Why a Decision Near the Boundary Needs More Than One Check
In reference-scaled bioequivalence, a result can look close to passing for different reasons. The estimated test-to-reference mean ratio may be near its permitted point-estimate limit; the estimated reference within-subject variability may be close to the threshold at which scaling begins; or the uncertainty in one or both estimates may be large enough that a point calculation gives a fragile impression. These are distinct features of the analysis, not interchangeable explanations for a decision.
The central interpretive task is therefore to separate three questions: does the mean ratio satisfy its constraint, does the variability-based criterion satisfy its rule, and how much uncertainty is present in the estimates used by that rule? The exact regulatory procedure is jurisdiction- and analysis-specific. The simplified calculations below clarify the geometry of a boundary; they are not substitutes for the prescribed test.
Three Components of the Interpretation
For a common point-estimate constraint, the estimated geometric mean ratio must lie within 0.80 to 1.25. On the log scale this is equivalent to \(\ln(0.80)\leq d\leq\ln(1.25)\); because \(\ln(0.80)=-\ln(1.25)\), it can also be written as \(|d|\leq\ln(1.25)\). This constraint is about the estimated mean ratio. It does not establish that the variability-based criterion passes.
The variability-based criterion uses reference within-subject variability according to a prespecified rule. To illustrate the boundary without reproducing any one jurisdiction's full procedure, consider a generic capped scaled width \(w(s_{WR})=\begin{cases}c,&s_{WR}\leq s_0\\ \min(c,ks_{WR}),&s_{WR}>s_0\end{cases}\), where \(k>0\) and \(c>0\) are fixed constants. The point-estimate geometry would compare \(|d|\) with \(w(s_{WR})\). In an actual regulatory analysis, the prescribed inferential procedure may use confidence bounds or a criterion involving both the contrast and variance estimate. A simple comparison of point estimates is not enough to declare regulatory bioequivalence.
The third component is uncertainty. The contrast \(d\) is estimated, and \(s_{WR}\) is estimated. Near a boundary, uncertainty in either may matter: uncertainty in \(d\) affects evidence about the treatment difference, while uncertainty in \(s_{WR}\) affects how much scaling the rule permits and whether the scaling trigger is supported. A threshold is not made certain merely because the estimated variability lies just above it.
The capped form makes two limits visible. Below the trigger \(s_0\), the applicable criterion does not expand with variability; above it, the permitted width may increase, but only up to a cap. The mean-ratio constraint remains in force even when the variability-based width is wider. The cap and trigger values, and the way uncertainty enters the decision, must come from the governing method rather than be chosen after seeing the data.
Why a Point Estimate Can Mislead at the Boundary
Suppose the relevant variability estimate is only slightly above its trigger. A plug-in calculation may then give a scaled width only slightly larger than the unscaled width. A small change in \(s_{WR}\)—from sampling variation, a model choice, or an influential observation—can move the estimate across the trigger. This does not mean that the true variability abruptly changes at the boundary. It means that a rule uses a threshold to determine which criterion applies, while the quantity compared with that threshold is estimated with uncertainty.
The distinction can be seen algebraically. For the illustrative rule, if \(s_{WR}<s_0\), increasing \(s_{WR}\) does not increase the permitted width if the rule remains unscaled below the trigger. If scaling applies and the cap has not been reached, the width is \(ks_{WR}\), so its local change is \(k\) times the change in \(s_{WR}\). Once \(ks_{WR}\geq c\), further increases do not widen the criterion. Thus, even in this simple geometry, the effect of a variability estimate depends on which side of the trigger and cap it lies.
This derivation explains why two studies with similar treatment contrasts can have different point-estimate geometry if their estimated reference variability differs. It does not establish the uncertainty of the actual decision: that requires the confidence-bound or other inferential procedure specified for the analysis. In particular, an interval for \(d\) alone cannot represent uncertainty in \(s_{WR}\), and an interval for \(s_{WR}\) alone cannot represent uncertainty in \(d\).
Worked Example: A Mean-Ratio Constraint Still Applies
Consider a hypothetical analysis using the illustrative capped geometry above. Let \(d=0.2220\), \(s_{WR}=0.296\), \(k=0.760\), and \(c=\ln(1.25)=0.22314\). Suppose, for this illustration only, that the scaling trigger is \(s_0=0.294\). Check the point-estimate constraints.
