Tree-based exploratory identification of predictive biomarkers in observational data

TL;DR

This paper introduces a tree-based method called predMOB for identifying predictive biomarkers in observational data, demonstrating its effectiveness with confounder adjustment strategies through simulations and real data application.

Contribution

It combines predMOB with confounder adjustment methods, showing how to accurately identify predictive factors in observational studies, especially when confounding is present.

Findings

Covariate adjustment correctly identifies predictive factors under confounding.

IPTW may fail when the predictive factor is linked to confounders.

Combining covariate adjustment and IPTW offers robustness in complex scenarios.

Abstract

The idea of "stratified medicine" is an important driver of methodological research on the identification of predictive biomarkers. Most methods proposed so far for this purpose have been developed for the use on randomized data only. However, especially for rare cancers, data from clinical registries or observational studies might be the only available data source. For such data, methods for an unbiased estimation of the average treatment effect are well established. Research on confounder adjustment when investigating the heterogeneity of treatment effects and the variables responsible for this is usually restricted to regression modelling. In this paper, we demonstrate how the predMOB, a tree-based method that specifically searches for predictive factors, can be combined with common strategies for confounder adjustment (covariate adjustment, matching, Inverse Probability of Treatment Weighting (IPTW)). In an extensive simulation study, we show that covariate adjustment allows the correct identification of predictive factors in the presence of confounding whereas IPTW fails in situations in which the true predictive factor is not completely independent of the confounding mechanism. A combination of both, covariate adjustment and IPTW performs as well as covariate adjustment alone, but might be more robust in complex settings. An application to the German Breast Cancer Study Group (GBSG) Trial 2 illustrates these conclusions.