Avoiding pitfalls when combining multiple imputation and propensity scores.

Area of Science:

• Epidemiology
• Biostatistics
• Health Data Science

Background:

• Observational data analysis faces challenges from confounding and missing data.
• Propensity scores address confounding, while multiple imputation handles missing data.
• Optimal methods for combining these techniques in statistical analysis remain unclear.

Purpose of the Study:

• To compare the accuracy and precision of two methods for combining propensity scores and multiple imputation.
• To evaluate the 'Across' versus 'Within' approaches in estimating exposure effects.
• To provide guidance on handling incomplete observational data in statistical modeling.

Main Methods:

• Two approaches, 'Across' and 'Within', were developed for integrating propensity scores and multiple imputation.
• The 'Across' method averages propensity scores before analysis; the 'Within' method analyzes each imputation separately.
• Methods were assessed via Monte Carlo simulations and application to the British Society for Rheumatology Biologics Register data.

Main Results:

• The 'Within' approach yielded unbiased exposure effect estimates with appropriate confidence intervals.
• The 'Across' approach resulted in biased estimates and overly narrow, unrealistic confidence intervals.
• Simulation and real-world data analyses consistently favored the 'Within' method.

Conclusions:

• The 'Within' approach is recommended for propensity score analyses involving incomplete observational data.
• Researchers should prefer the 'Within' method to ensure accurate and reliable effect estimates.
• This study clarifies best practices for complex statistical analyses in observational research.