Bayesian Methodology for Disclosure Limitation and Statistical Analysis of Large Government Surveys

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

Surveys with complex probability sampling designs involving clustering and stratification are a major source of empirical data for governmental and scientific use. The standard approach to survey inference is design-based, with statistical inferences being based on the sampling distribution with population values treated as fixed. This approach has been a powerful force for the development of objective statistical analysis of large surveys, with reliable operating characteristics under weak assumptions about the population. However, the design-based paradigm is too limited to handle: (i) the increased availability of data from a variety of sources, such as surveys, censuses and administrative records; (ii) increasing demands for analyses that go beyond simple descriptive information such as means and totals for large domains; (iii) the development and analysis of data masked to preserve confidentiality; and (iv) the analysis of data subject to unit and item nonresponse. These questions can be addressed by a model-based Bayesian approach, with models that capture the relevant features of the population under study and take into account important features of the sample design, and non-informative priors that limit subjectivity in the analysis. Bayesian methods are enjoying a resurgence in statistics, with the development of computational tools that make them practically feasible. However, the application of Bayesian methods to sample surveys remains very limited. The goal of this research is to develop useful, practical Bayesian methods for sample survey inference that have good design-based properties. The dissemination of public use data files is crucial to the research community in order to conduct research that forms the basis for rational policy decisions. This research will develop methods for disseminating detailed micro-data files that greatly reduce the risk of disclosure of the identity of respondents to a data intruder. Methods will be based on multiple imputation of key variables, an approach that allows for valid statistical inferences using existing software and limits the degree of information loss. The methods will be tested on large government surveys collected by the National Center for Health Statistics and other federal agencies. This research also will develop Bayesian methods for three topics in the statistical analysis of complex surveys: (i) the analysis of surveys where the sampled units have differential probabilities of inclusion; (ii) the handling of unit and item nonresponse in surveys; and (iii) the analysis of samples collected using rotating panel designs. This research is supported by the Methodology, Measurement, and Statistics Program and a consortium of federal statistical agencies under the Research on Survey and Statistical Methodology Funding Opportunity.