Explaining Discordant FDA and EMA Results

Why the Same Study Can Receive Different Decisions

Reference-scaled bioequivalence (BE) methods allow the acceptance criterion to reflect within-subject variability for the reference formulation. FDA and EMA approaches do not apply that idea in exactly the same way. They use different scaling constants and decision rules, and the EMA approach caps the expanded limits. Consequently, the same estimated treatment difference and reference variability can lead to different conclusions, even when both analyses are correctly conducted.

A discordant result is not, by itself, evidence that one analysis is wrong. The useful question is which difference in the rules or analysis inputs produced the divergence? The comparison is interpretable only when the analyses use the same endpoint, data, subject population, and appropriate variance model. Earlier material on estimating reference variability and the FDA linearized criterion supplies the model-specific context; here the emphasis is on tracing differences between decisions.

Definition: Write \(\delta=\mu_T-\mu_R\) for the difference between the test and reference population means on the log scale, and \(s_{WR}\) for the estimated within-subject standard deviation for the reference formulation. The geometric mean ratio (GMR) is \(\exp(\delta)\). Reference-scaled criteria use \(s_{WR}\) to set a variability-dependent boundary, subject to framework-specific rules.

The comparisons below assume a suitable replicate design that identifies reference within-subject variability, an appropriately specified analysis model, and the same prespecified endpoint and analysis population in each framework. They also assume the log-transformed endpoint and a correctly calculated confidence interval or confidence bound from that model. These assumptions matter: different data handling, variance models, or endpoint definitions can produce different inputs before either framework’s decision rule is applied.

The Decision Rules and Their Scaling

For the FDA reference-scaled average BE approach, the linearized criterion uses \(\theta=[\ln(1.25)/0.25]^2\), approximately \(0.7967\). The upper bound of a one-sided 95% confidence interval for the linearized quantity must be below zero. The FDA approach also requires the estimated GMR to be within 0.80 to 1.25. Thus a favorable plug-in estimate alone does not establish BE: the confidence-bound requirement and point-estimate restriction must both be met.

\delta^2-\theta s_{WR}^2 < 0, \qquad \theta=\left(\frac{\ln(1.25)}{0.25}\right)^2\approx 0.7967

The equation shows the central scaling idea: a larger reference standard deviation increases the allowable squared difference in log means. In an actual FDA analysis, the decision is not made merely by substituting point estimates into this expression; the specified one-sided confidence bound must be calculated using the analysis model and its uncertainty.

For the EMA approach, when the reference within-subject coefficient of variation exceeds 30%, the 90% confidence interval for the GMR is compared with expanded limits based on \(\exp(\pm 0.760s_{WR})\). The limits are capped at 0.70 and 1.43. EMA also requires the estimated GMR to remain within 0.80 to 1.25. If the variability does not exceed the scaling threshold, the conventional 0.80 to 1.25 limits apply rather than expanded limits.

L_{\mathrm{EMA}}=\max\left(0.70,\exp(-0.760s_{WR})\right), \qquad U_{\mathrm{EMA}}=\min\left(1.43,\exp(0.760s_{WR})\right)

The threshold can be related to \(s_{WR}\) through the lognormal conversion \(CV_{WR}=\sqrt{\exp(s_{WR}^2)-1}\). At a reference CV of 30%, this gives \(s_{WR}=\sqrt{\ln(1+0.30^2)}\approx0.2936\). The EMA limits expand continuously from the conventional limits at that threshold, but the cap prevents further expansion beyond 0.70 to 1.43.

Key distinction: FDA evaluates a confidence bound for a variance-adjusted function of the log-mean difference and separately constrains the GMR point estimate. EMA evaluates whether the 90% confidence interval for the GMR lies within variability-dependent, capped limits and separately constrains the GMR point estimate. These are different tests, not interchangeable calculations of one common interval.

Why the Scaling Rules Can Diverge

A short algebraic comparison clarifies one source of discordance. If the FDA linearized expression were assessed using only its point estimates, the condition would be

\delta^2-\theta s_{WR}^2<0 \quad\Longleftrightarrow\quad |\delta|<\sqrt{\theta}\,s_{WR} \approx 0.8926s_{WR}.

This algebra is a way to compare the nominal scale of the two rules, not a replacement for the FDA confidence-bound procedure. EMA’s uncapped log-limit scale is \(0.760s_{WR}\), whereas the FDA point-estimate boundary from the expression above is about \(0.8926s_{WR}\). But EMA tests whether an entire 90% confidence interval fits inside its limits, while FDA uses a confidence bound on a different function. The FDA point-estimate restriction can also be decisive. Therefore the larger FDA coefficient does not imply that FDA must always pass when EMA passes.

