Disclosure Avoidance in the Census Bureau’s 2010 Demonstration Data Product

Abstract

Producing accurate, usable data while protecting respondent privacy are dual mandates of the US Census Bureau. In 2019, the Census Bureau announced it would use a new disclosure avoidance technique, based on differential privacy, for the 2020 Decennial Census of Population and Housing [19]. Instead of suppressing data or swapping sensitive records, differentially private methods inject noise into counts to protect privacy. Unfortunately, noise injection may also make the data less useful and accurate. This paper describes the differentially private Disclosure Avoidance System (DAS) used to prepare the 2010 Demonstration Data Product (DDP). It describes the policy decisions that underlie the DAS and how the DAS uses those policy decisions to produce differentially private data. Finally, it discusses usability and accuracy issues in the DDP, with a focus on occupied housing unit counts. Occupied housing unit counts in the DDP differed greatly from 2010 Summary File 1 differed greatly, and the paper explains possible sources of the differences.

Supported by the Minnesota Population Center (R24 HD041023), funded through grants from the Eunice Kennedy Shriver National Institute for Child Health and Human Development.

Appendices

A Privacy Loss Budget Allocations

Table 2 lists the 7 person and 6 household queries that received direct allocations of the privacy loss budget. The allocations are shown in the $PLB_{frac}$ column. The bold rows are the queries with the largest PLB allocation.

Table 2. Privacy loss budget allocations and scale parameters for 2010 DDP queries.

The $Scale_{nation}$ and $Scale_{county}$ columns list the scale factors used to generate the statistical distributions from which noise injection values are drawn. The $Scale_{nation}$ values are used for the nation and state histograms, and the $Scale_{county}$ values are used for the county, tract group, tract, block group, and block histograms.

The $Hist_{size}$ column lists the number of cells in the particular query. This is the number of cells on each row of the histogram (i.e., for each geographic unit). The value is generated by multiplying together the number of categories for each variable in a query. For example, the Sex * Age (64 year bins) query has two categories for sex and two categories for age giving a histogram size of 4. Category counts for each variable are listed in frequently asked question 11 in  [7].

The variable names in the household queries are described in Table 3.

Table 3. Variable names and descriptions.

B Top-Down Algorithm flow diagram

Figure 3 depicts the flow of data through the noise injection and optimization steps of the Census Bureau’s Top-Down Algorithm.

Fig. 3.

Census DAS optimization flow diagram.