In Victoria injury surveillance data are drawn from hospital morbidity data. The accuracy and reliability of these data are often questioned. We aimed to ascertain the reliability of injury data in the Victorian inpatient minimum database. A random sample of 546 public hospital separations with principal diagnosis ICD-9-CM codes 800-999 was selected from four metropolitan hospitals. Medical records were reviewed, and the hospital coding was compared with the record content. The frequency of error in any coding field was 73 per cent (349/480); of diagnosis error, 61 per cent (292/480); of procedure error, 45 per cent (168/370); of error in the principal diagnosis, 19 per cent (93/480); and of error in external-cause codes (E-codes), 16 per cent (75/480). Ninety-four per cent of errors (87/93) in the principal diagnosis involved recoding within the same group of codes. Only 6 per cent (6/93) were recoded to principal diagnoses other than injury. Sixty-two per cent (181/292) were errors of omission of codes for comorbid conditions. Nearly half the errors in the principal diagnosis were minor, involving the last two digits. E-codes were more complete than diagnosis codes. The best predictors of error in the principal diagnosis were greater length of stay, type of injury code (poisonings and toxic effects were associated with lower error rates) and death as the outcome. While selection of data from secondary diagnosis fields may not provide complete data, the use of the principal-diagnosis code and E-codes for injury surveillance is feasible and reliable. The database is a valuable source of injury surveillance data, bearing in mind the limitations of coded hospital morbidity data.