The cap adds another difference at high reference variability. Without the cap, EMA limits would keep widening as \(s_{WR}\) increased. With the cap, they stop at 0.70 and 1.43. FDA’s linearized expression does not use that same EMA limit cap; its own point-estimate requirement remains in force. The effect of the cap depends on whether the confidence-interval endpoint would otherwise extend beyond the capped limit.

Worked Example: FDA Pass, EMA Fail

Consider a hypothetical replicate study with the same prespecified log endpoint in both analyses. Suppose the fitted model estimates \(s_{WR}=0.350\), a GMR of 1.24, and an EMA 90% confidence interval of [1.16, 1.32]. Suppose the FDA model-based one-sided 95% upper confidence bound for its linearized criterion is \(-0.0080\). Treat that confidence bound as a hypothetical model output; it cannot be reconstructed from these summary numbers alone.

First check the reference CV: \(CV_{WR}=\sqrt{\exp(0.350^2)-1}=\sqrt{\exp(0.1225)-1}=\sqrt{0.130319\ldots}\approx0.3610\), or 36.10%. The EMA scaling threshold is exceeded.

The uncapped EMA limits are \(\exp(\pm0.760\times0.350)=\exp(\pm0.266)\), approximately [0.7665, 1.3047]. These are within the cap, so the applicable expanded limits are approximately [0.7665, 1.3047]. The interval’s upper endpoint, 1.32, exceeds 1.3047. The EMA interval criterion fails, even though the point GMR of 1.24 is inside 0.80 to 1.25.

For FDA, the estimated log difference is \(\delta=\ln(1.24)\approx0.2151\). Its plug-in linearized expression is \(\delta^2-\theta s_{WR}^2\approx(0.2151)^2-0.7967(0.350)^2=0.04627-0.09760=-0.05133\). The plug-in value is negative, but the actual decision uses the stated upper confidence bound. Since \(-0.0080<0\), that bound meets the FDA criterion. The GMR of 1.24 also meets the FDA point-estimate restriction. On the supplied hypothetical outputs, FDA passes and EMA fails. The decisive difference is the EMA interval endpoint relative to its expanded upper limit, not a failure of the FDA point estimate.

Worked Example: EMA Pass, FDA Fail

Now consider another hypothetical analysis of 36 subjects in a suitable replicate design. Suppose \(s_{WR}=0.500\), the GMR is 1.06, the EMA 90% confidence interval is [0.82, 1.38], and the FDA model-based one-sided 95% upper confidence bound for the linearized criterion is \(+0.0060\). Again, the bound is a supplied hypothetical model output.

The reference CV is \(\sqrt{\exp(0.500^2)-1}=\sqrt{\exp(0.25)-1}=\sqrt{0.284025\ldots}\approx0.5329\), or 53.29%. The uncapped EMA upper limit would be \(\exp(0.760\times0.500)=\exp(0.380)\approx1.4623\), and the lower limit would be approximately \(1/1.4623=0.6839\). Applying the cap gives limits of 0.70 and 1.43.

The entire EMA interval, [0.82, 1.38], lies within [0.70, 1.43], and the point GMR of 1.06 is within 0.80 to 1.25. The EMA criteria are met.

For FDA, \(\delta=\ln(1.06)\approx0.05827\), so the plug-in expression is \((0.05827)^2-0.7967(0.500)^2\approx0.003395-0.1992=-0.1958\). This plug-in value is negative, but the supplied FDA upper confidence bound is \(+0.0060\), which is not below zero. FDA therefore fails its confidence-bound criterion. The GMR restriction is met, but both FDA requirements are necessary. This example demonstrates why a negative point estimate of the linearized expression cannot be substituted for its confidence-bound test. It also illustrates how an EMA interval can fit within capped limits while FDA’s different uncertainty calculation does not meet its criterion.

Worked Example: Tracing a Difference to the Variance Estimate

Suppose two documented model specifications for a hypothetical replicate dataset estimate \(s_{WR}=0.290\) and \(s_{WR}=0.300\). These are not alternative numbers to choose between after seeing the BE result; the example illustrates why the specified model and its diagnostics must be checked before comparing decisions.

For \(s_{WR}=0.290\), \(CV_{WR}=\sqrt{\exp(0.290^2)-1}=\sqrt{\exp(0.0841)-1}\approx\sqrt{0.0877}\approx0.2962\), or 29.62%. For \(s_{WR}=0.300\), \(CV_{WR}=\sqrt{\exp(0.0900)-1}\approx\sqrt{0.0942}\approx0.3069\), or 30.69%. Thus the first estimate is below, and the second above, the EMA 30% scaling threshold.

