This repository contains source code for the PH-SAFE disclosure avoidance application. PH-SAFE was used by the Census Bureau for the protection of individual 2020 Census responses in the tabulation and publication of the Supplemental Demographic and Housing Characteristics File (S-DHC). Previous source code releases have included the code for earlier data releases focused on demographic and housing characteristics respectively.

PH-SAFE combines 2020 Census response information for households and the individuals within them, infusing those statistics with statistical noise to create privacy-protected tabulations.

Because information about very large households can be highly disclosive, households above a certain size are truncated, removing members above the threshold.

The resulting truncated data is used to generate a preliminary tabulation of counts and ratios for characteristics (sex, race) of household occupants. Noise is then infused into the innermost detail cells of the preliminary tables to generate the final output of the PH-SAFE algorithm.

The resulting protected table is then statistically post-processed to improve accuracy (removing certain illogical results from the noise-infusion, such as negative counts or ratios with 0 as the denominator) and to produce credible intervals for the resulting statistics.

The PH-SAFE code itself can be found in the phsafe directory of this repository. PH-SAFE was built on Tumult's "Analytics" and "Core" platforms, whose source is found in the tumult subdirectory and makes use of customized CEF (Census Edited File) readers implemented by MITRE and included in the mitre subdirectory. All of these components are implemented in Python and the latest version of the platforms can be found at [[https://tmlt.dev/]]. The post-processing code can be found in the SDHC_Model_Based_Estimates subdirectory and is written in R.

In the interests of both transparency and scientific advancement, the Census Bureau committed to releasing any source code used in creation of products protected by formal privacy guarantees. In the case of the the Detailed Demographic & Housing Characteristics publications, this includes code developed under contract by Tumult Software (tmlt.io) and MITRE corporation. Tumult's underlying platform is evolving and the code in the repository is a snapshot of the code used for the production of S-DHC

The bureau has already separately released the internally developed software for the Top Down Algorithm (TDA) used in production of the 2020 Redistricting and the 2020 Demographic & Housing Characteristics products.

This software for this repository is divided into five subdirectories:

• configs contains the specific configuration files used for the production S-DHC runs, including privacy loss budget (PLB) allocations and the rules for adaptive table generation. These configurations reflect decisions by the Bureau's DSEP (Data Stewardship Executive Policy) committee based on experiments conducted by Census Bureau staff.
• phsafe contains the source code for the application itself as used to generate the protected microdata used in production.
• mitre/cef_readers contains code by MITRE to read the Census input files used by the SafeTab applications.
• tumult contains the Tumult Analytics platform. This is divided into common, analytics, and core directories. The core directory also includes a pre-packaged Python wheel for the core library.
• ctools contains Python utility libraries developed the the Census Bureau's DAS team and used by the MITRE CEF readers.
• SDHC_Model_Based_Estimates contains the R modeling code used for postprocessing the privacy-protected join table produced by PH-SAFE.