Official Seal, Everything Is Broken (2026)

Everything Is Broken

Report No. 1 · Nonprofit Hospital Executive Pay vs Charity Care · Annual Report, 2026 edition

Report No. 1 · 2026 edition · plain-text original

Reproduction manual

Everything Is Broken · Report No. 1 · Dataset as of Nov 1, 2026

This document is a complete, self-contained procedure for rebuilding the nonprofit hospital charity care and executive pay dataset from public sources, with no prior knowledge of this project. Follow it in order.

1. Purpose and scope

The dataset answers one question for every nonprofit hospital in the United States: how much free care did it give to patients who could not pay, and how much did it pay its chief executive, as stated in its own filings. From those two figures and the hospital's size, every other figure in the report is derived.

Scope: every organization that filed an IRS Form 990 with Schedule H (the hospital schedule) in the IRS's 2025 and 2026 e-file releases. That is 2,305 distinct filers (by EIN), which after roll-up are about 1,650 health systems. Public hospitals (county, state, VA) do not file Form 990 and are not included. For-profit hospitals do not file Form 990 and are not included.

The data rule, which governs everything below: every figure in the master dataset is either copied from a filing with its exact form, schedule, part, and line recorded, or computed from such figures with the formula stated. The only estimated figures are third-party ones (Lown Institute), and they are tagged by source so they can be excluded.

A second, smaller set of Chicago-specific sources (audited financial statements, Illinois Attorney General filings, Lown) was gathered first and remains in the dataset tagged by source_type. The national procedure in Section 5 does not depend on it.

2. Definitions

Every term below names the form line it comes from and the element name in the IRS e-file XML (namespace http://www.irs.gov/efile, schema version 2024v5.1 for tax year 2024 filings). Dollar amounts are whole dollars. Percentages in the XML are decimals (0.00728 means 0.728 percent).

Term Form and line XML element Meaning
Charity care Schedule H, Part I, line 7a IRS990ScheduleH/FinancialAssistanceAtCostTyp Free or discounted care given under the hospital's written financial assistance policy, valued at the hospital's cost, not list price. Columns: TotalCommunityBenefitExpnsAmt (gross cost), DirectOffsettingRevenueAmt (grants received for it), NetCommunityBenefitExpnsAmt (net), TotalExpensePct (net as share of total expenses, filed by the hospital). The report uses net and the filed percentage.
Medicaid shortfall Schedule H, Part I, line 7b UnreimbursedMedicaidGrp Cost of treating Medicaid patients minus Medicaid payments.
Community benefit, total Schedule H, Part I, line 7k TotalCommunityBenefitsGrp Sum of 7a through 7i: charity care, Medicaid shortfall, other means-tested programs (7c), community health improvement (7e), health professions education (7f), subsidized health services (7g), research (7h), cash and in-kind contributions (7i).
Community building Schedule H, Part II, line 10 TotalCommuntityBuildingActyGrp (sic, IRS spelling) Housing, economic development, workforce, and similar. Not part of 7k.
Bad debt Schedule H, Part III, line 2 BadDebtExpenseAmt Bills the hospital tried and failed to collect. Not community benefit on the IRS form; often added by hospitals in public statements.
Medicare shortfall Schedule H, Part III, line 7 MedicareSurplusOrShortfallAmt Medicare revenue (ReimbursedByMedicareAmt) minus Medicare cost (CostOfCareReimbursedByMedcrAmt). Negative is a shortfall. Not community benefit on the IRS form.
FPG free-care threshold Schedule H, Part I, line 3a Percent100Ind / Percent150Ind / Percent200Ind (checkbox = X) or FreeCareOthPercentageGrp/FreeCareOtherPct Family income, as a percent of the federal poverty guideline, below which care is free.
FPG discount threshold Schedule H, Part I, line 3b Percent200DiscountedInd ... Percent400Ind or DiscountedCareOthPercentageGrp/DiscountedCareOtherPct Income below which care is discounted.
Facility count and list Schedule H, Part V, Section A HospitalFacilitiesCnt; one HospitalFacilitiesGrp per facility with BusinessName, USAddress, StateLicenseNum, and type indicators (ChildrensHospitalInd, TeachingHospitalInd, CriticalAccessHospitalInd, EmergencyRoom24HrsInd) Which hospitals the filing covers.
Total revenue, expenses, net Form 990, Part I, lines 12, 18, 19 IRS990/CYTotalRevenueAmt, CYTotalExpensesAmt, CYRevenuesLessExpensesAmt Current-year totals.
Net assets, total assets Form 990, Part I, lines 22, 20 NetAssetsOrFundBalancesEOYAmt, TotalAssetsEOYAmt End of year.
Group return Form 990, Part I, line H(a) GroupReturnForAffiliatesInd true means one filing covers several hospitals. Figures are then system-wide.
Compensation, per person Form 990, Part VII, Section A One Form990PartVIISectionAGrp per person: PersonNm, TitleTxt, ReportableCompFromOrgAmt (col D), ReportableCompFromRltdOrgAmt (col E), OtherCompensationAmt (col F), and indicator flags OfficerInd, KeyEmployeeInd, HighestCompensatedEmployeeInd, FormerOfcrDirectorTrusteeInd Total compensation for the report is D + E + F.
Schedule O Schedule O IRS990ScheduleO/SupplementalInformationDetail with FormAndLineReferenceDesc and ExplanationTxt Free-text supplement. Some filers put every Part VII title here.
Charity care as % of expenses (system) computed sum of member 7a net divided by sum of member Form 990 line 18 Computed the same way for every row, single-filer or not. The filed TotalExpensePct is kept in filers_latest.csv for comparison; 24 single filers differ from the computed figure by more than half a point, mostly foundations and joint-venture partners whose expenses are not hospital operations.
CEO pay as % of charity care computed CEO total compensation divided by system charity care net, times 100 Over 100 means one person was paid more than the system gave in charity care.

What "nonprofit" means here. A 501(c)(3) organization: no shareholders, no federal income tax, and in nearly all states no property or sales tax. Surpluses stay inside and may go to buildings, investments, and salaries. The label does not imply charity.

