Why the Estimand Has Two Components
Conventional average bioequivalence focuses on the average difference in exposure between test and reference. For a highly variable drug, however, the reference product's own within-subject variability is also relevant to a scaled approach. The reference-scaled average bioequivalence estimand therefore includes both a treatment mean difference and a measure of reference variability. These components answer different questions: how far apart are the treatment averages, and how variable is repeated exposure under the reference treatment within the target population?
The distinction matters even before a decision rule is applied. A mean ratio alone does not describe reference repeatability, while a variability measure alone does not describe the average test–reference difference. The estimand states the population quantities of interest; it is not itself a test statistic, confidence interval, or decision rule.
The endpoint is ordinarily a positive exposure measure, such as AUC or \(C_{\max}\), analyzed on the natural-log scale. On that scale, \(\Delta\) is a difference in mean log exposure. Exponentiating it gives the ratio of geometric means on the original scale:
Thus, \(\Delta=0\) corresponds to a geometric mean ratio of 1. A positive \(\Delta\) means the test treatment has the higher geometric mean; a negative \(\Delta\) means it has the lower one. The variance component \(\sigma^2_{WR}\) is measured in squared log units, and its square root, \(\sigma_{WR}\), is a within-subject standard deviation on the log scale. Neither quantity is the variance of all observations pooled across participants.
What Within-Subject Reference Variance Represents
Within-subject reference variance describes how repeated log-endpoint values vary when the same participant receives the reference treatment under comparable conditions. It is a population characteristic of repeatability, not a comparison of test and reference means. Participants can differ substantially in their typical exposure, yet repeated reference measurements within each participant can still be quite consistent.
A simple model makes the distinction precise. Suppose a participant's \(j\)th reference-period log endpoint is
where \(b_i\) represents participant \(i\)'s stable deviation from the population mean and \(e_{Rij}\) represents the within-subject deviation for that reference occasion. If the occasion deviations have mean zero and variance \(\sigma^2_{WR}\), then \(\sigma^2_{WR}\) describes repeatability after accounting for stable between-participant differences. The model is a conceptual way to define the quantities; an actual analysis must use a design and model appropriate to its data.
The definition presumes that the repeated reference administrations represent the same treatment, endpoint, and relevant conditions in the target population. Changes in formulation, dose, sampling, adherence, or study conditions can make the repeated observations incomparable. The variance is also specific to the endpoint and scale: a within-subject variance for log AUC is not automatically the variance for log \(C_{\max}\).
Worked Example: Interpreting the Mean Difference
Suppose a hypothetical study targets the population mean log AUC under test and reference treatments. Assume the relevant population means are \(\mu_T=7.42\) and \(\mu_R=7.35\), with AUC recorded in consistent units before taking natural logs. Find the mean difference and its geometric mean ratio interpretation.
Mean difference. By definition,
Geometric mean ratio. Exponentiate the log-scale difference:
The test geometric mean AUC is therefore about 1.073 times the reference geometric mean AUC, or about 7.25% higher. This describes the relative population averages; it says nothing by itself about how variable repeated reference AUC values are within participants.
A Short Justification for the Variance Component
Under the simple model above, take two repeated reference observations from the same participant. Their difference removes the participant's stable deviation:
If the two occasion deviations have the same variance \(\sigma^2_{WR}\) and are uncorrelated, the variance of their difference is
This explains why repeated reference observations can reveal within-subject variability that a single reference observation per participant cannot separate from other sources of variation. The identity depends on the stated equal-variance and zero-covariance assumptions. If the occasions are correlated or have different variances, dividing the difference variance by two does not, in general, recover the common within-subject variance. The identity is a justification of the estimand's interpretation, not a complete estimation procedure.
Worked Example: Relating Repeat Differences to Reference Variability
Consider a hypothetical population model in which repeated log-reference AUC values for a participant have within-subject standard deviation \(\sigma_{WR}=0.25\). Assume the two reference occasions have equal variance and uncorrelated occasion deviations. What variance would their within-participant difference have?
First square the standard deviation to obtain the within-subject variance:
By the derivation, the variance of the difference between two reference observations is twice this value:
The variance is \(0.1250\) squared log-AUC units. The calculation concerns the distribution of within-participant repeat differences, not the spread of participants' average AUC values. It relies on the equal-variance and uncorrelated-occasion assumptions stated in the question.
How This Differs From Unscaled Average Bioequivalence
An unscaled mean-ratio comparison uses the treatment mean difference on the log scale, or equivalently its exponentiated geometric mean ratio, and evaluates it against prespecified fixed limits. Its estimand is about the average test–reference contrast. Reference-scaled average bioequivalence retains that mean difference but also includes the reference within-subject variance. The latter is the feature that allows a scaling approach to account for reference repeatability; the particular scaling rule and decision limits are separate matters.
