clinicians.build · interactive · September 19, 2026 · built on HHS‑OIG, Sep 17 2026
Pull Your Own Hundred
A hundred claims became twelve million dollars. Here is the entire machine that does that — 184,390 real Medicare claim reviews across 89 provider types — and the arithmetic that decides what your own hundred would be worth.
Primary source:“Methodist Hospital Received at Least $12.4 Million in Medicare Overpayments,” HHS Office of Inspector General, report A‑09‑23‑03001, issued Sep 14 2026, posted Sep 17 2026 — 100 claims worth $1,426,020 reviewed; 27 non-compliant; $256,926 in net overpayments found; an estimated $12.4M of the $62M Medicare paid. Data:CMS Comprehensive Error Rate Testing (CERT) claim-level review file via MIMI Labs, reporting year 2024 (claims submitted Jul 1 2022 – Jun 30 2023) — 185,349 medically reviewed claims, of which the 184,390 falling in the 89 provider types with at least 50 reviewed claims are plotted here.
of the 89 provider types sit outside the band that pure sampling luck could explain. Of the 8 biggest — every type with 10,000+ reviewed claims — it is all 8.
There is no national error rate
Every dot is one provider type. Horizontal is how many of its claims Medicare pulled and read; vertical is how many of those failed. The shaded funnel is the range you’d expect if every provider type in the country had exactly the same true error rate and the only thing moving the dots was which claims got drawn.
Almost nothing is inside it.
50
inside the funnel — consistent with luckoutside it — genuinely different
pooled rate—
types shown—
outside funnel—
highest rate—
Drag the minimum-sample slider up, the move that usually makes a spurious pattern evaporate. Here it barely helps. The spread across the small types is about as wide as the spread across the big ones, and the big ones are further from the centre line, not closer. That is the opposite of a small-n artifact: it is what real, structural difference looks like once you have enough claims to see it.
Switch the error definition and watch the map redraw. On documentation, the two biggest dots on the chart climb: clinical laboratory sits at 24.9% and medical supply companies at 20.2%, and for the lab that is 96% of everything it was flagged for. Switch to medical necessity and both drop to almost nothing — that column belongs to inpatient rehabilitation at 30.7%, where the dispute is whether the patient needed the bed at all.
the 80/20 lens
The single most expensive sentence in compliance is “the national improper payment rate is X%.” Nothing on this chart is the national rate. Your exposure is set by which of these dots you actually are — and the spread runs from nephrology at 45.4% down to federally qualified health centres at 2.1%, a twenty-fold gap, with three types where reviewers found nothing wrong at all. Benchmark against your segment, or don’t benchmark.
Now pull your own hundred
OIG reviewed 100 claims and found 27 bad. Then they estimated what the whole book was worth — and asked for the low end of that estimate, not the middle. Here is what that arithmetic does.
100
27%
$62M
what you observed—
90% interval on the true rate—
the demand (lower bound)—
if they used the midpoint—
Move the sample size and watch the interval breathe. At 25 claims it is nearly thirty points wide; at 500 it is under seven. The sample size is not an accounting detail. It is the single lever that decides how much of your book the government can credibly reach.
and this is the part that changed
Extrapolation exists because reading a claim is expensive, so you read a hundred and infer the rest. Every number on this page is downstream of that constraint. When checking a claim costs a thousandth of producing one, you stop inferring — you read all of them. The organisation that has already graded its own population has nothing left to extrapolate from: it has the census, not the sample.
where this is thin — read before you quote it
The simulator is not OIG’s method. OIG extrapolates dollars from a stratified random sample, weighting each stratum separately; this is a single-stratum Wilson interval on a claim count, then multiplied by paid dollars. It is the shape of the argument, not a reproduction. On the audit’s own inputs it happens to land close — see above — and you should treat that as instructive, not as validation.
Provider type is not a hospital. CERT samples claims nationally and labels them by the type of provider that billed. A dot is “all SNF claims reviewed in the country,” not one facility. Your own rate can sit anywhere inside a dot.
CERT's sample is not random across the program. It is a stratified design that deliberately over-samples claim types considered high risk, which is why a plain claim count here (15.8% overall) is far above CMS’s published dollar-weighted Medicare fee-for-service improper payment rate (6.55% for FY2025). Both come from this program. They are not the same statistic and neither is the one an auditor would apply to you.
The funnel assumes one shared true rate. That is the null hypothesis, and it is obviously false here — which is the point of drawing it. Being “outside the funnel” means “not explainable by sampling alone,” not “bad.” Some of these differences are policy: inpatient rehabilitation faces a medical-necessity standard that ophthalmology simply does not.
89 types, not 91. Types with fewer than 50 reviewed claims in 2024 are dropped — 959 claims in all — because a rate on 12 claims is not a rate. That threshold is a judgment call and it is the one most likely to move the picture.
One year. Reporting year 2024 only. The companion piece runs the same program back to 2011, where the overall rate is nearly twice as high.
The question to leave with
Not “what is our error rate.” You don’t know it and a hundred claims won’t tell you within fifteen points. The question is: if somebody hostile pulled a hundred of ours tomorrow and took the low end, what would they ask for?
You can compute that this afternoon. It is the cheapest number in the building and nobody owns it.