Four checks for low-fidelity synthetic data: recommendations for disclosure control and quality evaluation

Confidential administrative data is usually only available to researchers within a Trusted Research Environment (TRE). Recently, some UK groups have proposed that low-fidelity synthetic data (LFSD) be made available to researchers outside the TRE, to allow code-testing and data discovery. There is a need for transparency so that those who access LFSD know how it has been created and what to expect from it. Relationships between variables are not maintained in LFSD, but a real or apparent data breach can occur from its release. To be useful to researchers for preliminary analyses LFSD needs to meet some minimum quality standards. Researchers who will use the LFSD need to have details of how it compares with the data they will access in the TRE clearly explained and documented. We propose that these checks should be run by data controllers before releasing LFSD to ensure it is well documented, useful and non-disclosive. <b>Labelling</b> To avoid an apparent data breach, steps must be taken to ensure that the synthetic data (SD) is clearly identified as not being real data.<b>Disclosure</b> The LFSD should undergo disclosure risk evaluation as described below and any risks identified should be mitigated.<b>Structure</b> The structure of the SD should be as similar as possible to the TRE data.<b>Documentation</b> Differences in the structure of the SD compared to data in the TRE must be documented, and the way(s) that analyses of the SD expect to differ from those of data in the TRE must be clarified. We propose details of each of these below; but a strict, rule-based approach should not be used. Instead, the data holders should modify the rules to take account of the type of information that may be disclosed and the circumstances of the data release (to whom and under what conditions).