What we compute, and what we only cite

The honesty of a data publication is mostly a list of what it does not have. This is that list.

What we compute from

The Sanctuary warehouse holds public datasets at Census tract and ZIP level, staged with a checksum and a retrieval date for every file. As of this issue, the sources actually in use are:

• American Community Survey, 2024 five-year estimates (2020–2024). The base. Population, households, age, tenure, income, poverty, mobility, language, education. Twelve tables, six geography levels. Every proportion carries its margin of error, computed with the Census Bureau's published approximation.
• CDC PLACES, 2025 release. Small-area health estimates: chronic conditions, disability, preventive care. Modeled, not measured. We label it every time.
• CDC/ATSDR Social Vulnerability Index, 2022. Tract-level vulnerability across four themes. Also modeled.
• USDA SNAP Retailer Access Map, 2025. Food access, low-income low-access tracts, straight-line distance.
• IRS Exempt Organizations Business Master File. Nonprofit and religious organization counts, at ZIP level.
• LEHD LODES 8, 2022. Jobs and workers by tract, wage bands.
• Opportunity Atlas. Intergenerational mobility for children who grew up in a tract decades ago. Context about a place's past, never a prediction about the people living there now.
• Census ZCTA-to-tract relationship files, 2020. How a ZIP is assembled from tracts.

What that list cannot do

It cannot see your congregation. Nothing here is attendance, giving, membership, visitors, or return rates. This publication knows the neighborhood and is agnostic about the church.

It cannot see religion. There is no tract-level measure of belief, affiliation, or practice that would survive scrutiny. Anybody telling you the religious composition of your ZIP to two decimal places is selling you something.

It cannot predict. There are no outcome labels in this warehouse — no past conditions paired with past results. Descriptive statistics are what we have, and descriptive is where we stop.

It is thinner where the ACS is thin. Small ZIPs, rural tracts and institutional populations produce wide margins of error and awkward denominators. When the data is too thin to say anything, the honest answer is that the data is too thin.

What we only cite

Denominational numbers — membership, attendance, appointments, closures — are published by denominational bodies in their own formats and under their own terms. They are third-party data. We will quote them, interrogate them, and link them. We will not recompute them and we will not present them as if we had.

The same goes for reporting. When a journalist or an outlet has done the work, the proper move is to credit it and link it, not to launder it into an original finding. Copying a story and adding a paragraph is not analysis, and this publication does not do it.

Attribution and terms

• Census Bureau data is public domain. PLACES, SVI, LODES and SNAP data are public. The Opportunity Atlas carries terms that we confirm before publishing any figure derived from it.
• Where a source is modeled, we say "modeled."
• Where a source is licensed, we either clear it or leave it out.
• Where we quote reporting, we link the original and keep quotation to the minimum needed to make the point.

Re-running us

The warehouse code is not public, but the method is not secret either. Every figure names its query and its column, the receipts table under each issue lists them together, and the datasets are all downloadable from their publishers. If you want to check a number, you can rebuild it from the Census summary files and the table numbers given. If it comes out different, write to [email protected] and it will be treated as a correction, not a complaint.

More on the rules themselves: how the numbers work.