3. Environment and access

  • A Linux or macOS machine with Python 3.10 or later (standard library only; no packages required), curl, and unzip. The IRS zips use Deflate64 compression, which Python's zipfile cannot read; the system unzip can.
  • About 10 GB of free disk during the run (each batch unpacks to about 3 GB temporarily; the kept Schedule H filings total 1.5 GB; the master CSV is 170 MB).
  • Network access to apps.irs.gov (bulk XML and index files) and projects.propublica.org (API and rendered pages for cross-checks). If working inside a sandbox with a domain allowlist, both must be added. Everything else is optional.
  • No credentials, API keys, or accounts. All sources are public.

Time: the full national pull, parse, and load takes under 30 minutes of machine time. The manual review steps in Section 6 took about two hours for the systems above $1B in revenue.

Alternate route when bulk download is unavailable. ProPublica's Nonprofit Explorer renders every Schedule H as an HTML page at https://projects.propublica.org/nonprofits/full_text/<object_id>/IRS990ScheduleH (it redirects to a signed S3 URL that must be opened in a real browser; automated fetchers are refused). The text of that page can be saved and parsed with parse_schh.py. This is how the first nine Chicago filers were done and it produces identical figures, but it is one page per filing and does not scale. ProPublica's XML download (/nonprofits/download-xml?object_id=...) is behind a bot check and did not work from any route tried. ProPublica's JSON API (/nonprofits/api/v2/organizations/<EIN>.json) works without restriction and gives summary financials for every filing year but not Schedule H.

4. Data model

There is one source of truth and several derived tables. Never edit a derived table by hand; change the source or the rule and rebuild.

hospital_data.csv, the master. Long format, one row per figure, 364,682 rows at this version. Columns:

Column Content
row_id Sequential integer from 1, never reused. Contiguity is checked.
added_at UTC timestamp of the append
status active or ignore
ignore_reason Why a row is ignored and which row supersedes it
system Parent system name (filled for the Chicago rows; blank for bulk rows, where system membership is derived in the roll-up)
entity_name Filing entity exactly as on the filing
ein NN-NNNNNNN
fiscal_year_end YYYY-MM, or a range like 2020-2022avg
metric Standardized key; see the metric list below
value The figure, no thousands separators; text for text metrics
unit USD, pct, count, text
person, person_title For compensation rows
source_type IRS990, Audit, IL_AG, Lown, IHPA, IHA, ProPublica
source_doc Filing name, tax year, filing date, IRS object ID
source_location Form, schedule, part, line, column
source_url Where retrieved
retrieval_method IRS bulk XML parsed by parse_xml.py, built-in browser ..., API JSON, WebFetch summary, hand-keyed, computed
notes Caveats such as GROUP RETURN, officer flags on pay rows

The append-only rule. Rows are never edited or deleted. A wrong value is retired by appending a row with status = ignore whose ignore_reason names the bad row, then appending the corrected row. add_rows.py is the only writer and can only append. Keep a dated backup copy before any bulk load.

Metric keys written by the bulk loader, per filing: filer_city_state, group_return, total_revenue, total_expenses, net_income_990, total_assets, total_liabilities, net_assets, compensation_current_officers, other_salaries_wages, investment_income, program_service_revenue, contributions_and_grants, cash_non_interest_bearing_eoy, savings_temp_cash_eoy, investments_public_securities_eoy, investments_other_securities_eoy, employee_count, indiv_over_100k_cnt; per person: compensation_reportable (col D), compensation_related_org (col E), compensation_other (col F); per Schedule H line 7a to 7k: sch_h_<line>_<name>_total_expense, _offsetting_revenue, _net, _pct_total_expense (the _pct value is stored as a percentage, i.e. the XML decimal times 100); plus sch_h_community_building_total, sch_h_bad_debt, sch_h_bad_debt_fap_eligible, sch_h_medicare_revenue, sch_h_medicare_cost, sch_h_medicare_shortfall, sch_h_fpg_free_pct, sch_h_fpg_discount_pct, sch_h_facility_count, sch_h_facilities_list.

Derived tables (all under irs/, rebuilt by build_systems.py from the master in seconds):

  • filers_latest.csv: one row per EIN, latest fiscal year, with the headline fields, the identified CEO and confidence, and flags. 2,305 rows.
  • systems_rollup.csv: one row per system after roll-up, with summed figures, the CEO counted once, the computed ratios, childrens_hospital, zero_charity_care_flag, and publishable. About 1,650 rows.
  • ceo_unresolved.csv: filers where the rule found no CEO.
  • ceo_review_large.csv: for filers above $1B that needed review, the top six Part VII entries and any Schedule O text mentioning CEO, used for the manual pass.

Override files (hand-maintained, read by the build): ceo_overrides.csv and rollup_overrides.csv, described in Section 6. context_notes.csv (Section 6.5a) is read only by build_site.py and changes no figure.

Raw inputs kept on disk: irs/index_2025.csv, irs/index_2026.csv, irs/xml/<object_id>_public.xml (19,245 files), and for the Chicago page route raw_<ENTITY>_<FY>_SchH.txt. Every figure can be re-checked against these without network access.

5. Procedure

Run in this order. Scripts are in the project folder (hospital_data/ and hospital_data/irs/); the four that matter are quoted or described fully enough to rewrite from this section alone.

Step 1. Get the IRS index files. They list every e-filed return in the release with EIN, name, tax period, return type, and the object ID and batch zip that holds it.

mkdir -p irs && cd irs
curl -L -o index_2025.csv https://apps.irs.gov/pub/epostcard/990/xml/2025/index_2025.csv
curl -L -o index_2026.csv https://apps.irs.gov/pub/epostcard/990/xml/2026/index_2026.csv

Columns: RETURN_ID, FILING_TYPE, EIN, TAX_PERIOD, SUB_DATE, TAXPAYER_NAME, RETURN_TYPE, DLN, OBJECT_ID, XML_BATCH_ID. At this version: 748,906 rows for 2025 and 436,239 for 2026; RETURN_TYPE is 990, 990EZ, 990PF, or 990T. Only 990 returns carry Schedule H. Check that a few known hospital EINs appear (Northwestern 364724966, Rush 362174823) before continuing.

