Pliny said the half of the sentence everyone quotes. The half nobody quotes was a name. Here is every Medicare specialty drawn as a box, 1,044 real clinicians drawn as dots, and a measurement of how little the box tells you about the dot.
On a warship in the ash, with the helmsman telling him to turn back, Pliny the Elder said the line that ended up on four hundred thousand LinkedIn posts — and then he said the other half of it. Fortes fortuna iuvat. Pomponianum pete. Fortune favours the bold. Head for Pomponianus. A friend. By name. At an address, across the bay, at Stabiae.
Clinician-builders almost never run short of the first half. They run short of the second. They aim at ambient documentation, at care coordination, at primary care — and a category is not a place you can arrive.
That is a claim about attention. It is also, unusually, a claim you can measure. Medicare publishes what every billing clinician in the country actually did last year. Group them into the categories a product deck would use, and you can ask precisely how much the category tells you.
Below: every dot is one real clinician. Horizontal is how many Medicare beneficiaries they saw in the year. Vertical is how many distinct HCPCS codes they billed — a rough measure of how varied a day is. Pick a category and the red box appears: the true 10th–to–90th percentile of that whole category, computed from all of its members, not from the dots.
Select Family Practice. The box runs from 28 patients to 409 — a 14.6× spread — and from 8 distinct billing codes to 66. Both of those people are family practice. One of them is running a 400-patient panel across sixty-six kinds of encounter; the other bills eight codes and sees twenty-eight people. There is no product that is right for both, and “family practice” is the word that hides the difference.
That is not one bad category. Across all 87, the median category is 10.8× wide from its 10th to its 90th percentile on panel size alone. Run the formal version and you get the number in the second readout: sorting 1.24 million clinicians into 103 specialty categories accounts for 35.8% of the variance in how many patients they see, and 50.6% of the variance in how many distinct things they bill. Roughly two-thirds of what makes a clinician’s year what it is happens inside the category, not between categories.
A category buys you about a third of the answer. The other two-thirds is the person. That ratio is why interview-one-real-clinician beats survey-the-segment so reliably, and it is not a motivational claim — it is a variance decomposition on a census of the whole country. If you have picked a specialty and think you have picked a user, you have made about 36% of the decision and shipped as though you had made all of it.
Every dot you can see is a sample, and the sample lies worst about the categories you care most about. Each category contributes twelve randomly drawn clinicians. Compare the median of those twelve against the true median of the whole category and the error is not where you would guess. Medical Toxicology has 36 members in the entire country; twelve of them land its median within 1.5%. Internal Medicine has 93,627 members; twelve of them miss its median by 45%. Physician Assistant, 112,514 members, misses by 42%.
Across the 27 largest categories the median sampling error is 27.7%; across all 87 it is 19.5%. The big categories are harder to sample precisely because they are internally enormous — which is the same finding as the box, arriving through a different door. A dozen conversations characterises a tiny specialty well and a huge one badly, and the huge one is the one on your slide.
Now drag the min-category-size slider. Watch the fourth readout. With all 87 categories in, the sampled dots show a respectable-looking correlation between how many patients a category’s clinicians see and how old those patients are — —. Filter to the 15 categories with 20,000+ clinicians and the same correlation, computed the same way, falls to —. The trend was substantially an artifact of thinly-sampled small categories bouncing around. The relationship is real but weaker than it first looks — in the full population it holds at about +0.41 among the big categories rather than vanishing — and that gap between the sampled slope and the true one is the ordinary way a chart of a sample talks you into a story.
| category | clinicians | panel p10–p90 | spread | codes p10–p90 |
|---|
This is Medicare Part B fee-for-service only — no Medicare Advantage, which is now about half of all beneficiaries, and no commercial or Medicaid volume at all. A clinician with a young commercial panel appears here as a small dot because Medicare barely sees them, not because their day is small. Counts are distinct NPIs with at least eleven Medicare beneficiaries, so a part-time biller sits beside a full-timer with no way to tell them apart, and that alone inflates the within-category spread the whole piece is built on. Specialty is CMS’s own rendering provider type, derived from the specialty code on the claim — it is one label per NPI, so a clinician who does two jobs is filed under one of them.
“Distinct HCPCS codes” is a crude proxy for variety: a dermatologist billing forty codes and a hospitalist billing nine are not obviously in the order the axis implies, and specialties with narrow fee schedules will look narrow whatever their days are like. Categories with fewer than 30 clinicians are dropped entirely; 16 categories in the source did not clear that bar. And the variance decomposition is on logged values, which is the honest way to handle a distribution this skewed but does mean the 35.8% describes proportional differences, not raw headcounts.
None of that moves the finding much, because the finding is about magnitude rather than precision. You do not need the boxes to be exactly right to see that they are far too big to aim at.
Fortune favours the bold. It does not tell you where to sail. The second half of the sentence does, and the second half of the sentence is a name.