Controlling false match rates in record linkage using extreme value theory

Choosing the weight threshold that separates matches from non-matches is one of the hardest parts of probabilistic record linkage, especially without labelled training data. This paper applies extreme value theory to the distribution of comparison weights, modelling the tail behaviour to control the false-match rate in a statistically principled way. The result is a threshold-selection method that does not depend on a gold-standard training set, which is rarely available in practice. Published in the Journal of Biomedical Informatics, the approach gives registries and biomedical data managers a defensible basis for calibrating linkage decisions.