|
Abstract:
|
Statistical methods for exploring treatment-covariate interactions such as Subpopulation Treatment Effect Pattern Plot (STEPP) have been widely utilized in clinical oncology. The choice of parameters for constructing overlapping subgroups in STEPP is crucial to characterize the final graphical pattern and identify ranges of a continuous marker with potential treatment effects. However, this critical decision is made empirically without statistical interpretations. The search for "optimal" parameters can be translated into balancing the bias/variance trade-off. We use cross-validated (CV) profile likelihood for Cox model and grid search for parameters to determine the most appropriate number of subgroups in terms of prediction error minimization and complexity control. Our method works for both tail-oriented and sliding window approaches, and provides straightforward graphics for decision-making. As CV procedures for calibration of model parameters involve repeated but independent steps, our R package "cvSTEPP" incorporates parallel computations that utilize multicores to accelerate the process. Results on simulated datasets that mimic real biomarker data in oncology are presented.
|
