Fundprint: who owns America's autism therapy clinics

Private equity owns at least 1,621 autism therapy (ABA) clinics across 42 US states. This dataset names, locates and traces each one to a public source document that was fetched and content-hashed.

• Site: whofundsmytherapist.com
• Archived release and DOI: 10.5281/zenodo.23426864
• Pipeline code: github.com/fundprint/fundprint-data
• Methodology: github.com/fundprint/fundprint-methodology

Dataset version 2026.07-beta, methodology 2026.10-verification-v1.

At a glance

Clinics traced	1,757 in 42 states, held by 19 institutional owners
Private-equity owned	1,621, under 15 sponsors
Other financial owners	136 (a pension fund, a family office, two search funds)
PE share of all ABA sites (registry basis, a floor)	2.8% nationally; 15.8% in Minnesota, 13.2% in New Mexico, 9.5% in Arizona
Attested by the owner's own directory	1,575 of 1,757
Hand-verified accuracy	66 of 67 sampled PE clinics correct on site, address, brand and owner (98.5%, 95% CI 92.0% to 99.7%)
Platform coverage	21 of 32 known PE-backed ABA platforms

Load it

from datasets import load_dataset

clinics = load_dataset("gameloft737/fundprint", "clinics", split="train")
pe = clinics.filter(lambda r: r["firm_type"] == "private_equity")   # 1,621

Files

Path	What it is
data/clinics.parquet / .csv	One row per clinic (a physical site). The main table.
data/owners.parquet / .csv	Operating companies (brands) and the firm that owns each.
data/firms.parquet / .csv	Institutional owners, including former owners with 0 clinics.
derived/aba_market.json	National and per-state shares against the 21,088 ABA sites in the federal registry.
derived/coverage.json	The 32 in-scope PE-backed platforms and which ones this dataset has read.
derived/reconciliation.json	Comparison with the published JAMA Pediatrics and PESP estimates.
derived/state_files.json	HHS OIG Medicaid ABA audits paired with each state's footprint.
derived/verification.json, reviewed_3bf05f0a.csv	The hand-verification sample and its scored results.
derived/snapshot_meta.json	Versions and headline totals of this release.

In the CSVs, the sources list is joined with |.

Clinic fields

Field	Meaning
id	Stable clinic identifier.
name, address, city, state, zip	The site as its best source states it.
npi	A registry provider ID at the site, if any. Not an identity key: one NPI can cover many sites and one site can carry several NPIs.
lat, lng	ZIP-level, not street-level: the Census ZCTA centroid of the ZIP (or of its 3-digit prefix). One clinic is unplaced.
owner_name, owner_id	The operating company (brand).
firm_name, firm_id, firm_type	The controlling institutional owner. firm_type is private_equity, pension_fund, family_office or other (search funds).
confidence_overall	The weakest of the four dimensions below: owner_verified, registry_current, registry_aging or registry_stale.
confidence_open	Whether the site is still operating: owner_listed (the owner's directory lists it today) or a registry freshness grade.
confidence_address	owner_stated or registry_filed.
confidence_site_type	center (the owner lists it as a centre) or unverified.
ownership_basis	curated (a hand-verified acquisition source) or portfolio (the firm's own portfolio page).
registry_last_updated	When the registry record was last certified, if the site has one.
confidence_score, confidence_method	The resolver's name-match score and method.
sources	URLs of every source document behind the row.

How it was built

Four public source families: the CMS NPPES bulk registry, operators' own published location directories and rosters, PE firms' portfolio pages, and hand-verified acquisition announcements. Every source is fetched and content-hashed. Rules that matter most:

• A clinic is a physical site, keyed on canonicalized street and ZIP within an owner, never on an NPI.
• An owner's complete directory overrides the registry in both directions: it adds centres the registry never recorded and retires registry rows the owner does not run.
• A wrong street is worse than no street. Parsers refuse an address they cannot split safely.
• Corrections are quarantines, never deletes. Every decision is an append-only record.

Full rules: METHODOLOGY.md.

Limitations

• Both the count and the shares are floors. About 500 facilities sit at 11 PE-backed platforms this dataset has not yet read; they are named in derived/coverage.json.
• The 2.8% share uses the registry on both sides: 580 PE sites the registry can see, out of 21,088 registry ABA sites. Directory-only clinics are excluded from both sides, and closed centres can be removed from the numerator but not the denominator.
• Do not compute pe_clinics / clinics (about 91%). That is PE's share of this dataset, not of the market.
• 182 clinics rest on the registry alone. The registry never marks a closed site closed, so 30 of them are graded registry_stale (untouched 6+ years).
• The verification sample was reviewed by the author alone, so there is no inter-rater figure.
• No audit attributes any improper payment to private equity, and this dataset makes no such claim. The state audit files pair figures with footprints; they do not assert cause.

Citation

@dataset{doke_2026_fundprint,
  author    = {Doke, Atharva},
  title     = {Fundprint: a site-level census of institutionally owned ABA clinics},
  year      = {2026},
  version   = {2026.10-verification-v1},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23426864},
  url       = {https://doi.org/10.5281/zenodo.23426864}
}

License

CC BY 4.0: reuse freely, with attribution. Archived copies of third-party source documents are not included and are not covered by this license.

Corrections are welcome and are published with a changelog entry: open an issue on fundprint-data.