ENSURING BALANCED DISTRIBUTION OF PROGNOSTIC FACTORS IN TREATMENT OUTCOME RESEARCH

In comparative or matching research involving two or more treatments, the equivalence of the patient groups is of critical importance. In th e past, equivalence has either been imposed by matching or balancing, or has been assured statistically by randomization. Matching and balan cing, while useful in many contexts, nonetheless have important limita tions, as does simple randomization. In recent years, a new tool has b een developed that represents a compromise between balancing and rando mization. This method, um randomization, gives clinical investigators new options for improving the credibility of studies at a relatively m odest cost. Urn randomization is randomization that is systematically biased in favor of balancing. It can be used with several covariates, both marginally and jointly, producing optimal multivariate equivalenc e of treatment groups for large sample sizes. It preserves randomizati on as the primary basis for assignment to treatment and is less suscep tible to experimenter bias or manipulation of the allocation process b y staff than is balancing. Disadvantages include the fact that it is m ore difficult to implement, and that it violates the simple probabilit y model of simple randomization. A number of research studies on addic tions, including client-treatment matching trials, have used urn rando mization. A summary of the mechanics of urn randomization is presented , and guidelines for its use in treatment studies are discussed.