Rank-Minimization for balanced assignment of subjects in clinical trials

Abstract

Minimization (M) is the most popular algorithm for balancing large numbers of subject variables in treatment groups of small clinical trials. However, its use has been limited because of its complexity, vulnerability to selection bias and lack of a generally accepted method for statistical analysis of the data. Rank-Minimization (RM) is a promising new algorithm. It is less complex since it does not require unique programming for each clinical trial to convert continuous to categorical variables. In this study RM is compared to M for balance of variables and vulnerability to selection bias in 1000 simulated trials using 200 subjects with 15 continuous variables. With RM there were no instances of significant imbalance to cause rejection of the null hypothesis, i.e. a Student's t> or =2, although it occurred in 0.4% of the 15000 tests for M. For moderate imbalance, i.e. 1< or = t < 2, the figures were 3% (RM) and 12% (M). The probability of guessing the next assignment was 0.636 (RM) and 0.683 (M). The smaller figure is superior to that of restricted randomization in blocks of five per treatment group. Improvement in balance, a decrease in vulnerability to selection bias and ease of application along with improvements in the statistical analysis should result in the general acceptance of RM for assigning subjects to treatment groups in clinical trials.