Causal inference in longitudinal studies with history-restricted marginal structural models.

Area of Science:

• Causal Inference
• Longitudinal Data Analysis
• Statistical Modeling

Background:

• Marginal Structural Models (MSMs) are used for causal inference in longitudinal studies.
• Standard MSMs define causal effects based on the entire treatment history from study start to outcome collection.
• This approach may not always be optimal or practicable for public health research.

Purpose of the Study:

• Introduce and formalize the statistical framework for History-Restricted Marginal Structural Models (HRMSMs).
• Demonstrate how HRMSMs offer a more flexible and practical approach to causal analysis in longitudinal studies.
• Highlight the advantages of HRMSMs in terms of computational tractability and statistical power.

Main Methods:

• Develop three consistent estimators for HRMSM parameters: Inverse Probability of Treatment Weighted (IPTW), G-computation, and Double Robust (DR) estimators.
• Establish the formal statistical framework underlying HRMSMs.
• Adapt standard causal inference assumptions (counterfactuals, consistency, time-ordering, sequential randomization) for HRMSM parameter identification and estimation.

Main Results:

• HRMSMs allow causal effect analysis based on a user-specified, shorter exposure history.
• HRMSMs improve computational efficiency and mitigate statistical power concerns in study design.
• The proposed IPTW, G-computation, and DR estimators are consistent under sufficient model assumptions.
• Standard assumptions for MSM identification and estimation also ensure HRMSM parameter identification and estimation.

Conclusions:

• HRMSMs provide a valuable extension to MSMs for causal inference with longitudinal data.
• These models enhance flexibility, computational aspects, and statistical power in research.
• HRMSMs and their associated estimators are well-supported by established causal inference principles.