# Methods: three distinct 2025 housing cross-sections

## Source freeze and reproduction

The source is the U.S. Census Bureau’s 2025 American Community Survey one-year Detailed Tables, released October 29, 2026. Data files, group schemas, table shells, geography labels, subject definitions and accuracy/jam-value documentation were retrieved directly from Census on October 31, 2026. `source-manifest.json` records each exact URL, UTC retrieval time, byte length and SHA-256. Its relative paths resolve from the directory containing that manifest. It covers primary source material, not newsroom contacts.

Run `python reproduce.py --out public/reproduced` for a complete offline numerical rebuild and source-byte verification. Python 3 standard library is sufficient. Add `--charts` to rebuild the three PNG/SVG charts; this requires matplotlib (production: 3.10.8). The script resolves its inputs relative to itself, does not require network access, does not import private records, and never edits frozen raw files. The `--out` directory is created if needed. Charts are original visualizations of these calculations. The output claims and input CSVs contain exact source cells and unrounded results; prose rounds percentages to one decimal and approximate count MOEs to whole units.

## Geographies and universes

- Pittsburgh, PA Metro Area: `310M700US38300`.
- Buffalo-Cheektowaga, NY Metro Area: `310M700US15380`.
- Memphis, TN-MS-AR Metro Area: `310M700US32820`.

These are the areas in the frozen 2025 ACS geography file, not identically named cities. No county or metro estimates are combined. In particular, Memphis includes territory in three states; no value is presented as a city-only or Shelby County-only estimate.

Official frozen table shells establish the universes: B25126 and B25124 are occupied housing units; B25004 and B25131 are vacant housing units; B25130 is other vacant. Household counts here correspond to occupied housing units. They are population estimates, not respondent counts or survey sample sizes. Age is the age of the designated householder, not every occupant. Group-quarters residents are outside these household universes.

## Calculations

For every proportion, `p = numerator / denominator`; the numerator is a subset of the denominator. A multi-cell count is the sum of mutually exclusive cells, and its approximate 90% MOE is `sqrt(sum(cell_MOE^2))`. For a proportion the approximate MOE, in percentage points, is `100 * sqrt(MOE_num^2 - p^2 * MOE_den^2) / denominator`. If the radicand is negative, use the Census ratio fallback with a plus sign instead; `ratio_fallback` records this. No selected ratio needed that fallback.

These are approximations without the full survey covariance, not exact replicate-weight errors. [Census worked examples](https://www2.census.gov/programs-surveys/acs/tech_docs/accuracy/2025_ACS_Accuracy_Document_Worked_Examples.pdf) describe the approach and its limitations. Percentage error-bar endpoints are clipped to 0–100% for display only; stored MOEs remain unaltered. A large MOE on a small estimate is not evidence of zero cases.

- **Pittsburgh:** Within each of six age-by-tenure groups in B25126, sum the 1950–59, 1940–49 and 1939-or-earlier cells and divide by that group’s published total. The pre-1960 boundary follows available bins; it is not a safety, condition or accessibility standard. Construction year is not renovation year.
- **Buffalo:** In B25124, combine four-person and five-or-more-person renter households. Divide each building-size category by the sum of those two household-size totals (23,757). Separately describe one-person renter households using their own denominator (79,568). Five-plus-unit buildings combine the 5–19, 20–49 and 50-plus bins. One-unit includes detached and attached homes. Two-to-four-unit buildings cannot be separated into duplexes/triplexes/fourplexes. No bedroom, crowding, family-composition or unmet-demand result is inferred.
- **Memphis:** B25004 describes all 65,429 vacant units. Its other-vacant subtotal (31,603) exactly equals the B25130 total. Sum B25130 cells 7 and 8 for the repair-related description; report both components separately. “All remaining” groups are explicit disjoint sums, not residual estimates with unknown error. For-rent share among vacant units is not the rental vacancy rate. Repair categories include cleaning; neither repair nor abandonment labels establish inspection findings, cost, ownership or availability.

## Selection, missingness and multiple comparisons

This is exploratory, not preregistered. The discovery screen selected whole counties (summary level 050) and named metro areas (310, excluding micro areas), component 00, from the geography file: 1,256 candidate geographies, some overlapping. It generated 10,281 valid metric/geography rows across 18 descriptive metrics and five tables. In total 1,154 geographies had at least one screened metric. Availability differed sharply: B25126 had 1,107 candidate rows; B25124 380; B25004 649; B25130 219; B25131 911. Missing published rows were not converted to zero. The schema summary records those coverage counts. Initial inspection emphasized metros with at least 15,000 in the metric denominator and also inspected a named regional shortlist.

All original screen rows, code and missing-row exclusions are retained in the private production record. No tested hypothesis family, p-value filter or annual-change analysis was used to select stories. The three places were chosen for substantial interpretable groups, distinct research questions and relevant local reporting context. Published intervals are descriptive 90% intervals, not simultaneous or selection-adjusted intervals. No national rankings, definitive between-category ordering, equality claims, statistically certain majorities, annual increases or causal effects are claimed. Proximity or overlap of plotted intervals is not, by itself, a general significance test.

All selected estimate/MOE cells were checked as numeric, finite and nonnegative, with positive denominators and valid subset bounds. Blank, symbolic, negative-sentinel and nonfinite cells would stop selected reproduction rather than become zero. Raw files preserve any annotation/jam values; API schemas and the official 2025 jam-value guide are frozen. No present discovery row failed the numeric checks, but absent table/geography rows are separately recorded. These cross-sections do not repair missing coverage and should not be compared with earlier county or differently defined metro estimates as trends.
