Worked EMA Analysis With Expanded Limits

What the EMA Expanded-Limits Analysis Requires

For a highly variable drug, the within-subject variability of the reference product can be used to widen the acceptance limits for the test-to-reference comparison of Cmax under the EMA approach. The widening is not unlimited: it is capped, and the estimated test-to-reference geometric mean ratio must independently remain within 80.00% to 125.00%. Thus, an expanded interval alone does not establish bioequivalence.

The analysis uses log-transformed pharmacokinetic measurements. Let \(\theta\) be the model-estimated mean difference between test and reference on the log scale, so that \(\exp(\theta)\) is the test-to-reference geometric mean ratio. Let \(s_{WR}\) be the estimated within-subject standard deviation for the reference product. In a replicate design, repeated reference observations allow this component to be estimated; the design and variance model must support that interpretation.

The coefficient of variation formula follows from the lognormal relationship between a log-scale standard deviation and a coefficient of variation. The threshold is equivalent to a standard deviation of about 0.2936, because \(\sqrt{\exp(0.2936^2)-1}\) is approximately 0.300. The cap corresponds to a reference within-subject coefficient of variation of 50%; variability above that level does not widen the allowed limits further.

From Reference Variability to Capped Limits

The calculation has two distinct stages. First, use \(s_{WR}\) to calculate the candidate scaled limits. Second, if those limits exceed the permitted maximum range, replace them with the capped limits. The cap is applied to the limits, not to the estimated variance component.

The candidate limits are symmetric around zero on the log scale, so they are reciprocals on the ratio scale. The cap is applied symmetrically as well: the final interval cannot be narrower on the low side than 0.6984 or wider on the high side than 1.4319. These rounded values represent 69.84% and 143.19%.

Worked Analysis: Cmax Does Not Meet Both Requirements

Consider a hypothetical 24-participant, three-period partial replicate study of a test product and a reference product. Assume the study was randomized and conducted with an appropriate washout; the log-scale model is prespecified and fits the treatment and period structure; and residual diagnostics reveal no severe departure that would invalidate the model-based estimates. Suppose the fitted model estimates \(s_{WR}=0.5200\), with 18 degrees of freedom for the treatment contrast, and estimates \(\theta=-0.09431\). The model-based 90% confidence interval on the log scale is \(-0.09431\pm0.27152\).

The estimated reference variability and capped limits are those calculated above: the final range is 0.6984 to 1.4319. The geometric mean ratio is \(\exp(-0.09431)\approx0.9100\), or 91.00%. This point estimate is within the required range of 80.00% to 125.00%.

Exponentiating the confidence interval endpoints gives \(\exp[-0.09431\pm0.27152]=[\exp(-0.36583),\exp(0.17721)]\approx[0.6936,1.1939]\). The upper endpoint is below 1.4319, but the lower endpoint, 0.6936, is below the permitted lower limit, 0.6984. The entire interval is therefore not contained within the expanded limits.

The conclusion for this hypothetical Cmax analysis is that the EMA criteria are not met. The point estimate requirement is met, but the confidence-interval requirement is not. The point estimate being inside the limits cannot compensate for an interval that extends outside them.

Worked Analysis: Cmax Meets the Expanded-Limits Criteria

The comparison is endpoint-specific. A pass for Cmax does not imply that another pharmacokinetic endpoint passes, nor does it show that every clinical or regulatory question about the products has been resolved. Each endpoint must be evaluated under its own applicable criterion.

Why Endpoint and Model Checks Matter

Under the EMA approach described here, expansion is for Cmax when the reference within-subject variability qualifies. It should not be applied automatically to AUC simply because an estimate of reference variability is high. AUC is assessed against its applicable acceptance limits unless a relevant regulatory framework specifically permits another rule. The analysis plan should identify the endpoint and procedure before examining whether the results are favorable.

For example, suppose a hypothetical AUC analysis has a log-scale mean difference of \(-0.10536\) and a 90% confidence interval of \(-0.10536\pm0.12500\). The point estimate is \(\exp(-0.10536)=0.9000\), within the conventional range. But the interval is \([\exp(-0.23036),\exp(0.01964)]\approx[0.7942,1.0198]\). Since 0.7943 is below 0.8000, this interval does not fit within the conventional 80.00%–125.00% range. The point estimate alone does not establish a pass.

A sound decision also depends on whether the model estimates the intended reference variability and treatment contrast. The earlier material on residual diagnostics, influential observations, and variance-model specification is relevant here: unusual residual patterns, a single influential participant, or an ill-suited variance structure can make \(s_{WR}\) unstable or the confidence interval misleading. Check that the replicate design actually contains repeated reference observations, that the reference variance component is identified by the design, and that the model and missing-data handling are consistent with the prespecified analysis.

A boundary result deserves particular care. If an interval endpoint is close to a scaled limit, report adequate precision and compare unrounded values where available; premature rounding can reverse an apparent containment decision. Also report the estimated \(s_{WR}\), its degrees of freedom or estimation method as appropriate, the resulting coefficient of variation, the uncapped and final limits, the point estimate, and the confidence interval. This makes the path from model output to conclusion auditable.

Common Pitfalls and Alternative Criteria

• Using the point estimate as the whole decision. A point estimate within 80.00% to 125.00% is necessary but does not show that the 90% confidence interval fits inside the expanded limits.
• Forgetting the cap. The formula \(\exp(\pm0.760s_{WR})\) produces candidate limits. When these exceed the maximum range, use the capped limits of 0.6984 to 1.4319.
• Applying expansion to the wrong endpoint. Do not infer that high variability permits expansion for every pharmacokinetic endpoint. Apply the endpoint-specific regulatory rule.
• Substituting total variability for reference variability. The scaling calculation requires the reference within-subject component, not a pooled measure of variability that combines unrelated sources.
• Choosing the model after seeing which result passes. Variance structure, missing-data handling, and diagnostics should be considered under a prespecified analysis, not selected to obtain a preferred conclusion.

The EMA expanded-limits procedure and the conventional fixed-limit approach differ in how the acceptance range is determined: the expanded approach uses reference within-subject variability for eligible Cmax analyses, while conventional limits do not widen based on that estimate. The FDA reference-scaled procedure is a separate alternative with its own criterion. It should not be treated as interchangeable with EMA interval containment; its specific calculation is a different analysis.