Analysis of clinical trial outcomes: alternative approaches to subgroup analysis

Subgroup analysis is understood as the investigation of the influence of factors other than treatment factors on the response variables or treatment effects in clinical trials. Usual techniques are the Mantel-Haenszel technique and regression technique. The Mantel-Haenszel technique combines (pools) treatment effects (expressed by relative odds) in different subgroups to an "overall" effect. This is meaningful if all subgroups have the same treatment effects. In general, however, there will be interactions between treatment and subgroups and the Mantel-Haenszel estimates will be misleading. Regression models allow analysis of interactions if the interactions are modeled by product terms. If there are many subgroup factors the number of product terms necessary for an adequate modeling of the interactions may be higher than the number of observations and an analysis of the interactions is impossible. To overcome these difficulties as an alternative approach the "classification tree" technique is proposed. This technique combines the possible subgroup constellations to subgroup clusters with homogeneous response within the cluster. From the distribution of the subgroup constellation to the different clusters inferences about the interaction structure can be obtained. The technique is demonstrated with an example of a long-term clinical trial.