TL;DR
The 2020 Census Disclosure Avoidance System (DAS) uses a "geographic spine" to protect privacy while releasing data. This paper investigates how the choice of this internal spine impacts the accuracy of the released census data, proposing and evaluating alternative spines that better align with specific geographic entities like AIAN tribal areas.
What Problem Does This Paper Address?
The U.S. Census Bureau uses the Disclosure Avoidance System (DAS) to protect the privacy of individuals while releasing aggregated data. A core component of the DAS is the "geographic spine," a hierarchical structure of geographic units (like states, counties, tracts, etc.) that guides the privacy-preserving data generation process. The standard "conventional geographic spine" is defined by the U.S. Census Bureau's Geography Division. However, this conventional spine can be distant from certain legally defined geographic areas, such as American Indian/Alaska Native (AIAN) tribal areas, and other important entities like minor civil divisions (MCDs) and incorporated places. This distance can lead to less accurate data for these "off-spine entities" (OSEs). This paper addresses the problem of how the choice of the internal geographic spine within the DAS affects the accuracy of the released data, particularly for these OSEs, and proposes and evaluates alternative spines to mitigate this issue.
Methods
The paper introduces the concept of "geographic spines" as hierarchical structures used internally by the DAS. It details how the DAS converts microdata into histograms for these geographic units and then adds noise to ensure privacy. The core of the paper lies in exploring and defining "alternative spines" that are better aligned with OSEs.
Key methods include:
- AIAN Spine: This spine is created by splitting standard geographic units (state, county, etc.) into two branches: one representing areas within AIAN tribal lands and another for areas outside. This process is applied recursively down to the block level.
- Spine Optimization Routines: These routines aim to bring OSEs closer to the spine.
- Stage 1: Involves redefining geographic levels to create new units that are intersections of OSEs and existing spine units. This can involve creating "optimized block groups" or "tract groups."
- Stage 2 (Bypassing): This stage reduces the variance of the final estimates by removing redundant geounits from the spine. The paper describes a specific decision rule for bypassing parent geounits based on privacy guarantees and accuracy considerations, particularly for pure-DP mechanisms.
- Privacy Guarantees: The paper formally defines differential privacy (DP) and zero-concentrated differential privacy (zCDP) and shows how the DAS, with its various spine configurations, maintains these privacy guarantees. It leverages lemmas and theorems related to DP and zCDP composition to prove privacy.
- Accuracy Metrics: Mean Absolute Error (MAE) is used to quantify the accuracy of total population queries across different geographic levels and OSEs for various spine configurations.
Datasets
The paper uses the 2010 Census Edited File (CEF) as the input data for evaluating the impact of different spine settings. This file contains microdata that the DAS processes. The paper does not introduce new benchmark datasets but rather uses existing census data to evaluate the performance of the DAS with different internal spine configurations.
Key Insights and Ideas
- Spine Choice Matters for Accuracy: The internal geographic spine used by the DAS significantly impacts the accuracy of the released data, especially for off-spine entities.
- AIAN Spine Improves AIAN Data Accuracy: Creating an "AIAN spine" by splitting geographic units into AIAN and non-AIAN branches brings AIAN areas closer to the spine, leading to substantial accuracy improvements for these specific regions.
- Optimized Spines Balance Accuracy: Spine optimization routines can improve accuracy for targeted OSEs while potentially degrading accuracy for standard tabulation units (like block groups). The paper explores how to balance these trade-offs.
- Bypassing Geounits for Variance Reduction: A "bypassing" operation can reduce the variance of estimates by reallocating privacy loss budgets. The paper derives decision rules for when to bypass parent geounits to improve accuracy, particularly for pure-DP mechanisms.
- Privacy Guarantees Persist: Crucially, the paper demonstrates that these spine optimization techniques, including bypassing, do not compromise the formal differential privacy guarantees of the DAS.
Key Results
- AIAN Spine Performance: Table 3 shows that the AIAN spine significantly reduces the MAE for "Aggregated AIAN Areas" compared to the conventional spine (e.g., MAE for "Aggregated AIAN Areas in States" drops from 259.4 to 0.806).
- Optimized Spine Performance: The optimized spine generally shows improved accuracy for OSEs and most conventional geographic levels compared to the AIAN spine. However, it can lead to decreased accuracy for standard tabulation block groups (Table 2 shows MAE increasing from 1.407 to 15.99 for block groups).
- Bypassing Impact: The paper provides theoretical results (Theorem 2) showing that bypassing a parent geounit can decrease or leave unchanged the variance of estimates, especially when the parent has few children and specific PLB allocations.
- Privacy Preservation: The paper formally proves that the DAS, even with optimized spines and bypassing operations, maintains its e-DP and p-zCDP guarantees.
Limitations and Future Work
- Trade-offs in Accuracy: Spine optimization, while improving accuracy for targeted OSEs, can degrade accuracy for other geographic units (e.g., block groups in the optimized spine). The paper acknowledges this trade-off.
- Computational Complexity: While the paper describes algorithms for spine optimization, the general problem of minimizing OSEDs for non-disjoint OSEs is noted as potentially difficult, similar to set covering problems. Algorithm 2 provides a greedy approximation.
- Focus on Total Population Query: The summary metrics in Section 7 primarily focus on the total population query. While the authors suggest relative MAE comparisons are generally consistent across queries, a broader analysis of different query types could be beneficial.
- Future Work: The paper implicitly suggests further exploration of spine optimization heuristics and their impact on various types of queries and OSEs. The development of more efficient algorithms for complex OSE scenarios is also a potential area for future research.
Summary updated