We read all 9,852,180 records in the federal provider registry. One finding held up under testing, and two things we expected to find turned out to be artifacts of how the data is built — artifacts that will mislead anyone who analyses this file the obvious way.
Behavior technicians — paraprofessionals who deliver applied behaviour analysis therapy under supervision — went from 27,045 new registrations in 2018 to 114,556 in 2025. In 2025 they were 22.1% of every new individual provider registration in the United States. More than one in five.
The federal government publishes the National Plan and Provider Enumeration System — NPPES — as a monthly file. Every NPI ever issued, free to download, no registration. The October 2026 release is 1.1 GB compressed and 11.2 GB as CSV: 9,852,180 records, of which 7,512,094 are individual providers and 1,982,161 are organisations.
We streamed the whole thing once and aggregated it. Then we tried to break our own results, which is where the interesting part is.
Taxonomy code 106S00000X. The registry defines it as a paraprofessional who implements
behaviour-analytic treatment plans under the close supervision of a certified behavior analyst —
in practice, the people delivering ABA therapy, overwhelmingly for autism.
The taxonomy code did not exist in any meaningful way before 2016 — fewer than 500 registrations a year carry it before then — so the comparison that matters is 2018 to 2025, both well after the code was established. That is a 4.2x increase.
A fourfold rise could just mean the registry as a whole is inflating. So we measured it against taxonomies nobody graduates out of, over exactly the same window:
| Provider type | 2018 | 2025 | Change |
|---|---|---|---|
| Physical Therapist | 13,034 | 12,088 | 0.93x |
| Optometrist | 1,589 | 1,528 | 0.96x |
| Massage Therapist | 3,170 | 5,775 | 1.82x |
| Behavior Technician | 27,045 | 114,556 | 4.24x |
Physical therapy and optometry are flat. Massage therapy nearly doubled. Behavior technicians went up more than fourfold. Whatever is happening, it is specific to that category, not to the registry.
Our first pass appeared to show something far more dramatic. Physician registrations looked like they had fallen off a cliff — from 15.5% of all new individual registrations in 2012 to 1.9% in 2025. A 71% collapse.
That is not true, and it took about ten minutes to work out why.
▬ Physicians ┄┄ Students in an organized health care education program
As physicians fall, students rise by almost exactly as much. That is not a workforce trend. It is the shape of a data-modelling decision.
In NPPES the specialty is current, but the registration date is frozen. A doctor who was issued an NPI as a medical student in 2024 is still carrying the student code today. They will re-declare as a physician in a few years — and when they do, that change will be attributed to 2024, because that is when their number was issued and that never changes.
So recent years are full of people who have not finished becoming what they will be. Every recent cohort is over-weighted towards trainee and entry-level codes and under-weighted towards terminal ones, and the effect gets stronger the closer you get to today.
Anyone who charts specialty mix by registration year without knowing this will publish a false number. It is an easy mistake to make: the data is clean, the trend is smooth, and the result looks like a story. The behavior technician finding survives only because we checked it against categories that cannot mature into something else.
The obvious companion question is which specialties are losing people. The registry contains 357,925 deactivated NPIs — 3.6% of all records — so the question looks answerable.
It is not. A deactivated NPI keeps exactly two populated fields out of 330: the number itself and the deactivation date. Entity type, specialty, address, everything else is cleared.
There is nothing left to group by. You can count deactivations, and that is all. CMS also publishes a separate monthly deactivation file, which has the same two columns.
One wrinkle worth knowing if you work with this file: a record carrying both a deactivation date and a later reactivation date is live again and keeps its full data. Those have to be counted as active or the deactivation total comes out too high.
We expected to find solo practice collapsing. It isn’t, quite. Among individual registrations that declared a status, the sole-proprietor share fell from about 38% in 2007 to about 29% by 2018 — and then stopped falling, holding between 30% and 32% ever since.
What is striking is the geography. The share of 2020–2026 registrations declaring sole-proprietor status varies roughly threefold between states:
Ohio and Florida register solo practitioners at roughly two and a half times the rate of Michigan and Minnesota. We do not have an explanation for that, and we are not going to invent one. It is in the data if someone wants to chase it.
Treat this number carefully. Sole-proprietor status is self-declared when the provider registers and is almost never updated afterwards. It describes somebody at the moment they got their number, not how they practise now.
NPPES is a registry of identifiers, not a census of working clinicians. Specifically:
Every claim on this page is therefore about registrations, not about practices or practitioners. That distinction is doing a lot of work, and dropping it is how this data gets misreported.
Source: CMS NPPES Data Dissemination V.2, 12 October 2026, downloaded from download.cms.gov/nppes. Public domain, free, no registration required. Specialty labels come from the NUCC Health Care Provider Taxonomy, version 25.1.
One streaming pass over the compressed archive; the CSV is never extracted. Primary specialty is the taxonomy flagged primary, falling back to the first populated slot. Records with a deactivation date and no reactivation date are excluded from active counts.
Both files are CSV, no email required. Use them for anything you like — if you publish something from them we would be glad to see it.
Essential Medical Billing & Consulting, “The fastest-growing provider in American healthcare
isn’t a doctor or a nurse”, October 2026.
https://essentialmbandc.com/research/npi-registrations-2026/
Honestly? Because it is our market and nobody had measured it. We bill for independent practices, so who is entering the profession, in what form, and in which states is not an abstract question for us.
There is also a practical reason the behavior technician number matters to anyone doing ABA billing: that work is supervision-heavy, unit-based and authorisation-dependent, and a category growing fourfold in seven years is a category where a lot of people are learning to bill it as they go.
By NPI registrations, it is the behavior technician - a paraprofessional who delivers applied behaviour analysis therapy under the supervision of a certified behavior analyst. Registrations grew from 27,045 in 2018 to 114,556 in 2025, and behavior technicians were 22.1 percent of all new individual NPI registrations in 2025.
The CMS NPPES file of 12 October 2026 contains 9,852,180 records: 7,512,094 individual (NPI-1) and 1,982,161 organisational (NPI-2). 357,925 are deactivated, about 3.6 percent.
They are not. In NPPES the taxonomy is current but the enumeration date is fixed, so a doctor who registered as a medical student is still coded as a student until they re-declare years later. Recent registration years are therefore over-weighted to trainee codes. Physicians appear to fall from 15.5 percent of new registrations in 2012 to 1.9 percent in 2025 while students rise from 1.1 to 10.1 percent - a mirror image, and an artifact.
Not by specialty or state. A deactivated NPI retains only two populated fields out of 330: the NPI itself and the deactivation date. Entity type, taxonomy and address are all cleared, so there is nothing left to group by. The separate monthly deactivation file has the same two columns.
Among individual registrations that declared a status, the sole-proprietor share fell from about 38 percent in 2007 to about 29 percent in 2018 and then stabilised at 30 to 32 percent. It varies roughly threefold by state, from 56.5 percent in Puerto Rico and 50.7 percent in Ohio down to 18.3 percent in Alaska. The field is self-declared at registration and rarely updated afterwards.