Philippe Boileau, Ning Leng, Nima Hejazi, Mark van der Laan and Sandrine Dudoit

Current approaches for uncovering treatment effect modifiers are limited to low-dimensional data or data with weakly correlated confounders, or are restricted to simple data-generating processes. We develop a general framework for defining model-agnostic treatment effect modifier variable importance parameters applicable to high-dimensional data with arbitrary correlation structure, deriving nonparametric estimators of these parameters, and establishing these estimators’ asymptotic properties. We showcase this framework by deriving absolute and relative treatment effect modifier variable importance parameters for data-generating processes with continuous, binary and time-to-event outcomes with binary exposures and potentially high dimensional confounders. One-step, estimating equation and targeted maximum likelihood estimators for each parameter are provided. Certain estimators are proven to be double-robust under non-stringent conditions. All are asymptotically linear under reasonable entropy constraints on the data-generating process and consistency-rate requirements on the nuisance parameter estimators. Numerical experiments with moderate- and high-dimensional confounders demonstrate that these estimators’ asymptotic guarantees, like false discovery rate control, are approximately achieved in realistic sample sizes for observational and randomized studies alike.

• A flexible approach for predictive biomarker discovery by Boileau et al. (2022)
• A nonparametric framework for treatment effect modifier discovery in high dimensions by Boileau et al. (2023 )

• uniCATE
• unihtee