Step 2. Download every batch and keep only Schedule H filings. fetch_all.sh, quoted in full:

#!/bin/bash
set -u
cd irs; mkdir -p xml logs
B2025="01A 02A 03A 04A 05A 05B 06A 07A 08A 09A 10A 11A 11B 11C 11D 12A"
B2026="01A 02A 03A 04A 05A 05B 06A 07A 08A"
for y in 2025 2026; do
  if [ "$y" = 2025 ]; then L="$B2025"; else L="$B2026"; fi
  for b in $L; do
    z="${y}_TEOS_XML_${b}.zip"
    [ -f "done_${z}" ] && continue
    curl -sS -m 900 -L -o "$z" "https://apps.irs.gov/pub/epostcard/990/xml/$y/$z" || { echo "DL FAIL $z"; continue; }
    rm -rf tmp && mkdir tmp
    unzip -q "$z" -d tmp || { echo "UNZIP FAIL $z"; continue; }
    find tmp -name '*.xml' -print0 | xargs -0 grep -l -m1 "IRS990ScheduleH" > "logs/schH_${z}.txt"
    xargs -a "logs/schH_${z}.txt" -I{} mv {} xml/
    rm -rf tmp "$z"; touch "done_${z}"
  done
done

The batch list is read from the IRS download page (https://www.irs.gov/charities-non-profits/form-990-series-downloads) and will differ in later years. Each batch is 100 to 520 MB and takes 15 to 40 seconds to download and filter. Result at this version: 19,245 files in irs/xml/, named <object_id>_public.xml. The done_ marker files make the script resumable.

Step 3. Parse each filing. parse_xml.py exposes parse_file(path) which returns a dict with the header fields, Form 990 summary, a people list from Part VII (with Schedule O titles attached where Part VII says "See Schedule O" and Schedule O uses the "Part VII, Section A, \, Additional Positions Held" format), and a sch_h dict with the Part I lines, Part II total, Part III lines, FPG thresholds, facility count and list. It uses only xml.etree.ElementTree and the element names in Section 2. Run it on one file to see the structure: python3 parse_xml.py xml/202601349349304525_public.xml.

Step 4. Load into the master. load_all.py walks irs/xml/, parses each file, writes one row per figure to hospital_data.csv through add_rows.add(), records the object ID in loaded_object_ids.txt so a re-run skips it, and writes parse_errors.txt for any file that fails. Each row's source_doc carries the object ID and source_url gives both the IRS bulk path and the ProPublica page for the same filing. Expect about 50 filings per second once add_rows.py caches the next row ID (the first version re-read the whole file per append and ran at 2 per second; the cached version is what is in the folder now).

Step 5. Build the analysis tables. python3 build_systems.py reads the master, keeps the latest fiscal year per EIN, applies the CEO rule, applies ceo_overrides.csv, rolls filers into systems, applies rollup_overrides.csv, computes the flags and ratios, and writes filers_latest.csv, systems_rollup.csv, and ceo_unresolved.csv. Seconds to run; rerun after any change to an override file.

Step 6. Manual review of the large filers, per Section 6. This is the only step that is not a script.

Step 7. Validate per Section 7.

Optional: the page route for a single filing. Open https://projects.propublica.org/nonprofits/full_text/<object_id>/IRS990ScheduleH in a browser, save the page's full text, and run python3 save_and_parse.py <rawfile> <system> <entity> <ein> <fye> <object_id> <filed_date> <notes> <fpg_free> <fpg_discount>. parse_schh.py reads the Part I table from the rendered text by finding the line letters and the trailing percentage cell on each row; blank cells are treated as zero. Used for the first nine filers and for any filing the IRS zips omit.

6. Decision rules

These are the judgment calls. Each is a fixed rule in build_systems.py, and each has a hand-maintained override file for the cases the rule gets wrong. Change the rule only if a better one is found; otherwise add an override with its reason.

6.1 Latest filing per EIN. The releases contain two or three fiscal years per hospital (late and amended filings). The analysis takes the row with the highest fiscal_year_end per EIN. Earlier years stay in the master for the trend. Where the releases hold more than one return for the same EIN and fiscal year (38 such pairs at this version, 23 of them at the latest year), only one is used: the one filed last, by the filing date in source_doc, then the highest object ID. Every figure for that year, financial and personal, comes from that one return. (Added Oct 22, 2026, after an outside audit found the build could mix a title from one return, pay from another and financials from a third; at Parkview Medical Center, CO, line 7k differed between versions.)

6.2 CEO identification. Among the people in Part VII of a filing:

  1. Exclude anyone flagged former (FormerOfcrDirectorTrusteeInd), or whose title contains FORMER, FRMR, FMR, RETIRED or EMERIT (RETIRED and the abbreviations added Oct 21, 2026, after a retired chief executive listed as 'Senior Advisor (Retired Pres./CEO)' was taken as the CEO), or whose title contains vice president, VP, assistant, associate, regional, division, foundation, auxiliary, medical staff, coach, or "see Schedule" (when no Schedule O title was recovered), or that serves the chief executive rather than being one: advisor to, chief of staff, "to the President/CEO" (added Oct 22, 2026: Dana-Farber's 'Sr. Advisor to President & CEO' had been taken as its CEO over Benjamin Ebert). Also exclude anyone whose title carries a start date after the fiscal year end (AS OF, EFF, BEG, START followed by a month and year): that person was not yet chief executive in the filing year (added Oct 22, 2026: Weirton's 2024 return lists 'COO (CEO AS OF 1/2025)', and the year's CEO was the predecessor, John C. Frankovitch).
  2. If the title contains CEO or CHIEF EXEC(UTIVE): tier 0.
  3. Else if the title contains PRESIDENT: tier 1.
  4. Else if the title contains ADMINISTRATOR or EXECUTIVE DIRECTOR: tier 2.
  5. Take the lowest tier present. Within that tier, order is: a candidate whose title carries neither an end date nor an interim marker; then an interim or acting candidate (current at year end); then a candidate whose title carries an end date (termed, term, thru, through, until, til, end, ended, ending, resigned, departed). So a named successor, including one whose title carries a start date (eff., beg., as of, start), stands for the year; an interim stands over a departed executive; and if no current successor is listed, the departing CEO stands. (Rule widened Oct 21, 2026: the first version matched only TERMED, THRU <digit>, THROUGH <digit>, UNTIL <digit>, so titles like 'THRU JUNE 2024', '(TERM 12/24)' and 'END 04/24' were not recognized and the departing executive's final-year payout was taken over the successor; 40 rows changed.) Among what remains, the highest total pay (columns D + E + F). A person listed on more than one Part VII line (one line per role) is counted once, taking the maximum single entry per column rather than the sum, because filers repeat the same totals on each line. Confidence is recorded as ceo_title, president_title, or administrator_title, with _multiple appended when more than one person matched that tier. No match: unresolved.