Mean-ratio constraint. The estimated ratio is \(\exp(0.2220)=1.2486\) (approximately), which is below 1.25 and above 0.80. Thus the point estimate satisfies this constraint, though it is close to the upper limit.
Illustrative variability-based geometry. Since \(0.296>0.294\), the estimate is above the illustrative trigger. The uncapped width is \(ks_{WR}=0.760(0.296)=0.22496\), which exceeds \(c=0.22314\). The capped width is therefore \(w(s_{WR})=0.22314\). Since \(|d|=0.2220\leq0.22314\), the point-estimate geometry also lies within this illustrative width.
This is not a regulatory pass conclusion. It is a point-estimate check under stipulated numbers. A real conclusion requires the correct jurisdiction-specific rule and its inferential calculation. The example also shows why “the scaled criterion passes” does not make the mean-ratio constraint redundant: both must be assessed separately.
Worked Example: A Small Change Across the Trigger
Use the same illustrative constants \(s_0=0.294\), \(k=0.760\), and \(c=0.22314\), with a fixed contrast \(d=0.2220\). Compare two hypothetical estimates of reference variability, \(s_{WR}=0.292\) and \(s_{WR}=0.296\). This comparison illustrates boundary sensitivity, not an actual regulatory decision.
At \(s_{WR}=0.292\), the estimate is below the trigger, so the scaled branch is not activated. Under a simple unscaled-width comparison using \(c\), \(0.2220\leq0.22314\), so the contrast is within that width. At \(s_{WR}=0.296\), the estimate is above the trigger; the capped width is \(\min(0.760(0.296),0.22314)=0.22314\), and the contrast is again within it. These numbers deliberately show that crossing a trigger need not change the binary point-geometry result.
Now take \(d=0.2238\) instead. The mean-ratio constraint fails because \(0.2238>0.22314\), regardless of whether the variability estimate is 0.292 or 0.296. This illustrates why a change in scaling status cannot rescue a failed point-mean constraint in a method that retains that constraint.
Finally, suppose an approximate interval for \(s_{WR}\) were reported as 0.247 to 0.345. It spans \(s_0=0.294\), so the data do not locate the variability estimate clearly on one side of the trigger. Such an interval should be calculated using a method appropriate to the variance estimator and its degrees of freedom; it should not be inferred from a symmetric normal approximation by default. The appropriate conclusion is that the boundary classification is uncertain, not that the point estimate itself is invalid.
Uncertainty in the Treatment Contrast
A confidence interval for a log-scale contrast is often translated to the ratio scale by exponentiating both endpoints. This is a monotone transformation: if \(L\leq d\leq U\), then \(\exp(L)\leq\exp(d)\leq\exp(U)\). The resulting ratio-scale interval is generally asymmetric around the estimated ratio, even when the log-scale interval is symmetric around \(d\). Keep full precision through the endpoint calculations and round only the reported results.
Worked Example: Translating a Log-Scale Interval Near the Limit
Suppose a hypothetical estimated log treatment contrast is \(d=0.03922\), with an approximate standard error \(SE=0.0900\). For an illustrative 90% normal interval, use the critical value 1.645. The standard error and critical value are assumed appropriate for this example; this is an interval illustration, not a complete scaled bioequivalence test.
The estimated ratio is \(\exp(0.03922)=1.0400\) (approximately). The log-scale interval is \(0.03922\pm1.645(0.0900)\). The margin is \(1.645(0.0900)=0.14805\), so the endpoints are \(0.03922-0.14805=-0.10883\) and \(0.03922+0.14805=0.18727\). Rounded to four decimal places, the log-scale interval is \([-0.1088,\,0.1873]\).
Exponentiating the unrounded endpoints gives \(\exp(-0.10883)\approx0.8969\) and \(\exp(0.18727)\approx1.206\). The approximate ratio-scale interval is therefore \([0.8969,\,1.206]\). It contains 1.00 and lies within 0.80 to 1.25. In the usual confidence-interval interpretation, the method used to construct this 90% interval has 90% coverage in repeated sampling under its assumptions; it is not correct to say there is a 90% probability that the fixed parameter lies in this particular interval.
