Providing bespoke synthetic data for the UK Longitudinal Studies and other sensitive data with the synthpop package for R1

Abstract

Abstract. Synthetic data methods were designed to address the conflicting demands placed on data holders to unlock the research and policy potential of microdata while at the same time preserving the confidentiality of individuals. Recently, these methods have become more widely recognized in the UK and the provision of bespoke synthetic data has been approved to expand the use of one of the UK Longitudinal Studies. The process of producing useful synthetic data involves, however, a substantial investment of research time, as it always requires some customising for the characteristics of an individual data set. At the same time, a substantial part of it can be automated and this is essential when the process has to be conducted rapidly and on a regular basis. This paper describes the application of synthetic data to the UK Longitudinal Studies, details implementation process for the Scottish Longitudinal Study and presents methods used in an R package synthpop that has been developed to facilitate production of non-disclosive entirely synthetic data. A reproducible example using open data is given to illustrate the synthesising procedure and to provide insights into quality of synthetic data generated using different automated approaches.

General and Specific Utility Measures for Synthetic Data

“…When the generative model that informs the synthesis adheres too closely to the proposed utility model, the validity checks such as the existence of other interactions will not be apparent in the synthesized data; see Nowok et al . () and Raab et al . () for examples.…”

“…This scenario has been implemented by some researchers and agency staff; see Nowok et al . () or Reiter et al . () for further examples.…”

“…We consider expanding the propensity score models to include CART models (Breiman et al ., ). These models have proved useful for generating synthetic data (Reiter, ) and have been shown to outperform other machine learning techniques (Drechsler and Reiter, ; Nowok, ) and parametric models (Nowok et al ., ) for this purpose. Additionally, boosted tree models have been found to be useful for estimating propensity scores in causal inference applications; see McCaffrey et al .…”

“…When the synthesizer does have some knowledge of the models that the analyst has in mind and bases the synthesis on these models, this may falsely reassure the analyst that their model is the correct model. When the generative model that informs the synthesis adheres too closely to the proposed utility model, the validity checks such as the existence of other interactions will not be apparent in the synthesized data; see Nowok et al (2017) and Raab et al (2017a) for examples. Thus general measures of utility could be more helpful in allowing an assessment of how well the final inference might agree with what would have been obtained if the user had access to the unchanged data for all of the analyses, rather than just at the final stage of a gold standard analysis.…”

Comments on four papers on synthetic data in Volume 32 Issue 1 the Statistical Journal of the IAOS

“…'s overview of international developments [14]. More details of our methods and our synthpop package for R are available [15][16][17].…”