Common Errors in Applying Scaled Limits

Why Small Rule Errors Change the Decision

Reference-scaled bioequivalence (BE) methods allow the acceptance criterion to depend on within-subject variability of the reference product for specified endpoints and circumstances. That flexibility is not a general permission to widen limits whenever a study is noisy. Scaling has eligibility rules, a qualifying variability threshold, a defined calculation, and—in some approaches—a cap. It can also leave a separate point-estimate condition in force. Missing any one of these requirements can change a pass into a fail, or produce an unjustified pass.

The errors considered here are especially easy to make when several regulatory approaches are treated as interchangeable. The worked examples use the EMA-style expanded-limit approach to illustrate threshold, endpoint, cap, and point-estimate checks. The FDA reference-scaled approach is different: it uses a linearized criterion rather than simply applying those expanded limits. The applicable product-specific guidance and jurisdiction determine which procedure and endpoints are permitted.

The CV conversion is worth checking explicitly. A CV is expressed on the original scale, whereas \(s_{WR}\) is on the log scale. The conversion is \(s_{WR}=\sqrt{\ln(1+CV_{WR}^{2})}\), with the CV entered as a proportion. For example, a CV of 42% is entered as 0.42, not 42. Rounding the CV or standard deviation too early can also shift a result close to a threshold or limit.

A Sequence of Checks Before Applying Limits

A sound scaled analysis begins by confirming the governing regulatory procedure and the endpoint’s eligibility. Next, verify that the design and model supply the reference-specific within-subject variability required by that procedure. Only then compare the estimate with the qualifying threshold, calculate any expanded limits, apply a cap where required, and check every additional decision condition. The conditions are cumulative: passing the interval criterion does not excuse failure of a point-estimate rule.

For the EMA-style procedure, the expanded limits are applied to the 90% confidence interval for the test-to-reference ratio on the original scale, with the confidence interval constructed from the log-scale analysis. The interval must lie entirely within the applicable limits. The point-estimate condition is separate: the GMR itself must lie between 0.80 and 1.25. The limits are not a replacement for that condition.

Worked Examples of Common Errors

Why the Cap and Separate Checks Matter

In the EMA-style approach, the cap prevents limits from widening without bound as estimated reference variability increases. On the original ratio scale, the capped limits are approximately 69.84% and 143.19%. One precise way to express the rule is:

The interval condition is that the full required confidence interval lies within \([L,U]\). The point-estimate condition is assessed separately against \([0.80,1.25]\). These checks protect against different problems: the cap limits the allowed expansion, while the point-estimate rule constrains the estimated average difference even when the confidence interval satisfies the scaled criterion.

The upper cap is not the same as the qualifying threshold. The threshold answers whether expansion may begin; the cap answers how far permitted limits may widen. Confusing those two roles can lead either to unjustified expansion below the threshold or to over-expansion at high variability.

EMA-Style Limits and the FDA Criterion Are Not Interchangeable

A frequent source of error is to call any variability-dependent procedure “scaled limits” and then transfer formulas from one regulatory approach to another. The EMA-style procedure described above uses expanded ratio limits, a cap, a confidence-interval rule, and a separate GMR check. The FDA reference-scaled average BE approach instead uses a linearized criterion based on the test-reference difference and the reference within-subject variance. It does not amount to applying the EMA expanded limits with a different cap.

In the FDA approach, the analysis uses the prescribed replicate design and model to form an upper confidence bound for a quantity of the form \((\mu_T-\mu_R)^2-\theta\sigma_{WR}^2\), where \(\mu_T-\mu_R\) is the log-scale treatment difference, \(\sigma_{WR}^{2}\) is the reference within-subject variance, and \(\theta\) is the scaling constant specified by the method. The upper bound must meet the method’s criterion, and the required point-estimate restriction still applies. The confidence-bound calculation cannot be replaced by checking whether a 90% interval falls between EMA-style expanded limits.

Both approaches depend on reference variability, but their estimands, uncertainty calculations, and decision rules differ. Use the method specified for the product and jurisdiction. If an analysis report gives a single set of expanded limits without identifying the governing approach, the endpoint, and the variability estimate, the basis of the decision is not adequately clear.

Diagnostics and Failure Modes

Before accepting a scaled result, check that the analysis has not confused reference within-subject variability with total variability or with a pooled treatment variance. Confirm that the replicate observations and treatment coding support estimation of the intended reference component, and that the model and missing-data handling are appropriate for the design. An unusual estimate should prompt review of the data and model; it is not, by itself, a reason to substitute a more favorable variance estimate.

Also inspect values near the 30% threshold and decisions near either limit. Recalculate the CV-to-standard-deviation conversion from the unrounded estimate, and verify whether the regulatory rule uses a strict threshold. Check the limits after applying the cap, not just the raw exponential formula. Finally, record the confidence-interval result and GMR result as separate findings. A conclusion based only on “the interval fits” is incomplete when a point-estimate check is required.

• Wrong reasoning: “The study was variable, so wider limits are appropriate.” Correction: Expansion requires the specified endpoint, reference-based variability estimate, and qualifying threshold.
• Wrong reasoning: “The endpoint’s CV is high, so it can be scaled.” Correction: Variability does not make an ineligible endpoint eligible.
• Wrong reasoning: “The calculated limits are very wide, so the interval passes.” Correction: Apply the regulatory cap before comparing the confidence interval.
• Wrong reasoning: “The confidence interval is inside the limits, so BE is established.” Correction: Verify the GMR point-estimate condition and every other required criterion.
• Wrong reasoning: “EMA and FDA both use reference variability, so their limits are interchangeable.” Correction: Use the specific procedure’s design, model, uncertainty calculation, and decision rule.