Using severity rather than randomization for assignment in clinical trials: Monte Carlo tests and practical illustrations of the Robbins-Zhang method.

In randomized clinical trials, patients are not differentially assigned to treatments by severity because available methods of data analysis are sensitive to regression to the mean and yield biased estimates of treatment effect. However, a method proposed by Robbins and Zhang provides consistent estimates of treatment effect in clinical trials, even when the more severely ill are assigned to the active treatment and the less severely ill are assigned to a placebo. This method was assessed by Monte Carlo trials. All combinations of two models of drug effect, five true score distributions, four magnitudes of error variance, and four sample sizes were assessed. The method works sufficiently well that its use should be considered. Further, the method gives correct results even when the regression discontinuity method fails. The method was also compared with a standard analysis of variance of difference scores using real psychiatric data from two ordinary randomized trials; similar results were obtained by both methods.