[Comparing synthetic data generation methods in pharmacoepidemiology: reconciling reproducibility with privacy...

Area of Science:

• Pharmacoepidemiology
• Health Data Science
• Privacy-Preserving Technologies

Background:

• Real-world (RW) data is crucial for pharmacoepidemiology but faces privacy restrictions for secondary use.
• Synthetic data generation offers a solution to balance privacy protection with RW evidence generation for regulatory decision-making.

Purpose of the Study:

• Compare synthpop and Conditional Tabular-Generative Adversarial Networks (CT-GANs) for generating tabular synthetic data.
• Evaluate the ability of both methods to preserve statistical structure and reduce re-identification risk.

Main Methods:

• A comparative study using a large anonymized RW dataset of COVID-19 patients.
• Synthetic data generated using synthpop and CT-GAN methods.
• Evaluation using general utility, specific utility, and disclosure/re-identification measures.

Main Results:

• Synthpop demonstrated superior general and specific utility compared to CT-GAN.
• Synthpop exceeded acceptable thresholds for standardized mean differences (SMD) in over half the variables.
• Both methods reduced the already negligible identity disclosure risk of the original dataset.

Conclusions:

• Synthpop offers a better balance of statistical accuracy and privacy protection for synthetic data generation.
• Synthpop provides greater methodological transparency and requires less computational time.
• Standardized metrics and acceptable thresholds for disclosure risk are needed for synthetic data in research and regulatory contexts.