What an NTI Bioequivalence Result Does—and Does Not—Say
For a narrow therapeutic index (NTI) drug, differences in exposure may have important clinical consequences, so bioequivalence decisions can use tight, endpoint-specific acceptance limits. A result is meaningful only in relation to the endpoint, analysis, limits, and jurisdiction that were specified for the assessment. The central interpretive discipline is to state exactly what the result supports—and not to turn an average comparison into a claim about every patient.
In a conventional average bioequivalence assessment, the comparison concerns a population-level average for a prespecified exposure endpoint, such as AUC or Cmax. It does not directly demonstrate that each individual would have the same exposure on the two formulations, nor does it establish that every jurisdiction classifies the drug or applies the same decision criteria in the same way.
Interpreting a Confidence Interval Against Tight Limits
For an exposure endpoint analysed on the logarithmic scale, the average test-to-reference contrast is commonly summarized by a geometric mean ratio (GMR). A confidence interval for the mean log contrast is transformed back to the ratio scale by exponentiating both endpoints. The decision is then made by comparing the whole interval with the prespecified acceptance limits. Tight limits make this comparison more demanding; they do not change what the interval estimates.
The example below uses an explicitly hypothetical rule: the interval must lie within 0.95 and \(1/0.95\), or approximately 1.0526. These values are chosen only to illustrate interpretation. They are not presented as the requirements of any particular jurisdiction. The confidence interval and its limits must refer to the same endpoint and analysis scale.
Under the usual two one-sided test formulation, a 90% two-sided confidence interval corresponds to two one-sided tests at \(\alpha=0.05\), provided the specified model and inference are appropriate. That connection explains the decision rule; it does not mean there is a 90% probability that the fixed true GMR lies in the observed interval. The interval procedure has a long-run coverage interpretation under its assumptions.
Worked Example: A Tight-Limit Interval That Passes
Suppose a hypothetical, prespecified analysis of log AUC produces a log-scale estimate of \(0.01193\), with \(SE=0.01200\) and \(df=19\). Assume the analysis model is appropriate for the randomized crossover data, the stated standard error and degrees of freedom are valid, and the required interval is a 90% t confidence interval. For this illustration, the prespecified ratio limits are 0.9500 and \(1/0.9500=1.0526\).
For a two-sided 90% interval with 19 degrees of freedom, use \(t_{0.95,19}=1.729\). The log-scale interval is
Exponentiating the endpoints gives the interval on the GMR scale:
Both endpoints are inside the hypothetical limits \([0.9500,1.0526]\). Under the stated assumptions and this illustrative rule, the interval supports average bioequivalence for AUC. The appropriate conclusion is limited to that endpoint and criterion. It does not assert that every individual has equal AUC on the two formulations.
Average Equivalence Is Not Individual Interchangeability
An average contrast can be close to zero on the log scale even when individual contrasts vary substantially around that average. This is not a contradiction: the average and the spread of individual differences are distinct features of a distribution. An average bioequivalence interval addresses uncertainty about the average contrast under the specified analysis; it does not, on its own, place a bound on every person’s formulation-specific difference.
This distinction matters particularly for NTI medicines. A small average difference may be reassuring about the population mean and still leave questions about variability, clinical monitoring, or an individual patient’s response. Those questions require evidence and estimands that address them directly. They cannot be answered simply by rephrasing an average bioequivalence pass as “the products are interchangeable for everyone.”
Worked Example: Opposing Individual Differences Can Average to Zero
Consider a deliberately simplified hypothetical population in which half of individuals have a log test-to-reference exposure contrast of \(+0.20\), and half have a contrast of \(-0.20\). This construction is used to distinguish an average from individual differences; it is not a claim about any real drug or study.
The mean log contrast is
The corresponding geometric mean ratio is \(\exp(0)=1.000\). But the two individual-level ratios in this simplified example are
Thus, the population average contrast can be zero while the individual contrasts are not zero. A sufficiently precise estimate of the average could yield a narrow interval around a GMR of 1, yet that finding would not erase the differences represented in this example. The correction is to report an average-equivalence result as an average-equivalence result, and not to infer individual interchangeability from it alone.
Regulatory Criteria Are Jurisdiction-Specific
A clinical concern about a drug’s narrow therapeutic index and a regulator’s operational classification or bioequivalence requirements are related but are not the same statement. Regulatory definitions, lists, endpoint choices, study requirements, and acceptance limits can differ across jurisdictions and can change over time. Therefore, “this is an NTI drug” does not, by itself, establish a single worldwide set of bioequivalence rules.
A defensible interpretation names the jurisdiction and the applicable, current requirement. It also identifies the endpoint or endpoints to which the criterion applies. A finding under one jurisdiction’s specified criterion should not be reported as automatic evidence that another jurisdiction’s requirements have been met. Nor should a jurisdiction-specific regulatory decision be presented as a universal clinical statement about every patient.
Worked Example: The Same Interval Under Two Hypothetical Rules
Suppose a hypothetical study estimates a GMR of 1.0100 for an endpoint, with a 90% confidence interval of \([0.9480,1.0560]\). For illustration only, compare this fixed interval with two hypothetical acceptance ranges: Rule A uses \([0.9500,1.0526]\), while Rule B uses \([0.9000,1.1111]\). These ranges are not attributed to any real regulator.
