|
Abstract:
|
Data producers should collect high-quality data and protect confidentiality when releasing data. Protection of confidentiality necessitates altering values of identity information and sensitive attributes. Therefore, it is important to find the data alteration method with high confidentiality that provides satisfactory data quality. This paper focuses on developing methods of measuring data quality when the distribution of data is not assumed to be normal. We treat data utility as a problem of evaluating similarities of original data structure to masked data structure. The data utilities we present are based on the cumulative distribute function, clustering, and propensity score approaches. When the distribution departs from normal, simulations for a variety of data structures show how these measures can be used for evaluating disclosure limitation procedures.
|
