A Bayesian Record Linkage Methodology for Multiple Imputation of Missing Links

Probabilistic record linkage can be an effective research technique even if available records lack strong personal identifiers or if identifying fields contain many errors or omissions. Traditional methodologies typically select a single set of linked record pairs for research based on a match weight test statistic and clerical review of marginal pairs. However, missing links (false negatives) can make such datasets unrepresentative of the total population of true linked pairs. The methodology described here addresses this problem. A full Bayesian model is developed for the posterior probability that a record pair is a true match given observed agreements and disagreements of comparison fields. Observed-data posterior distributions for model parameters and true match status are estimated simultaneously through MCMC data augmentation with parallel chains. This process gives multiple complete representative sets of imputed linked record pairs. Population estimates can be obtained from each imputation and consolidated using established techniques. Application of linkage imputation by a consortium of traffic safety researchers is described