Uncertainty in Propensity Score Estimation: Bayesian Methods for Variable Selection and Model Averaged Causal...

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Causal inference from observational data often uses propensity scores (PS) to control for confounding.
• High-dimensional covariate sets in big data present challenges for selecting variables in PS models.
• Current methods for PS model selection often ignore the inherent uncertainty.

Purpose of the Study:

• To propose novel Bayesian methods for propensity score variable selection and model averaging.
• To address uncertainty in variable selection for propensity score models.
• To improve the estimation of causal treatment effects using observational data.

Main Methods:

• Developed three Bayesian approaches for propensity score (PS) variable selection.
• Implemented model averaging to estimate causal effects, weighting by data-driven model support.
• Utilized simulation studies and applied methods to a real-world dataset of brain tumor treatments.

Main Results:

• The proposed Bayesian methods effectively select relevant variables for PS models.
• Model averaging provides a robust approach to estimating causal effects by incorporating uncertainty.
• The methods were successfully applied to compare surgical versus nonsurgical brain tumor treatments.

Conclusions:

• Bayesian variable selection and model averaging offer a principled way to handle uncertainty in propensity score analysis.
• These methods enhance the reliability of causal effect estimation from complex observational data.
• The approach is applicable to real-world health research, such as comparing treatment effectiveness.