A Bayesian Procedure for File Linking to Analyze End-of-Life Medical Costs

Area of Science:

• Health Economics
• Biostatistics
• Data Linkage

Background:

• End-of-life medical expenses represent a substantial portion of overall healthcare spending.
• Accurate analysis requires linking patient cost data with mortality information.
• Existing datasets (Medicare claims, death certificates) lack unique identifiers for direct linkage.

Purpose of the Study:

• To develop and validate a statistical method for linking Medicare claims and death certificate data.
• To accurately estimate end-of-life medical expenditures by overcoming data linkage challenges.
• To provide a robust methodology applicable to other large-scale data linkage problems.

Main Methods:

• Utilized a novel imputation technique involving sampling from the joint posterior distribution of model parameters and file permutations.
• Developed linking models incorporating regression of location of death and cost measures on cause of death (CoD) and demographics.
• Employed exact distribution for small cells and the Metropolis algorithm for large cells, with sparse matrix structures for efficiency on ≈1.7 million cases.

Main Results:

• Successfully linked only 33% of deaths uniquely using common variables.
• The imputation procedure generated 'm' datasets with imputed matches between Medicare claims and death certificates.
• The methodology demonstrated efficient computation for a large dataset, facilitating analysis of imputed data using multiple imputation rules.

Conclusions:

• The proposed imputation method effectively links disparate health datasets (Medicare claims, death certificates) when unique identifiers are absent.
• This approach enables more accurate estimation of end-of-life medical costs and can be generalized to other file-linking applications.
• The study provides a valuable tool for health economic research and policy-making.