On models for statistical disclosure risk estimation

Risk estimation in the context of statistical disclosure control is usually model based. Many models and related methods have been proposed in recent years, and some programs are distributed to implement them. However, the literature and the program manuals rarely discuss statistical questions such as the sensitivity or robustness of the estimates relative to the assumptions of the models, goodness of fit tests for the validity of the model, and variance estimates of the risk measures proposed. In order to deal with these issues in concrete terms we chose here to discuss two well-known models and methods for risk estimation. We study variations on a model of Bethlehem et. al. (1990), to which we shall refer as the B model, and on a model due to Benedetti and Franconi (1998), to which we shall refer as the ARGUS model or the A model for brevity, since it forms the basis for risk estimation in the ARGUS program. In particular the goodness of fit tests proposed here are based on an embedding of the A model into the B model which is demonstrated in this paper. We come to analyze and contribute to these models, not to bury nor to praise them.