Abstract:
|
For the task of predicting differential treatment response based on biomarkers in clinical development, combination(s) of multiple biomarkers may be necessary to achieve better performance. For continuous biomarkers, existing parametric methods rely on a priori model assumptions; on the other hand, some non-parametric approaches, e.g., splines, quickly lose interpretability when modeling joint effects of multiple predictors. Original recursive partitioning algorithms identify cutoffs on individual predictors, while sometimes index variables of small (but unknown) numbers of candidate biomarkers can better capture the boundaries of subgroups. Building on sparse sufficient dimension reduction, we present a useful approach to identify such index variables of linear combinations without making assumptions on the functions that link them to the treatment outcome. This approach allows co-existence of both prognostic and predictive indices, and the resulting variables can be further adopted by the different approaches mentioned above as new predictors. Simulation studies will be presented to evaluate its performance under various scenarios.
|