Why this order: CEO is unambiguous. President is usually the same person at a single hospital but is the wrong person at a university or a system parent, so it is a lower tier and is reviewed. The exclusion list exists because "Regional President" and "Vice President" match PRESIDENT; "Head Coach, Football" was the top-paid person at the University of Miami; and "See Schedule O" tells you nothing.

6.3 ceo_overrides.csv. Columns: ein, fye, ceo_name_override, reason, resolved_by, date. A row with a name forces that person (the name must match the filing's PersonNm exactly, including merged title text such as DONNA HYLAND PRESIDENTCEO); a row with an empty name excludes the filer from any CEO-based figure (manual_excluded). The reason column is mandatory and should say what in the filing supports the decision. Three overrides rest on public knowledge rather than filing text (Mass General Brigham, Houston Methodist, UHS Binghamton); each now carries a dated public citation in its reason column (added Oct 30, 2026). The PeaceHealth override has its citation too (Oct 30, 2026). An override with an empty name excludes the filer from every CEO-based figure; this is used when the only chief-executive titles on a filing belong to a parent or regional organization's executives (whose pay is the parent's, not the hospital's), when a successor's reported pay is for a prior role elsewhere, or when the filing covers a closure or wind-down year. Where the filing also lists the hospital's own chief executive under a CEO title, the override names that person instead (McLaren Greater Lansing; Piedmont Atlanta, Piedmont Athens, Millcreek and Corry Memorial, added Oct 30, 2026). Seventy-three rows at this version (Oct 31, 2026). The independent audit of Oct 31, 2026 (Section 7, item 12) added 23: two name the chief executive at year end where the rule took someone else (Banner Health: Amy Perry, President & CEO, whom Wyoming Medical Center's return shows succeeding Peter Fine, 'CEO (THROUGH JUNE 30, 2024)'; Brattleboro Memorial: Christopher Dougherty, President & CEO, whose acting co-CEO successors were appointed after the year ended); four exclude a chief executive who is not the hospital's own (Alice Hyde and Elizabethtown, whose 'MG PRESIDENT & CEO' heads the UVM Health Network Medical Group; UCSF Health Community Hospitals, set against UCSF Health's system CEO, paid by UCSF; St Mark's Medical Center, a wind-down year after its Oct 12, 2023 closure); and seventeen exclude filers whose chief executive at year end reports $0 in Part VII, where load_all.py had dropped the $0 successor and the departing chief executive was taken (Section 8). These seventeen are excluded rather than named because the successor is not in the master; each reason names the successor and when they took over. resolved_by is independent_audit_2026-10-31 on all 23. Each row is keyed to a fiscal year, so check after every rebuild that each row's fye matches the filer's latest year in filers_latest.csv; on Oct 30, 2026 the Emory row had gone stale and was renewed. Two rows are data checks rather than CEO decisions: filers whose Schedule H line 7a cannot be reconciled with their own Form 990 totals (Cdt Gmsp, PR, at 97 percent of expenses; UCF Academic Health, whose charity care exceeds its total expenses) are shown as filed and excluded from CEO-based figures.

Parent-executive review (Oct 30, 2026). Every ranked row in the top 100 by CEO pay, and every headline row, where the CEO's column E (related organizations) is more than 80 percent of the total was read. Kept where the executive is the system's own chief executive on a row that holds that system's hospitals (Advocate, Montefiore, Kaiser, AdventHealth Sunbelt, RWJBarnabas, MemorialCare, OhioHealth, Northwell, CommonSpirit Washington, Froedtert, Wellstar Kennestone, Sanford, University Hospitals, Aspirus, Norton, Adventist Health, Prisma, Excela, Luminis, Covenant Parkwest, WellSpan York, ProMedica Toledo). Excluded where a parent or regional executive is set against one hospital: St Luke's Maumee (McLaren), Baylor Institute for Rehabilitation (Baylor Scott & White), Hartford Hospital (Hartford HealthCare), Carroll Hospital Center (LifeBridge); Piedmont Atlanta and Athens set to their own CEOs. Six headline rows with mostly related-organization pay were reviewed. Five are kept (ProHealth, Medical Center of Central Georgia, Tarzana, USC Verdugo Hills, Holyoke): ProHealth's CEO is the system's over both its hospitals; Tarzana, USC Verdugo Hills and Holyoke report their own chief executive, paid through the parent; the Medical Center of Central Georgia reports Navicent's own president, paid through Atrium Health. Millcreek is changed: its CEO-titled entry, John M Ferretti DO, is LECOM's chief executive (titled 'CHAIR, LECOM CEO' on Corry Memorial's filing), and the filing lists the hospital's own President/CEO, Mary L Eckert, paid by the hospital; with her pay ($723,538 against $796,878) Millcreek leaves the headline list. Look up the EIN in filers_latest.csv; do not type it from memory (two of the first drafts had wrong EINs).

