Balancing Inferential Integrity and Disclosure Risk via Model Targeted Masking and Multiple Imputation.

Area of Science:

• Statistics
• Data Privacy
• Health Research

Background:

• Open data sharing is crucial for research reproducibility but raises privacy concerns.
• Multiply imputed (MI) synthetic datasets are used to protect identity, but can lead to information loss.
• Existing methods may weaken or invalidate inferences from synthetic datasets.

Purpose of the Study:

• To investigate a novel masking framework with data augmentation (DA) and a tuning mechanism.
• To balance identity disclosure protection with data utility preservation.
• To evaluate the effectiveness of the DA-MI strategy on a restricted-use Canadian Scleroderma Research Group (CSRG) dataset.

Main Methods:

• Utilized a new masking framework incorporating data augmentation (DA) and multiply imputation (MI).
• Employed a tuning mechanism to balance data utility and identity protection.
• Applied the DA-MI strategy to analyze work-disability and interstitial lung disease outcomes within the CSRG dataset.

Main Results:

• The DA-MI strategy achieved 0% identity disclosure risk.
• All inferential conclusions were preserved.
• High confidence interval (CI) overlap (98.5% and 95.5% on average, minimum 91%) was maintained compared to original data.
• Conventional methods showed significantly lower CI overlap (73.9%-91.8%, minimum 28.1%).

Conclusions:

• The DA-MI masking framework effectively protects participant identities.
• This method facilitates the sharing of valuable research data.
• DA-MI preserves data utility, ensuring reliable research conclusions.