Regression-Assisted Bayesian Record Linkage for Causal Inference in Observational Studies with Covariates Spread Over...

Area of Science:

• Statistics
• Econometrics
• Epidemiology

Background:

• Observational studies often involve data from multiple sources.
• Linking these datasets can reduce bias by including more covariates.
• Probabilistic record linkage is common but doesn't account for linkage uncertainty.

Purpose of the Study:

• To develop a method for causal inference that accounts for uncertainty in probabilistic record linkage.
• To improve the accuracy of causal effect estimation when using linked observational data from multiple files.
• To integrate Bayesian record linkage with causal inference techniques.

Main Methods:

• Fusing regression-assisted, Bayesian probabilistic record linkage with causal inference.
• Utilizing a Markov chain Monte Carlo sampler to generate multiple plausible linked datasets.
• Applying causal inference estimators, specifically those based on propensity score overlap weights.

Main Results:

• The proposed method propagates uncertainty from imperfect linkages into causal inferences.
• It leverages variable relationships to enhance record linkage quality.
• Simulations and real-world data analysis demonstrate improved accuracy in estimated treatment effects.

Conclusions:

• The integrated approach offers a more robust framework for causal inference with linked observational data.
• Accounting for linkage uncertainty is crucial for reliable causal effect estimation.
• This method enhances the validity of findings from multi-source observational studies.