6.4 System roll-up, and what a row is. Filers whose identified CEO has the same normalized name are one system. The name is normalized to uppercase letters, credentials (MD, DO, PhD, RN, FACHE and similar) are dropped, and the words are sorted, so 'Reed MD Monica' and 'REED MONICA', or 'Wandersleben Jennifer' and 'JENNIFER WANDERSLEBEN', are the same key; single-letter words (middle initials) are dropped too, so 'KEVIN B MAHONEY' and 'MAHONEY KEVIN' match (middle initials added Oct 30, 2026; this joined Princeton HealthCare to Penn Medicine, UVA Culpeper to UVA Prince William, and Berkshire Medical Center to Fairview Hospital and North Adams, and would have joined HealthPartners RC to Hutchinson Health through Nathan Pulscher, who moved between them in 2024 and is split by override) (credentials and word order added Oct 22, 2026; this joined Adventist GlenOaks, Schuyler/Cayuga and the AdventHealth Heartland filers, and one move between hospitals that it would have joined, Hunter Nostrant at MyMichigan, is split by override). A row in systems_rollup.csv is therefore a filing, or a set of filings, with the chief executive those filings report; it is not always the whole corporate system. Where a system's hospitals file under their own presidents, each hospital is its own row under its own president and the parent's executives are attributed to no hospital. A shared-officer test (two or more non-clinical Part VII officers in common between filers in different rows) finds 129 such clusters at this version, Providence, Piedmont, Indiana University Health, Bon Secours, Sutter, Novant and Orlando Health among them; assembling them is the main improvement planned for the next edition and must be done filer by filer with the evidence recorded, because the same test also pairs unrelated organizations through common names and joint-venture board seats. Members' revenue, expenses, net assets, employees, charity care, community benefit, bad debt, Medicare shortfall, and facility count are summed; the CEO is counted once; the lead filer is the member with the highest revenue. Known failure modes: two unrelated systems with a CEO of the same name merge (none found above $1B); a system whose subsidiaries list regional presidents does not merge with its parent (handled by exclusions in 6.3 and merges in 6.5); a person who moved between systems during the period can merge them (CHOC and Rady were first split on this reasoning, then joined on Oct 21, 2026 once the filings showed CHOC had become part of Rady Children's Health); the same person spelled differently on different filings fails to merge (nine such joins at this version: Mount Sinai alone was five rows because Brendan Carr is spelled three ways).

6.5 rollup_overrides.csv. Columns: ein, fye, action, system_key_override, reason, date. Actions: split (put this filer under the given key instead of its CEO's), merge and join (same mechanics, used when a parent or a subsidiary should join a named system), parent (join the named system and supply its CEO; used for a parent that files no Schedule H and was loaded for Part VII only), drop_duplicate (exclude the filer because another filer already contains it). The system's lead filer for naming is the highest-revenue member; the CEO comes from the parent member when one is flagged, else from the lead. Sixty-one rows at this version (Nov 1, 2026). The independent audit of Oct 31, 2026 added four: The Medical Center of Peach County joins the Navicent / Medical Center of Central Georgia row (join, key DELVECCHIO FINLEY; Navicent Health controls it per the parent's Part III and Schedule R); St Nicholas and St Clare Memorial join St Vincent and St Mary's under Robert Erickson (join, key ROBERT ERICKSON; HSHS's Aug 23, 2024 release names him Wisconsin Market President and CEO over all four hospitals and John Wagner, their Part VII CEO, Market COO); Winner Regional (SD) is split from Labette County (KS) under its own key (split, BRIAN WILLIAMS WRHC), because their chief executives are two different people named Brian Williams. The existing Wyoming Medical Center row was re-keyed from PETER S FINE to AMY PERRY so that it stays joined to Banner after Banner's CEO override; a join keyed to a chief executive's name must be re-keyed whenever an override changes that chief executive. The headline scope check of Nov 1, 2026 (Section 7, item 13) added three: HSHS St Joseph's Hospital, Highland, and HSHS Holy Family, Greenville, join Chris Klay's row (key CHRIS KLAY), because HSHS's July 23, 2024 release names Klay President and CEO of the four-hospital Southern Illinois Market and Michael Janis, the CEO-titled officer on their returns, the market's COO; and Pawnee Valley Community Hospital joins Hays Medical Center (key EDWARD HERRMAN), which manages it and lists it on its own Schedule H. Ten assemble Parkview Health (IN), whose parent filer alone had shown $84,340 of charity care on $1.2B of revenue; the evidence for joining its hospitals, which list local presidents rather than the system CEO in Part VII, is that every one of them lists the Parkview Health CFO in Part VII. When a system's hospitals file under local presidents, a shared officer in Part VII is the filing-based test for membership. Six assemble Advocate Health: the NC parent (EIN 88-4157429, loaded from batch 2025_11B into irs/xml_nonH/) plus Advocate Health and Hospitals, North Side, Condell, Sherman, and the Aurora (WI) group return. Atrium Health is a North Carolina public authority and files no Form 990, so the system's Southeast charity care is not in the data.

6.5a context_notes.csv. Columns: ein, note, source, source_url, date_added. Background that a reader needs and the filing cannot show, such as a merger after the fiscal year. build_site.py prints each note as a "Context:" line near the top of the page of whichever system contains that EIN; {source} in the note becomes a link to source_url labeled with source (publication and date), and [text](ein:XX-XXXXXXX) becomes a link to the page of the system containing that EIN. Notes are context only: build_systems.py does not read the file and no note changes a figure, ranking or flag. Two rows at this version (Oct 30, 2026): FHN (Freeport Memorial Hospital, EIN 36-2181997, Freeport IL) became part of Mercyhealth on Jan. 1, 2026 (Rock River Current, Dec. 31, 2025), after both systems' filed fiscal years, so FHN's page notes the merger and Mercyhealth's page (Mercy Health System Corporation, EIN 39-0816848) notes that FHN is not in its figures.

6.6 Group-return deduplication rule. Within a system, if a filer's name with "Group Return" or "Group" removed equals another member's name, and that other member is the group return, the non-group filer is dropped as contained. This removes the double-count of Cleveland Clinic Foundation inside its group return, and the same at OhioHealth. Tanner Medical Center is not deduplicated: its parent filing and its group return list disjoint facilities (Carrollton; Higgins and Villa Rica), so both are kept and summed. Ashtabula likewise.

6.7 Flags on every system.

  • childrens_hospital: yes if every member's name contains CHILDREN, PEDIATRIC, SHRINERS, or ST JUDE; mixed if some do. Children's hospitals are reported as a separate category because Medicaid and CHIP cover most low-income children, leaving little uncompensated care to classify as charity; six of the nine systems where the CEO out-earns charity care are children's hospitals, which is a feature of the category, not of those hospitals.
  • zero_charity_care_flag: on a filer, yes if 7a net is zero but 7k is positive. On a system (changed Oct 22, 2026), yes if the system's summed 7a is zero with 7k positive, or if flagged members hold more than 10 percent of the system's revenue; the flagged members are listed in zero_7a_members. A small member at zero no longer bars a system with positive charity care (Navicent Health Baldwin, 5.5 percent of its system's revenue, whose line 7a cost is fully offset by revenue). Filer-level cases: These are classification cases (Grady, Bethesda, McLeod, UMC New Orleans, Adena, Mount Nittany, Variety Children's), not zero-charity hospitals, and are excluded from rankings until each filing's Part VI narrative is read to find where charity care was booked.
  • publishable: yes only if the CEO confidence is ceo_title, ceo_title_multiple, or manual_override, charity care is positive, and members flagged zero hold no more than 10 percent of the system's revenue. Nothing with publishable = no goes in a ranking.

