Common Methods for Handling Missing Data in Marginal Structural Models: What Works and Why.

Area of Science:

• Causal inference
• Longitudinal data analysis
• Missing data methods

Background:

• Marginal structural models (MSMs) are vital for estimating causal effects in longitudinal observational studies.
• Incomplete confounder data presents a significant challenge, potentially biasing intervention effect estimates.
• Existing literature offers limited practical guidance on effective missing-data handling for MSMs.

Purpose of the Study:

• To review and assess existing methods for handling missing data within MSMs.
• To evaluate the performance of five common missing-data methods through simulations.
• To provide practical recommendations for choosing appropriate methods based on data characteristics and scientific context.

Main Methods:

• Review of literature on missing-data techniques for MSMs.
• Realistic simulations assessing bias of complete-case analysis, last observation carried forward, missingness pattern approach, multiple imputation, and inverse-probability-of-missingness weighting.
• Application to a sleep apnea cohort using electronic health record data.

Main Results:

• Simulation results quantify the bias introduced by different missing-data methods under various missingness mechanisms.
• The study highlights the strengths and limitations of each evaluated method in practice.
• The application demonstrates the practical implications of method choice on real-world data.

Conclusions:

• No single method is universally superior; the choice depends on the specific study.
• Careful consideration of missingness reasons, its impact on observed data relationships, and the data source is crucial.
• Informed selection of missing-data handling strategies is essential for reliable causal inference using MSMs.