Utility-Based Analysis of Statistical Approaches and Deep Learning Models for Synthetic Data Generation With Focus on Correlation Structures

This study compares statistical and deep-learning synthetic-data generators from a utility perspective, focusing on how faithfully each preserves the correlation structure among variables — a property many downstream analyses depend on but that generators often distort. Using utility-oriented evaluation, it examines where deep models and classical statistical methods succeed or fail at reproducing multivariate dependencies. The findings help practitioners choose methods according to the analyses the synthetic data must support, rather than by fidelity scores alone. Published in JMIR AI, the work sharpens the criteria for evaluating synthetic data in health applications.