Propensity Score Analysis with Partially Observed Baseline Covariates: A Practical Comparison of Methods for Handling...

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Propensity score methods are widely used in non-interventional clinical studies.
• Missing data in baseline covariates is a common challenge in observational datasets.
• Effective handling of missing data is essential for reliable propensity score analysis.

Purpose of the Study:

• To compare the performance of statistical methods for handling missing data in propensity score analysis.
• To identify the most effective methods for achieving covariate balance in the presence of missing data.

Main Methods:

• Evaluated methods that account for missing data during estimation and methods based on imputation (e.g., multiple imputation).
• Applied methods to a prospective registry dataset for unprotected left main coronary artery disease treatment.
• Assessed method performance based on the overall balance of baseline covariates.

Main Results:

• Methods explicitly addressing missing data demonstrated superior performance compared to complete case analysis.
• The best covariate balance was achieved using a propensity score estimation method that incorporates missing data via stochastic approximation of the expectation-maximization algorithm.

Conclusions:

• When a missing at random mechanism is plausible, methods that utilize missing data for propensity score estimation or imputation are recommended.
• Sensitivity analyses are advised to assess the impact of chosen missing data handling and propensity score estimation methods.