At 0.290, the EMA rule uses conventional limits; at 0.300, it uses expanded limits, which are approximately \(\exp(\pm0.228)\), or [0.7961, 1.256]. This expansion is small because the estimate is just above the threshold. FDA’s commonly used log-SD threshold is about 0.294, so these two estimates also fall on opposite sides of that threshold. The calculation identifies a possible source of decision change: a variance estimate near the scaling boundary can change which rule applies. It does not establish which estimate is correct. That requires checking the prespecified model, design, data, and variance estimation.

How to Investigate a Discordant Result

A disciplined comparison separates the analysis into inputs, rule, and decision. First verify that both results refer to the same endpoint, log transformation, analysis population, and replicate data. Then check that the reference variance is estimated from the appropriate within-subject contrasts under each specified model. A difference in model assumptions or variance estimation can change both the estimated scale and its uncertainty; it should not be described as a difference in regulatory criteria until the inputs have been reconciled.

Align the analysis inputs.
Confirm the endpoint, records, subject population, period handling, treatment coding, and model specification. Establish whether the analyses use comparable estimates of \(\delta\) and \(s_{WR}\).

Check the scaling region.
Calculate the reference CV from \(s_{WR}\) and identify whether each framework applies scaling. Check whether EMA limits are capped and whether a variance estimate is close to the threshold.

Apply each complete decision rule.
For FDA, check the upper confidence bound and the GMR restriction. For EMA, check both confidence-interval endpoints against the applicable limits and check the GMR restriction.

Locate the first point of divergence.
Compare estimates, uncertainty calculations, boundary values, and final criteria separately. Report the specific step that changes the decision rather than attributing the outcome vaguely to “different scaling.”

Diagnostics help distinguish a genuine framework difference from an unstable analysis. Review residual diagnostics, influential observations, missing-period handling, and variance-model specification as relevant to the design. A result near a scaling threshold or decision boundary deserves particular scrutiny because modest changes in an estimate or its uncertainty can change the applicable rule or the pass/fail outcome. Such sensitivity is a reason to report it transparently, not to select whichever model yields a preferred conclusion.

The main alternative to reference scaling is conventional average BE, which compares a 90% confidence interval for the GMR with 0.80 to 1.25. That is not a substitute for comparing the FDA and EMA scaled procedures when the question concerns discordant scaled decisions. Nor can the two scaled procedures be reduced to a single common “scaled interval”: their constants, uncertainty criteria, and boundary restrictions differ. Both remain subject to the design and model assumptions needed to estimate reference within-subject variability appropriately.

Common Pitfall and Limits of the Comparison

A frequent mistake is to calculate \(\delta^2-\theta s_{WR}^2\) from point estimates, see a negative number, and conclude that FDA has established BE. The correction is that the FDA criterion requires the prescribed upper confidence bound to be below zero, as well as a GMR point estimate within 0.80 to 1.25. Another mistake is to compare an EMA interval with uncapped limits when the cap applies; the correction is to calculate the variability-based limits and then enforce the 0.70 to 1.43 bounds.

A cross-framework comparison is limited by the quality and comparability of its inputs. Near-boundary results may be sensitive to the reference variance estimate, and different model specifications can change the estimate or its uncertainty. The frameworks’ outcomes alone do not identify the cause; the calculations and model outputs must be examined. Reporting the estimates, confidence results, applicable limits, and point-estimate checks makes the explanation auditable.

Key takeaway: FDA and EMA discordance can arise from different scaling constants, uncertainty rules, EMA’s limit cap, or different variance estimates. Trace each analysis from shared inputs through its own complete decision rule; do not infer the cause from pass/fail labels alone.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A hypothetical FDA plug-in value of \(\delta^2-\theta s_{WR}^2\) is negative, but its one-sided 95% upper confidence bound is \(+0.002\). Does the FDA scaled criterion pass?
For \(s_{WR}=0.400\), calculate the uncapped EMA limits and state whether the cap changes them.
An EMA 90% confidence interval is within the expanded limits, but the GMR point estimate is 1.27. Does the EMA analysis meet all stated requirements?
Two analysts obtain different reference standard deviations from the same study. What should be checked before calling the resulting decision difference a jurisdictional difference?

Solutions. (1) No. The required upper confidence bound is not below zero, even though the plug-in value is negative. (2) The limits are \(\exp(\pm0.760\times0.400)=\exp(\pm0.304)\), approximately [0.7379, 1.3553]. Neither crosses the 0.70 or 1.43 cap, so the cap does not change them. (3) No. The point estimate 1.27 is outside 0.80 to 1.25, so the separate GMR restriction fails. (4) Check whether both analyses use the same endpoint, data, analysis population, treatment and period coding, missing-data handling, and prespecified variance model; then inspect diagnostics and the variance estimation. Only after aligning or explaining those inputs can the remaining difference be attributed to the frameworks’ decision rules.