Advances in computer technology have made it possible to store large amounts of observational data concerning treatment of patients for medical disorders. It has been suggested that these data banks might replace randomized clinical trials as a means of evaluating the efficacy of therapies. A review of the methodological problems likely to arise in analyzing such data for the purpose of comparing treatments suggests that sound inferences would not generally be possible because of difficulties with bias in treatment assignment, nonstandard definitions, definitions changing in time, specification of groups to be compared, missing data, and multiple comparisons.