Nursing home ownership crosswalk: 2,192 Medicare- and Medicaid-certified facilities in California, Arizona, Florida and Washington resolved to their ultimate controlling owners (2026)

Published Aug 10, 2026 on Dryad. https://doi.org/10.5061/dryad.nvx0k6f7j

Data files

Aug 10, 2026 version files 9.26 MB

00_read_me.csv

735 B
01_ownership_crosswalk.csv

3.95 MB
02_cms_quality_reference.csv

908.39 KB
03_methodology_audit_trail.csv

4.32 MB
04_confidence_summary.csv

2.85 KB
05_owner_label_normalization_log.csv

12.41 KB
06_codebook.csv

50.11 KB
README.md

23.69 KB

Abstract

This dataset resolves each Medicare- or Medicaid-certified nursing home in four US states to the party that ultimately controls it, using only sources that are free to access. Federal law requires nursing facilities to disclose their owners, but in practice the disclosure often stops at a single-purpose entity created to hold one building. Reading the federal record alone, it is usually not possible to tell who sets a facility's budget or whether the facility belongs to a larger network. This crosswalk closes that gap for a complete state population: 2,192 facilities, 2,181 of them resolved to 759 distinct controlling owners, with 11,995 free public source citations attached row by row.

Coverage and structure. The dataset covers all 2,192 nursing homes in California (1,164), Florida (694), Washington (194), and Arizona (140) listed in the CMS Nursing Home Provider Information file (NH_ProviderInfo), from the nursing-homes theme of the CMS Provider Data Catalog, snapshot of May 27, 2026, retrieved June 9, 2026. By CMS certification type these are 2,061 dually Medicare- and Medicaid-certified facilities, 103 Medicare-only and 28 Medicaid-only; the first two groups are skilled nursing facilities in the Medicare sense and the 28 Medicaid-only facilities are nursing facilities, which is why their certification numbers carry a letter. It is therefore a complete state population rather than a sample, and the frame was verified as an exact set match against that file on all 2,192 certification numbers. Of the 2,192 facilities, 2,181 (99.5 percent) are resolved, naming 759 distinct controlling owners. The frame was also cross-checked against the July 1, 2026 provider release, which contains 2,192 four-state facilities as well but differs by one in each direction: one facility in the May frame had left the federal file by July, and one facility present in July had not yet entered it in May. Seven comma-separated files accompany the deposit. Four are keyed to the CMS Certification Number (CCN) and join on it directly: the ownership crosswalk (2,192 rows x 13 columns), a CMS quality and staffing reference passed through unmodified (2,192 x 67), a per-row methodology audit trail (2,192 x 10), and a log of every owner-label and citation change made after initial resolution (140 x 4). Three are documentation rather than facility-level data: an aggregate confidence summary (50 x 5), a variable-level codebook defining all 107 columns across all files (107 x 7), and a single-line pointer to the README.

Data values. Each facility carries a resolved owner, an ownership-type classification, a confidence level, a resolution-depth flag, and its own list of source URLs. Ownership types are: individually or family-owned multi-facility network, 1,400 facilities (63.9 percent); non-profit corporation, 318 (14.5 percent); publicly traded chain or REIT, 307 (14.0 percent); independent single facility, 99 (4.5 percent); and private equity backed, 68 (3.1 percent). Confidence is high on 1,781 rows, medium on 381 and low on 30, where "high" requires independent corroboration in addition to a cited source. Resolution depth records whether a controlling entity above the facility was identified (2,034 rows), whether the facility's own entity is the terminal owner (147), or whether resolution failed (11). Provenance consists of 11,995 individual source citations drawn from 1,594 distinct internet domains, a median of five per facility and a maximum of eleven, including CMS ownership and provider files, state business registries, SEC filings, state Attorney General transaction approvals, non-profit tax filings, court records and company disclosures.

