Weighting in marginal structural models with intermediate confounder-mediators in longitudinal studies

In longitudinal settings, conventional regression methods that adjust for confounding may not be directly applicable when treatment and confounders are both time-varying, and some confounders are also in the treatment pathway. Marginal structural models (MSMs) are designed to address this problem and have been widely used in epidemiologic studies. MSMs consist of two parts: a treatment model and an outcome model. The treatment model is used to construct weights. However, complex longitudinal data structures allow multiple ways to proceed, affecting both the causal parameter of interest and the validity of statistical estimation and inference. We evaluate two approaches to the treatment models that estimate two different causal effects of the treatment: the marginal effect and the controlled direct effect. We also explore the impact of treatment model misspecification on treatment effect estimation. We illustrate the concepts by estimating the effect of hormonal contraceptives on HIV-1 infection risk using prospective studies in African HIV-1 discordant couples. Through different treatment model specifications, MSMs can be used to estimate both the marginal and the controlled direct effect of a time-varying treatment in settings with time-varying confounders. Treatment model misspecifications may lead to bias in estimation, especially when time-varying confounders are missing.