Simultaneous record linkage and causal inference with propensity score subclassification.

Area of Science:

• Observational studies
• Causal inference
• Data linkage

Background:

• Estimating treatment effects from observational data often relies on merging disparate datasets.
• Probabilistic record linkage is frequently used but introduces potential errors.
• These linkage errors can bias causal effect estimates, particularly in propensity score subclassification methods.

Purpose of the Study:

• To develop and evaluate methodology for causal inference in observational studies using propensity score subclassification with probabilistically linked data.
• To address scenarios where covariates and treatment assignments are in one file, and outcomes are in another.
• To improve the accuracy of additive treatment effect estimation despite linkage errors.

Main Methods:

• Developed methodology for causal inference with propensity score subclassification on data constructed via probabilistic record linkage.
• Assumed linkage errors are independent of other variables.
• Proposed and evaluated algorithms for selecting record pairs for causal effect estimation.

Main Results:

• Demonstrated conceptually how linkage errors can impact causal estimates in subclassification.
• Simulation studies showed that case selection procedures improve the accuracy of treatment effect estimates.
• Selected cases yielded more accurate estimates compared to using only confirmed true links.

Conclusions:

• The proposed methodology offers a robust approach to causal inference with linked observational data.
• Case selection algorithms can mitigate the negative impact of linkage errors on treatment effect estimates.
• This work enhances the reliability of causal inference when merging datasets with probabilistic record linkage.