Cumulative Probability Models for Semiparametric G-Computation

Abstract

Time-varying confounding is a commonly encountered challenge in longitudinal observational studies that seek to evaluate the causal effect of a time-dependent treatment. Because a time-varying confounder is influenced by prior treatment while simultaneously serving as a cause of later treatment, simple approaches to account for confounding such as regression adjustment are insufficient for such scenarios. G-computation (a longitudinal generalization of standardization) can be implemented to estimate the total causal effect of the treatment. While g-computation can accommodate challenges such as censoring and truncation by death, it sometimes gets criticized for its reliance on parametric models and possible non-robustness to model misspecification. In this work, we explore semi-parametric cumulative probability models (CPMs) for use within g-computation. We use simulation techniques to evaluate the finite-sample properties of this approach. We further apply this approach to a fully-simulated data set that mimics properties of a SEER (Surveillance, Epidemiology, and End Results)-Medicare linked database of women with endometrial cancer. Specifically, we implement a nested g-computation approach in this data set to estimate mean three-month cumulative costs for specified longitudinal treatment trajectories. Our results suggest that the CPM approach has desirable finite-sample properties including low bias, and that the CPM is a promising tool for causal inference in longitudinal studies.