overture-coverage-equity

Code and small derived data for an entry to the Zindi Bias Bounty Mapping Equity Challenge (Humane Intelligence, 2026). The repository does two things:

1. Leaderboard estimate. A label-free estimate of the tract-level coverage_gap_score for all 9,794 scored census tracts, built only from the challenge's own Overture, ACS and strata files.
2. Bias Discovery: "Children on dialysis are invisible on the open map". We compared every CMS-certified dialysis facility in the scored tracts with Overture places. Pediatric units are missing from a dialysis_clinic category query far more often than adult units: 8 of 12 versus 127 of 851, odds ratio 11.4, Fisher p = 8.4e-5. Under a stricter test, only 1 of the 12 is findable as a pediatric clinic. Houston and Dallas-Fort Worth have no pediatric unit that the query returns. The nearest one it returns is in Austin, 232 km from Houston and 273-290 km from Dallas-Fort Worth.

Write-ups: see the Zindi discussion threads "Best Bias Discovery: …" and "Best Documentation: …".

Compliance statement

• The scored submission (outputs/submissions/structural_v1.csv) uses only challenge files that remained published after 2026-09-25: Overture extracts, ACS housing, strata tables and tract boundaries. It never reads TIGER/Line roads, Microsoft Building Footprints, HIFLD/USGS facilities, County Business Patterns, any *-coverage-gap* file, or other copies of them. src/common.py::assert_permitted refuses those file names, and src/download.sh never fetches them. The estimator uses Overture's own attributes, mainly road routes and place names and categories. We have asked the organisers to confirm this is in scope.
• The Bias Discovery analysis uses only Overture's own attributes (place names, categories, phone numbers) from the challenge files, plus openly licensed external data, which the rules allow for this prize: the CMS dialysis facility listing, the US Census Bureau geocoder, and OpenStreetMap Nominatim as a fallback geocoder. It uses none of the withdrawn reference layers. It does not estimate or calibrate any score.
• No supervised training on labels. No labels were ever published. A fixed seed (SEED in src/train.py) makes train.py deterministic.

Reproduce

Tested on Windows 11 with Python 3.13 on CPU. The commands work in any bash, including Git Bash.

python -m venv .venv
source .venv/bin/activate          # Windows Git Bash: source .venv/Scripts/activate
pip install -r requirements.txt    # pinned versions; see note below

goal	command	download	time
Bias Discovery tables and figures	bash run_all.sh discovery	~340 MB	~2 min
Leaderboard file from committed features	bash run_all.sh submission	none (needs data/SampleSubmission.csv)	seconds
Everything from raw data	bash run_all.sh all	~3 GB	~10 min

data/SampleSubmission.csv comes from the competition's Zindi Data tab. It is not redistributed here.

Discovery step by step (what run_all.sh discovery runs):

MODE=dialysis bash src/download.sh       # Overture places, tract polygons, sample IDs, national strata
python src/discovery_dialysis.py         # -> outputs/dialysis_summary.md, dialysis_facilities.csv, dialysis_pediatric_units.csv
python src/audit_pediatric_matches.py    # -> outputs/pediatric_match_audit.csv (strict pediatric test)
python src/fig_dialysis.py               # -> outputs/figures/dialysis_unseen_rates.png, dialysis_pediatric_texas.png

The external inputs are committed in data/external/, so the discovery step makes no calls to CMS, Census or Nominatim. src/get_external.py documents and re-runs how they were made: it downloads CMS, rebuilds the geocoder input, and calls the Census batch geocoder.

A clean run on these committed files reproduced the committed outputs byte-for-byte: structural_v1.csv, dialysis_facilities.csv and dialysis_pediatric_units.csv.

Note: pins such as duckdb 1.4.1, pandas 2.3.3 and pyarrow 21.0.0 are deliberate. Newer wheels were blocked by Windows Smart App Control on the development machine. DuckDB installs its spatial extension on first use, which needs internet access once.

Layout

run_all.sh                       one-command reproduction (discovery | submission | all)
src/common.py                    paths, permitted-file guard, CRS helper
src/download.sh                  public challenge data from Source Cooperative (resumable)
src/features.py                  per-tract features from Overture roads/buildings/places/infra + ACS
src/estimate.py                  label-free estimator + submission writer (exact Zindi columns)
src/train.py                     spatial GroupKFold CV pipeline (proxy target; real labels if ever published)
src/bias.py                      coverage indicators by stratum, regressions, maps
src/discovery_dialysis.py        Bias Discovery: CMS dialysis facilities vs Overture places
src/audit_pediatric_matches.py   strict check of the 12 pediatric matches
src/fig_dialysis.py              discovery figures
src/get_external.py              how data/external/ was produced (CMS download, Census geocoder)
data/external/                   CMS snapshot (6 states), geocoder output, Nominatim cache
data/features/                   per-tract features (2.4 MB), so the estimator runs without the 3 GB download
outputs/                         results committed for inspection (regenerated by run_all.sh)

Data sources and licences

source	used for	licence	URL	retrieved
Challenge bundle (Overture 2026-08-19.0 extracts, ACS housing, strata, tracts)	everything	CC BY-SA 4.0 (challenge), with the upstream licences below	https://source.coop/humane-intelligence/bias-bounty-mapping-equity-challenge	2026-10-02
Overture Maps places	estimator, discovery	CDLA-Permissive-2.0	https://overturemaps.org	via bundle
Overture Maps transportation, buildings	estimator	ODbL 1.0	https://overturemaps.org	via bundle
CMS Dialysis Facility – Listing by Facility (dataset 23ew-n7w9, modified 2026-06-16)	discovery	US Government work, public domain	https://data.cms.gov/provider-data/dataset/23ew-n7w9	2026-10-02
US Census Bureau Geocoder, batch geographies/addressbatch, Public_AR_Current / Current_Current	discovery (795 of 863 facilities)	public domain	https://geocoding.geo.census.gov/geocoder/	2026-10-02
OpenStreetMap Nominatim (fallback geocoder)	discovery (68 of 863 facilities, none pediatric)	ODbL 1.0, © OpenStreetMap contributors	https://nominatim.openstreetmap.org	2026-10-02

data/external/cms_dfc_facility_2026-06-16_6states.csv is an unmodified row subset of the CMS file, covering AZ, CA, NM, OK, TX and WA. It is kept because CMS refreshes the listing; the next update is announced for 2026-10-28.

Licence

Code: MIT (see LICENSE). Derived data keeps the licence of its source. Files derived from Overture and the challenge bundle are under ODbL / CDLA-Permissive-2.0 / CC BY-SA 4.0. Files derived from Nominatim coordinates are under ODbL. Files derived only from CMS or Census data are public domain.