The randomized clinical trial (RCT) has become generally accepted as the most valid study mechanism for assessment of therapeutic regimens. Because random allocation of treatment guards against any systematic association of treatment assignment with prognosis, differences in outcome that are observed among treatment groups in an RCT will tend to result from treatment effects and not from inherent differences among the groups. Studies in which treatment allocation is by design associated with patient characteristics—eg, where treatment is selected individually for each patient, or according to the clinic to which the patient comes, or according to the time period in which the patient is treated—are vulnerable to selection bias. This bias is incurred whenever patients with differing prognoses have differing probabilities of receiving the treatments to be compared.
In an article in this issue (p 2455), Horwitz and Feinstein explore the potential usefulness of another approach to the evaluation of therapeutic