An empirical comparison of record linkage procedures

We consider the problem of record linkage in the situation where we have only non-unique identifiers, like names, sex, race etc., as common identifiers in databases to be linked. For such situations much work on probabilistic methods of record linkage can be found in the statistical literature. However, although many groups undoubtedly still use deterministic procedures, not much literature is available on deterministic strategies. Furthermore, there appears to exist almost no documentation on the comparison of results for the two strategies. In this work we compare a stepwise deterministic linkage strategy with a probabilistic strategy, as implemented in AUTOMATCH, for a situation in which the truth is known. The comparison was carried out on a linkage between medical records from the Regional Perinatal Intensive Care Centers database and educational records from the Florida Department of Education. Social security numbers, available in both databases, were used to decide the true status of each record pair after matching. Match rates and error rates for the two strategies are compared and a discussion of their similarities and differences, strengths and weaknesses is presented.