15%, 19%, 36%, 43%, 48%, 49%, 57%, 66%, 67% — all published in the last eighteen months, all about the same question, all currently quoted to benefits committees deciding whether to keep covering these drugs in 2027. I ran every one of them through my own checker. Nine came back green. Two of the nine survive the question that comes first, and only one of those pins an actual number.
The short version. The spread between 15% and 67% is mostly not a disagreement about employers. It is a disagreement about which employers were counted — and in four of the ten cases, about whether anything was counted at all. Two of these figures are estimates of a national population, not tallies of respondents; a weighted numerator is a sum of fractional weights, so no whole number produces it. Three more are printed beside a number that is not their denominator. One is printed beside no denominator at all.
Each row is the figure exactly as it circulates, the base printed next to it, and what happens when you try to recover the count. Press the button and your browser enumerates every integer.
| Source | Figure | Population the figure is over | Base printed beside it | Verdict |
|---|
A reachability check asks: does some whole number of employers, out of the stated base, round to the stated percentage? For 66% out of 2,010 the answer is yes — twenty different integers work, 1,317 through 1,336. Every one of them is fiction. Mercer's figure is not "1,327 of the 2,010 employers we surveyed." It is a weighted estimate for the roughly 134,000 US employers with 50 or more staff, built from a stratified probability sample plus a convenience sample of Mercer's own current and prospective clients, with statistical calibration applied to offset the bias that introduces. And it is not over the 2,010 anyway: it is over the employers in the sample with 20,000 or more staff, a count that does not appear in the release, in the prior year's release, or anywhere else I could find.
So the checker was answering the wrong question — confidently, in green. That is a defect in my tool, not in Mercer's survey, and I have fixed it: the denominator checker now asks whether the figure is a count before it does any arithmetic, and refuses to produce integers for a weighted one.
Business Group on Health publishes "67% cover for weight management" beside "105 employers." That pair is checkable, and it resolves to exactly one integer:
70 / 105 = 66.6667% → 67%
No other whole number out of 105 displays as 67%. The base is real, it is the base of that question, and the count is recoverable by any reader in a second. It is also the smallest survey of the ten. WTW's 57% also clears the check, but four different counts out of 417 display as 57%, so it narrows to a range rather than a number — and the 417 is supplied by the tracker, not by the sentence that carries the figure. Everything else fails before the arithmetic starts. The reader can pin the one number nobody would think to doubt, and cannot pin any of the eight from the larger, more respected institutions.
Ordered by how large and how advised the employers are, the numbers line up almost perfectly: 15 → 19 → 43 → 49 → 66 → 67. There is no contradiction between these surveys to resolve. There is one gradient, sliced at six different places, being read as six competing estimates of one quantity.
The tracker that collects them says the figures "reflect each organization's own definitions and are not directly comparable." That is true and it names the smaller problem. The larger one is that some of these cells are estimates of America and others are tallies of a room, and no footnote about definitions tells a reader which is which.
None of them require access to anyone's data:
KFF answers all three in public, including a 13% response rate it had no obligation to advertise. That is why its numbers are the ones a careful reader can actually use, and it is worth saying plainly, because the checking on this page otherwise reads as an attack on the people who disclose most.
My checker told me nine of ten figures were fine. I nearly published that. The check was correctly implemented and answered its question correctly; the question was the wrong one. An instrument that can only return "reachable / unreachable" will return one of those two for every input, including inputs where neither is meaningful — and it will look most confident exactly where it is least applicable. The fix was not better arithmetic. It was adding a step in front of the arithmetic and a control that proves the new step can override a green. Both are live on the checker now; press Run controls there and C6 is the one that would catch this.