We are hosting a one day symposium on New Developments in Statistical Disclosure Control on 7 May 2004. The meeting will consist of 4 expert lectures delivered by Steve Fienberg, Jon Forster, Chris Skinner and Yosef Rinott, followed by discussion and questions.

Registration fee for the symposium is £25, which includes lunch and refreshments.

For further details contact Danny Pfeffermann (D.Pfeffermann@soton.ac.uk), Phone: +44 (23) 8059 6689 Fax: +44 (23) 8059 3846

To register, contact Jane Schofield; email: (jms1@socsci.soton.ac.uk), Phone: +44 (23) 8059 5376 Fax: +44 (23) 8059 3846

Registration Form

Expert Lecture Abstracts

• Steve Fienberg - Statistical Disclosure Limitation: Releasing Useful Data for Statistical Analysis
  
  Disclosure limitation has often been viewed by statistical agencies solely as a mechanism for "protecting" confidentiality, and not in terms of providing data that are useful for statistical analysis. A true statistical approach to disclosure limitation needs to assess the tradeoff between preserving confidentiality and the usefulness of the released data, especially for inferential purposes. In this presentation we discuss these issues, illustrate them with some recent disclosure limitation approaches for categorical data linked to log- linear model analysis, and we describe some of the statistical research challenges that remain. We also articulate some general principles that can help serve as a foundation for disclosure limitation methodology in the future.
   
• Jon Forster
  
  We propose a method for assessing the risk of individual identification in the release of categorical data. The approach is Bayesian and provides posterior summaries of identification risk. For per-record risks, these summaries are typically predictive probabilities, and for global measures, full posterior distributions. By utilising a Bayesian approach to estimation under model uncertainty, known as model-averaging, we can provide more realistic estimates of disclosure risk for individual records than are provided by methods which ignore the multivariate structure of the data set. For categorical data, we consider log-linear models. Computation is more straightforward if we restrict consideration to those log-linear models which are also decomposable graphical models. However, we also present an example where the full flexibility of the log-linear model class is exploited.
   
• Yosef Rinott - On Information and Disclosure Risk Measures and Estimation
  
  Click here for the abstract of Yosef's lecture in PDF format.
   
• Chris Skinner - Record-level Measures of Disclosure Risk for Survey Microdata
  
  It is of interest to estimate measures of disclosure risk at the record level for survey microdata. Skinner and Holmes (JOS, 1998) proposed one approach for a measure defined in terms of the probability of population uniqueness. Skinner and Elliot (JRSS’B’, 2002) proposed a file-level measure of disclosure risk in terms of the probability that an observed match is correct and discussed its advantages compared to a measure defined in terms of population uniqueness. In this paper we extend the measure proposed by Skinner and Elliot to a record-level measure. We discuss the estimation of this measure for certain compound log-linear Poisson models. The properties of estimators for different models are evaluated using data from the General Household Survey. There is little evidence of gains from including random effects in the model. The measure obtained by omitting random effects is easier to estimate than that of Skinner and Holmes. It is found to discriminate effectively between records with different levels of risk and the average value of the estimated measure within different subpopulations is found to track well the average ‘true value’ of the measure.

Symposium Programme (Microsoft Word format)