Non-inferiority margin and equivalence trials

A non-inferiority trial does not ask whether the new option is better. It asks whether it is no worse than the standard by more than a pre-specified margin, usually because the new option is shorter, cheaper, less toxic or easier to deliver. The margin is written into the protocol before anyone is enrolled and is expressed either as an absolute difference (for example a few percentage points of five-year recurrence) or as a limit on the hazard ratio (for example the upper end of the confidence interval must stay below 1.25). The trial succeeds if the whole confidence interval for the difference sits on the acceptable side of the margin. An equivalence trial goes one step further and requires the interval to sit inside a margin on both sides, which is how biosimilars are shown to match their reference antibody.

Worked examples in the corpus are the de-escalation trials. PERSEPHONE randomised 4,088 women to six or twelve months of adjuvant trastuzumab and met its non-inferiority margin on four-year disease-free survival (89.4 percent against 89.8 percent). FAST-Forward randomised 4,096 women to 26 Gy in five fractions over a week or 40 Gy in fifteen fractions and met its margin on five-year local relapse (1.4 percent against 2.1 percent). PACE-B did the same for five-fraction stereotactic radiotherapy in prostate cancer, TAILORx for omitting chemotherapy in mid-range recurrence scores, and DYNAMIC for using a blood test to halve chemotherapy use in stage II colon cancer. In each case the margin, not the point estimate, is what the guideline committees debated.

The traps are worth knowing. A wide margin makes almost anything look non-inferior, so a stated margin should be justified from what the standard treatment adds over no treatment and from what patients would accept losing. A sloppy trial (poor adherence, wrong population, patients lost to follow-up) drifts towards no difference, which in a superiority trial is a failure but in a non-inferiority trial looks like success, so per-protocol analyses are reported alongside intention-to-treat. And a superiority trial that misses is not a non-inferiority result: CONVERT was designed to show once-daily radiotherapy was superior in limited-stage small-cell lung cancer, did not, and is often quoted as if it had proved the two schedules equivalent, which it was not powered to do.