Guided Bayesian imputation to adjust for confounding when combining heterogeneous data sources in comparative...

Area of Science:

• Biostatistics
• Epidemiology
• Health Services Research

Background:

• Comparative effectiveness research often uses large observational data to estimate average causal effects.
• Missing confounders in main studies necessitate advanced methods, as standard imputation may fail with limited validation data.

Purpose of the Study:

• To propose a Bayesian approach for estimating average causal effects (ACE) in main studies with missing confounders.
• To improve confounding adjustment by borrowing information from a smaller validation study.
• To integrate Bayesian model averaging, confounder selection, and missing data imputation.

Main Methods:

• Developed a Bayesian framework combining model averaging, confounder selection, and imputation.
• Allowed for differential treatment effects between main and validation studies.
• Propagated uncertainty from imputation and confounder selection in ACE estimation.

Main Results:

• Simulations compared the proposed method against existing approaches.
• Applied to Medicare data on brain tumor surgical resection and survival (10,396 main, 2220 validation).
• Incorporating SEER-Medicare validation data reduced the estimated ACE by 30%.

Conclusions:

• The Bayesian approach effectively enhances confounding adjustment in observational studies with missing covariate data.
• Leveraging validation data significantly impacts causal effect estimates, highlighting the importance of comprehensive covariate information.
• The method provides a robust framework for handling missing data and confounder uncertainty in comparative effectiveness research.