Primary Submission Category: Weighting
A Robust Sequential Covariate Balancing Approach for Estimating the Effects of Time-Varying Treatments on Survival Outcomes
Authors: Yige Li, José Zubizarreta,
Presenting Author: Yige Li*
In longitudinal studies, treatments or exposures can vary across time and depend on covariates responding to previous exposures. For such studies, we propose Robust Sequential Covariate Balancing (RSCB), a flexible and stable weighting technique, designed to estimate the effects of time-varying treatments on a general class of outcomes, including survival outcomes. RSCB utilizes a backward covariate balancing procedure for identification and estimation. RSCB has a product form parallel to inverse probability weighting (IPW) but utilizes modellable trends in the covariates and meanwhile minimizes weights variation. Unlike IPW methods, whose estimates can converge at slow rates because of inadequate covariate overlap, RSCB gains efficiency by prioritizing outcome-relevant forms of covariate balance. In comparison to the g-computation formula, RSCB does not extrapolate and maintain robustness to misspecification of covariates and outcomes models. In contrast with longitudinal stable balance weighting (LSBW), RSCB can accommodate multiple types of outcomes, longer time courses, and higher covariate dimensions. We illustrate this new method in a study of the effects of peer antisocial behaviors on drug relapse of adolescents.