A survey of 105 employers reports that 67% of them cover GLP‑1s for weight management, and that 72% of those covering intend to keep the coverage in 2027. The first two numbers fix the covering group at exactly seventy employers. No whole number of employers out of seventy rounds to 72%. Read the very same figures under truncation and the covering group becomes seventy‑one rather than seventy — and 72% is unreachable there too. Three conventions, three covering groups, no 72% in any of them. The base of the 72% is therefore not the group the sentence names, and a reader has no way to learn what it is. Everything below is arithmetic, and it runs in your browser.
The population, from the release’s own methodology note:
“The survey, which examined coverage of GLP-1s specifically for weight management, was completed in February-March 2026 by 105 employer members of Business Group on Health.”businessgrouphealth.org — 2026 GLP-1 survey press release
The first rate, over that population:
“67% of surveyed employers currently cover GLP-1s for weight management…”same release, body and subhead
The second rate, over the subgroup the first rate just defined:
“Of those covering GLP-1s for weight management, only 72% said they were likely to continue that coverage in 2027, while 10% said they likely would not…”same release
And this is how the pair travels. An independent tracker sets the rate beside its base, in one row of one table, so that readers can compare it against KFF, Mercer, WTW and SHRM:
“Business Group on Health — 2026 — 67% cover for weight management — Large, mostly self-funded member employers — 105 employers; fielded Feb–Mar 2026”therxindex.com/research/glp1-employer-coverage-tracker/ — GLP-1 Employer Coverage Tracker
Putting the base next to the rate is the right thing to do, and it is what makes the defect visible. The two rates in that row cannot both be about the populations they name.
The covering group is not stated as a count anywhere. It has to be recovered from the 105 and the 67%. That recovery is exact: only one whole number of employers out of 105 displays as 67%.
Now the exhaustive part. Take the covering group recovered above and walk every possible answer — nought employers, one employer, two, all the way to the whole group — and display each as a percentage the way the release does. If 72% is a real answer from that group, it appears in the list. Here is the whole list, so nothing is taken on trust.
Run the question backwards. Of every possible base from 1 to 105, which ones can produce both of the subgroup figures the release prints — a 72% and a 10%? The answer is a short list, and the recovered covering group is not on it. The closest admissible base sits just below it.
There is exactly one ordinary explanation, and it is the one the release does not give. If some of the covering employers skipped the follow-up question, the base of the 72% is smaller than the covering group, and the arithmetic closes. The step above prints how much smaller it would have to be. That is a fact about the survey the reader is not told: no per-question base is published, the release links no report, and the underlying survey is not public.
So the honest statement is not that the number is wrong. It is that the number cannot be recomputed by anyone outside the survey, and that the sentence naming its base names a base it cannot have. Anyone placing that rate in a comparison table beside sample sizes is putting a weight on a figure whose denominator is unknown to them.
An instrument that answers “no” to everything proves nothing. These controls check the instrument itself: each one names the answer it must give, including the four other rates in the same release — which all reconcile cleanly against 105, as they should.