Reuse potential. Because the crosswalk is keyed to the CCN in the same six-character zero-padded form CMS uses, it joins directly and without normalization to the federal files that share that key: quality and staffing measures, inspection and enforcement records, and Medicare cost reports, including the related-organization payment schedule on Form CMS-2540-10. That makes it possible to ask questions that previously required licensed commercial data, such as whether financial extraction tracks the ownership categories that federal disclosure policy is organized around. Two properties are intended to support extension rather than only citation: every source is a free public URL, so any single row can be independently checked or disputed, and the full per-row reasoning trail is published alongside the data rather than withheld. One loading note, because it is the single most common way to break a join of this kind: read the CCN column as text, not as a number. Twenty-eight of the 2,192 values contain a letter, such as 05A024, because that format encodes Medicaid-only certification, and 757 more carry a leading zero, so a numeric read empties the first group and silently drops the second. The README gives the one-line fix for Python and R.

Limitations that affect interpretation. All counts derived here are four-state counts and are therefore floors rather than national totals; the one exception is Number of Facilities in Chain, which is passed through from CMS and counts each chain nationally, so it must not be compared to a within-sample count without first verifying the overlap. The private equity share of 3.1 percent is a floor, since private equity ownership is systematically under-disclosed in federal data. The distinct-owner count of 759 is a slight over-count, because groups remain in which one owner label is a token subset of another, for example Generations Healthcare alongside Life Generations Healthcare, LLC, and these require case-by-case judgment rather than a rule; no count of such groups is asserted, because that number proved to depend on how a group is defined, and the direction of the error understates rather than overstates ownership concentration. Confidence is lowest where disclosure obligations are weakest, 99 percent high for publicly traded chains against 52 percent for independent single facilities, which is treated in the documentation as a finding rather than a defect. No external accuracy rate is claimed; a small comparison against commercial databases was withdrawn as too underpowered to support one, and a properly powered validation is future work. Ownership turns over continuously, so this is a dated snapshot and not a live register. Where a facility's real estate is held by a REIT and operations are leased to an operator under a triple-net master lease, the operator is recorded as the controlling party, because operating control determines staffing, budget and admissions while the landlord sets rent. The REIT landlord is therefore not recorded in a structured field, though it frequently appears in the per-row reasoning; the publicly traded chain and REIT category counts publicly traded owners and operators rather than landlords. Analyses of rent-driven financial pressure will need to add a landlord column, which is a planned extension.

Legal and ethical considerations. No subscription-gated commercial deal database was used to build or verify this dataset, because their licenses bar redistribution and would make an open release impossible; citations to PitchBook, Preqin and Irving Levin Associates were removed and each removal was verified not to be load-bearing before it was made. Ninety-nine of the 2,192 rows (4.5 percent) cite freely viewable commercial business-information aggregators, and every such domain is enumerated with its row count in the README so that a reuser applying a stricter source standard can exclude them. All source records are public. Where an individual is named, the dataset is recording a structural fact about the public disclosure record; it contains no personal information beyond what those public filings disclose, and it measures ownership and control only. It contains no measure of care quality attributable to any owner, and no inference about any named individual's or firm's conduct is supported by it.

Internal project name: Healthcare Ownership Architecture Resolution Dataset, v1.0.

This dataset resolves each Medicare- or Medicaid-certified nursing home in four US states to the party that ultimately controls it. Facilities are required to disclose ownership to the US federal government, but the disclosure commonly terminates at a limited-liability company created to hold a single building, so the federal file does not identify the party that sets a facility's operating budget and cannot be aggregated to the level of a controlling owner.

The dataset covers 2,192 facilities, which is every Medicare- or Medicaid-certified nursing home in California, Arizona, Florida and Washington listed in the CMS Nursing Home Provider Information file (NH_ProviderInfo), snapshot of May 27, 2026 (retrieved June 9, 2026), and is therefore a complete state population rather than a sample. The frame was verified as an exact set match against that file: all 2,192 certification numbers, and the state counts CA 1,164 / FL 694 / WA 194 / AZ 140. It was then cross-checked against the July 1, 2026 provider release, which also contains 2,192 four-state facilities but differs by one in each direction, since one facility in the May frame had left the federal file by July and one facility present in July had not yet entered it in May. Of the 2,192 facilities, 2,181 (99.5%) are resolved, naming 759 distinct controlling owners. (The Ultimate Parent Entity column holds 761 distinct values; two are sentinels rather than owners, UNRESOLVED on 11 rows and a STALE placeholder on 9 rows, so the owner count is 759 and not the column's cardinality.) Each facility carries an ownership-type classification, a confidence level, a resolution-depth flag, and its own source citations. There are 11,995 individual citations drawn from 1,594 distinct internet domains, a median of 5 per facility. Every citation is a publicly accessible URL rather than an extract from a licensed database.

