Random forests for generating partially synthetic, categorical data
Several national statistical agencies are now releasing partially synthetic, public use microdata. These comprise the units in the original database with sensitive or identifying values replaced with values simulated from statistical models. Specifying synthesis models can be daunting in databases that include many variables of diverse types. These variables may be related in ways that can be difficult to capture with standard parametric tools. In this article, we describe how random forests can be adapted to generate partially synthetic data for categorical variables. Using an empirical study, we illustrate that the random forest synthesizer can preserve relationships reasonably well while providing low disclosure risks. The random forest synthesizer has some appealing features for statistical agencies: it can be applied with minimal tuning, easily incorporates numerical, categorical, and mixed variables as predictors, operates efficiently in high dimensions, and automatically fits non-linear relationships.
