# Data guide

Source data were frozen November 9, 2026. All CSV estimates retain the precision of the audited calculations; article figures are rounded for readability. These are population estimates of renter households, not survey sample sizes or individual renters.

## Files

- focal_estimates_2025.csv: the seven estimates featured in the article, in article order.
- headline_metrics.csv: focal and contextual records from both 2024 and 2025.
- metrics_2024.csv and metrics_2025.csv: all 728 potential geography-and-metric rows per year. There are 529 and 491 computable rows, respectively. The 13 metrics cover overall burden and severe burden, four householder-age burden groups and seven household-income burden groups.
- roster_2024.csv and roster_2025.csv: the three states and 53 full metropolitan areas. Cross-state metros retain their full areas.
- coverage_2024.csv and coverage_2025.csv: table availability for each geography. False means the table is unavailable, not zero burden.
- original_comparison_families_2024.csv and original_comparison_families_2025.csv: complete original within-year comparison families. The article relies on 2025 results.
- focused_year_changes.csv: seven focal share-change tests and three North Carolina metro count-change tests, treated as separate adjustment families. None passes its stated Holm adjustment.
- all_share_changes_exploratory.csv: the broader screen of computable and unavailable annual share changes. The 483 computable tests form one adjustment family. These are exploratory results, not a geography ranking or a set of established trend headlines.
- original_2025_ratio_sensitivity.csv: full Florida age-gap and Georgia/Atlanta majority families under the plus-sign ratio diagnostic. It preserves the original results alongside the alternative uncertainty calculation.
- all_share_ratio_sensitivity.csv: all 483 computable exploratory share changes under the plus-sign diagnostic. The NC and Durham youth results do not pass this sensitivity after the same full-family adjustment.
- source_manifest.json: exact Census source URLs, retrieval timestamps, byte counts and SHA-256 hashes for the ten 2024 and 2025 source files. Hashes identify the frozen versions, not a guarantee that a later download has not been revised.

## Reading metric columns

- geoid and geography identify the full Census geography; geography_type and states describe its scope.
- reference_year is the survey reference year, not the retrieval or publication year. product identifies ACS one-year estimates.
- table, numerator_variables, total_variable and excluded_variables identify the source table cells. source_url is the exact raw-data URL.
- total_estimate is the full relevant group. excluded_not_computed is the number without a computed ratio. denominator_estimate is total minus excluded cases. numerator_estimate is the estimated burdened count.
- percent equals 100 times numerator divided by denominator. percent_moe90 is its approximate 90% margin in percentage points; percent_ci90_low and percent_ci90_high are the corresponding displayed bounds.
- Columns ending in moe90 are 90% margins of error, in the unit of the estimate. Household counts are uncertain even when printed as whole numbers.
- status distinguishes computed from unavailable results. A blank or unavailable value is not zero. quality_flags includes ratio_moe_fallback when the subset-proportion formula requires the plus-sign fallback, and precision flags where applicable.

## Reading comparison columns

Each family names the set used for Holm adjustment. a_value minus b_value gives difference; majority tests use a fixed b_value of 50%. For share comparisons, differences and margins are in percentage points; for count comparisons, they are households. p_two_sided is the unadjusted approximate normal-test p-value. p_holm is adjusted within the family, whose size is family_tests. significant90_holm indicates whether the adjusted p-value is below 0.10.

Sensitivity columns beginning with ratio use the plus-sign ratio approximation. This is a conservative diagnostic relative to the chosen subset method, not an official replacement or a guaranteed design-based upper bound. Read the methodology before interpreting it.

## Limits that apply to every file

All burden shares exclude not-computed cases. Household income bands are nominal, not individual wages, AMI bands or fixed real-income cohorts. Age refers to the householder and cannot identify retirement or severe burden by age. Annual comparisons are descriptive, use approximate uncertainty and do not follow the same households over time. Atlanta has a documented cross-year boundary-transfer limitation with an unquantified effect. No row establishes causality, eviction or homelessness risk, current asking rent or a population affordability ranking.
