Comparison of methods for handling covariate missingness in propensity score estimation with a binary exposure

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Observational data analysis for causal effects is prone to confounding bias.
• Propensity scores are commonly used to control for confounding.
• Handling missing covariate data in propensity score estimation remains a challenge.

Purpose of the Study:

• To evaluate the performance of various methods for handling missing covariate data in propensity score estimation.
• To compare established methods like multiple imputation (MI, MIMP) and treatment mean imputation with less-evaluated approaches including single imputation (SI) variants and Generalized Boosted Modeling (GBM).

Main Methods:

• A simulation study was conducted to assess different missing data handling techniques.
• Methods evaluated included multiple imputation (MI, MIMP), treatment mean imputation, single imputation with prediction error (SI+PE), single imputation with prediction error and parameter uncertainty (SI+PE+PU), and Generalized Boosted Modeling (GBM).

Main Results:

• SI+PE, SI+PE+PU, MI, and MIMP demonstrated comparable performance and superiority over treatment mean imputation and GBM regarding bias.
• MI and MIMP additionally accounted for the uncertainty introduced by imputing missing values.
• Direct application of GBM to incomplete data using surrogate splits resulted in significant bias.

Conclusions:

• Imputation of missing covariates prior to applying Generalized Boosted Modeling (GBM) is recommended.
• Multiple imputation methods (MI, MIMP) and single imputation variants (SI+PE, SI+PE+PU) are effective strategies for handling missing covariate data in propensity score estimation.