Comparison of anonymization techniques regarding statistical reproducibility

Anonymization opens up innovative ways of using secondary data without the requirements of the GDPR, as anonymized data does not affect anymore the privacy of data subjects. Anonymization requires data alteration, and this project aims to compare the ability of such privacy protection methods to maintain reliability and utility of scientific data for secondary research purposes.

The French data protection authority (CNIL) defines anonymization as a processing activity that consists of using methods to make impossible any identification of people by any means in an irreversible manner. To answer project's objective, a series of analyses were performed on a cohort, and reproduced on four sets of anonymized data for comparison. Four assessment levels were used to evaluate impact of anonymization: level 1 referred to the replication of statistical outputs, level 2 referred to accuracy of statistical results, level 3 assessed data alteration (using Hellinger distances) and level 4 assessed privacy risks (using WP29 criteria).

87 items were produced on the raw cohort data and then reproduced on each of the four anonymized data. The overall level 1 replication score ranged from 67% to 100% depending on the anonymization solution. The most difficult analyses to replicate were regression models (sub-score ranging from 78% to 100%) and survival analysis (sub-score ranging from 0% to 100. The overall level 2 accuracy score ranged from 22% to 79% depending on the anonymization solution. For level 3, three methods had some variables with different probability distributions (Hellinger distance = 1). For level 4, all methods had reduced the privacy risk of singling out, with relative risk reductions ranging from 41% to 65%.

None of the anonymization methods reproduced all outputs and results. A trade-off has to be find between context risk and the usefulness of data to answer the research question.