The Randomized Clinical Trial
Harrington, David P., Journal of the American Statistical Association
The randomized clinical trial is among the most important methodological tools in biostatistics. Some have conjectured that it could be the most significant advance in scientific medicine in the twentieth century (Smith 1998). In the statistical community, it is the method of choice for controlling confounding in medical studies where two or more interventions are compared and where the choice of intervention by either the subject or the study investigator could lead to selection bias. Among physicians, it is central to evidence-based medicine and is the major step along the way from treatment choice based on opinion and experience to that based on scientific finding.
Randomization in experimental designs is, of course, not limited to studies of human subjects. Fisher (1926, 1935) advocated it in the study of biological problems, and randomization became widely used in agricultural experiments. But randomized clinical trials in medicine have special features that set them apart from other experimental designs: the experimental units are human subjects, sometimes diagnosed with a potentially fatal disease, and the study subjects and their physicians cede treatment choice to a random allocation scheme. A methodologic device for statisticians supplants one of the central aspects of the patient--physician relationship--personal choice of treatment. This intrusion into the caregiving role in medicine will always spark controversy (Hellman and Hellman 1991).
It is impossible to estimate the frequency with which randomized trials have been used in medicine, but some estimate that the number of randomized trials runs into the hundreds of thousands (Chalmers 1998). The Cochrane Collaboration (1999) lists review teams in more than 50 diseases that maintain libraries of reviews of published trials, and that archive now indexes more than 250,000 trials.
Close relatives of the randomized clinical trial were conducted as early as 1747 (Lind 1753), when James Lind studied six strategies for treating scurvy on the British ship HMS Salisbury. Although Lind appears to not have assigned his treatments randomly, he recognized that bias can result when responses are correlated more with, disease characteristics or environmental features than with a putatively successful treatment. He surely sought to avoid the shoals described by his contemporary, Thomas Jefferson, in an 1807 letter to Dr. Caspar Wistar: "The patient, treated on the fashionable theory, sometimes gets well in spite of the medicine. The medicine therefore restored him, and the young doctor receives new courage to proceed in his bold experiments on the lives of his fellow creatures" (Jefferson 1984). Lind's study design attempted to hold constant the factors within the control of the investigator (other aspects of diet, the environment of the patient) while varying only the treatment. By today's standa rds, Lind's study was flawed for two reasons: He assigned only two patients to each treatment (he would not be the last investigator to conduct an underpowered study), and he apparently failed to recognize that unmeasured patient characteristics beyond his observation and control could influence results.
The modern randomized trial emerged in the midtwentieth century. Some believe that the 1948 report in the British Medical Journal (Medical Research Council 1948) of a randomized trial comparing streptomycin with a no-treatment control is the first detailed published account of randomized treatment assignments. The use of randomization in that trial was due in large part to the pioneering work of Bradford Hill. In addition to controlling possible confounding, randomization apparently also provided an acceptable way to ration the small supply of streptomycin (Doll 1998). Within 20 years, randomization in medical research had become widely adopted.
In one of the delightful ironies of modern science, the randomized trial "adjusts" for both observed and unobserved heterogeneity in a controlled experiment by introducing chance variation into the study design. …