Fig. 3

Simplified illustration of Bayesian matching. This process omits some of the methods of the full algorithm, and shows only a three-state comparison for names (name match, metaphone match, or no match). Comparison settings are as in Table 2. A Identified version. Calculations are shown left to right across identifiers: the log odds are updated with every successive identifier comparison. B De-identified version, showing identifier probabilities explicitly (e.g. the population probability of SMITH as a surname and of the corresponding metaphone SM0 arising from other surnames). Hashed identifiers are shown as four hexadecimal characters but in practice would be much longer. Some hashed representations are more complex than shown (e.g. “fuzzy” DOB representations; see text). C Winners are declared or not based on two decision thresholds, θ and δ (see text)