Individual Disclosure Risk Measures Based on Log-Linear Models

Dissemination of microdata files should be constrained to the confidentiality pledge under which a statistical agency collects survey data. To protect the confidentiality of respondents, statistical agencies perform a two-stage statistical disclosure control procedure. In the first stage, with respect to a disclosure scenario, the risk of disclosure of each unit is estimated. After the removal of direct identifiers, e.g. name and address, other indirect identifiers, called key variables, could still allow the disclosure of some confidential information about a unit. Usually, most of the key variables registered in social microdata files are categorical. An important problem in statistical disclosure control (SDC) is the estimation of the (number of) sample uniques that are also population uniques, i.e. units at risk of disclosure. In this paper, extensions of the Poisson-loglinear model to estimate a disclosure risk measure in contingency tables are presented. The main contribution is the development of smoothing strategies based on a penalized likelihood approach and on graphical log-linear models decomposition. Results of several tests performed on Italian 2001 census data will be presented.