Hierarchical Bayesian Record Linkage Theory

In record linkage, or exact file matching, one compares two or more files on a single population for purposes of unduplication or production of an enhanced, merged database. Record linkage has many applications, including in population enumeration efforts, to create databases for epidemiological investigations, and to improve survey sample frames. Latent class and mixture models have been used to implement computerised record linkage of large databases. Probabilities that pairs of records, one record from each of two files, pertain to the same person (a match) or to different people (a nonmatch) are estimated based on model parameters and Bayes’ theorem. In some settings, there is experience with similar record linkage operations that can inform prior opinions concerning model parameters. In this paper, Bayesian record linkage alternatives are developed and compared through simulation. A hierarchical Bayesian model allows parameters to vary by file blocks, which are similar to geographical blocks in census applications. Techniques for incorporating one-to-one matching between files into the likelihood itself and computing posterior distributions of parameters and linkage indicators are presented.