6.8 Computed ratios. charity_care_pct_computed = system charity care net divided by system total expenses, times 100. ceo_pay_pct_of_charity_care = CEO total compensation (D + E + F) divided by system charity care net, times 100. Both are reported alongside the inputs, never alone.

7. Validation

Run all of these before using any figure. Each has been run at this version with the result shown.

  1. Row integrity. row_id values in the master are contiguous from 1 with no gaps. Result: 364,682 rows, contiguous. Script: load the CSV, assert ids == range(1, n+1).
  2. Parser against an independent route. For filings pulled by both the browser page route and the bulk XML route, every Schedule H figure must match to the dollar. Result: UChicago, Loyola, Holy Cross, 133 of 133 values matched. If the IRS schema changes in a later year, redo this on at least three filings before loading.
  3. Parser against the hospital's own audit. Northwestern's audited statements report cost of charity care of $122,632 thousand for FY2025; Schedule H line 7a gross is $130,076,665 and net $114,908,722. The audit figure sits between gross and net, consistent with a different offset treatment; the Illinois AG filing's $114,908,722 equals Schedule H net exactly. Use the AG or Schedule H figure, not the audit, when comparing hospitals.
  4. Parser against the state filing. Northwestern's FY2025 Illinois AG community benefit report matches Schedule H line by line: charity care 114,908,722; education 102,305,423; research 72,736,172; subsidized services 61,602,066; bad debt 60,235,509; community building 6,218,881. Any future parse change should preserve this.
  5. Parse errors. parse_errors.txt must be empty. Result: empty.
  6. Override integrity. Every EIN in both override files must exist in filers_latest.csv. Result: all present.
  7. Spot checks against ProPublica summary pages. Total revenue, expenses, and net assets for a sample of filers agree with the ProPublica organization page for the same object ID. Done for the nine Chicago filers.
  8. Independent pay check against CEO Paycheck (run 18 Oct; snapshot of their 200 visible rows saved as ceopaycheck_snapshot_2026-10-18.csv, results in ceopaycheck_crosscheck.csv). Of 129 EINs present in both, 121 agree on the CEO and the pay figure to the dollar. The eight differences are each explained: two where they count an officer the filing marks Former (Intermountain's Allen, Seattle Children's Sperring) and we exclude; two where they count a departing CEO with an end date in the title (Texas Children's Wallace, Centra's Carrier) and we take the named current successor; one where both choices are defensible (UChicago: they take the highest-paid officer, Mark Anderson, the medical dean; we take Thomas Jackiewicz, President of the health system); and three small systems where the parent's CEO does not appear on the hospital filing. The 71 CEO Paycheck rows absent from our table are parent EINs whose hospitals file under subsidiary EINs. The check found two real errors on our side, both fixed: a person listed twice in Part VII had their pay summed (now the maximum single entry per column is taken; this affected Mount Sinai NY and Rochester Regional), and BJC's group return resolved to a physician rather than Richard Liekweg, whose name field carried a prefix.
  9. Outside audit before publication (Oct 21, 2026; fixes Oct 22). An independent read-only audit of the draft reproduced its headline of 75 and found: Weirton's CEO wrong (start date after year end); Navicent excluded by one zero-7a member; Grant Memorial joined to Uniontown because Uniontown's CEO sits on Grant's board ex officio (override excludes her as Grant's CEO); three name-format splits (GlenOaks, Schuyler, AdventHealth Heartland); duplicate returns per EIN and year mixed in the build; cancer centers in the large-system median; and the word "paid" for a figure that includes deferred and other compensation. A review of the audit found one more wrong CEO (Dana-Farber). All are fixed by the rules above, one CEO override (Grant), one rollup override (Nostrant), and two headline exclusions (Memorial Sloan Kettering, Dana-Farber, also now out of the large-system medians). Headline after fixes: 77 of 672. Still open: 7 President-only rows that would exceed 100 percent if the president is the chief executive; the Mass General Brigham override needs a dated citation.
  10. Full-table audit (Oct 21, 2026). Eight checks run across every ranked row: shared non-clinical officers across rows (129 clusters, recorded as a limitation); the same person split by spelling (nine joins); parent or regional executives on subsidiary returns (title referencing another organization, or identical pay on a larger filer with an executive title; 11 exclusions and 8 joins); end-dated and interim titles (rule widened); line 7a small next to lines 7c and 7g (64 narratives read, none reports charity care elsewhere); mixed year-ends (12 systems, documented); expenses far from revenue (two closures excluded; foundations and joint-venture partners kept out of rate-based lists); filed versus computed percentage (24 gaps, explained); and 50 filings re-parsed from XML with no mismatches.
  11. Sanity distribution. Median charity care percent across all filers should be near 1 percent and no filer should exceed about 25 percent (Shriners, which is almost entirely charity, is the ceiling at 21 percent). A value far outside that range is a parse error, not a finding. Result at Oct 30, 2026: the filers above 25 percent are foundations, a behavioral-health provider, joint-venture partners and wind-down entities whose Form 990 expenses are not hospital operations (Methodist Healthcare Ministries of South Texas 57 percent, St David's Foundation 40 percent and The Rapides Foundation 26 percent are ranked, and Centerstone of Florida, 37 percent, was until it was excluded from CEO figures on Oct 31, 2026 (item 12); they are kept out of rate-based lists by the operating-ratio test but remain in the CEO-pay ranking), plus two filings whose Schedule H cannot be reconciled with their Form 990 (Cdt Gmsp, UCF Academic Health), excluded by override (Section 6.3). None of them is on the headline list.