Under Rule A, the lower interval endpoint, 0.9480, is below 0.9500, and the upper endpoint, 1.0560, is above 1.0526. The interval is not wholly inside the range, so it does not meet Rule A’s stated criterion. Under Rule B, both 0.9480 and 1.0560 fall inside the wider limits, so the interval meets that illustrative criterion.
The data and interval have not changed; the decision differs because the prespecified rules differ. In practice, the applicable jurisdiction’s actual current requirement must be identified and applied. The example does not license choosing whichever rule gives a preferred answer after seeing the result.
Observed Variability Does Not Authorize Wider Limits
Acceptance limits are part of the decision rule, not a tuning parameter to adjust after examining the study. A study may show more variability than anticipated, resulting in a wider confidence interval or reduced precision. That outcome can make it harder to demonstrate that the full interval lies within a tight range. It does not, on its own, justify widening that range.
This is especially important because limits chosen after looking at the results can make the conclusion depend on the observed data in an uncontrolled way. If variability is high, the appropriate response is to report it, investigate whether the design and analysis assumptions were met, and follow the prespecified decision criteria. Any legitimate change to regulatory requirements must be justified independently and prospectively under the relevant process, not improvised to rescue a particular result.
Worked Example: Why a Post Hoc Limit Change Does Not Rescue a Result
Use the interval from the preceding example, \([0.9480,1.0560]\), and suppose Rule A’s prespecified limits are \([0.9500,1.0526]\). The interval fails Rule A because
Now suppose the study team observes substantial variability and proposes changing the limits after seeing the interval to \([0.9400,1.0600]\). The interval would fit within those newly chosen limits, but that arithmetic does not make the modified rule valid. The result remains a failure under the original prespecified Rule A. The post hoc comparison may be described as an exploratory calculation, but it cannot replace the planned decision or support a confirmatory pass.
The correct interpretation separates two facts: the observed interval did not meet the prespecified criterion, and the study showed the variability and uncertainty reflected in that interval. Neither fact is removed by changing the limits after the analysis.
Diagnostics, Alternatives, and Limits of the Conclusion
Before interpreting an NTI bioequivalence decision, check that the comparison being reported is the one the protocol specified. Confirm the endpoint, analysis scale, confidence level, acceptance limits, model, and handling of incomplete data. Verify that the interval is calculated from the stated analysis and that its transformed endpoints are correctly compared with the ratio-scale limits. If more than one endpoint must pass, a pass for one endpoint alone does not establish that the joint requirement has been satisfied.
Warning signs include reporting only the point estimate, giving limits without naming the jurisdiction or endpoint, switching from the planned interval to a different analysis after seeing results, or treating an interval that overlaps a limit as a pass. Another warning sign is a claim about each patient based only on an average contrast. Such statements exceed what the reported average analysis establishes.
The main alternative to an average-level question is not a looser interpretation of the same average result, but an analysis designed to answer a different question. Individual-level interchangeability, subject-specific differences, or clinical outcomes under monitoring are distinct targets and may require different evidence. They should not be substituted for the prespecified average bioequivalence estimand after results are known. Conversely, an average bioequivalence assessment remains useful for its intended population-average question; its limitation is scope, not that the calculation is meaningless.
Common Misinterpretations
- “The GMR is near 1, so bioequivalence is established.” A point estimate alone is insufficient; the required interval must satisfy the prespecified limits.
- “Average bioequivalence means every patient will have equivalent exposure.” The average result concerns a population-average contrast, not a guarantee about each individual.
- “NTI status means the same limits apply everywhere.” The clinical concern does not eliminate jurisdictional differences in regulatory classification and requirements.
- “The study was variable, so wider limits are reasonable now.” Variability may widen uncertainty, but does not authorize changing the planned limits after observing the data.
- “A pass under one rule proves a pass under every rule.” Each conclusion must be evaluated against the applicable jurisdiction’s specified criteria.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A GMR estimate is inside its limits, but the confidence interval extends beyond the lower limit. Has the interval criterion been met? Explain.
- In a hypothetical population, half the log exposure contrasts are \(+0.10\) and half are \(-0.10\). Find the mean log contrast and explain what it does—and does not—say about individuals.
- Why can an NTI drug’s regulatory classification or bioequivalence requirements not be assumed to be identical in every jurisdiction?
- A study fails its prespecified tight limits, and the team widens them after observing high variability. What is the valid conclusion under the original rule?
Solutions. 1. No. The whole required interval must lie within the acceptance limits; an in-range estimate does not compensate for an interval endpoint outside them. 2. The mean is \(\frac{1}{2}(0.10)+\frac{1}{2}(-0.10)=0\), so the corresponding average ratio is \(\exp(0)=1\). Individual contrasts remain \(+0.10\) or \(-0.10\); the mean does not establish individual equality. 3. Jurisdictions can define classifications and specify requirements differently, so the applicable current rule must be identified for each regulatory context. 4. The study did not meet the original prespecified criterion. The post hoc wider limits do not change that result; variability and uncertainty should be reported and investigated, not used to rewrite the decision rule.