Missing Data in Marginal Structural Models: A Plasmode Simulation Study Comparing Multiple Imputation and Inverse...

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Marginal structural models (MSMs) are increasingly used in epidemiology to address time-varying confounding.
• Contradictory recommendations exist for handling missing data within MSMs.
• This study addresses the need for clear guidance on missing data methods in MSMs.

Purpose of the Study:

• To compare the validity and precision of MSMs estimates using complete case analysis (CC), multiple imputation (MI), and inverse probability weighting (IPW).
• To evaluate these methods under various missing data scenarios, including different missing mechanisms, percentages, and confounder types.

Main Methods:

• A plasmode simulation study was conducted using data from the Osteoarthritis Initiative.
• 81 scenarios were simulated, varying missing mechanisms (MCAR, MAR, MNAR), missing percentages (10-30%), confounder types (time-independent/varying), and analytical approaches (CC, IPW, MI).
• Performance was assessed using relative bias, mean squared error, and empirical power.

Main Results:

• Multiple imputation (MI) generally yielded less biased estimates (1.2%-6.7%) with better precision (0.17-0.18) than inverse probability weighting (IPW) (relative bias: -5.3% to 8.0%; precision: 0.19-0.53).
• MI demonstrated consistent empirical power across all simulated scenarios.
• Complete case analysis (CC) was not explicitly detailed in results but implied to be less performant.

Conclusions:

• Multiple imputation (MI) demonstrates a clear advantage over inverse probability weighting (IPW) for handling missing data in marginal structural models (MSMs).
• These findings provide practical guidance for researchers using MSMs in the presence of missing confounder data.