Evaluating Synthetic Data for Health Economic Modelling: A Comparison of Three Methods for Emulating the UKPDS Outcomes Model 1 Equations

Synthetic data may reduce barriers to reproducible research when access to sensitive patient-level datasets is constrained, but evidence remains limited on whether synthetic datasets can replicate the complex equation-estimation and prediction tasks used in health economic microsimulation models. This Stage 1 Registered Report prespecifies a task-specific comparison of three synthetic data-generation approaches for emulating the United Kingdom Prospective Diabetes Study (UKPDS) Outcomes Model 1 equation set. Collaborating teams across three institutions will independently develop code to generate synthetic data using a shared training dataset that preserves the structural properties of the original data but contains no original data. A separate team, with exclusive access to the original UKPDS data, will apply these methods to generate synthetic datasets and evaluate their fidelity by re-estimating UKPDS event, mortality, and risk-factor equations and comparing the estimates with those from the original data. Evaluation metrics include process-level mean squared error, Z-score-like standardised differences, confidence-interval coverage, prediction-level mean absolute differences, and model-level mean absolute differences in estimated quality-adjusted life years. We have completed only the data preparation, training data generation, and code transfer; we have not conducted analyses of the restricted UKPDS data. All model specifications, evaluation criteria, and analytic procedures are fixed in advance.