Reporting a Reference-Scaled Bioequivalence Analysis

Why Transparent Reporting Matters

A reference-scaled bioequivalence conclusion depends on more than a point estimate and a confidence interval. The reader needs to know which replicate design produced the reference variability estimate, how that estimate was obtained, which jurisdiction-specific rule was prespecified, and how the rule was applied to the endpoint. Without those details, a reported “pass” cannot be checked or interpreted.

This tutorial focuses on reporting the analysis, not choosing a regulatory pathway. EMA expanded limits and the FDA linearized criterion are different procedures; their numerical results should not be substituted for one another. In either case, state the analysis population, endpoint, model, variance estimate, relevant decision quantities, diagnostics, and conclusion. A conclusion for one pharmacokinetic endpoint does not establish a conclusion for another.

Definition: In a reference-scaled bioequivalence analysis, the treatment contrast is evaluated using a rule that incorporates the within-subject variability of the reference product, estimated from replicate reference observations. A complete report identifies the rule and the reference-variability estimate used by that rule, as well as any additional point-estimate or confidence-bound requirements.

What to Specify Before Reporting a Result

Begin with the design and analysis population. Name the replicate sequence design—for example, a partial replicate design such as TRR, RTR, and RRT—and give the number of subjects randomized and analyzed, the numbers completing each sequence, and any exclusions or missing periods. In a partial replicate design, the repeated reference administrations provide within-subject reference information even though the test product is not necessarily repeated. State which endpoint and analysis population the results describe.

For a pharmacokinetic endpoint such as AUC or C_max, the analysis is typically conducted on log-transformed values. Give the model’s fixed effects, random effects, and variance structure in enough detail to identify how the treatment contrast and reference within-subject variance were estimated. A representative model may include sequence, period, and treatment as fixed effects, and subject nested within sequence as a random effect, with a covariance or variance structure that permits estimation of the reference within-subject variability. This is not a universal specification: the report must describe the model actually prespecified and fitted, including how the repeated reference measurements contribute to its variance estimate.

Report the method used to estimate reference within-subject variance, such as REML, and report the estimate on the log scale as \(s_{WR}\), or the variance as \(s_{WR}^2\), with units made clear. These are log-scale quantities and are not percentages. If a coefficient of variation is also reported, give its definition. For a lognormal within-subject distribution, the usual conversion is \(CV_{wR}=\sqrt{\exp(s_{WR}^2)-1}\); state whether the value is expressed as a proportion or percentage.

The report should identify the applicable guideline or prespecified procedure and its version, the scaling threshold, any cap, and any separate point-estimate condition. The scaling calculation is endpoint-specific: a reference-variability estimate for AUC does not supply the estimate for C_max.

Two Different Scaling Rules

Under the EMA expanded-limits approach, scaling is triggered when the reference within-subject standard deviation exceeds the threshold \(s_{WR}=0.294\), approximately corresponding to a within-subject coefficient of variation of 30%. When the threshold is exceeded, the limits are \(\exp(\pm 0.760s_{WR})\), subject to the specified cap \(s_{WR}=0.5\). The 90% confidence interval for the test/reference geometric mean ratio must lie within the applicable limits, and the point estimate must also lie within 0.80 to 1.25. When the threshold is not exceeded, the usual 0.80 to 1.25 limits apply.

Under the FDA reference-scaled average bioequivalence approach for highly variable drugs, the analysis uses a linearized criterion. With \(\delta=\mu_T-\mu_R\), where the means are on the log scale, the criterion is based on \(\delta^2-\theta\sigma_{WR}^2\), where \(\sigma_{WR}\) is the population reference within-subject standard deviation and \(\theta=[\ln(1.25)/0.25]^2\), approximately 0.7967. The decision uses a one-sided 95% upper confidence bound for this criterion, which must be no greater than zero; the point-estimate requirement must also be met. Report the bound and the point estimate, not merely the plug-in value of the criterion.

The difference is consequential. EMA compares a confidence interval with variability-dependent limits and separately constrains the point estimate. FDA evaluates an upper confidence bound for a linearized criterion and also has a point-estimate requirement. A report should name the procedure explicitly rather than use the ambiguous phrase “scaled limits.”

