A1962
Title: Robust and efficient g-estimation of structural nested mean models in randomized trials with partial compliance
Authors: Tomohiro Shinozaki - The University of Tokyo (Japan) [presenting]
Abstract: Treatment switching and concomitant therapies often make randomized trials only partially compliant with allocation, complicating causal interpretation. G-estimation of structural nested mean models (SNMMs) in such trials is revisited and an estimator is proposed that combines two sources of information: randomization-based estimating equations and equations based on no-unmeasured confounding for observed treatment histories. The equations are solved jointly for a common SNMM using generalized method of moments (GMM). Compared with conventional randomized g-estimation, the approach can include heterogeneous causal effects of non-randomized treatments and use covariates to improve precision. Compared with observational g-estimation, it retains randomization as an instrumental source of variation and may be less sensitive to unmeasured confounding. This perspective also connects to recent work on orthogonal and debiased estimating equations, although the focus is on classical g-estimation and GMM rather than machine-learning implementation. The method is evaluated in simulations covering non-compliance patterns, proportions of non-compliers, unmeasured confounding, and misspecification of auxiliary outcome and propensity-score models. An application uses data from the SELECT BC Study, a randomized trial of first-line chemotherapy for metastatic breast cancer in which concomitant therapies and treatment switching were frequent.
