Propensity score analysis with missing data.

Area of Science:

• Biostatistics
• Epidemiology
• Observational Studies

Background:

• Observational studies are prone to bias due to confounding variables.
• Propensity score analysis aims to reduce this bias by balancing covariates between treatment and control groups.
• Missing data in covariates or outcomes complicates propensity score analysis.

Purpose of the Study:

• To review statistical theories and methods for propensity score analysis with incompletely observed covariates.
• To compare traditional and modern machine learning approaches for estimating propensity scores with missing data.
• To illustrate and evaluate these methods using an empirical example.

Main Methods:

• Review of propensity score analysis theory.
• Exploration of estimation methods for propensity scores with missing covariates, including logistic regression, random forests, and generalized boosted modeling.
• Description of balance diagnostics and equating methods for incomplete data.

Main Results:

• Various statistical methods exist for handling missing covariates in propensity score analysis.
• Machine learning methods offer advanced techniques for estimating propensity scores with incomplete data.
• Empirical examples demonstrate the application and comparison of these methods.

Conclusions:

• Effective methods exist for propensity score analysis even with missing covariate data.
• Careful application of these methods can improve the validity of causal inference from observational studies.
• The choice of method depends on the specific dataset and research question.