A comparison of five major categories of stroke in 651 patients revealed significant differences in the frequencies of diagnoses at the Beth Israel and Massachusetts General hospitals in Boston, Mass. (P less than 0.001 by chi-square test). To analyze these differences, we modeled the diagnostic process at each hospital with a Bayesian procedure and performed a crossover study in which each patient was rediagnosed by the model from the opposite hospital. The results indicate that the differences in the frequency of lacune and subarachnoid hemorrhage were associated with the patient population, whereas the differences in the frequency of embolism and atherothrombosis were associated with the diagnostic process. There was a marked difference in the use of arteriograms on the two stroke services, but no difference in morbidity or mortality. The modeling procedure described can be used to compare clinical processes when the allocation of patients is thought to be biased.