NISS Bayesian Data Editing for Continuous Microdata With Application to the 2007 Census of Manufactures

We present a full Bayesian, joint modeling approach to simultaneous editing and imputation for continuous microdata under the linear constraints. To provide high-quality data, statistical agencies spend large amounts of resources to detect erroneous values in the collected datasets and correct them. Several automatic data editing systems are based on the Fellegi-Holt method which uses a two-stage process of first finding the erroneous data items by logical conditions, called edits, and then imputing new values for those detected items, usually using relatively simple imputation methods such as hot deck imputation or ratio imputation. Our approach replaces the two step process with a single probability based, data-driven approach in which we (i) specify a flexible joint probability model for the continuous variables, which can capture complex associations, (ii) stochastically identify erroneous items, which, unlike Fellegi-Holt routines, can reflect the information of individual records, and (iii) impute new values from the model in ways guaranteed to satisfy all ratio edits as well as balance edits without deriving implied edits. In this paper, we describe this integrated editing approach with simulation study and application to the 2007 U.S. Census of Manufactures data. We compare the approach against the Fellegi-Holt based approaches, showing how the joint model-based approach can offer improved accuracy.