Fast Bayesian Record Linkage With Record-Specific Disagreement Parameters

Abstract

Researchers are often interested in linking individuals between two datasets that lack a common unique identifier. Matching procedures often struggle to match records with common names, birthplaces, or other field values. Computational feasibility is also a challenge, particularly when linking large datasets. We develop a Bayesian method for automated probabilistic record linkage and show it recovers more than 50% more true matches, holding accuracy constant, than comparable methods in a matching of military recruitment data to the 1900 U.S. Census for which expert-labeled matches are available. Our approach, which builds on a recent state-of-the-art Bayesian method, refines the modeling of comparison data, allowing disagreement probability parameters conditional on nonmatch status to be record-specific in the smaller of the two datasets. This flexibility significantly improves matching when many records share common field values. We show that our method is computationally feasible in practice, despite the added complexity, with an R/C++ implementation that achieves a significant improvement in speed over comparable recent methods. We also suggest a lightweight method for treatment of very common names and show how to estimate true positive rate and positive predictive value when true match status is unavailable.

Keywords:

• Bayesian method
• Census data
• Data matching
• Fellegi-Sunter
• Gibbs sampler
• Scalability

Supplementary Materials

FBRSRL.R: R script containing functions, including main function FBRSRL(), for constructing comparison data, sampling the posterior and producing a matching. Includes notes on use. (R file)

gibbs_c.cpp: C++ code for Step 3 of the Gibbs sampler. To be compiled by Rcpp package before running FBRSRL().

gibbs_c_asym.cpp: C$++$ code for Step 3 of the Gibbs sampler when $n_B/n_A>200$. To be compiled by Rcpp package before running FBRSRL().

Acknowledgments

The author thanks Jiaying Gu and Shari Eli, as well as Martin Burda, En Hua Hu, Jamie Uguccioni, participants of the 2020 NBER-NSF SBIES Conference, and others for their help and suggestions. The comments of two anonymous referees were extremely helpful. The support of Early Indicators of Later Work Levels, Disease and Death, Dora L. Costa, principal investigator is gratefully acknowledged. This article is the responsibility of the author and does not necessarily represent the views of the NIH/NIA.

Additional information

Funding

This research was supported by the Social Sciences and Humanities Research Council of Canada, and by NIH grant number P01 AG10120.