Key reporting principle: State the chosen rule before presenting the result. For an EMA analysis, report the applicable limits and the 90% confidence interval. For an FDA analysis, report the linearized-criterion upper confidence bound and its decision threshold. In both cases, report the point estimate and every additional requirement separately.

Worked Reporting Examples

Worked Example: Reporting an FDA Linearized Criterion for AUC

Suppose a hypothetical partial replicate study estimates \(s_{WR}=0.350\) for log AUC. The model estimates a test/reference geometric mean ratio of 1.025 and a 90% confidence interval of 0.972 to 1.081. Assume the analysis plan specifies the FDA reference-scaled average bioequivalence procedure, and that the model's one-sided 95% upper confidence bound for the linearized criterion is \(-0.0180\). This bound is an assumed model output.

Identify the variability estimate. The estimated reference within-subject standard deviation is \(s_{WR}=0.350\). The FDA decision is based on the linearized criterion and its model-based confidence bound, not EMA expanded limits.

Apply the FDA criterion. The assumed one-sided 95% upper confidence bound for the linearized criterion is \(-0.0180\), which is less than or equal to zero.

Apply both requirements. The upper confidence bound is less than or equal to zero, and the point estimate, 1.025, lies within 0.80 to 1.25. On these hypothetical results, the AUC endpoint meets the stated FDA reference-scaled average bioequivalence criteria. A clear report would also give the model, the variance-estimation method, the numbers analyzed, and the diagnostics.

Worked Example: Reporting an FDA Linearized Criterion

Suppose a hypothetical FDA analysis for log C_max estimates a reference within-subject standard deviation of \(s_{WR}=0.400\) and a test/reference geometric mean ratio of 1.040. For illustration, suppose the prespecified model’s one-sided 95% upper confidence bound for the linearized criterion is \(-0.0180\). This bound is an assumed model output; it cannot be recovered from the point estimate and \(s_{WR}\) alone.

Show the point plug-in calculation without confusing it with the decision bound. The log treatment contrast is \(\delta=\ln(1.040)=0.0392207\). Using \(\theta=0.7967\), the plug-in criterion is

\delta^2-\theta s_{WR}^2 =(0.0392207)^2-0.7967(0.400)^2 =0.00153826-0.127472 =-0.1259337\approx -0.1259.

This negative plug-in value is descriptive; it is not itself the FDA confidence-bound decision. The reported one-sided 95% upper bound is \(-0.0180\), which is less than or equal to zero, so the linearized criterion is met. The point estimate 1.040 is also within 0.80 to 1.25. Therefore, under the stated hypothetical procedure and assuming all other prespecified requirements are satisfied, the C_max endpoint meets those FDA criteria. The report must identify the model-based bound and its degrees-of-freedom or uncertainty method, rather than imply that the plug-in calculation establishes the result.

Worked Example: Reporting Different Decisions by Endpoint

Suppose a hypothetical FDA analysis reports AUC results as in the first example, while an EMA analysis for C_max estimates \(s_{WR}=0.310\), a point estimate of 1.135, and a 90% confidence interval of 1.012 to 1.270.

For C_max, \(0.310>0.294\), so the scaling trigger is met. The expanded-limit exponent is \(0.760(0.310)=0.2356\). The limits are \(\exp(-0.2356)\approx0.7901\) and \(\exp(0.2356)\approx1.266\). The point estimate 1.135 meets the separate 0.80 to 1.25 point-estimate condition. However, the upper confidence limit 1.270 exceeds the expanded upper limit of approximately 1.266. The C_max endpoint therefore fails the stated confidence-interval requirement.

The accurate summary is that AUC meets its stated FDA criteria and C_max does not meet its stated EMA criteria. Whether the study as a whole meets a regulatory decision rule depends on the applicable endpoint requirements; it cannot be inferred by reporting only the AUC result.

Justifying the Scaling Calculation

The expanded-limit calculation follows directly from the rule on the log scale. If the permitted log-scale distance from no treatment difference is \(0.760s_{WR}\), the corresponding ratio-scale limits are obtained by exponentiation:

-0.760s_{WR}\leq \mu_T-\mu_R\leq 0.760s_{WR} \quad\Longrightarrow\quad \exp(-0.760s_{WR})\leq \exp(\mu_T-\mu_R)\leq\exp(0.760s_{WR}).

