Paper 1224-2017: A Moment-Matching Approach for Generating Synthetic Data in SAS®

Disseminating data to potential collaborators can be essential in the development of models, algorithms, and innovative research opportunities. However, it is often time consuming to get approval to access sensitive data such as health data. An alternative to sharing the real data is to use synthetic data, which has similar properties to the original data but does not disclose sensitive information. The collaborators can use the synthetic data to make preliminary models or work out bugs in their code while waiting to get approval to access the original data. A data owner can also use the synthetic data to crowdsource solutions from the public through competitions like Kaggle and then test those solutions on the original data. This paper implements a method that generates fully synthetic data in a way that matches the statistical moments of the true data up to a specified moment order as a SAS macro. Variables in the synthetic data set are of the same data type as the true data (for example, integer, binary, continuous). The implementation uses the linear programming solver within a column generation algorithm and the mixed integer linear programming solver from the OPTMODEL procedure in SAS/OR. The COFOR statement in PROC OPTMODEL automatically parallelizes a portion of the algorithm. This paper demonstrates the method by using the Sashelp.Heart data set to generate fully synthetic data copies.