychoses differs according to the sets of diagnostic criteria of different classification systems.