Methods and limits
How the 1902 records were built, what each tier means, where the model is involved, and what has not been checked. Read this before citing a number.
Lineage
| Authors' map | 186 enclosed-mall points from Ellen Dunham-Jones and June Williamson's public ArcGIS "Interactive Mall Map" (web map ce33ab26…, last edited 20 January 2021): title, one-line description, coordinates, strategy layer. These are the "180" records marked from the authors' map. They have not been reviewed by the authors for this site. |
|---|---|
| Wikipedia | English Wikipedia categories for defunct, demolished and abandoned shopping malls in the United States (717 pages) and the by-country equivalents (133 pages): summary text, coordinates, page image, Wikidata id. Pages that are not about a specific retail site (lists, cities, companies, events) are filtered by title and description; some still slip through and are flagged as the review finds them. |
| Wikidata | 1,821 US shopping-centre items and 498 worldwide items with an end date, demolition state or closure: opening and dissolution dates, area, owners and operators (undated statements), images, English article links. Matched to the authors' points by name and distance (85 of 201). |
| Esri 2014 layer | Esri StoryMaps "Closed_US_Malls" (152 points, Directory of Major Malls data as of 2014): GLA, store count, opening year, anchors. 92 matched to existing records, 60 added as new. |
| Model extraction | For records with a Wikipedia article, the full article text was read by an open-weight model (NVIDIA Nemotron 3 Super via API, Qwen 3 8B on a Mac mini) with a fixed schema. Every non-null field must be supported by a sentence copied from the text; those sentences are stored in sources.csv and shown on the page. Curated values are never overwritten by the model. Fields filled this way carry the origin "extracted, unreviewed". |
| Research sets | Planning facts for 25 flagship retrofits (rezoning, approving body, year, subsidies with amounts, plan documents) and 59 transactions, listings, loans and foreclosures since 2024 were compiled from news, municipal documents and rating-agency notes by research agents with a source URL and a quote per row, then matched to records by name and state. |
| County parcels | Public ArcGIS REST layers of county assessors (13 counties so far) queried with a 40 m buffer around the site point: owner of record, acreage, use, zoning, values, last sale where published. Adjacent parcels can be included; county terms of use apply. |
| Census | Geocoder for state, place and county; ACS 5-year 2018 and 2023 county population, median gross rent, median household income and vacancy, used only in the multifamily screen. |
| Photos | Wikimedia Commons: the article's page image plus up to four files from the item's Commons category or a name search; only free licences; author and licence shown per image. |
Tiers
| From the authors' map | Taken as published in 2021. Nothing reviewed by the authors for this site. |
|---|---|
| Imported, unreviewed | Everything else. Fields have an origin tag; model-extracted ones have a quote. No person has checked them yet. |
| Contributed | Submitted through the form and reviewed by the editor before publication. None yet. |
Assessment layer
Each record with enough data carries an assessment written by a language model (NVIDIA Nemotron 3 Super via API, or Qwen3 on a Staffbox Mac mini) from the record alone: the fields with their origins, parties, deals, county parcels, context layers and timeline. The model is told to use no outside knowledge and to say when something is unknown. The output is structured (what happened, why, where it stands, notes for a planner, notes for a developer, risks and unknowns, a data-quality grade) and is regenerated when the record changes; a hash of the record it was written from is stored with it. It is unreviewed and is labelled as such on the page. It is a reading aid, not a finding; errors in it trace back to errors in the record.
Place layers for new urbanists
Three layers answer the questions a Congress for the New Urbanism reader asks first. Walkability: the EPA National Walkability Index for the site's census block group (1 to 20), with its inputs: weighted intersection density, distance to the nearest transit stop, employment and household mix, jobs reachable within 45 minutes by car and by transit, and the share of households with no car (EPA Smart Location Database, 2021 release). 15-minute neighbourhood: counts of grocers, schools, parks, health care, civic buildings and food within 800 m from OpenStreetMap, plus cycleway-tagged and sidewalk-tagged ways; a site with five of the six daily-needs categories is tagged a 15-minute neighbourhood. OpenStreetMap coverage varies by town, so a zero may mean unmapped. Opportunity Zone: whether the point falls in a designated federal Opportunity Zone tract (HUD layer). All three are US only, point-based, and retrieved on the date shown in contexts.csv. Two more: housing permits for the site's county from the Census Building Permits Survey (2023 and 2024 annual county files: units in 1, 2–4 and 5+ unit buildings), and National Register of Historic Places points within 500 m (National Park Service). Also the CDC/ATSDR Social Vulnerability Index 2022 for the tract (overall and four theme percentiles, with poverty, no-vehicle and unemployment shares), a computed cluster count of other atlas sites within 10 and 25 km, and a web layer: pages with photos and ULI, planning or academic references found through a search engine (Google and Bing results via a self-hosted metasearch), stored as links with the page title. We never copy or host those images; the page must mention the site's name to be kept, and nobody has reviewed the matches.
