Hospital morbidity data in the form of International classification of diseases, 9th revision, clinical modification codes are often used for epidemiological studies and disease surveillance. We aimed to evaluate the reliability of the Victorian In-patient Minimum Database for use in epidemiological studies and disease surveillance. Data from 1993-94 were collected, as part of a coding audit of public hospitals in Victoria, from 7052 randomly selected records. The frequency of discrepancy in any coding field was 53 per cent, and of discrepancy in the principal diagnosis, 22 per cent. New Australian national diagnosis-related group (ANDRG) codes were assigned as a result of discrepancy in 13.6 per cent of cases. Discrepancy rates increased with increasing rarity of ANDRG, from 50 per cent to 56 per cent. Predictors of change in ANDRG assignment were discrepancy in the principal diagnosis, ANDRG frequency of over 0.6 per cent, more than three diagnoses, medical ANDRGs, length of stay over five days and rural hospitals. Rates of any discrepancy increased from 36 per cent in patients with one diagnosis to 94 per cent in patients with 12 diagnoses. The discrepancy rates were consistent with those of other studies. Coding discrepancy is likely to be caused by universal difficulties associated with the coding of hospital records, rather than any unique local problems. The predictors of discrepancy suggest that more complex cases are more prone to coding discrepancy. In areas where the database is less reliable, use of a supplementary data source, such as link-age studies, would improve reliability.