Data Smearing: An Approach to Disclosure Limitation for Tabular Data

Abstract

Statistical agencies often collect sensitive data for release to the public at aggregated levels in the form of tables. To protect confidential data, some cells are suppressed in the publicly released data. One problem with this method is that many cells of interest must be suppressed in order to protect a much smaller number of sensitive cells. Another problem is that the covariates used to aggregate and level of aggregation must be fixed before the data is released. Both of these restrictions can severely limit the utility of the data. We propose a new disclosure limitation method that replaces the full set of microdata with synthetic data for use in producing released data in tabular form. This synthetic data set is obtained by replacing each unit’s values with a weighted average of sampled values from the surrounding area. The synthetic data is produced in a way to give asymptotically unbiased estimates for aggregate cells as the number of units in the cell increases. The method is applied to the U.S. Bureau of Labor Statistics Quarterly Census of Employment and Wages data, which is released to the public quarterly in tabular form and aggregated across varying scales of time, area, and economic sector.

Generating Partially Synthetic Geocoded Public Use Data with Decreased Disclosure Risk by Using Differential Smoothing

“…Thus, in some instances, it may be attractive to take a partially synthetic approach in which only a collection of values or variables are replaced with imputed values (e.g. Little (1993), Kennickell (1997), Reiter (2003Reiter ( , 2004, Abowd and Woodcock (2004), Little et al (2004), An and Little (2007) and Toth (2014)). Here, we consider the case where only a (log-)normally distributed Y is to be protected; extensions of our approach to generalized linear models and multivariate analyses are described in the Web appendix B.…”

“…Little (), Kennickell (), Reiter (), Abowd and Woodcock (), Little et al . (), An and Little () and Toth ()). Here, we consider the case where only a (log‐)normally distributed Y is to be protected; extensions of our approach to generalized linear models and multivariate analyses are described in the Web appendix B.…”

“…In some cases, agencies may generate fully synthetic data (Rubin, 1993;Reiter, 2002Reiter, , 2005Raghunathan et al, 2003;Quick et al, 2014), in which the released datasets are comprised entirely of simulated records. We, however, take a partially synthetic approach in which only a collection of values/variables are replaced with imputed values (Little, 1993;Kennickell, 1997;Abowd and Woodcock, 2004;Reiter, 2003Reiter, , 2004An and Little, 2007;Toth, 2014). Specifically, we assume the data consist of exact geographic locations and covariate information for each individual, as well as a continuously varying response which will be multiply imputed.…”