Comparison of anonymization techniques regarding statistical reproducibility

Area of Science:

• Data Science
• Biostatistics
• Privacy Engineering

Background:

• Anonymization enables secondary data use by removing personal identifiers, bypassing GDPR requirements.
• Data alteration is inherent in anonymization, necessitating evaluation of its impact on data reliability and utility.
• This study compares anonymization techniques for their effectiveness in maintaining scientific data integrity for secondary research.

Purpose of the Study:

• To evaluate the impact of different anonymization methods on the reliability and utility of scientific data.
• To compare the performance of anonymization techniques across various statistical analyses.
• To assess the trade-off between privacy protection and data usefulness in secondary data analysis.

Main Methods:

• Four anonymization solutions were applied to a cohort dataset.
• Analyses were reproduced on anonymized data to assess replication (Level 1) and accuracy (Level 2).
• Data alteration was measured using Hellinger distances (Level 3), and privacy risks were quantified (Level 4).

Main Results:

• Replication scores varied from 67% to 100%, with regression and survival analyses being challenging.
• Accuracy scores ranged from 22% to 79%, indicating significant data utility loss with some methods.
• All methods reduced privacy risks (41%-65%), but some altered variable distributions.

Conclusions:

• No single anonymization method perfectly reproduced all original statistical outputs and results.
• A critical trade-off exists between the level of privacy protection and the utility of anonymized data for research.
• Selecting appropriate anonymization techniques requires careful consideration of the specific research context and data needs.