OBJECTIVE: To assess the accuracy of clinicians in predicting violence in mental patients. Specifically, to determine if clinicians can predict violence when variation in rates of violence attributable to age, race, and sex is controlled. DESIGN: Two samples of psychiatric patients, matched on age, race, sex, and admission status, were followed up in the community during a 6-month period. One group included individuals assessed by psychiatric emergency department clinicians as likely to be violent to another person during the follow-up period; the other was a comparison group. Patients provided self-reports of violent incidents, and a "collateral," ie, an individual with detailed knowledge of the patient's life, provided this same information. Official records were also reviewed. SETTING: Patients were recruited in the emergency department of a metropolitan psychiatric hospital. Patients and collaterals were interviewed in their homes or in public places in the community. PATIENTS OR OTHER PARTICIPANTS: A consecutive sample of individuals coming into a psychiatric emergency department during daylight and evening shifts was obtained. A total of 2452 patients were approached for consent and 1948 consented. A final sample of 357 patients whom clinicians assessed as likely to be violent and their matched comparison patients were included. MAIN OUTCOME MEASURES: Patients', collaterals', and official records' reports of incidents in which the patient laid hands on another person or threatened someone with a weapon. RESULTS: Violence during the follow-up period was reported in approximately 45% of the cases: 36% in the comparison group and 53% in the cases predicted to be violent. Overall clinical accuracy was significantly better than chance, but predictions of female patients' violence were not better than chance. CONCLUSIONS: The level of patient violence reported using self-reports and collateral reports was higher than has been obtained using other methods. Clinical judgment adds to predictive accuracy, but overall accuracy was modest and particularly low for female patients.