How Regulatory Margins Are Chosen and Prespecified

Why Margin Choice Must Precede the Results

An equivalence analysis asks whether the observed evidence is sufficiently compatible with differences inside prespecified bounds. The bounds are not merely a technical setting for a statistical test: they define what counts as acceptable equivalence for the particular regulatory question. Choosing them after examining study results risks turning the decision into a search for a boundary that makes the result pass.

The conventional 80–125% acceptance limits are an established default for common bioequivalence settings, but they are not a universal rule for every product, endpoint, jurisdiction, or study purpose. Applicable regulations and guidance may provide general criteria, product-specific criteria, or a route for a scientifically justified proposal. The investigator’s task is to identify which requirements govern the planned study and settle any uncertainty before the comparative results are available.

A margin should be tied to the regulatory question, endpoint, and analysis scale. It is not a quantity estimated from the observed treatment difference, the study’s observed variability, or the width of its confidence interval. Those data describe the study; they do not retrospectively determine the standard against which the study is judged.

Establishing Which Criterion Applies

Margin selection begins with the governing requirements, not with the study output. The relevant jurisdiction, product, dosage form, route, endpoint, and proposed study purpose can affect which guidance applies. Product-specific guidance may add or modify criteria relative to a general standard. Where requirements are ambiguous, conflicting, or silent on a material feature, the appropriate response is to resolve the issue prospectively, potentially through regulatory consultation, rather than to select whichever plausible criterion later gives the preferred result.

A defensible selection process records the source and version of the governing requirement and explains why it applies. It also checks that the required criterion is being used for the correct endpoint and scale. A boundary expressed for one endpoint or transformation cannot automatically be transferred to another. Likewise, a general criterion should not be treated as permission to ignore a product-specific requirement.

If the applicable guidance does not directly settle a proposed margin, a sponsor may need to provide a scientific and regulatory justification and obtain agreement through the appropriate process. The justification should be developed independently of the trial’s comparative outcome. A margin motivated only by the desire to increase the chance of passing is not a defensible rationale.

What to Prespecify

Prespecification should be specific enough that an independent reviewer could reproduce the decision without guessing which rule was intended. The protocol and, where appropriate, the statistical analysis plan should identify the applicable guidance; the endpoint and analysis scale; the lower and upper bounds; the confidence level or test procedure; the analysis population and method relevant to the decision; and how the result will be judged. The chosen criterion should be consistent across these documents.

In the usual interval-based formulation, equivalence is concluded only when the prespecified confidence interval lies wholly within the prespecified margins. If the margins are \(M_L\) and \(M_U\), and the interval is \([L,U]\), the decision condition is:

This condition is meaningful only if the interval procedure is suitable for the design and estimand, and if its confidence level and analysis were also fixed in advance. It does not license changing \(M_L\) or \(M_U\) after the interval is known.

At a minimum, the planned decision relies on these conditions:

• The governing regulatory criterion has been identified correctly, including any applicable product-specific requirement.
• The margins refer to the endpoint and scale used by the prespecified analysis.
• The statistical procedure and its assumptions are appropriate for the study design and analysis population.
• The decision rule and margins are fixed before comparative results are examined.

Prespecification does not make a poorly chosen criterion valid. It makes the criterion transparent and prevents the study’s observed results from determining the standard after the fact.

Why Result-Driven Changes Are Not Harmless

A post hoc change can alter the question the analysis answers. If a confidence interval falls outside the prespecified margins, widening the margins until it fits does not provide new evidence about the treatment difference. It changes the acceptance standard after the data have revealed which direction would help.

This is also a statistical selection problem. A simple illustration uses two decision procedures that each have false-positive probability \(\alpha=0.05\) under a null condition. If the procedures were independent and a researcher selected whichever one declared success, the probability of at least one false positive would be:

Thus, in this simplified illustration, the chance is 9.75%, not 5%. Real candidate margins and analyses need not be independent, so this calculation is not a formula for every post hoc choice. It demonstrates the general issue: choosing a decision rule in response to results can invalidate the error-rate interpretation that would apply to a rule fixed in advance.

Amendments, Documentation, and Review

Not every prospective change is improper. New information may arise before results are examined—for example, a clarification of which guidance applies, a regulatory response, or discovery that the planned endpoint was mapped to the wrong criterion. A change made for a documented reason before access to comparative results can be different from a change made to rescue an unfavorable outcome. The timing, reason, approvals, and affected analyses should be recorded, and any regulatory notification or agreement required for the particular study should be followed.

If the comparative results have already been examined, changing the primary margins cannot restore the original prespecification. The original analysis remains the confirmatory assessment under the stated rule. An alternative margin may be explored only with clear disclosure of its post hoc status and rationale; it should not be presented as though it had governed the original decision.

Useful diagnostics for a margin-selection process are procedural as well as statistical. Reviewers can check whether the cited guidance actually applies, whether the protocol and analysis plan agree, whether the endpoint and scale match the margin, and whether the dates of decisions precede access to comparative results. An unexplained discrepancy, a late amendment with no independent rationale, or a margin chosen because the confidence interval nearly passes is a warning sign.

Alternatives and Limits

The principal alternatives are not interchangeable margins to try against the same result. A general regulatory standard provides a default where it governs; product-specific guidance may tailor the criterion to a particular product or question; and, where neither clearly resolves the issue, a sponsor may seek agreement on a justified proposal. The right route depends on the applicable regulatory framework. A statistical convenience, such as a margin that is easier to meet, is not an alternative regulatory basis.

Prespecification has limits. It cannot correct an inapplicable guidance choice, an unsuitable endpoint, a flawed design, or an invalid analysis. Nor does a result inside the chosen bounds establish every form of therapeutic similarity or guarantee an individual patient response. It supports the regulatory equivalence conclusion defined by that criterion, subject to the study’s assumptions and requirements.

A common mistake is to reason, “The confidence interval misses the chosen boundary by very little, so the margin should be widened.” The correction is that closeness does not change the prespecified rule: report whether the interval satisfies the original bounds, and treat any later proposal transparently as a new or exploratory question. A margin is a standard for judging evidence, not a target to be moved toward the evidence.