POET II (NEJM 2026): non-inferiority sample size only reproduces at one-sided alpha 0.05, but the paper states a one-sided 97.5% CI. Am I missing something?
I'm preparing POET II for a journal club (open-label RCT of response-tailored vs standard antibiotic duration in left-sided endocarditis, NEJM, Aug 2026) and can't reconcile the sample size with the stated confidence level.
Stated in the paper and supplement: safety endpoint, non-inferiority, assumed event rate 8%, margin 7.5 points, 90% power gives 450 patients, "with a one-sided confidence interval of 97.5%". Per pathogen: margin 12.5 points, 80% power gives 118 patients.
My check, using n per arm = (Zα + Zβ)² × 2p(1−p) / Δ², with p = 0.08 and no true difference:
| Overall (Δ 7.5, 90% power) | Per pathogen (Δ 12.5, 80% power) | |
|---|---|---|
| One-sided alpha 0.05 | 224 per arm (~450) | 59 per arm (118) |
| One-sided alpha 0.025 | 275 per arm (~550) | 74 per arm (~148) |
Only alpha 0.05 reproduces the stated numbers. The 2020 design paper (margin 5%, 80% power, "alpha 5%") gives 730 patients, which also fits one-sided 0.05 (about 924 at 0.025). The results report two-sided 95% CIs (equivalent to one-sided 97.5%), so the analysis is stricter than the calculation. By my numbers, power is then about 83% overall (about 88% with the 508 randomised) and about 70% per pathogen.
Questions:
- Is there a standard variant (continuity correction, variance under the null, software default) that gives 450 and 118 at one-sided 0.025? I couldn't find one.
- If it is just a labelling slip, would you treat it as a minor wording issue or a real power issue in an appraisal?