The big one: the directory just changed more in one release than in the last four

The US government publishes a list of everyone who provides health care. On 20 August it published a new version, and this one is not a routine refresh.

We reloaded all of it and compared every number against the previous release. Here is what actually moved.

Where a doctor works: the record count more than doubled

The most important fact in the directory is where each clinician works. Without it, software cannot find your records.

That link is stored in a record called a PractitionerRole. There were 7.0 million of them. Now there are 16.5 million.

Nearly ten million new records. That is the largest single change we have seen since we started measuring.

But the share of clinicians covered went up only five points

Here is the part that surprised us.

Adding 6.9 million active records moved the share of clinicians who have a workplace at all from 27% to 31% nationally. In Pennsylvania, where we track this most closely, 38.1% to 43.7%.

Ten million records. Five points.

The reason is that most of the new records went to people who already had one. The average clinician with a workplace listed went from about 2 records to about 4.7. The directory got much more detailed about the people it already described, and only somewhat better at describing new ones.

That is not a criticism. Knowing a doctor works at four named places is more useful than knowing they work at two. It is just a different thing from covering more doctors, and the headline number hides which one happened.

The gaps closed most where they were worst

We published a finding last week showing that the directory describes some professions well and others barely at all, and that the split tracks who bills Medicare. Here is that same table, before and after.

Every single profession improved. The largest multiples are at the bottom of the table: pharmacists went from 1 clinician with a workplace to 526, dentists nearly tripled, counselors and social workers went up half again.

The gradient is still there. A nurse practitioner is still twenty times more likely to have a workplace listed than a pharmacist. But it is a much smaller gradient than it was three months ago.

We want to be careful here. We published that finding a few days before this release. We have no reason to think the two are connected, and we are not claiming credit for someone else's work.

Organization hierarchy: from unusable to complete

In the May release, the directory had a field for saying that one organization is part of a larger one. It was populated 148,834 times, and every single one of those pointed at an organization that was not in the file. The field resolved to nothing at all.

We said so publicly, and told several people building on this data not to rely on it.

In this release it works. 140,017 references, 43,551 parent organizations, none of them missing. CMS published the parents.

If you built around that field being useless, rebuild. We were right about the old release and we are glad to be wrong about this one.

Health insurers appeared

The directory now carries health plans, which it did not before. 233 of them, across 27 insurers, and those insurers are now filed under a "payer" category that did not exist in May.

This matters because the directory was always meant to cover insurers as well as providers, and until this release it simply did not.

One caveat. The insurers are named, but none of them publishes a working address for software. So you can now look up who an insurer is, and still not reach them. Identity arrived. Reachability did not.

Endpoint attribution went backwards

We track how many of the directory's web addresses say who they belong to. In May, 19,334 of 114,071 did, about one in six. Now it is 16,262 of 110,973.

That is 16.9% down to 14.7%.

We do not know why, and we would rather say that than guess. The total number of addresses fell too, so some of this is addresses being removed rather than names being lost. It is the one number in this release that moved the wrong way, and we are flagging it rather than burying it.

Everything else that moved

Half the organization file is now tax records rather than actual organizations, up from 41%. If you count organizations in this directory, check which kind you are counting.

More quality measures, re-run against this release

Social Security numbers: 46, then 41, then zero

Last year the Washington Post reported that the government's provider file contained doctors' Social Security numbers. We checked it ourselves and found 46 in the April release, then 41 in May. CMS had scrubbed some and missed others.

In this release there are none. Zero in 7,373,232 practitioner records, zero in the organization records.

We were careful about this one, because a search that finds nothing looks exactly the same as a search pointed at the wrong place. So we checked that the scan still works: 7,371,126 of those records still contain the field the numbers used to sit in. The field is there. The numbers are not.

46, then 41, then 0.

Duplicate organizations: 13, not two thirds

We have been reporting for months that about two thirds of organizations in the directory appear more than once, and warning people not to count rows and call it a count of organizations.

That was true, and the reason was not what we implied. Half the organization file is tax records rather than descriptions of organizations, and CMS files one of each under the same ID. So an organization shows up twice because the file holds two different kinds of record about it, not because anyone entered it twice.

Once you separate the two kinds: 13 organizations out of 2,202,028 are genuinely listed twice.

The advice does not change, because the extra rows are really there and will really break your count. What changes is who has a problem to fix. We were describing a data entry mess. It is a modeling decision, and it needs a sentence of documentation rather than a cleanup.

Phone numbers and coordinates

Places in the directory carried no phone numbers at all in May. They now carry one for 79.0% of the 2,535,686 of them. Coordinates went from 93.9% to 98.3%. Clinician phone coverage is unchanged at 99.97%.

Organization phone coverage fell, 99.9% to 95.6%. That tracks the organization file being half tax records now.

Cross-references

References between records that point at nothing: none in the workplace or place records. 724 remain, all of them web addresses pointing at an organization the file does not contain, out of 33.4 million references.

Method notes and corrections

Four things worth recording, briefly, because they change what you can trust.

The parent-organization field works now. We said publicly it was broken and told people not to rely on it. Rebuild if you designed around it.

We reported in June that places carry no phone numbers. They now carry one for four in five. That report is still up.

Our comparison against the federal provider registry gave a false positive. The public copy we compare against stops in February; against an August release, 245,374 recently registered providers looked like they were missing. We sampled eight and all eight were real and active. That signal is now scored at zero while our reference data is behind, and the affected records are labeled needs-review rather than high-risk. The state pages drew their "verify these yourself" samples from the same signal, so those now come from the federal exclusion list instead, which we download fresh.

One number we published moved down. Vendors can name 65% of the addresses the directory cannot, not the 76% we reported in May. It would still take attribution from 14.7% to 70%.

If you build on this data, the transferable lesson is that a source changing shape rarely crashes the code that reads it. It returns nothing, and nothing looks like an answer. Three of our own checks did exactly that this week and each has been changed to fail loudly instead.

Check our work

• Every number above, as data: https://ainpi.dev/api/v1/release-deltas.json
• Profession-by-profession change: https://ainpi.dev/api/v1/role-gap-delta.json
• What the two new record types contain: https://ainpi.dev/findings/ndh-new-resource-types
• Plain-language guide to all of this: https://ainpi.dev/primer
• Every published finding, all re-measured against this release: https://ainpi.dev/findings

We reloaded the whole thing, 45 GB, and re-ran every measurement. The scripts are open and the data is public, so you can check any of it.

— Eugene Vestel, FHIR IQ