Half the price rides on a number whose denominator is missing

Sword Health ties 50% of program cost to a member-reported outcome. This page recomputes the published evidence behind that outcome — in your browser, from the sources, with no data from me. Free. Nothing to install, nothing to buy.

Built 19 Aug 2026 · sources: swordhealth.com, api.crossref.org, mdpi.com · every check below runs on your machine and can be repeated by hand

1. What is being priced

“In this whitepaper, we outline how Sword links 50% of program cost to a clinically-validated, member-reported outcome.” swordhealth.com/reports-and-guides/outcome-pricing-claims-validated · 3 Mar 2026

The pricing page that announces the model backs it with five outcome claims and five footnotes. Two of those footnotes carry the surgery and painkiller claims — the two that map most directly onto an employer’s downstream spend:

“47% decrease in the use of prescribed painkillers³ · 50% reduction in costly surgeries⁴”
³ Janela D, Costa F, Molinos M, et al. Long-term outcomes of a fully remote digital musculoskeletal care program. Healthcare (Basel). 2022;10(8):1595. https://doi.org/10.3390/healthcare10081595
⁴ Janela D, Costa F, Molinos M, et al. Impact of digital musculoskeletal care on surgery intent and healthcare utilization. Healthcare (Basel). 2022;10(8):1595. https://doi.org/10.3390/healthcare10081595 swordhealth.com/newsroom/outcome-pricing · footnotes 3 and 4, retrieved 19 Aug 2026

Two different titles. One DOI. Press the button: your browser asks Crossref what that DOI actually is, and what the other three are.

2. The responder rate cannot be over the population it names

Follow the DOI to the paper. Its abstract reports the headline that a pay-for-outcome trigger would most resemble — the share of people who improved enough to count:

“A completion rate of 74.2% (396/534) … with 66.8% treatment responders considering pain.”

And in the results, the same figure with a qualifier the abstract drops:
Among completers, 66.8% of participants surpassed the MCIC of 30% reduction in pain.” Janela et al., Healthcare 2022;10(8):1595 — abstract and §3.2.2

534 people started. 396 finished. So “among completers” should mean 66.8% of 396. It cannot. No whole number of people out of 396 rounds to 66.8% — the fraction jumps 66.7% → 66.9% and skips it. Same for 534. Your browser can enumerate every possibility in a millisecond:

The third denominator in that test is not a guess pulled from air. It is printed in the same section of the paper, in the degrees of freedom of the correlation run on the same pain variable — r(191), which for a Pearson correlation means n = 193. And 129 of 193 is 66.84%, which rounds to 66.8%.

So the number that anchors “meaningful improvement” is computed over roughly 193 people out of the 534 who started — about 36% of them. The other ~64% are not failures in the figure. They are not in the figure.

What the payer is actually buying, at each denominator

DenominatorPeopleResponder shareWho is excluded
as published~19366.8% everyone who dropped out, plus everyone missing a pain reassessment
completers396≈ 32.6% the 138 who dropped out
everyone enrolled534≈ 24.2% nobody

Lower two rows carry the same ~129 responders over the larger populations — the arithmetic the button prints. If a share of the missing did respond, the true figure sits between; the point is that the published 66.8% is the ceiling, not the estimate, and the page that sets the price does not say so.

3. “50% reduction in costly surgeries” is not what the cited paper measures

The paper behind footnote 4 does not count surgeries. It asks a question:

“Self-reported surgery intent assessed by the question ‘How likely are you to seek surgery to address your condition in the next 12 months?’ (range 0 (not likely)–100 (extremely likely)).” Janela et al. 2022, §2 Methods — secondary outcomes

The reported movement on that question is 70.1%, in a single-arm cohort with no control group, of people with chronic hip pain. There is no 50% in it, no surgery count in it, and no comparison group in it. An employer reading “50% reduction in costly surgeries” is reading a claim about operations. The citation offered leads to a claim about willingness.

4. The controls — because an instrument that cannot fail proves nothing

Both buttons above would be worthless if they printed a verdict regardless of input. Each carries a control that must come out the other way, and the page marks itself RED if a control agrees with it:

5. If you pay on this number

If you are a benefits leader, a stop-loss carrier, a health plan or a consultant with 50% of a contract riding on a member-reported outcome, the question to put in writing is one line long:

Over which population is the outcome rate computed — everyone enrolled, everyone who completed, or everyone who answered the survey — and what happens to a member who stops answering?

A vendor whose engine is honest answers it in a sentence. A number that only exists over its survivors will not survive the question. That gap is the whole of what I do: I take a published number somebody is paid on, recover the population it was computed over, and hand back the recomputation with the arithmetic exposed — so the person writing the cheque can repeat it without me.