Missing data in the exposure of interest and marginal structural models: a simulation study based on the Framingham...

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Missing data are prevalent in longitudinal studies, complicating causal effect estimation.
• Marginal structural models (MSMs) are crucial for addressing time-dependent confounding in observational studies.
• Limited research exists on the impact of missing exposure data within MSMs.

Purpose of the Study:

• To investigate the impact of missing exposure data on causal effect estimates using MSMs.
• To evaluate the performance of different missing data handling methods in realistic simulation settings.
• To assess the influence of unmeasured variables on missing data and subsequent bias.

Main Methods:

• Simulations based on the Framingham Heart Study data.
• Application of marginal structural models (MSMs) to estimate causal odds ratios.
• Comparison of four missing data methods across seven missing data structures.
• Utilized causal diagrams to interpret results.

Main Results:

• Complete case analysis demonstrated the least bias across all tested missing data structures.
• Censoring individuals at the first missing exposure, with a censorship and propensity model, performed favorably.
• Unmeasured variable prediction of missing data minimally increased bias, except under specific high-impact scenarios.

Conclusions:

• Complete case analysis is a robust method for handling missing exposure data in MSMs.
• Careful consideration of missing data mechanisms, including unmeasured confounding, is essential for valid causal inference.
• Causal diagrams are valuable tools for planning and interpreting analyses involving complex missing data patterns.