Resolution used three stages. First, an exact join on the CMS Certification Number against two federal files covering 247,632 ownership records. Second, a constrained web research pass run independently over every facility, required to corroborate industry and geography before accepting any name match and to return "unresolved" rather than choose between candidates. This second pass was applied to the full population rather than only to gaps, so that each facility carried two independently sourced determinations that could be compared. Third, the two were reconciled: agreement supplies the corroborating signal the confidence rules require, and every disagreement or internally inconsistent owner group was read and adjudicated by hand rather than resolved by rule.

Description of the data and file structure

Six data files. All are UTF-8 encoded, comma-separated, with a single header row and no byte-order mark. They are flat tables with no nested structure.

File	Rows	Columns	Contents
`01_ownership_crosswalk.csv`	2,192	13	The primary file. One row per facility: identifiers, resolved controlling owner, ownership type, confidence level, source citations, resolution depth.
`02_cms_quality_reference.csv`	2,192	67	Quality, staffing, inspection and enforcement measures for the same facilities, passed through unmodified from the CMS provider file so they can be joined to resolved ownership.
`03_methodology_audit_trail.csv`	2,192	10	Per-row reasoning, disambiguation notes, structural detail, and the federal owner records each resolution was checked against.
`04_confidence_summary.csv`	50	5	Aggregate counts: confidence, resolution depth, ownership type and state, plus two cross-tabulations. Derived entirely from `01`; contains no new information.
`05_owner_label_normalization_log.csv`	140	4	Every change made to owner labels and citations after the initial resolution, one row per change, so this release is reproducible from the prior state.
`06_codebook.csv`	107	7	Variable-level definitions for every column in every file, including data type, meaning of blank cells, and non-blank counts. Read this first.

00_read_me.csv is a single-row pointer file. It names the dataset, directs the reader to this README, and repeats the one warning that matters most for reuse: the CMS Certification Number must be read as text, never as a number. It carries no figures of its own, deliberately, so that no count in this deposit can go stale in two places at once.

Relationships between files

All files share the column CMS Certification Number (CCN) as the join key. Every CCN appears exactly once per file, and the same 2,192 CCNs appear in 01, 02 and 03. 04 is an aggregate of 01. 05 and 06 are documentation.

The one thing most likely to break an analysis

CMS Certification Number (CCN) must be read as text, not as a number. The values in these files are correct: always six characters, zero-padded, matching the CMS convention exactly, so this deposit joins to any CMS provider or ownership file with no normalization required. The risk is downstream, in how your software reads the column. A numeric read breaks the join in two ways at once, and neither raises an error. Twenty-eight of the 2,192 values contain a letter, for example 05A024 and 50A261, and a numeric read converts those to empty values and drops the facilities entirely. It also discards the leading zero on the 757 purely numeric values that carry one, so 035003 becomes 35003, which no longer matches the CMS file. The result is a silent partial join: 1,435 of 2,192 rows match and the rest disappear without warning. All 28 are Medicaid-only facilities. The letter is not an anomaly; it encodes Medicaid-only certification, which is why the exclusion is structural rather than random. Losing them removes the facilities least likely to be privately owned, which biases ownership estimates toward private forms.

In Python: pd.read_csv(path, dtype={'CMS Certification Number (CCN)': str}). In R: readr::read_csv(path, col_types = readr::cols(.default = "c")).

Controlled vocabularies

Institutional Ownership Type, exactly five permitted values:

Value	Facilities
Individually/Family-Owned Multi-Facility Network	1,400
Non-Profit Corporation	318
Publicly Traded Chain / REIT*	307
Independent / Single Facility	99
Private Equity Backed	68

*REIT: Real Estate Investment Trust.

Confidence Level Final, three permitted values: high (1,781), medium (381), low (30). A value of high requires an independent corroborating signal and a cited authoritative source.

Resolution Depth, three permitted values: parent_identified (2,034), facility_is_terminal (147), unresolved (11).