  12. Second independent audit before publication (Oct 31, 2026). A second independent audit of the draft checked the chief executive taken for each ranked row against the year-end date and the organization the title refers to, and the roll-up joins against each pair of filings. Findings and fixes: two rows had a chief executive who had left or not yet taken over (Banner Health, Brattleboro Memorial; named by override); four set an executive who is not the hospital's own against its charity care (Alice Hyde and Elizabethtown, UCSF Health Community Hospitals) or covered a wind-down year (St Mark's Medical Center); seventeen had taken a departing chief executive because the successor reports $0 in Part VII and load_all.py drops $0 officers (Section 8; all seventeen excluded); Peach County belongs in the Navicent row; St Nicholas and St Clare belong with St Vincent and St Mary's under HSHS's Wisconsin market CEO; and two different people named Brian Williams had joined Labette County (KS) and Winner Regional (SD). Overrides in 6.3 and 6.5. Results: headline 76 of 668 before, 74 of 656 after (Navicent at 85.4 percent and the St Nicholas row leave the list; Brattleboro stays at 120.4 percent; Labette alone is 149.7 percent); ranked systems 1,052 before, 1,034 after; systems 1,626 before, 1,629 after. The large-system medians (229 non-children's general systems over $1B) are unchanged at 1.23 percent charity care and 7.5 percent CEO pay to charity care. Check after every rebuild that no join key has been orphaned by a CEO override (a key that matches no member's own chief executive).

  13. Headline scope check before publication (Nov 1, 2026). Every headline row (74 after the Oct 31 audit) was checked against one fixed checklist: (a) who was chief executive at fiscal year end; (b) what that person's title covers (this hospital, or a market, region, medical group or parent); (c) whether the row holds every hospital that person runs, and nothing else. Evidence came from the filings and from the hospitals' and systems' own leadership announcements, bios and local news. One line per row, with the verdict, evidence and source URLs, is in headline_scope_check_2026-11.csv. Result: 72 ok, 2 fix, 0 unclear. Fixes: the HSHS Southern Illinois row under Michael Janis (153.1 percent) joins Chris Klay's market row (71.6 percent) and leaves the headline list; Pawnee Valley joins Hays Medical Center (355.6 to 328.1 percent, still listed). Seven of the ok rows rest on thinner evidence (the filing plus directory or undated pages): Thorek and Artesia (filing only, consistent across years), St Croix, Fairchild, Anaconda, Auburn and Welia; none showed a sign of a different chief executive. Tarzana (100.5 percent) is kept under its own chief executive although Providence also has a service-area chief executive over it, because the sister hospitals file inside another row's group return. Results: headline 73 of 655; 1,033 ranked; 1,627 systems. Large-system medians unchanged.

8. Known limits and gaps

  • The IRS zips and index disagree at the edges. Some filings listed in the index are absent from the zips (Rush and Mount Sinai FY2025 among them). The page route fills these. Check any specific hospital you intend to name.
  • Group returns report system totals. Northwestern's 1.07 percent covers nine hospitals; Northwestern Memorial Hospital alone is not separable from the IRS filing. Facility-level charity care exists in state filings (Illinois AG Form AG-CBP-1) but the posted PDFs carry those tables as images, not text, and were not captured at this version.
  • Systems split across EINs are reassembled by the CEO-name rule and overrides. The rule was verified by hand for every multi-filer system above $1B. Below $1B it has not been reviewed.
  • 198 filers have no identifiable CEO by the rule, all below $1B in revenue. They are in the dataset with ceo_confidence = unresolved and excluded from any CEO-based figure.
  • Fiscal years differ (June, August, September, December year ends). "Latest filing" spans calendar 2024 to mid-2025. Twelve multi-filer systems sum members with different year-ends (Northwell, University of Maryland, Gundersen, Aspirus among them); their figures mix two fiscal years.
  • Compensation is the Form 990 figure, which includes deferred compensation when paid out. A departing executive's final year is often inflated (Northwestern's Harrison, $10.83M in FY2025, marked Former). The multi-year rows in the master allow a three-year average where needed.
  • Charity care at cost depends on each hospital's cost-to-charge method, which the IRS lets them choose. The filed percentage is comparable across hospitals only to the extent their cost accounting is; this is an industry-wide limitation shared by every study using Schedule H.
  • Children's hospitals structurally report low charity care and are reported as their own category.
  • Zero charity care on 7a with large 7k means a classification choice, not zero charity. Ten large filers; excluded until read.
  • The Chicago-only sources (audit, Illinois AG, Lown, IHPA, IHA) exist only for the nine original filers. IHPA is an advocacy source of unknown funding. Lown's fair-share figures are estimates and are the only estimated numbers in the dataset.
  • Officers who report $0 are not in the master. load_all.py line 47, if tot == 0: continue, drops officers who report $0, so a successor reporting $0 is invisible to CEO selection; the affected filers found by the Oct 31 audit carry overrides (Section 6.3). Other filers where a new chief executive took over late in the year without pay on that return may still show the departing chief executive; the end-date and interim rules in 6.2 limit but do not remove this.
  • Public and for-profit hospitals are absent because they do not file Form 990. Comparisons are among nonprofits only.

9. Results as of this version

The table below is from the first full draft and predates the pre-publication checks in Section 7, items 9 to 13. After the Oct 22, 2026 fixes (item 9): headline 77 of 672 general hospital systems over $100 million; 1,057 ranked systems; 229 non-children's general systems over $1B with median charity care 1.23 percent and median CEO compensation 7.6 percent of charity care. After the Oct 30, 2026 parent-executive review, data checks and middle-initial rule (Section 6.3, 6.4): headline 76 of 668; 1,052 ranked systems; 1,626 systems; median CEO compensation 7.5 percent of charity care for the same large-system group. After the Oct 31, 2026 second audit (Section 7, item 12): headline 74 of 656; 1,034 ranked systems; 1,629 systems. After the Nov 1, 2026 headline scope check (item 13), the published figures: headline 73 of 655; 1,033 ranked systems; 1,627 systems; large-system medians unchanged throughout.

The median nonprofit hospital system spends about one percent of expenses on charity care, and that figure is flat across every size of system. All figures from the hospitals' own filings, latest fiscal year (mostly 2024 and 2025), after the roll-up and review in Section 6.