This interval informs uncertainty in the mean contrast. It does not itself incorporate uncertainty in the reference variance estimate or replace the scaled procedure's prescribed decision rule. A reviewer should therefore report the interval with its scale and method, then interpret it alongside—not in place of—the variability-based criterion.
Assumptions, Diagnostics, and Failure Modes
Interpretation is meaningful only if the analysis corresponds to the design and model. The endpoint should be analyzed on the specified scale, typically the log scale for pharmacokinetic measures such as AUC or \(C_{\max}\). The replicate design must provide the repeated reference observations needed to estimate reference within-subject variability. The model must represent the treatment, period, sequence, and subject structure required by the design, and the variance component used for scaling must be the reference within-subject component—not total variability or an unrelated residual quantity.
The inferential procedure also depends on its distributional and model assumptions. Independent subjects, an appropriate covariance structure, and suitable residual behavior matter; the relevant degrees of freedom must correspond to the variance estimator and method. Review residual diagnostics and influential-observation checks, and examine whether conclusions are sensitive to defensible variance-model choices. A visually unusual observation is not automatically grounds for exclusion, but a result that changes materially when one observation or reasonable model specification changes deserves explicit investigation.
Near a scaling boundary, inspect and report the estimate of \(s_{WR}\), its uncertainty or the relevant inferential output, the scaling trigger, and the estimate's distance from that trigger. Also report the mean-ratio estimate and its uncertainty separately. A boundary-near estimate is a reason for careful interpretation, not permission to select whichever model or criterion gives the preferred result.
| What is near a boundary? | What it concerns | What it does not establish |
|---|---|---|
| GMR near 0.80 or 1.25 | The point-estimate constraint on the treatment mean ratio | Whether the variance-based test passes |
| \(s_{WR}\) near its trigger | Which scaling branch may apply and how stable that classification is | That the treatment contrast is acceptable |
| Contrast near a scaled limit | The relationship between the estimated contrast and the permitted criterion | That uncertainty has been handled by the prescribed test |
A common mistake is to reason, “The reference variability estimate is just above the trigger, so the scaled limits apply and the study passes.” This skips the mean-ratio constraint, treats an estimated variance as certain, and may substitute a point comparison for the required inferential rule. The correction is to state each component separately: identify the governing method and trigger, check the point-estimate constraint, evaluate the variability-based criterion using its prescribed uncertainty procedure, and describe how close the estimates are to their relevant boundaries.
Unscaled average bioequivalence is a useful conceptual comparison because its limits do not widen with reference variability. A scaled method instead permits a variability-dependent criterion under specified conditions, while retaining its own restrictions and inference. The detailed implementation differs across regulatory approaches; the introductory material on reference-scaling decisions explains why scaling can help and where its limits lie. Sensitivity analysis can assess robustness, but it is not an alternative decision rule and cannot replace the prespecified method.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- A hypothetical estimated GMR is 1.26, while the reference variability estimate is above its scaling trigger. Can the variability-based criterion override the 1.25 point-estimate limit?
- In an illustrative capped rule with \(k=0.70\), \(c=0.2231\), and \(s_{WR}=0.25\), calculate the capped width.
- An uncertainty interval for \(s_{WR}\) spans the scaling trigger, although its point estimate is just above the trigger. What is the appropriate interpretation?
- Why should a 90% confidence interval for the log treatment contrast not be described as the complete scaled bioequivalence analysis?
- What is wrong with reporting only that the estimated contrast is inside the scaled width?
Solutions. (1) No. The mean-ratio point-estimate constraint is a separate requirement and 1.26 exceeds 1.25. (2) The uncapped width is \(0.70(0.25)=0.1750\); since this is less than 0.2231, the capped width is 0.1750. (3) The point estimate is above the trigger, but its location relative to the boundary is uncertain; interpret it using the appropriate variance-estimation inference and regulatory rule. (4) It describes uncertainty in the mean contrast but not necessarily uncertainty in reference variability or the prescribed scaled decision calculation. (5) That statement omits the mean-ratio constraint and may confuse a point-estimate comparison with the required inferential procedure.