State, four permitted values: CA (1,164), FL (694), WA (194), AZ (140).

Aggregating to the owner or network level

Filter to Resolution Depth == "parent_identified" before any owner-level or network-level aggregation. The 147 facility_is_terminal rows name no party above the facility because there genuinely is none; including them in a network rollup would create 147 spurious single-facility networks. The 11 unresolved rows name no owner at all.

Two kinds of number in this dataset, which are not comparable

Counts derived from these rows are four-state counts and are floors, never national totals. An owner operating outside these four states is invisible to them.

One column is different. Number of Facilities in Chain is a national count, passed through from CMS unmodified. Across the 144 chains present here, the CMS figure exceeds the number of rows in this dataset for 88 chains, equals it for 56, and is never lower. The Ensign Group is reported as 338 facilities of which 128 fall inside these four states; Genesis Healthcare as 187 of which 24 are here. That column must not be compared to a count derived from these rows without first checking the overlap.

Blank cells

Blank means "not recorded," never zero. 06_codebook.csv states what a blank means for every column individually. The largest blocks are Chain Name and Number of Facilities in Chain, blank on 538 rows where CMS records no chain affiliation, and Federal Direct/Indirect Owners, blank on 467 rows where CMS discloses no owner holding 5 percent or more.

Known limitations

Eleven facilities are unresolved. For every one of them the federal file itself reports that ownership data is not available, so there is no withheld answer that better method would recover. Resolving them requires California state licensing and business-registry records facility by facility. Seven of the eleven carry alphanumeric CCNs, so the unresolved rows concentrate in the same population as the identifier problem described above.
Four taxonomy gaps. No category cleanly fits employee-owned operators, government agencies, joint ventures or closely held investment companies. Those rows are typed by operating control, confidence-capped, and the reason is recorded per row in 03.
Chain labels record history, not control. Twenty-four facilities carry the CMS chain label "Genesis Healthcare" and represent three different arrangements. A bankruptcy court approved the sale of Genesis's 175 facilities to a NewGen-backed entity in January 2026 for approximately $991 million, and the federal chain label had not changed as of the July 2026 file. Where a pending change of control is known, it is recorded in the affected rows rather than substituted for the federal value.
Nine rows are flagged stale following a 2025 transaction in which Centre Partners exited Covenant Care.
One private-equity classification rests on an unverified entity link. CCN 555753 is typed Private Equity Backed on the basis that Vivra Specialty Partners and Vivra Specialty Care, LLC are the same firm. No primary filing in the assembled evidence establishes that link, and the row's Reasoning field says so. The row is therefore held at medium confidence. A reuser who disagrees can exclude it; doing so moves the private-equity share from 3.1% to 3.1%.
Owner-label groups remain unmerged where one label is a token subset of another, for example Generations Healthcare alongside Life Generations Healthcare, LLC. Some of these are genuine duplicates and some are genuinely distinct ownership groups, for example David Johnson alone versus David Johnson and Frank Johnson, so they were left as recorded rather than merged by rule. No count of these groups is asserted here, because the count is not well defined. It varies with how a group is delimited and with whether subset relations are chained transitively, so the same data yields materially different totals under different criteria. Rather than publish a figure that changes with the definition, the limitation is stated qualitatively. The consequence is that the distinct-owner count of 759 is a slight over-count and network concentration is correspondingly under-stated. The error direction is therefore conservative with respect to any claim about ownership concentration. The criterion actually applied is documented row by row in 05_owner_label_normalization_log.csv, so a reuser can impose a different criterion and derive their own count.
No external accuracy rate is claimed. A comparison against commercial databases was run on a small number of owners and then removed from this release, because sixteen owners is too small a sample to support an accuracy percentage. A properly powered validation is future work.
This dataset measures control, not care quality. Nothing in it supports an inference about the conduct or quality of care of any named owner. The measures in 02 are CMS's, unmodified, and are provided for joining only.
Snapshot. The facility frame is the May 27, 2026 CMS provider file; ownership research and cross-checks ran through July 2026. Ownership turns over continuously, so treat this as a dated snapshot rather than a live register.

