Revisiting g-estimation of the Effect of a Time-varying Exposure Subject to Time-varying Confounding

Abstract Marginal Structural Models (MSMs), with the associated method of inverse probability weighting (IPW), have become increasingly popular in epidemiology to model and estimate the joint effects of a sequence of exposures. This popularity is largely related to the relative simplicity of the method, as compared to other techniques to adjust for time-varying confounding, such as g-estimation and g-computation. However, the price to pay for this simplicity can be substantial. The IPW estimators that are routinely used in applications make inefficient use of the information in the data, and are susceptible to large finite-sample bias when some confounders are strongly predictive of exposure. Moreover, the handling of continuous exposures easily becomes impractical, and the study of effect modification by time-varying covariates even impossible. In view of this, we revisit Structural Nested Mean Models (SNMMs) with the associated method of g-estimation as a useful remedy, and show how this can be implemented through standard software.