Optimal balancing of time-dependent confounders for marginal structural models

Abstract Marginal structural models (MSMs) can be used to estimate the causal effect of a potentially time-varying treatment in presence time-dependent confounding via weighted regression. The standard approach using inverse probability weighting (IPTW) sensitive model misspecification and lead high-variance estimates due extreme weights. Various methods have been proposed partially address this, including covariate balancing propensity score (CBPS) mitigate misspecification, truncation stabilized-IPTW (sIPTW) temper In this article, we present kernel optimal (KOW), convex-optimization-based that finds weights for fitting MSMs flexibly balance confounders while simultaneously penalizing weights, directly addressing above limitations. We further extend KOW control informative censoring. evaluate performance simulation study, comparing it with IPTW, sIPTW, CBPS. demonstrate use studying initiation on time-to-death among people living human immunodeficiency virus negative advertising elections United States.