Why Total Variability Cannot Replace Reference Variability

Why the Choice of Variance Matters

Reference-scaled average bioequivalence allows the acceptance criterion to depend on the within-subject variability of the reference formulation. The logic is specific: when repeated reference administrations show substantial within-subject variation, the reference variance provides the scale used by the applicable procedure. It does not follow that any large estimate of variability is an acceptable substitute.

This distinction matters because different variance components answer different questions. The total spread of observed concentrations or log exposure measures can include differences between subjects. A pooled residual can combine variation from test and reference administrations. Test-treatment variability describes repeatability of the test formulation, not the reference formulation. Substituting any of these can change the scaled criterion, sometimes substantially, without changing the treatment mean difference.

The details of the estimate depend on the replicate design and model. The earlier material on Estimating Reference Variability in Full Replicate Designs, Estimating Reference Variability in Partial Replicate Designs, and REML Estimation of the Reference Variance Component addresses those design-specific calculations. Here the emphasis is on what the target variance means and why other quantities cannot stand in for it.

What the Scaled Criterion Uses

The agencies’ reference-scaled procedures do not use identical decision rules, but the variance input has a shared interpretation: it is within-subject reference variability. For example, the FDA linearized criterion includes a term of the form \((\mu_T-\mu_R)^2-\theta\sigma_{WR}^2\), where \(\mu_T-\mu_R\) is the difference in population mean log outcomes and \(\theta\) is a specified scaling constant. The applicable procedure uses an estimate of the reference variance, along with its other requirements.

This expression illustrates why the variance’s identity matters: increasing the variance inserted into the expression makes the inequality easier to satisfy, all else equal. In practice, a procedure has additional conditions, such as requirements on the point estimate or on whether scaling is permitted. The expression alone is not a complete regulatory decision rule. EMA expanded limits likewise use reference within-subject variability within the applicable limits and restrictions. Neither approach licenses replacing that quantity with whichever variance is largest or easiest to calculate.

The interpretation is also conditional on the model. The reference variance is a model-based within-subject component on the log scale, not a universal property of a drug independent of design, conditions, or population. Its estimate can be uncertain, especially in small samples; a large point estimate is not by itself evidence that a different variance component is appropriate.

Why Repeated Reference Measurements Identify It

A simple full-replicate argument shows what information is needed. Suppose a subject receives the reference formulation twice, and let \(Y_{R1}\) and \(Y_{R2}\) be the subject’s two log outcomes. Under a model with a subject effect \(b_i\) and independent within-subject errors \(e_{i1}\) and \(e_{i2}\), write \(Y_{Rj}=\mu_R+b_i+e_{ij}\), with each error having variance \(\sigma_{WR}^2\). Assume the two reference administrations have the same within-subject variance, and that their errors are independent conditional on the subject effect.

The subject effect and common reference mean cancel in the difference. If the sample variance of the within-subject reference differences is \(s_D^2\), the corresponding simple estimate is \(s_D^2/2\). More elaborate crossover designs and mixed-effects models account for their design structure, but the identifying information remains repeated reference observations within subjects.

This justification relies on its stated assumptions: the reference measurements are on the log scale; the relevant within-subject reference variance is common across the replicate administrations; and the errors have no unmodelled covariance that changes the variance of their difference. If, for example, replicate errors are correlated, then \(\operatorname{Var}(e_{i1}-e_{i2})=2\sigma_{WR}^2-2\operatorname{Cov}(e_{i1},e_{i2})\), so simply dividing the difference variance by two need not estimate the intended component. Model specification and design therefore matter as well as arithmetic.

Why Other Variances Do Not Answer the Same Question

A total variance across individual outcomes combines sources of variation. In a simple random-intercept model, a reference observation can be written \(Y_{Ri}=\mu_R+b_i+e_{Ri}\), where \(b_i\) is the subject effect and \(e_{Ri}\) is within-subject error. If these components are independent, then \(\operatorname{Var}(Y_R)=\operatorname{Var}(b_i)+\operatorname{Var}(e_{Ri})\). The first term represents differences between subjects; the second is the within-subject reference variance. Scaling with the total would count between-subject heterogeneity as if it were within-subject reference variability.

A pooled residual variance has a different problem. Pooling test and reference residuals imposes, or at least behaves as though, they share a common variance. If the true reference and test variances differ, the pooled quantity estimates neither reference-specific component. It can be a suitable estimate of a common residual variance under an explicitly justified equal-variance model, but that is a different model and a different target. Equal variances should not be assumed merely because software reports a pooled residual.

Test-treatment variability is also not interchangeable with reference variability. The test and reference formulations may have different within-subject variances. A test variance can be larger or smaller than the reference variance; either direction can distort the criterion. Repeated test observations can estimate test variability, but they do not identify the variance of repeated reference outcomes without an additional equality assumption.

Assumptions, Diagnostics, and Failure Modes

Before interpreting a reference-scaled result, verify that the variance estimate corresponds to the intended component. The replicate allocation must supply information about repeated reference administrations, and the model must represent the design correctly. The variance scale must match the criterion: a variance in log-units squared cannot be replaced by a standard deviation, a coefficient of variation on the original scale, or a variance from a different transformation without the prescribed conversion.

Diagnostics can reveal when a model-based reference variance is questionable, though they cannot make the wrong variance component appropriate. Inspect residuals and fitted variance components for influential observations, nonconstant spread, unusual replicate differences, and design or period patterns. Check whether reference replicates are represented as such in the model and whether missing periods or unbalanced data have been handled in a way consistent with the design. The earlier material on Residual Diagnostics for Reference-Scaled Models, Instability of Reference Variance Estimates in Small Samples, and Sensitivity to Variance-Model Specification treats these concerns in depth.

A boundary-like estimate, a very wide uncertainty interval, or a large change under reasonable model specifications should prompt scrutiny, not a switch to total variance or the test variance. Likewise, a pooled estimate is not a remedy for an unstable reference estimate: it may conceal the instability by imposing a different variance assumption. If the variance model is not supported, report and assess that limitation using appropriate sensitivity analyses rather than relabeling a different quantity as reference variability.

Common Mistake and Practical Consequences

Wrong reasoning: “The total variance is larger, so it is a conservative choice for scaling,” or “the pooled residual is what the software reports, so it must be the reference variance.” Neither argument identifies the parameter required by the method. In a variance-scaled rule, a larger substituted variance can make the allowable criterion more permissive; it is not necessarily conservative. A smaller substitute can make it more restrictive. The direction depends on the values, but either substitution changes the target.

Correction: Identify the variance component by its definition and by the data that estimate it. Reference replicate contrasts target within-subject reference variability under the relevant model. Total spread, pooled residual spread, and test repeatability have different meanings. The distinction is not bookkeeping: it protects the scaled decision from being driven by subject heterogeneity or by a formulation other than the reference.

The main alternative is a procedure that does not scale limits using reference variability. Such a procedure avoids this particular variance substitution, but it answers the decision question under different acceptance criteria; it does not make total or test variance a valid input to a reference-scaled procedure. FDA and EMA reference-scaled approaches also differ in their rules, so apply the specified jurisdictional method rather than borrowing a variance definition or decision step from another approach.