Methods for estimation and evaluation of marker-guided treatment rules based on multivariate marker panels

Abstract

Due to vast heterogeneity in patients' responses, a uniformly preferred treatment is often not available. In such cases, clinical practice may be enhanced by use of person-level information that could guide treatment choice and lead to better outcomes for both treated individuals and for the population. The scientific challenge is to identify those factors that can be used to target treatment, and to accurately quantify the expected treatment benefit as a function of candidate markers. The proposed research develops statistical methodology for the generation and evaluation of a single index score, that estimates the expected treatment benefit associated with patient characteristics as a linear combination of the markers in the context of continuous and survival outcome. Our methods specifically decouple the model used to generate the treatment benefit score from non-parametric methods that are adopted to evaluate the score. Cross-validation methods ensure honest evaluation of the score performance at the population level, if it was used to guide treatment.