Non-inferiority trials explained: what the margin means
Non-inferiority trials explained: what the margin is, how to read the confidence interval, and why "not worse" is a narrower claim than it sounds.
By Manouchehr Hessabi, MD, MPH
A non-inferiority trial is a randomized trial designed to show that a new treatment is not worse than an established one by more than a prespecified amount, called the margin. It does not show that the two treatments are the same, and it does not show that the new one is better. The entire conclusion depends on where that margin was set and on assumptions the trial itself cannot check.
Readers often hear "non-inferior" as "just as good." This explainer sets out what the claim actually is, how the margin is chosen and read, and why the design asks more of the reader than a conventional trial. It is educational and not a substitute for personal medical advice.
What is a non-inferiority trial?
A few terms first. The comparator, or active control, is the established treatment the new one is tested against. A superiority trial asks whether one treatment is better than another, or better than a placebo (an inactive treatment). A non-inferiority trial asks a different, one-sided question: is the new treatment unacceptably worse than the comparator?
The US Food and Drug Administration's final guidance on the design, Non-Inferiority Clinical Trials to Establish Effectiveness (CDER and CBER, November 2016), puts the distinction plainly. Such trials were once called clinical equivalence trials, but their intent "is not to show that the new drug is equivalent, but rather that it is not materially worse than the control."
Why run such a trial at all? The FDA guidance gives the usual reason as an ethical one. When an effective treatment already exists that prevents death or irreversible harm, it would not be ethical to give some participants a placebo instead. The new treatment therefore has to be compared with the existing one. The same guidance notes that a somewhat larger margin may be justified if the new treatment offers some other important advantage, for example in safety.
What is the non-inferiority margin, and who chooses it?
The margin is chosen by the investigators before the trial begins, and it is the single most important number in the report. The FDA guidance distinguishes two versions of it:
- M1 is "the entire effect of the active control assumed to be present in the NI study," estimated from earlier trials of the control, usually against placebo.
- M2 is "the largest clinically acceptable difference (degree of inferiority) of the test drug compared to the active control."
The logic runs like this. If a trial rules out a loss larger than M1, it shows the new treatment has at least some effect. But most investigators want more than that: they want to show the new treatment preserves most of the benefit of the established one. So the margin actually used, M2, is usually smaller than M1. The guidance is explicit that M2 is "a matter of clinical judgment" but "can never be greater than M1."
Because M1 is estimated from historical trials rather than measured in the current one, it rests on what the guidance calls the constancy assumption: that the control works about as well today, in this population, as it did in the earlier studies. If standard care around the control has changed, that assumption may not hold, and the guidance notes that M1 sometimes has to be "discounted" to a smaller value.
How do you read the confidence interval against the margin?
A confidence interval is the range of treatment differences reasonably compatible with the data. Most trials report a 95% interval. In a non-inferiority trial, the question is simple: does the whole interval stay on the acceptable side of the margin?
Picture the margin as a line drawn on a number line of possible differences. If the end of the interval nearest to "new treatment worse" stays clear of that line, non-inferiority is shown. If the interval crosses it, the trial has not ruled out an unacceptable loss.
The direction of the arithmetic varies between reports, which is a common source of confusion. The FDA guidance writes the difference as control minus test, so non-inferiority depends on the upper bound. Many trials write it as new minus established, so the same logic applies to the lower bound. Readers should always check which way the difference is defined before reading the numbers.
A worked example: the AcT trial
The AcT trial, reported in The Lancet by Menon and colleagues (2022), compared two clot-dissolving drugs for acute ischaemic stroke at 22 Canadian stroke centres. Its primary outcome was the proportion of patients with a modified Rankin Scale score of 0 or 1 at 90 to 120 days, meaning no symptoms or no significant disability; the modified Rankin Scale is explained separately.
The prespecified rule was that non-inferiority would be met if the lower bound of the 95% confidence interval for the difference (tenecteplase minus alteplase) was above minus 5 percentage points. The result:
- 296 of 802 patients (36.9%) on tenecteplase reached the good outcome.
- 266 of 765 patients (34.8%) on alteplase reached it.