Because \(\exp(\mu_T-\mu_R)\) is the test/reference geometric mean ratio, the resulting limits are multiplicative and symmetric around 1 on the ratio scale. The cap prevents the expanded limits from widening without bound. The confidence interval is still checked against those limits, while the separate point-estimate condition prevents the scaling rule from allowing an excessively displaced estimated ratio.

The FDA criterion also combines treatment difference and reference variability, but does so through a different function. Its point plug-in form subtracts a variability allowance, \(\theta s_{WR}^2\), from the squared log treatment contrast. The confidence-bound procedure accounts for estimation uncertainty in the criterion. This is why substituting a point calculation for the upper confidence bound is not a valid shortcut.

Diagnostics and Reporting Checks

A transparent result includes enough diagnostic information to judge whether the fitted model and variance estimate are credible. Report convergence status, whether estimates were on a boundary, and relevant residual checks for the log-scale model. Examine unusual observations, influential subjects, sequence or period patterns, and whether missing periods or exclusions could change the repeated-reference information. State any sensitivity analyses that materially affect the treatment contrast, reference variance, or decision.

An implausibly small or large reference variance, strong sensitivity to one subject, poor model convergence, or a variance estimate near a boundary warrants investigation rather than a bare pass/fail statement. Diagnostics do not replace the prespecified decision rule, and a favorable diagnostic plot does not prove all model assumptions. Report notable limitations and explain whether the primary conclusion is robust to reasonable analysis choices.

A useful presentation table can include, for each endpoint, the model-based point estimate, confidence result, \(s_{WR}\), applicable rule, and decision. Add units for the endpoint, usually such as hours·concentration for AUC and concentration for C_max. Give ratios without physical units, but identify them as test/reference geometric mean ratios. Include the sample sizes and the method for variance estimation in the table notes or surrounding text.

Common Reporting Pitfalls

Reporting only “bioequivalent.” This omits the endpoint, procedure, estimate, and decision quantities. Name the endpoint and show the values that establish the conclusion.
Calling the EMA limits and FDA criterion interchangeable. They use different decision procedures. Identify the jurisdiction-specific approach and report its corresponding confidence result.
Treating the plug-in FDA criterion as the FDA decision. A negative plug-in value does not establish that its upper confidence bound is at most zero. Report the bound from the specified analysis.
Omitting the point-estimate requirement. A confidence result alone may not satisfy the complete rule. State the point estimate and whether the separate condition is met.
Reporting a variance without its scale or source. Distinguish the reference within-subject variance from total variability and identify the log scale and estimation method.
Generalizing across endpoints. A conclusion for AUC does not automatically apply to C_max. Present endpoint-specific results and conclusions.

Key takeaway: A defensible reference-scaled report makes the decision reproducible: it identifies the replicate design and model, reports the reference within-subject variability and its estimation method, specifies the exact scaling rule, presents the required confidence result and point estimate, summarizes diagnostics, and states a conclusion for each endpoint.

Check Your Understanding

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

In an EMA expanded-limits analysis, why must a report give both the 90% confidence interval and the point estimate?
A reference within-subject standard deviation is 0.280. Under the EMA rule described here, does scaling apply? What limits are used?
Why is a negative FDA point plug-in criterion insufficient to claim that the FDA linearized criterion is met?
What should a report say if AUC meets its stated scaled criteria but C_max does not?

Solutions. (1) The confidence interval must lie within the applicable limits, and the point estimate must separately satisfy the point-estimate condition; one does not imply the other. (2) No. Since \(0.280<0.294\), the unscaled 0.80 to 1.25 limits apply. (3) The FDA decision uses the specified one-sided 95% upper confidence bound, which incorporates uncertainty; the point plug-in value does not establish that this bound is at most zero. (4) State the endpoint-specific outcomes separately and avoid claiming that both endpoints meet the criteria. Any overall conclusion must follow the applicable prespecified requirements.