An empirical evaluation of easily implemented, nonparametric methods for generating synthetic datasets

When intense redaction is needed to protect data subjects’ confidentiality, statistical agencies can release synthetic data, in which identifying or sensitive values are replaced with draws from statistical models estimated from the confidential data. Specifying accurate synthesis models can be a difficult and labor intensive task with standard parametric approaches. We describe and empirically evaluate four easy-to-implement, nonparametric synthesizers based on machine learning algorithms—classification and regression trees, bagging, random forests, and support vector machines—on their potential to preserve analytical validity and reduce disclosure risks. The results suggest that synthesizers based on regression trees can provide high utility with low disclosure risks.