The authors assessed physicians' probability estimates of coronary artery disease (CAD) in 250 patients undergoing a screening exercise stress test. True likelihood of disease (prevalence) was derived from the literature. Discrimination and calibration were assessed by comparing physicians' probability estimates and prevalence using pairwise comparisons, rank correlation, and linear regression. There were differences in the discriminative abilities of the physicians based on patient characteristics. For example, the physicians had better discriminative ability for patients with typical cardiac chest pain compared with atypical chest pain. The physicians were able to predict the prevalence of CAD in broad groups of patients. However, they overestimated probabilities for patients with low prevalence of disease and underestimated probabilities for patients with high prevalence of disease. The authors conclude that physicians make consistent errors in the use of probability estimates. The quality of these estimates depends on patient characteristics such as type of chest pain and true likelihood of disease.