- A reanalysis of the US Census Bureau's Disclosure Avoidance System
- Project Year:
2022
- REU Student (s):
Leah Ghazali | University of Richmond VA
- Student 1 Institution:
University of Richmond
- Project Mentor:
Ruobin Gong
- Project Mentor Area:
Statistics
- Project Abstract:
Differential privacy is a promising mechanism for privacy protection, providing a way to calculate the maximum amount of privacy risk when an individual's data is included in a study. The US Census Bureau recently incorporated differential privacy to protect census data using the Disclosure Avoidance System (DAS). The DAS essentially consists of 2 steps: noise injection and post-processing. Researchers have conducted analyses on the DAS to examine its effectiveness and found numerous biases, which they specifically credited to the post-processing step. However, the US Census Bureau did not release the census data following the noise injection but before the post-processing, so there is no way to differentiate the effects of each step. Through the production of noisy data and replication of research by Kenny et al. (2021, Science Advances), we found that the biases found were due to post-processing. More importantly, we found that the analyses we replicated presented misleading results. When analyzing the error of the DAS, the researchers used a fitted error, which they calculated using a generalized additive model (GAM). The model's predictor variables included parameters plotted on the x-axis of their figures, essentially creating the illusion of stronger trends.We found that using the fitted model resulted in deceptive images that exaggerate the effect of the biases. By replicating their study, we found that the biases are still present, though less intense. We hope that our findings will provide a more accurate depiction of the biases of the DAS.
