Discriminating powers of partial agreements of names for linking personal records. Part I: The logical basis

Machines have difficulty when using people's names to link medical and other records pertaining to the same individuals because of nicknames, ethnic synonyms, truncations, misspellings and typographical errors. Present algorithms used to compute the discriminating powers (or ODDS) associated with partial agreements of names are based, inappropriately, on the degrees of outward similarity alone. They are particularly ineffective in dealing with names that look alike but are unrelated, and with related names that have little apparent similarity. A fundamentally different rationale is, therefore, proposed which, like the human mind, assesses the relatedness of two alternative forms of a name in terms of how often they are used, interchangeably in practice. This must be taken into account if the associated discriminating powers (ODDS) are to be correctly computed. A way of implementing this more precise approach is described and illustrated, using the given names on linked records from an earlier epidemiological study. This first study of two describes the logical basis for record linkage, a second one the empirical test.