Measure Value
Distinct filers (EINs) 2,305
Systems after roll-up 1,653
Charity care as % of expenses, all systems with a figure, median 1.03
Systems with revenue of $1B or more (excluding zero-7a classification cases) 306
Of those: charity care median 1.20
Of those: under 1 percent 41%
Of those: under 2 percent 73%
Publishable $1B+ systems (confident CEO, positive charity care), non-children's 224
Non-children's: charity care median / CEO pay median / CEO pay as % of charity care median 1.23% / $2.43M / 8.4%
Non-children's: CEO paid more than 50% of system charity care 10
Non-children's: CEO paid more than 100% 3 (Parkview, IN; Eisenhower, CA; Northern Light, ME)
Publishable children's systems 25
Children's: charity care median / CEO pay median / ratio median 0.48% / $2.71M / 27.6%
Children's: ratio over 100% 6

Size does not predict charity care. Median charity care by revenue band: $10B+ 1.11%, $3B to $10B 1.13%, $1B to $3B 1.28%, $500M to $1B 1.23%, $200M to $500M 1.24%, $50M to $200M 0.63%. Dollar-weighted, the $10B+ tier is 1.23% against 1.60% for the sector. CEO pay rises steeply with size (median $11.3M at $10B+, $0.5M at $50M to $200M) while the charity rate does not, so the ratio falls with size and is a small-hospital statistic; it is used case by case, not as a sector ranking.

Highest-paid CEOs among publishable systems: Grossman, NYU Langone, $30.3M (charity care 1.44%, ratio 21%); Corwin, NewYork-Presbyterian, $26.3M (1.11%, 24%); Ozuah, Montefiore, $16.7M (1.05%, 27%); Balser, Vanderbilt, $16.5M (2.82%, 7%); Lassiter, Dignity/CommonSpirit, $14.2M (1.68%, 6%). Advocate Health, assembled from the NC parent and five Midwest filers: $17.8B revenue, $114.9M charity care, 0.66 percent of expenses, the lowest rate of any system above $10B. The parent lists two chief executives for 2024, Eugene Woods (CEO, $25.78M) and James Skogsbergh (Co-CEO, $30.86M, retired at year end); Woods is taken for the ratio (22.4 percent), and the report should show both, since together they were paid 49 percent of what the system's roughly 30 Midwest hospitals gave in charity care.

Chicago in the national frame. Northwestern's nine-hospital group: 1.07% charity care, CEO Chrisman $8.17M, ratio 7.1%, at the national median on every measure. Advocate's six Illinois hospitals: 0.58%. UChicago Medical Center: 0.73%. Endeavor: 1.11%. Rush: 1.20%. Loyola: 2.24%. The three safety-net comparison hospitals: Mount Sinai 9.81%, Holy Cross 9.51%, St. Bernard 3.96%.

State medians range from 0.28% (North Dakota) and 0.29% (Maine) to above 1.2% (Alabama, Arizona); Illinois is 0.94% across 99 filers.

10. Annual re-run

The whole point of the pipeline is that next year costs almost nothing. What changes:

  1. Batch list. Read the new year's batch names off the IRS download page and add them to fetch_all.sh. Keep the prior years' done_ markers so only new batches download.
  2. Schema. Each tax year has a schema version (2024v5.1 at this version). Element names in Section 2 have been stable since 2020; check three filings against the ProPublica page route (Section 3) before loading, per Validation step 2.
  3. Fix the $0-officer filter first. Before the next load, remove or change load_all.py line 47 (if tot == 0: continue) so officers who report $0 are kept (the compensation rows can be written as 0, or the person recorded with a title-only row), and make sure build_systems.py can select a $0 chief executive. Then re-check the seventeen $0-successor overrides added Oct 31, 2026 (Section 6.3): with the successor in the master, each can name the successor or be dropped.
  4. Load. load_all.py skips object IDs already in loaded_object_ids.txt, so it appends only the new filings to the master. Back up the master first.
  5. Rebuild. build_systems.py picks the new latest year per EIN automatically. Prior years remain in the master for the trend.
  6. Review. Re-run the large-filer review (Section 6). CEO changes, mergers, and new group returns are where last year's overrides go stale; each override row carries its fiscal year, so a new year gets a fresh decision rather than inheriting one.
  7. Publish the same table on the same date, with last year's column beside it.

What an LLM or analyst needs to have read to do this: this tab, the four scripts named in Section 5, and the two override files. Nothing else in the project is required.

11. The site

The site is generated from systems_rollup.csv by one script, site/build_site.py, and rebuilt in seconds whenever the table changes. There is no database, no server code, and nothing to log into.

Inputs: irs/systems_rollup.csv and irs/filers_latest.csv. Two constants at the top of the script: REPORT_TITLE (placeholder at this version: Nonprofit Hospital CEO Pay Report) and PUBLISHER (Everything Is Broken).

Outputs, in site/out/: index.html (the ranked top 100 with search, sort, and a children's hospital toggle); all.html (every ranked system); system/<slug>.html, one per system with a page for every system that has either a pay figure or a charity care figure (1,614 at this version), each showing the three numbers, the filing link, a community benefit versus charity care bar, and the member filings; state/<XX>.html for every state, the District of Columbia and Puerto Rico where a system has a hospital (a system is listed in every state where a member filing reports a hospital facility in Schedule H Part V, with its filing state noted where different; filers with no US hospital are on state/outside-us.html), each opening with a generated sentence giving the state's headline count, its median charity care against the national median, and its highest-paid chief executive; states.html; methodology.html; data.html. All pages share one stylesheet written into each file, so the site is plain HTML with no build step beyond the script.

What goes on a system page: CEO total compensation (Part VII D+E+F), charity care as percent of expenses (Schedule H 7a column f), and CEO pay as a percent of charity care, each labelled with its form line; the filing's revenue, expenses, net assets, employees, facility count, bad debt and Medicare shortfall; a link to the exact filing on ProPublica by EIN and object ID; the CEO identification confidence as recorded; and for multi-filer systems a table of every member filing with its own link. A callout sentence appears automatically where the ratio is 25 percent or more. Systems not in the ranking (unidentifiable CEO, or zero on line 7a) get a page that says why.

Ranking rule: publishable = yes and a pay figure, sorted by CEO pay descending. Children's hospitals are ranked but shown in grey with a toggle, for the reason in Section 6.7.

Deploy: copy site/out/ to any static host (Cloudflare Pages, Netlify, GitHub Pages, or an S3 bucket) and point the domain at it. Rerun build_site.py after any change to the tables or the override files, and redeploy. For the annual edition, change EDITION, rebuild, and keep the prior year's out/ under a dated path so old links resolve.

Not yet built: per-system share images for social posts; the data download links (the files exist, the hosting does not); a prior-year column on each row once a second edition exists.