Identification of responders and non-responders in clinical trials using the "Bump Hunting" procedure
by
Kurt Ulm
Technical University Munich, Germany
Coauthors: Victoria Kehl (Tehcnical University Munich, Germany)

We propose a method for identifying responders and non-responders, which uses as a building tool the Bump Hunting procedure defined by Friedman and Fisher (1999). In the first step of the suggested method, a Cox-PH model is developed on one of the treatment groups (e.g. placebo) in order to adjust for prognostic factors. This model is then applied to the second treatment group and its Martingale residuals are used as the target variable of the Bump Hunting analysis. Bump Hunting creates a system of rules involving predictive factors, which identify two "bumps, " i.e. groups of patients, with large positive and large negative values of the residuals. Those are the non-responders and the responders respectively. The two treatment arms are then compared in each of the bumps.

The proposed method was applied to a controlled clinical trial with about 1200 patients, in which the verum and the placebo groups showed no significant difference in patient survival. We did, however, find two groups of patients who showed differences in the survival due to treatment. One "bump" containing patients with large positive residuals in the verum group (the non-responders, n = 115) lead to a significantly lower survival rate (p = .011) than the corresponding placebo group. A second bump of patients with large negative values was found (the responders, n = 89), who showed a significant benefit from verum (p = .024).

Friedman, J. H. & Fisher, N. I. (1999). Bump Hunting in High-Dimensional Data, Statistics and Computing, 9 (2), 123-162.

Date received: September 10, 2001