Marc Buyse, ScD IDDI, San Francisco, USA and I-BioStat, University of Hasselt, Belgium

Generalized pairwise comparisons in randomized trials of experimental interventions

Abstract:

Generalized pairwise comparisons extend the Mann-Whitney form of Wilcoxon’s test, or Gehan’s generalization of this test for two samples of (possibly censored) observation.  The test uses all pairwise comparisons between two patients, one in the treatment arm and one in the control arm, in terms of one or several prioritized outcomes. Each pair favors treatment, control, or neither. The treatment benefit is the difference between the proportion of pairs in favor of treatment less the proportion of pairs in favor of control; this statistic is called the “net treatment benefit”. For a single variable, this measure of treatment benefit has a simple relationship with traditional measures of benefit such as the risk difference (for binary outcomes), the mean difference (for continuous outcomes), or the hazard ratio (for right-censored outcomes). The pairwise comparison approach easily incorporates a threshold of clinical relevance in the analysis. For instance, if survival is the outcome of interest, a threshold of m months can be defined to estimate the net chance of a longer survival by at least m months. The pairwise comparison approach also allows several prioritized outcomes, for instance time to tumor progression and overall survival, to be analyzed simultaneously. More generally, this approach allows an overall benefit risk assessment to be performed, thus paving the way towards personalized medicine.