In using a regression tree to identify subgroups with differential treatment effects, the first requirement is that the method be free of bias in the selection of variables to split the nodes. A second requirement is that the method be able to differentiate between the effects of prognostic and predictive variables. A third requirement is that it does so with good accuracy. This talk will discuss a recent extension of GUIDE that satisfies these requirements and compares its performance with that of other existing regression tree methods.