Responder Identification in Clinical Trials with Censored Data

We present a newly developed technique for identification of
positive and negative responders to a new treatment which was
compared to a classical treatment (or placebo) in a randomized
clinical trial. This bump-hunting-based method was developed for
trials in which the two treatment arms do not differ in survival
overall. It checks in a systematic manner if certain subgroups,
described by predictive factors do show difference in survival due
to the new treatment. Several versions of the method were discussed
and compared in a simulation study. The best version of the
responder identification method employs martingale residuals to a
prognostic model as response in a stabilized through bootstrapping
bump hunting procedure. On average it recognizes 90% of the time
the correct positive responder group and 99% of the time the
correct negative responder group.