Introduction
The U.S. Census Bureau recently transitioned from traditional "data swapping" to "differential privacy" (DP) to protect individual identities in decennial census data. This shift has sparked significant debate among data users, particularly those involved in local infrastructure planning. This paper evaluates the comparative accuracy of these two methods by simulating their effects on single-year-of-age counts within synthetic school district data.
Problem Statement
Under Title 13, the Census Bureau is legally required to prevent the re-identification of individuals. Historically, this was achieved through swapping household records between geographic areas. In 2020, the Bureau adopted Differential Privacy, which introduces mathematical noise. Critics argue that this noise compromises the utility of data for critical local tasks, such as determining faculty hiring and classroom expansion based on the number of school-aged children (e.g., 4-year-olds) in a district.
Methodology
To compare these methods, the researchers developed representative variants of both algorithms:
- Swapping Implementation: A household-level swap prioritizing "unique" households while maintaining constant counts for populations over and under 18.
- Differential Privacy Implementation: A geometric mechanism adding noise to single-year-of-age histograms without the post-processing typically used by the Census Bureau, representing a "worst-case" accuracy scenario.
The primary evaluation metric was the Mean Absolute Percentage Error (MAPE), calculated against a "ground truth" synthetic dataset.
Data Sources
The study utilized synthetic microdata reconstructed to match 2010 Decennial Census tables. The analysis focused on seven specific census tracts in Rhode Island and Alabama, chosen to represent a spectrum of population sizes relevant to school district planning.
Discussion
The analysis revealed several key insights regarding the behavior of de-identification mechanisms:
- Universal Small Group Penalty: Both swapping and DP exhibit higher relative error as population size decreases. The "small group penalty" is not unique to DP.
- Predictability vs. Chaos: DP errors follow a predictable decaying exponential curve relative to the privacy budget ($\epsilon$). In contrast, swapping produces "jagged" and unpredictable error rates, making it difficult for users to assess data reliability.
- Algorithmic Transparency: Unlike the proprietary "swap rate," the DP privacy loss budget ($\epsilon$) can be publicly disclosed, allowing planners to quantify uncertainty for the first time.
- Outlier Vulnerability: Tracts with atypical age distributions (e.g., those dominated by nursing homes) showed spiked error rates across both methods due to the low counts of children.
Results
The study found that DP generally performed as well as or better than swapping at privacy parameter values likely used by the Census Bureau. Accuracy was highly sensitive to population size; for instance, the smallest tract (pop. 1,158) showed significantly higher MAPE than larger tracts. Furthermore, DP maintained better trend consistency for specific age subsets (like "Ages 4-5") compared to the total population, which is often skewed by small counts in elderly age groups.
Conclusion
While the use of synthetic data and simplified algorithms are limitations, this research suggests that Differential Privacy offers a viable and more transparent alternative to traditional swapping. Future work should focus on scaling these findings to more diverse geographies and developing methods to communicate statistical uncertainty to non-technical stakeholders, such as school board members, to ensure informed decision-making.
Summary updated