Sharing/Access information

Data were derived from the following public sources:

CMS Skilled Nursing Facility All Owners file: https://data.cms.gov/provider-data/dataset/y2hd-n93e
CMS Nursing Home Provider Information file: https://data.cms.gov/provider-data/dataset/4pq5-n9py
State business registries, including the California Secretary of State and Florida Sunbiz
State Attorney General transaction approvals, including California Attorney General healthcare transaction notices
California Department of Health Care Access and Information facility records
US Securities and Exchange Commission filings via EDGAR
Company websites, press releases, and court and bankruptcy filings

Every citation used for every row is recorded in the Combined Source Summary column of 01_ownership_crosswalk.csv as a resolvable URL. Readers are encouraged to examine individual rows and to contact me with corrections or about potential collaboration.

A precise statement about commercial data sources

No subscription-gated commercial deal database was used, because their terms bar redistribution,
which is incompatible with an open release. An earlier working version carried PitchBook or Preqin
citations on 48 rows, and 7 further rows cited Irving Levin Associates, a subscription publisher of
seniors-housing deal data. All 55 were removed. Each was checked first to confirm the citation was
not load-bearing: every one of the 55 retains at least two independent freely accessible sources, so no
confidence level changed. Every removal is logged row by row in 05_owner_label_normalization_log.csv.

The Irving Levin removal is worth naming rather than burying, because this release criticizes prior work
for depending on licensed data, and Irving Levin is one of the two sources that prior work used.
Citing it while making that criticism would have been indefensible.

Being precise about what does remain. 99 of the 2,192 rows, 4.5 percent, cite a commercial
business-information aggregator whose pages are freely viewable on the open web. Counted by domain, with
overlap because some rows cite more than one: 42 rows cite bizprofile.net, 39 cite dnb.com, 9 cite
CorporationWiki, 5 Buzzfile, 3 Bloomberg company profiles, 2 Crunchbase, 2 Tracxn, 2 RocketReach, 1
Bizapedia, 1 ZoomInfo. None is load-bearing. Every one of the 99 rows retains at least two independent
non-aggregator citations, verified row by row, so removing all of them would push no row below two sources
and change no confidence level. They are named here rather than glossed as "public records" so that a
reuser applying a stricter source standard than mine can find and exclude them.

A note on what is deliberately not in that list. Several third-party sites republish CMS data itself,
including ProPublica's Nursing Home Inspect, nursinghomereport.org, nursinghomedatabase.com, pbj320.com
and npidb.org. Those are free republications of the federal record rather than proprietary
business-information products, so they are not counted above. The distinction is between a site that resells
its own research and a site that mirrors a public file.

What the CMS quality reference does and does not carry

02_cms_quality_reference.csv carries 67 of the 99 columns in the CMS provider file, passed through
unmodified so that resolved ownership can be joined to care and enforcement measures without leaving this
deposit. Included: all star ratings and their chain averages, case-mix-adjusted and reported staffing hours,
nursing and administrator turnover, inspection and deficiency results, and enforcement actions including
fines, payment denials and penalty counts.

Not carried, and why: six columns that already appear in 01_ownership_crosswalk.csv (address, city, state,
ZIP, chain name, facilities in chain); twelve CMS footnote and suppression flags; the five case-mix
expected-value columns used as inputs to the staffing adjustment, since the adjusted series itself is
retained; three reported staffing components (nurse aide, LPN and RN separately, where the reported totals
and the full adjusted series are retained); latitude, longitude and the geocoding footnote; SSA county code;
whether the provider resides in a hospital; and the file's processing date.

This subsetting is a convenience, not a substitute. The complete provider file is public and is cited
below, so any omitted column can be joined back on the CCN with no normalization. A reuser who needs the
case-mix expected values, the geographic coordinates, or the hospital-based flag should go to the CMS file
directly.

What kind of facility is in this dataset

The population is the 2,192 nursing homes in California, Arizona, Florida and Washington listed in the CMS
Nursing Home Provider Information file (NH_ProviderInfo), published under the nursing-homes theme of the
CMS Provider Data Catalog. CMS's own theme name is "Nursing homes including rehab services," and these
facilities provide skilled nursing and rehabilitation care.