Before and after fields
A second extraction pass reads each record's text (the Wikipedia article, the narrative, planning and after notes, and the quoted sentences from sources) and fills before/after building facts: parking spaces and type, enclosed or open-air, levels, anchor count, original and retrofit architects, landscape architect, master planner, contractor, construction and retrofit costs, public investment and its kind, housing units built and planned, affordable units, hotel rooms, office and retail square footage after, jobs, whether a street grid was cut through, blocks, park acres, transit added, demolition year, phasing and stated density. Every value must carry a verbatim quote that appears in the text, and numbers must appear in the quote; anything else is dropped at load. Planned figures are kept apart from built ones. Nobody has reviewed these yet.
Tract change: the same ACS 5-year tract measures for 2013 and 2023 (population, median household income, renter share, median home value), with the change. Income and value changes are nominal. 2013 data use 2010 tract boundaries; where a 2020 tract has no 2013 match the page says so rather than comparing different areas.
Hazard and housing subsidy: the FEMA National Risk Index for the tract (composite risk and expected annual loss ratings, community resilience, and each hazard rated Relatively High or Very High) and HUD Low-Income Housing Tax Credit properties within 1 km (count, units, the largest projects and their placed-in-service year). Photo checks: a vision model (Qwen 3.5 9B on a Staffbox Mac mini) looks at every key photo and the gallery candidates and answers whether the image shows a mall, shopping center or its redevelopment, the view type and a quality grade; photos that fail are removed from the hero slot and a passing gallery image is promoted, or the record shows no photo. Captions on verified photos are the model's neutral one-line description.
County market: retail trade establishments and employment for the county in 2012 and 2023 (Census County Business Patterns, NAICS 44-45) and the county's typical home value and typical rent with one- and five-year changes (Zillow Research ZHVI and ZORI, used under Zillow's research-data terms with attribution). County figures are context, not the site's own.
Prices and traffic: the FHFA annual house price index for the tract (2000 = 100, all-transactions) with one-, five- and ten-year changes, and the busiest counted road within 600 m from the FHWA Highway Performance Monitoring System (annual average daily traffic, trucks, functional class, lanes, latest state file).
Overture, websites and attention
Each site is matched to an Overture Maps place (CDLA-Permissive 2.0) by name and distance within about 600 m, which gives a stable GERS id, a category, the website, social accounts and phone numbers that Overture aggregates from Meta and other providers, and a confidence score; the building footprint under the point gives floor area, height and a second GERS id. Both link back to the Overture explorer. The website is checked for liveness (an HTTP request) and, when the Internet Archive is reachable, its first and last Wayback captures are recorded. Attention is the monthly Wikipedia page-view series since July 2015 (total, last 12 months, change against the prior 12, peak month), a public proxy for interest in a site; it says nothing about the mall's own traffic. Follower and like counts are not collected: the platforms do not publish them without an app token, and we do not scrape them.
Profile elements, trends and lists
Each site carries a set of elements, the way a technology profiler tags a website: strategy, new uses, policy tools, finance and distress signals, ownership type, former anchors, place context, scale and era. They are derived by rule from the record itself (regular expressions over the after, planning, zoning and subsidy fields; deal rows; parcel owners of record; the context layers; anchors), and each tag stores the matching snippet as evidence and the origin of the field it came from. Elements inherit the tier of the record and are not separately reviewed; a wrong field makes a wrong tag. Trends count records with the field filled, so they describe this database, not the country. Lists are fixed combinations of elements, regenerated on each publish.
Known error modes
- Wrong article attached to a name (five Westgates, three Northgates): mitigated by name-token overlap and distance checks; some remain.
- Status from Wikipedia categories can be stale: an "abandoned" mall may have reopened. Category-only status is shown as "not verified" and scores zero in the screen.
- Years from Wikidata sometimes refer to the successor project. Conflicts between a year and the text are flagged for review, not resolved automatically.
- Model extraction can copy a plausible sentence that does not support the field. Quotes are shown so a reader can judge.
- Parcel buffers catch neighbours; the owner of record shown is the largest parcel's.
- Owner statements from Wikidata are undated and may be historical.
Not yet done
- No manual accuracy audit. A sample of 100 random records checked by a person is the next step; the error rate will be published here.
- No licence; no DOI; no versioned release.
- The authors' own 2,500-row database (an Excel workbook) is not yet imported.
- No API; the static files are the interface.
Build 2026-10-20. Pipeline source is in the Staffbox pilot folder; the schema is at schema.sql.