# Housing services analysis methods

## Scope and source vintage

These are three distinct cross-sectional analyses of the 2025 American Community Survey one-year estimates, retrieved from Census on October 31, 2026. They are not annual changes, national rankings, address-level diagnoses or causal estimates. The downloaded table-based Summary File contains published estimates and margins of error. The corresponding Census API group JSON freezes variable labels and concepts. `source-manifest.json` lists exact URLs, UTC retrieval times, byte lengths and SHA256 hashes. All file paths in that manifest resolve from its own directory.

`python3 reproduce.py` from any working directory verifies every frozen source hash, reads original data, and recreates the CSV, JSON and original SVG figures in this directory. It requires only the Python standard library and no network. Outputs replace derived files only, never raw inputs. Do not interpret the script's success as independent audit or permission to publish.

## Geography and universes

- Maine is state FIPS 23, GEO_ID `0400000US23`.
- New Mexico is state FIPS 35, GEO_ID `0400000US35`.
- Mississippi is state FIPS 28, GEO_ID `0400000US28`.
- Alaska, `0400000US02`, was inspected as a discovery alternative and is not a fourth article.

The `Geos20251YR.txt` source verifies state names and identifiers. No metro, city, county or tribal boundary is substituted for a state. Table B25117 (tenure by heating fuel), B25049 (tenure by plumbing) and B25053 (tenure by kitchen) have occupied-housing-unit universes. An occupied housing unit corresponds to a household, not a person. B28002 has a household universe. Vacant homes and group quarters do not belong in the analytical denominators.

Allocation table B992511 uses occupied housing units. B99259 and B992510 include **all housing units, including vacant**; their statewide allocation rates are contextual quality indicators, not rates specifically for the article's occupied-unit subset. B99281 and B99283 use households. Allocation count tables do not publish margins of error. Subscription allocation categories in B99283 overlap and must not be summed.

## Cells and arithmetic

The raw format is pipe-delimited. `B25117_E005`, for example, corresponds to API variable `B25117_005E`. The schema lookup is asserted before using a cell. `cell-audit.csv` retains every used estimate/MOE value and its label. Negative numeric sentinels, blanks and nonnumeric flags cause a hard failure; none is treated as zero. A genuine numeric zero with a positive MOE remains zero. The table-based files have no separate annotation columns; the script rejects special values rather than inferring meaning from missing annotations.

Maine: owner denominator B25117_002E; renter denominator B25117_012E. Owner fuel cells are 003–011, renter fuel cells 013–021, in matching order. Each fuel percent is its count divided by its own tenure denominator, not by all households. Categories identify the fuel used most, not every fuel or equipment installation. No heat-pump adoption, efficiency, price or burden finding is inferred.

New Mexico: B25049 and B25053 use owner denominator 002, renter denominator 005, incomplete owner count 004 and incomplete renter count 007. Complete plumbing means hot/cold running water and a bathtub/shower; it does not measure toilet presence, water quality, sewer connections or septic performance. Complete kitchen means a sink with faucet, stove/range and refrigerator. Counts of missing components do not identify the severity, repair cost, legal condition or assistance eligibility of an individual home.

For the missing-plumbing set P and missing-kitchen set K within a tenure group of N households, union bounds from the point estimates are `max(P,K)` to `min(N,P+K)`. Intersection bounds are `max(0,P+K-N)` to `min(P,K)`. The statewide bounds sum the separate owner and renter bounds. The resulting New Mexico union range, 10,302–18,339, is an algebraic range given the published point estimates, **not a confidence interval, certain population minimum or observed linkage**. No union MOE is asserted because the overlap and covariance are not supplied.

Mississippi: B28002_001E is the household denominator. The five-group partition is broadband other than cellular-only (004 minus 006), cellular-only (006), dial-up-only (003), access without subscription (012), and no access (013). Their integer point estimates sum exactly to total households. Cable/fiber/DSL (007) is shown as context but is not added to the partition; it overlaps other service types. Broadband here is the ACS subscription category, not a speed test, fixed-wireless inventory, service-availability map or FCC infrastructure classification. “No subscription” combines 012 and 013, but “no access” is 013 alone.

## Margins of error and comparisons

Published MOEs are 90% MOEs. For a subset proportion p=X/Y, approximate MOE is `100 * sqrt(MOE_X² - p²*MOE_Y²) / Y`. If the expression under the square root is negative, use the corresponding plus-sign ratio formula and flag the fallback. None of the selected article proportions needs that fallback. For broadband excluding cellular-only, the derived numerator MOE uses root-sum-squares of both component MOEs, ignoring unavailable covariance; it is an approximation, not an official Census-published MOE. Derived totals/differences similarly require caution.

For tenure-rate differences, approximate 90% MOE is the root-sum-square of the two proportion MOEs. Standard error equals MOE/1.645, with a normal two-sided p-value. This ignores covariance unavailable in the detailed tables. Census's frozen 2025 worked examples and testing instructions are included; margins do not capture every nonsampling error.

Discovery displayed heating, facilities and internet values in Maine, New Mexico, Mississippi and Alaska, and inspected age and income internet schemas. B28005 describes people in households by age; B28004 describes households by income. They cannot be joined to infer age-specific access within income strata and are not used in these articles. No national search for the most extreme rate or annual-change test was conducted.

To avoid selecting apparent tenure differences after inspection, `comparison-family.csv` retains all **44 tests**: nine heating-fuel comparisons and two facilities comparisons in each of the four observed states. Bonferroni familywise 5% uses 0.05/44; adjusted p-values are retained. Maine oil, electricity, utility gas and wood contrasts pass. Neither New Mexico facilities rate contrast passes. The New Mexico article therefore describes point estimates and counts, not an established higher/lower tenure-rate result. Mississippi's partition is descriptive; no interstate or temporal significance test is claimed. Individual 90% chart intervals are not simultaneous familywise intervals.

## 2025 collection and release qualifications

The Census Bureau's October 2026 user note reports that a six-week suspension of collection affected August–November panels and impacted sample size, response rates and margins of error. A separate release note says household record swapping ended for the 2025 one-year tabulations and some legacy release rules were removed; geographic population thresholds and quality rules remained. These analyses use only the single 2025 vintage and make no trend assertion across those changes.

- [Collection interruption](https://www.census.gov/programs-surveys/acs/technical-documentation/user-notes/2026-04.html)
- [2025 release rules](https://www.census.gov/programs-surveys/acs/technical-documentation/user-notes/2026-06.html)
- [Subject definitions](https://www2.census.gov/programs-surveys/acs/tech_docs/subject_definitions/2025_ACSSubjectDefinitions.pdf)
- [Worked examples](https://www2.census.gov/programs-surveys/acs/tech_docs/accuracy/2025_ACS_Accuracy_Document_Worked_Examples.pdf)

Allocation can fill missing or inconsistent responses; the supplied allocation figures do not prove bias or validate each household's answer. Sampling error, reporting error and imputation all counsel against interpreting small estimates as exact cases. Raw-source preservation establishes what was analyzed, not that all nonsampling error has been eliminated.