By CMS certification type the 2,192 break down as 2,061 dually Medicare- and Medicaid-certified, 103
Medicare-only, and 28 Medicaid-only. The first two groups are skilled nursing facilities in the Medicare
sense. The 28 Medicaid-only facilities are nursing facilities rather than skilled nursing facilities, and
that distinction is visible in the identifier: a letter in the certification number encodes Medicaid-only
certification. This is the reason the alphanumeric-CCN coercion described above is a structural rather
than a random exclusion. 95 of the 2,192 facilities reside within a hospital.

The phrase "skilled nursing facility" is therefore accurate for 2,164 of the 2,192 rows and slightly
imprecise for 28. Where precision matters, this documentation says nursing home.

OpCo and PropCo: who is recorded when a REIT owns the building

Where a facility's real estate is held by a REIT and operations are leased to an operator under a triple-net
master lease, the operator is recorded as the controlling party, not the landlord. Operating control
determines staffing, budget and admissions; the landlord sets rent. This follows the same operating-control
convention used for the Institutional Ownership Type column, and it means the Publicly Traded Chain / REIT
category counts publicly traded owners and operators rather than REIT landlords.

The practical consequence: there is no structured landlord field in this release. For example, the 30
Florida facilities resolving to Aviata Health Group are recorded as Aviata; Welltower, which acquired 48
skilled nursing facilities nationally in February 2025 and leased them back to Aviata's predecessor under a
long-term triple-net master lease, appears in the per-row reasoning on 9 of those 30 rows and as a cited
source on 8, but not as an owner value. Anyone studying rent-driven financial pressure, which is a central
mechanism in this literature, will need to add a landlord column from SEC filings. That is a planned
extension rather than an oversight, and it is stated here so no one mistakes the omission for an absence.

Source data and how to cite it

The federal inputs are US Government works, published by CMS for public access and reuse. CMS asks that users cite
their data sources. Cite them as:

Centers for Medicare & Medicaid Services. Skilled Nursing Facility All Owners. Provider Data Catalog. Snapshot
May 27, 2026, retrieved June 9, 2026. https://data.cms.gov/provider-data/dataset/y2hd-n93e

Centers for Medicare & Medicaid Services. Nursing Home Provider Information. Provider Data Catalog. Snapshot
May 27, 2026, retrieved June 9, 2026; cross-check release July 1, 2026.
https://data.cms.gov/provider-data/dataset/4pq5-n9py

Federal source vintage: CMS baseline snapshot 2026-05-27, retrieved 2026-06-09. Federal cross-check files dated 2026-07-01.

Acknowledgments

This dataset was built during the University of California, Berkeley Summer Undergraduate Research Fellowship (SURF), 2026. My faculty mentor was Professor Shreeharsh Kelkar, who shaped the research question and set the standard of evidence the project is held to. Alexandra Gessesse, my SURF graduate student mentor, advised on presenting the work to non-specialist audiences and pushed me to explain what the underlying public records actually are rather than list them. The fellowship was supported by the Landau sponsorship, and I am grateful to the donors who fund undergraduate research at Berkeley. I carried out the collection, resolution, adjudication and verification. Errors are my own.

The web research stage used Clay. I am a Clay campus ambassador and received platform credits in that capacity. Clay had no role in the design, execution, analysis or reporting of this work and did not review it before release.

Code/Software

No code is required to read these files. They are plain comma-separated text and open in any spreadsheet application, or in Python via pandas.read_csv and in R via readr::read_csv, subject to the CCN-as-text requirement described above.

The processing pipeline had three stages. The deterministic federal join, the four owner-label consistency tests, the party-set normalizer, the confidence audit and the export verification were written in Python 3.10.12 using pandas 2.3.3 and openpyxl 3.1.5. The intermediate web research stage used a commercial research-agent platform. The prompt governing it required independent corroboration of both industry and geography before accepting any name match, enumerated its output categories so the model could not emit a variant or a sixth ownership type, and instructed the agent to return "unresolved" rather than choose between candidate entities. That platform is proprietary, so the prompt is available from the author on request rather than published here. Those deterministic components are being prepared as a public code repository, which will be linked from this record. Because the research-agent stage cannot be rerun by a third party, end-to-end reproduction from the raw inputs will not be possible; what is reproducible is the deterministic majority of the pipeline and, through the per-row citations, every individual resolution.

