The directory got smaller, and almost nobody fell out of it
Correction, 17 October
Two figures and one statement in this update were wrong.
Insurer directories. We said 50 addresses in the file look like insurer provider directories. That counted only addresses whose wording names a directory. Counting addresses on servers that host an insurer's directory as well, there are at least 194, up from one in August. None of them is linked to an insurer or to any other organization, and no insurer points at one. The finding page has the breakdown.
The directory we test against. We said the insurer directory we test is not in the file. That address is not, but the same insurer has another directory address in the file.
Pennsylvania records access. We said we could connect 15.8% of Pennsylvania clinicians to a web address where software could request patient records, down from 18.7% in August, and that the drop came from the data. It did not, entirely: we built part of the August figure on May data, so the two figures did not describe the same pair of releases. We also now count every provider ID number a software vendor lists beside a web address, not only the first. Measured the same way on both releases, the share is 22.3% in October against 22.9% in August. Counting only links the directory itself records, or a vendor list naming 10 or fewer providers, it is 2.9% against 3.5%. Either way the share fell by 0.6 points, not the 2.9 points we reported. The comparison and how we ran it are public.
The US government publishes a list of everyone who provides health care, the National Provider Directory. On 7 October it published a new version. We reloaded all of it, re-ran every check we publish, and compared each number with the August release.
The short version: the file shrank by a quarter, and the share of clinicians it describes did not shrink with it. Health plans arrived in large numbers, and the directory gives none of them a web address.
A quarter smaller
| August | October | Change | |
|---|---|---|---|
| All records | 32,468,908 | 24,761,939 | −23.7% |
| Clinicians | 7,373,232 | 7,481,907 | +1.5% |
| Where-they-work records | 16,545,158 | 11,056,751 | −33.2% |
| Organizations | 4,402,671 | 2,058,500 | −53.2% |
| Places | 2,535,686 | 2,558,069 | +0.9% |
| Web addresses | 1,128,169 | 1,140,836 | +1.1% |
| Links between organizations | 483,992 | 465,876 | −3.7% |
Two lines carry nearly all of the drop.
Organizations halved because tax records left. In August, 2,199,519 of the organization records were labelled only as tax-ID records ("ein"), with no provider type. In October there are 79,721. Organizations labelled as providers fell too, but by about a tenth: 2,202,028 to 1,976,702.
Where-they-work records fell by 5.5 million, and coverage held. In August the directory more than doubled these records. Most of the new ones went to people who already had one. October takes much of that back. The average clinician with a workplace listed now has about 3.2 records, down from about 4.7.
What matters is whether people lost their workplace. They mostly did not. In Pennsylvania, where we track this most closely, the share of clinicians with a workplace listed went from 43.7% to 44.0%.
Health plans arrived. None has a web address
In August the directory named 27 health insurers for the first time. In October it names 997, and 7,575 individual health plans, each pointing at the insurer that runs it.
None of the 997 insurers has a web address in the directory.
There are now 50 addresses in the file that look like insurer provider directories, up from one. We checked both ways: none of them is attached to an insurer, and no insurer points at one. So the file now says who the insurers are and, separately, lists some of their directories, without connecting the two.
Federal rules have required Medicare Advantage, Medicaid and CHIP plans to publish these directories since 2021. We test one that answers a plain request today. It is still not in the directory.
Two things got worse
Phone numbers. In August, all but about 1,900 clinicians had a phone number listed. In October, 168,778 have none anywhere. On the clinician's own record, 176,084 have no phone: about 111,800 list only a fax number and 64,243 list nothing at all. For 7,306 of those, a workplace record supplies one. Coverage is still 97.7%, so the drop is small.
Reaching patient records through a clinician's practice. We can connect 15.8% of Pennsylvania clinicians to a web address where software could request patient records from their practice. In August it was 18.7%. That covers every route we know, including the vendors' own public lists. We used the same method for both releases, so this drop comes from the data, not from a change in how we count.
Pennsylvania, profession by profession
| Job | August | October |
|---|---|---|
| Nurse practitioners and PAs | 82.0% | 77.2% |
| Doctors | 74.3% | 74.8% |
| Foot doctors | 69.3% | 66.5% |
| Eye doctors | 67.7% | 63.6% |
| Chiropractors | 46.2% | 40.9% |
| Counselors and social workers | 22.9% | 27.9% |
| Physical and speech therapists | 24.6% | 25.7% |
| Dentists | 13.3% | 15.4% |
| Pharmacists | 4.2% | 6.7% |
| Nurses | 4.8% | 6.2% |
The professions at the bottom of the table kept rising. Four near the top fell: nurse practitioners and physician assistants lost almost five points, chiropractors more than five, eye doctors about four and foot doctors almost three. We do not know why yet, and we are not going to guess in print.
Two new kinds of record, one of them empty
The October file again carries two record types we do not load into our warehouse. One describes health plans. Every one of the 7,575 names its insurer, but only 233 say what kind of plan it is or which network it uses.
The other is meant to describe services, such as a clinic offering vaccinations. There are 54,445 of these. None says what the service is: no name, no type, no specialty. Fewer than 1 in 100 says where it is offered, and one names a provider. As shipped, it records that two parties are linked and nothing about the service itself.
What held steady
- Of the web addresses software can actually use, 14.6% say which organization they belong to. It was 14.7% in August.
- The software vendors' own public lists name the owner of 64.4% of the rest. One vendor moved its list during our run; we found the new location and used it.
- Social Security numbers in the file: still zero.
- Malformed provider ID numbers: zero, out of 9.5 million checked. August had four.
A correction
For two days before this update, our federal exclusion check reported no excluded providers at all, and our high-risk list dropped every provider flagged only by that check. That was wrong. The government changed how its exclusion file marks an empty field, from a run of zeros to a single zero, and our check then read every exclusion as lifted. It returned an empty answer instead of an error, which is the failure we most try to guard against.
It is fixed, and the loader now checks the file before loading it and refuses one in which most exclusions look lifted. The October answer: 7,799 of the 8,924 excluded providers who have an NPI appear in the directory.
We kept the August release
The government serves only the latest version of the directory. When October arrived, August disappeared from its site. We keep every release since May in one free, public archive on Databricks Marketplace, so anyone can check the comparisons above against the original records.
Check our work
- The measures we track, release by release: https://ainpi.dev/api/v1/measures.json
- Record counts, August against October: https://ainpi.dev/api/v1/release-deltas.json
- Insurers and their web addresses: https://ainpi.dev/findings/ndh-payer-endpoint-coverage
- The Pennsylvania map, now on October data: https://ainpi.dev/states/pa/connectivity
All of it is public data, and our code is open.
— Eugene Vestel, FHIR IQ