We analyzed the psychological process by which physicians solve complicated diagnostic problems, such as those posed in clinicopathological exercises. The challenge of differential diagnosis is to select the most probable cause of a patient's condition, yet the size of the problem, the nature of medical information, and the notorious inability of human beings to manipulate probabilities in their heads all conspire against the diagnostician to make it virtually impossible to employ Bayes' theorem in routine diagnosis. Unable to estimate the desired probabilities explicitly, physicians recast the problem into a form that uses one of their most effective mental skills--that of comparing patterns. A study of 50 clinicopathological conferences published in the Journal suggests that the following six steps are taken to arrive at a diagnosis: aggregation of groups of findings into patterns, selection of a "pivot" or key finding, generation of a cause list, pruning of the cause list, selection of a diagnosis, and validation of the diagnosis. Although the clinicopathological conference differs in some important ways from real-life diagnostic problems, we believe that the principles described here closely resemble those used in practice. Properly selected clinicopathological conferences are excellent windows through which to study diagnostic reasoning.