Private equity owns at least 1,621 autism therapy (ABA) clinics across 42 US states. Each one is named, located, and traced to a public source document that was fetched and content-hashed. This repository is the pipeline that builds that dataset, and the dataset itself.
The first peer-reviewed national estimate counted 574 PE-owned sites from a deal database. Restricted to what the federal provider registry can see, this dataset finds 580, six apart, by a method sharing none of its data. The other 1,041 clinics appear only in the operators' own published directories: centres opened after the deal that no deal list or registry records.
| Clinics traced | 1,757 in 42 states, held by 19 institutional owners |
| Private-equity owned | 1,621, under 15 sponsors |
| 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; the 11 not yet read are named |
Three findings the data supports:
- Counts sort by unit, not by rigor. Deal-based and registry-based counts see a company as it was when bought or registered (574, 580). Platform-based and directory-based counts see it today (1,521, 1,642 over the same 16 platforms). The gap between the kinds is close to threefold.
- The federal registry's blind spot differs by operator, from none to complete. One 18-centre operator is invisible to it, so no single correction factor can fix a registry-based count.
- In Wisconsin, the registry can see 2 of the 48 PE-owned centres, in a state whose Medicaid ABA spending federal auditors found at least $18.5 million in improper payments. No audit attributes any payment to ownership, and none is claimed here.
- Download the release: doi.org/10.5281/zenodo.23426864 (dataset, market shares, coverage, reconciliation, verification sample, code).
- Browse it: whofundsmytherapist.com, with a search by clinic, a map, and a page per audited state.
- The single file behind every number:
snapshot.jsonin the release (and in fundprint-dashboard). Every clinic carries its sources and a four-part confidence grade (open, address, site type, owner). Field reference:docs/schema.md. - Derived tables:
data/holds the market shares, platform coverage, reconciliation and state audit files, regenerated by the pipeline, never hand-edited.
Five stages: acquire, store, resolve, validate, publish. The rules that matter most:
- A clinic is a physical site, keyed on canonicalized street and ZIP within an owner. Never an NPI: one NPI can hold dozens of addresses, and one address can hold six NPIs.
- An owner's complete directory overrides the registry in both directions. For the largest operator, the registry missed 212 Texas centres and listed 105 addresses the operator does not run.
- A wrong street is worse than no street. Parsers refuse an address they cannot split safely rather than guess.
- Corrections are quarantines, never deletes. Every decision is an append-only record.
- Sources must be fetchable and content-hashed, or the claim does not ship.
The full, versioned rules are in
fundprint-methodology, and the
paper draft is in paper/. Design notes: docs/architecture.md.
git clone https://github.com/fundprint/fundprint-data.git
cd fundprint-data
pip install -e ".[dev]" # Python 3.12+
cp .env.example .env # set DATABASE_URL
psql "$DATABASE_URL" -f supabase/migrations/20250510000001_initial_schema.sql
pytestThe registry input is the free CMS NPPES monthly dissemination archive. The command
sequence that regenerates every published figure from it is in the paper (Section 7).
The verification sample is seeded and can be redrawn exactly from samples/.
Doke, A. (2026). Fundprint: a site-level census of institutionally owned ABA clinics (Version 2026.10-verification-v1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23426864
Found an error? Corrections are welcome and are published with a changelog entry: atharva.doke737@gmail.com, or open an issue.
- fundprint-methodology: what counts as a clinic, an owner and a PE-backed relation, and why.
- fundprint-dashboard: the public site, a static build from one JSON file.
The code is licensed under the MIT License. The data is licensed under CC BY 4.0: reuse it freely, with attribution. Archived copies of third-party source documents are not covered by either license.