On Discovering Treatment-Effect Modifiers Using Virtual Twins and Causal Forest ML in the Presence of Prognostic Biomarkers

The recent years have seen a rapid development in the general Machine Learning area and a similar strong trend has taken place within the drug development domain, known as Precision Medicine. Traditionally, the main focus in clinical trials has been to estimate the overall treatment effect, but the reality of treatment effect heterogeneity has led to an interest in data-driven approaches for automatically identifying treatment effect modifiers and subgroups with enhanced treatment effect. These techniques are different from general purpose supervised learning due to its causal inference flavour and the notorious presence of prognostic effects in clinical trials. In this work, we focus on the popular method Virtual Twins; despite being among the earliest method, relatively little attention has been paid to the specifics of its implementation, in particular to the relative merits of arm-specific modelling versus a ’common-surface’ approach. Since Virtual Twins is a generic algorithm (and not limited to any specific predictive base model such as the originally proposed Random Forest) we choose to base it on XGboost. We also study the forest-based Casual Forest, recently popular due to its (by-design) unbiased estimates of individual treatment effects. We compare the performance of these methods regarding their ability to disentangle prognostic and predictive effects. This is of critical importance in precision medicine, since only the latter are related to enhanced treatment effects; also a certain tendency to mistakenly declare prognostic biomarkers as predictive has been reported in the ML literature.