Guidelines for Producing Useful Synthetic Data

Abstract: We report on our experiences of helping staff of the Scottish Longitudinal Study to create synthetic extracts that can be released to users. In particular, we focus on how the synthesis process can be tailored to produce synthetic extracts that will provide users with similar results to those that would be obtained from the original data. We make recommendations for synthesis methods and illustrate how the staff creating synthetic extracts can evaluate their utility at the time they are being produced. We discuss measures of utility for synthetic data and show that one tabular utility measure is exactly equivalent to a measure calculated from a propensity score. The methods are illustrated by using the R package $synthpop$ to create synthetic versions of data from the 1901 Census of Scotland.