Resolution proceeded in three stages, with a documented disposition for the rows that none of them could resolve. The first two stages were run independently over the full population so that each facility received two determinations; the third reconciled them.

A deterministic federal join. Each facility's CCN was joined against two files from the CMS Provider Data Catalog: the Skilled Nursing Facility All Owners file (NH_Ownership, 247,632 national owner and manager records across 18 distinct disclosed roles) and the Nursing Home Provider Information file (NH_ProviderInfo). Because the CCN is a shared key, this is an exact match rather than a name comparison. CMS mixes those eighteen disclosure roles into a single column, including lenders, auditors, and corporate officers alongside genuine owners, so ownership interests were separated by role before use.

A constrained AI-assisted research pass, applied to every facility as an independent second determination. Every facility, including those the federal join had already resolved, was also passed to a commercial research-agent platform driven by a large language model. This was deliberate: the purpose was not to fill gaps but to produce a second, independently sourced determination for each facility that could be compared against the federal record. The agent operated under a prompt that required independent corroboration of both industry and geography before accepting any name match, and that returned "unresolved" rather than choosing between candidate entities. Its outputs were treated as candidate identifications requiring a citation, never as findings, and it was not used to assign confidence levels, which are rule-governed. Because both stages ran on every row, the source citations in 01_ownership_crosswalk.csv are pooled rather than attributed by stage. The platform is proprietary and cannot be redistributed; the prompt is available from the author on request.

Comparison and manual adjudication. The two determinations were then compared. Agreement between the federal record and the independent research pass is the corroborating signal that the confidence rules require; disagreement was adjudicated by hand rather than resolved by rule. Every internally inconsistent owner group was pulled and read individually rather than sampled. Owner labels were then tested for drift as sets of named parties rather than as strings, normalizing letter case, middle initials and word order, which found fourteen groups in which one owner had been recorded under two or more labels; 67 rows were relabeled. Before any group was merged it was screened for the opposite error, a false match between unrelated parties sharing a name, by requiring shared state, shared CMS chain affiliation and at least one shared source domain. Every change is recorded row by row in 05_owner_label_normalization_log.csv, so this release is reproducible from the prior state.

Disposition of rows that both stages failed to resolve. Both determinations can fail, and for 11 of the 2,192 facilities both did. Those rows are retained in the dataset and flagged unresolved in the Resolution Depth column rather than dropped from the population or assigned a guessed owner, so the denominator stays intact for anyone computing rates. Their Ultimate Parent Entity value is the literal string UNRESOLVED, which is a sentinel and not an owner name, and all 11 are held at low confidence. They still carry the evidence of the search that failed, between four and seven citations each.

Two properties of these 11 are worth stating rather than burying. First, there is no withheld federal answer the method missed: each of the 11 appears in the CMS ownership file as exactly one record whose role field reads "Ownership Data Not Available" with a blank owner name, which is CMS itself reporting the absence. Second, the failures are not randomly distributed. Seven of the 11 carry an alphanumeric CCN, so 25.0 percent of the 28 alphanumeric-CCN facilities are unresolved against 0.18 percent of the numeric-CCN facilities, a difference of roughly two orders of magnitude. All 11 are in California and all are typed as independent single facilities. By CMS certification type they are seven Medicaid-only, the same seven that carry an alphanumeric CCN, plus three dually certified and one Medicare-only, and none resides within a hospital; several are behavioral health centers by name. The facilities that the federal identifier scheme handles badly are therefore largely the same facilities that federal ownership disclosure does not reach, and both failures concentrate on Medicaid-only providers. Resolving them would require California state licensing and Secretary of State records facility by facility.

These 11 should not be conflated with the 147 rows flagged facility_is_terminal, which are a different outcome: there, an owner is named and no further indirect owner exists above it, which is usually a genuinely standalone facility. Anyone computing network- or chain-level statistics should filter to parent_identified first, and the column exists so that they can.

Software. The deterministic join, the owner-label consistency tests, the party-set normalizer, the confidence audit and the export verification were written in Python 3.10 using pandas and openpyxl.