Instrumental Variable Estimation of Marginal Structural Mean Models for Time-Varying Treatment

Area of Science:

• Epidemiology
• Biostatistics
• Longitudinal Data Analysis

Background:

• Marginal Structural Models (MSMs) address time-varying confounding in longitudinal studies.
• Identification of MSM parameters relies on the Sequential Randomization Assumption (SRA), which assumes no unmeasured confounding.
• Unmeasured confounding can violate SRA, necessitating alternative identification strategies.

Purpose of the Study:

• To develop sufficient conditions for identifying Marginal Structural Mean Models (MSMMs) parameters when SRA is violated due to unmeasured confounding.
• To utilize time-varying instrumental variables as an alternative to SRA for causal inference.
• To propose and evaluate a weighted estimator for MSMMs under these new conditions.

Main Methods:

• Developed identification conditions for MSMMs using time-varying instrumental variables.
• Required that no unobserved confounder predicts compliance type for time-varying treatments.
• Proposed a simple weighted estimator and assessed its finite-sample properties via simulation.

Main Results:

• The proposed weighted estimator demonstrated finite-sample properties in simulation studies.
• The method was applied to investigate the effect of delivery hospital type on neonatal survival.
• Sufficient conditions for identification were established even with unmeasured confounding.

Conclusions:

• The study provides a viable method for causal inference in longitudinal studies with time-varying confounding when SRA does not hold.
• Time-varying instrumental variables offer a powerful tool for addressing unmeasured confounding in MSMMs.
• The findings have implications for estimating treatment effects in observational health research, such as evaluating hospital quality on neonatal outcomes.