Creating Synthetic Data for Complex Surveys Using the Research and Development Survey: A Comparison Study

Synthetic data has been gaining popularity in many fields as an approach to retain data utility (the validity of inference using synthetic data) and protect confidentiality. However, creating synthetic data for complex surveys remains a challenge.

This research compared three approaches to incorporate survey design information (stratification, clustering, and sampling weights) during the synthetic data-generating process using the Research and Development Survey (RANDS), a series of primarily web surveys conducted by the National Center for Health Statistics, Centers for Disease Control and Prevention. Both parametric (logistic and linear regression models) and nonparametric (classification and regression trees [CART]) methods were used to create synthetic data. Data utility and disclosure risk were evaluated via confidence interval overlap, propensity score measurement, and average matching probability for re-identification.

Using the original survey design information as predictors during the synthesis process improved data utility for the parametric method. However, the nonparametric method yielded results with better data utility but slightly higher disclosure risk.