Check the Direction of a Noninferiority Margin

Write the unfavorable direction first

Noninferiority asks whether a new intervention is not worse than a comparator by more than a prespecified amount. Which interval endpoint matters depends on whether larger outcomes are beneficial or harmful and how the contrast is coded. Draw the scale before calculating a decision.

FDA’s noninferiority guidance addresses margin selection and conditions needed for an interpretable trial. A numerical margin cannot be justified solely because it produces a feasible sample size or contains the observed confidence interval.

Connect the margin to the estimand

Record population, outcome, assessment time, intervention comparison and effect scale. Explain the historical evidence and clinical reasoning supporting the tolerated loss, including whether the active comparator’s effect is expected to persist in the current setting.

Specify missing-outcome and nonadherence handling because analyses that dilute differences can make an unfavorable intervention look similar. Preserve the planned analysis populations and sensitivity analyses. A noninferiority label does not remove the need to assess assay sensitivity or trial conduct.

Compare hypothetical benefit outcomes

Suppose higher response probability is beneficial, and the effect is new treatment minus control. The prespecified margin is −5 percentage points. A compatible confidence interval of [−3, 4] percentage points stays above the unfavorable boundary and meets that interval-based noninferiority criterion. An interval of [−7, 2] does not.

Now recode the contrast as control minus new treatment. The corresponding intervals become [−4, 3] and [−2, 7], and the unfavorable boundary becomes +5. The same scientific conclusion now uses the upper endpoint. Keeping the old negative margin after reversing the contrast would change the question by mistake.

Report beyond a binary conclusion

State the confidence level and its relationship to the prespecified one-sided test. Explain that noninferiority does not mean equal effects, superiority or acceptable safety. A new intervention can satisfy an efficacy margin while introducing other harms that require separate evaluation.

If the criterion is not met, describe the evidence as insufficient for the planned noninferiority conclusion rather than proof that the new treatment is worse. Keep the effect estimate and full interval visible. The final result should make the direction, margin justification, analysis assumptions and unresolved limitations understandable without requiring readers to infer which endpoint was compared with which threshold.

Source verification: 6 November 2026. Examples are hypothetical.