Comparison of the Cox model and the regression tree procedure in analysing a randomized clinical trial

In a clinical trial comparing different treatments the patients may be rather heterogeneous with regard to their natural prognosis. Simple overall comparison of the treatment groups may lead to a biased estimate of the treatment effect even in a well-balanced randomized study, at least when survival time is the outcome. An adequate analysis of the treatment effect is only feasible in a multivariate framework where the important prognostic factors are accounted for and, additionally, treatment-covariate interactions may be evaluated. Analyses using the Cox model are compared with alternative approaches based on the Classification and Regression Tree (CART) technique. Basic differences between these approaches are outlined and discussed in the context of a randomized clinical trial of chemotherapy in patients with brain tumours.