Abstract
Marginal structural models (MSMs) estimate the causal effect of atime-varying treatment in the presence of time-dependent confounding viaweighted regression. The standard approach of using inverse probability oftreatment weighting (IPTW) can lead to high-variance estimates due to extremeweights and be sensitive to model misspecification. Various methods have beenproposed to partially address this, including truncation and stabilized-IPTW totemper extreme weights and covariate balancing propensity score (CBPS) toaddress treatment model misspecification. In this paper, we present KernelOptimal Weighting (KOW), a convex-optimization-based approach that findsweights for fitting the MSM that optimally balance time-dependent confounderswhile simultaneously controlling for precision, directly addressing the abovelimitations. KOW directly minimizes the error in estimation due totime-dependent confounding via a new decomposition as a functional. We furtherextend KOW to control for informative censoring. We evaluate the performance ofKOW in a simulation study, comparing it with IPTW, stabilized-IPTW, and CBPS.We demonstrate the use of KOW in studying the effect of treatment initiation ontime-to-death among people living with HIV and the effect of negativeadvertising on elections in the United States.