The distinction can be seen by comparing two hypothetical products with the same \(\Delta\), but different \(\sigma^2_{WR}\). An unscaled mean-ratio description gives the same average ratio for both. Their reference-scaled estimands differ because their reference repeatability differs. This does not, by itself, say whether either product meets a regulatory criterion: that conclusion requires the applicable rule, its assumptions, and the required analysis.
| Feature | Unscaled average comparison | Reference-scaled estimand |
|---|---|---|
| Average treatment contrast | Uses \(\Delta\), or \(\exp(\Delta)\) | Includes \(\Delta\), or \(\exp(\Delta)\) |
| Reference repeatability | Not a component of the mean-ratio estimand | Includes \(\sigma^2_{WR}\) |
| Interpretation | Relative population averages | Relative averages together with within-subject reference variability |
| What the estimand alone decides | No equivalence decision without margins and an inferential procedure | No scaled-equivalence decision without a specified scaling rule and procedure |
Worked Example: Same Ratio, Different Reference Variability
Two hypothetical drug comparisons have the same log-scale mean difference, \(\Delta=0.0400\). The first has \(\sigma_{WR}=0.15\); the second has \(\sigma_{WR}=0.30\). Compare their estimands without making a bioequivalence decision.
For both comparisons, the geometric mean ratio is
So each has a test geometric mean about 4.08% above its reference geometric mean. For the first comparison, the reference variance is
For the second, it is
The unscaled mean-ratio description is identical, but the reference-scaled estimands are \((0.0400,0.0225)\) and \((0.0400,0.0900)\). The second has four times the reference variance, while its within-subject standard deviation is twice as large. These facts distinguish the estimands; they do not establish that the second comparison qualifies for a particular scaled limit.
Assumptions, Diagnostics, and Limits
Before interpreting either component, specify the population, endpoint, log transformation, treatment conditions, and target study context. For \(\Delta\), the test and reference means must refer to the same endpoint definition and comparable conditions. For \(\sigma^2_{WR}\), the reference observations must be repeated within participants, and the intended variability must be distinguished from period effects, treatment changes, or other systematic sources of variation. A model may need to account for design features such as sequence and period; the estimand does not excuse an inappropriate analysis model.
Useful diagnostic questions include whether the repeated reference observations were obtained under the same dose and protocol, whether there are enough replicate observations to support the target variance, and whether plots or model summaries suggest unusual period patterns or influential observations. If each participant receives the reference only once, between-participant variation and within-subject reference variation are not separately identified by those observations alone. If the repeated administrations are not comparable, the resulting variability may describe more than reference repeatability.
The estimand is specific, not universal. A different endpoint, population, dosing condition, or log-scale definition can yield different quantities. Further, a variance component is subject to sampling uncertainty: reporting a point estimate alone does not make it known exactly. The next analytical steps concern how replicate designs identify and estimate this component; the definition here should not be confused with any one estimator.
Common Mistake
A common error is to use the variance of all reference measurements across all participants as \(\sigma^2_{WR}\). That pooled variance includes stable differences between participants as well as variation between repeated observations within a participant. It can therefore be much larger than the target within-subject variance. The correction is to define the variance at the within-participant level and use a design and analysis capable of separating it from between-participant variation. A second error is to treat a large \(\sigma^2_{WR}\) as evidence that the test and reference means are close; variability and mean difference are distinct estimand components.
Check Your Understanding
Answer each question. Solutions follow so that the definitions and calculations can be checked directly.
- If \(\mu_T=5.90\) and \(\mu_R=5.85\), what is \(\Delta\), and what geometric mean ratio does it imply?
- Why does one reference observation per participant generally fail to identify within-subject reference variance separately from between-participant variation?
- Two comparisons have the same \(\Delta\) but different \(\sigma^2_{WR}\). Which component of their estimands differs, and what cannot be concluded from this fact alone?
- Under the equal-variance, uncorrelated-occasion model, what is the variance of the difference between two repeated reference log measurements if \(\sigma_{WR}=0.20\)?
Solutions. (1) \(\Delta=5.90-5.85=0.0500\), so the geometric mean ratio is \(\exp(0.0500)=1.0513\), about 5.13% above 1. (2) With only one reference value per participant, there is no within-participant repeat contrast to separate occasion-level variation from stable participant differences. (3) The reference variance component differs; without the applicable scaled criterion and inferential analysis, no equivalence decision follows. (4) \(\sigma^2_{WR}=(0.20)^2=0.0400\), so the difference variance is \(2(0.0400)=0.0800\) squared log units.