- The unadjusted risk difference was 2.1 percentage points, with a 95% confidence interval from minus 2.6 to 6.9.
The lower bound, minus 2.6, sits above the margin of minus 5, so the trial met its non-inferiority threshold. Read carefully, that says the data are compatible with tenecteplase being up to 2.6 points worse or up to 6.9 points better, and that a loss as large as 5 points was ruled out. Because the interval also includes zero, the result does not show tenecteplase was better. The clinical context, including the other trials and current guidelines, is covered in tenecteplase versus alteplase trials; this example is used here only to show how a margin is read, not as treatment guidance.
Why are non-inferiority trials harder to interpret than superiority trials?
The FDA guidance identifies the central problem. A well-run superiority trial that finds a difference is "entirely interpretable without further assumptions (other than lack of bias)." A non-inferiority trial is not. It depends on something the trial does not measure: that the control actually had its expected effect in this study. The guidance calls this assay sensitivity, the ability of the trial to have shown a difference from placebo of a specified size. If the control did not work in the trial, then even a very small difference between the two arms "is meaningless and provides no evidence that the test drug is effective."
The second problem is that poor conduct pushes in the wrong direction. In a superiority trial, sloppiness tends to blur the arms together and make it harder to find a difference, a bias toward the null. In a non-inferiority trial, the same blurring makes the arms look alike, which is exactly what the trial is trying to show. The guidance calls this a bias "toward the alternative" and lists the usual culprits: poorly implemented entry criteria, poor adherence, overlapping concomitant treatments, inadequate measurement, errors in delivering assigned treatments, high attrition and poor follow-up. As the guidance puts it, "In a superiority trial, sloppiness can lead to study failure. In contrast, poor quality in an NI trial can sometimes lead to an apparent finding of non-inferiority that is incorrect."
This affects how the analysis population is chosen. In an intention-to-treat analysis, every participant is analyzed in the group they were randomized to, whether or not they received that treatment. That approach is conservative in a superiority trial. The FDA guidance cautions that intention-to-treat approaches "are not necessarily conservative in an NI study and can lead to an incorrect finding of non-inferiority." The logic of intention-to-treat analysis is covered in more depth separately.
What should a well-reported non-inferiority trial tell you?
Because so much depends on choices made before the data arrive, reporting matters. The CONSORT statement (Consolidated Standards of Reporting Trials) is the standard checklist for reporting randomized trials, and its primary focus is two-arm superiority trials. A dedicated extension for non-inferiority and equivalence trials, updated to the 2010 version of CONSORT by Piaggio and colleagues in JAMA (2012), exists so that readers can "assess the reliability of their results and conclusions."
Drawing on that extension's purpose and on the FDA guidance, a careful reader can look for a few things in any non-inferiority report:
- The margin and its rationale, stated in advance. A margin chosen or changed after the data are seen cannot support the conclusion.
- Where the margin came from. Ideally the historical evidence for the control's effect, and some reasoning about whether that effect still holds.
- The confidence interval shown against the margin, with the direction of the difference made clear.
- Which analysis populations were used, and whether the conclusion holds across them.
- Signs of trial quality: adherence, losses to follow-up and protocol deviations, since these bias toward non-inferiority.
Questions readers ask about non-inferiority trials
Does non-inferior mean the treatments are the same? No. It means the new treatment was not shown to be worse than the comparator by more than the prespecified margin.
Can a trial be both non-inferior and superior? Yes, if the confidence interval clears both the margin and the line of no difference. The FDA guidance notes such a finding can be interpreted without adjustment for multiple testing, and that beating an active drug is "much more difficult than demonstrating superiority to placebo."
Who sets the margin? The investigators, before the trial, from the estimated effect of the established treatment and clinical judgment about how much loss is acceptable.
Why not just run a superiority trial? Sometimes the goal is not to beat the existing treatment but to offer an alternative with other advantages, and a placebo comparison would be unethical when an effective treatment already exists.
Readers interested in how study design shapes the evidence behind clinical and public-health decisions can find related methodological work on the peer-reviewed publications page.