Clinical and administrative databases are increasingly used for clinical research. Descriptive studies and analyses of prognostic factors are established research uses of databases, but using them to compare therapies remains controversial. Such comparisons may be possible when a validated model of prognosis can account for the effect of clinically recognized variables on outcome. Models meeting these criteria have been developed using the Duke Cardiovascular Disease Database. They were applied to compare medical therapy to surgical therapy for coronary artery disease. Predictions from these models agreed well with the results of the three major randomized trials of bypass surgery versus medical therapy. These findings indicate that when data is reliable and understanding of factors affecting prognosis is good, researchers can use statistical techniques to minimize the bias due to non-random treatment assignment. Carefully performed analyses of observational clinical data can complement and extend randomized studies.