Reference-Replicate Contrasts for Estimating Variability

Why Contrast Repeated Reference Doses?

In a replicate bioequivalence design, some subjects receive the reference product more than once. Those repeated observations make it possible to learn about within-subject reference variability: how much a subject’s exposure to the reference product varies across administrations. The central idea is to compare the repeated reference observations within each subject, rather than treating all observations as if they came from different subjects.

For this tutorial, let \(Y_{i1}\) and \(Y_{i2}\) denote subject \(i\)’s two log-transformed exposure measurements after reference administrations. A log-transformed endpoint could be log AUC or log \(C_{\max}\). Define the signed reference-replicate contrast as \(D_i=Y_{i1}-Y_{i2}\). The sign records the order of subtraction; it does not alter the squared contrast \(D_i^2\).

The contrast is informative because stable differences between subjects can be large without indicating high within-subject variability. A subject who tends to have higher exposure than other subjects may have two high reference measurements. Subtracting those measurements removes that subject’s additive level, leaving a comparison of the subject’s repeat-to-repeat deviations.

Model and Assumptions

A simple model makes the contrast’s meaning precise. Suppose the two reference observations for subject \(i\) satisfy

Here \(\mu\) is an overall mean, \(b_i\) is the subject-specific additive effect, \(\pi_j\) is the mean effect of the period in which the observation is made, and \(\varepsilon_{ij}\) is the within-subject residual. Assume the residuals are independent across subjects and occasions, normally distributed with mean zero and common variance \(\sigma_{wR}^2\). The subject effect can vary across subjects, but it is the same for a given subject’s two reference observations in this model.

Subtracting the two observations gives

The terms \(\mu\) and \(b_i\) cancel. If the two periods have the same mean effect, then \(E(D_i)=0\). More generally, a common period difference can shift every contrast by the same amount; centering the contrasts removes that common shift when estimating their variance. Subject to the model assumptions, the contrast variance is

The factor of two matters: each contrast contains residual variation from two reference administrations. Consequently, the variance of the contrasts is twice the within-subject reference variance, not the within-subject variance itself.

Estimating Variability From Squared Contrasts

With \(m\) subjects who each contribute one contrast, let \(\bar D\) be the sample mean of the contrasts and \(s_D^2\) their sample variance. The centered variance estimate is

Each subject contributes a squared deviation, \((D_i-\bar D)^2\). Larger deviations contribute more to the estimate because they are squared; the estimate is not the average absolute difference. Under independent normal contrasts with common variance, the centered sum of squares divided by the contrast variance has a chi-square distribution with \(m-1\) degrees of freedom. Thus the estimator above is unbiased for \(\sigma_{wR}^2\) under those assumptions.

There is a related formula when the contrast mean is known to be zero, rather than estimated from the data: \(\sum_i D_i^2/(2m)\). In practice, an observed contrast mean can reflect sampling variation or a period difference. Unless the zero-mean condition is justified, centering the contrasts is the safer variance calculation. The two formulas need not give the same result in a finite sample.

A Short Derivation and an Uncertainty Example

The division by two follows directly from the variance of a difference. If the two residuals are independent and each has variance \(\sigma_{wR}^2\), then

For independent normally distributed contrasts with a common mean and variance, \((m-1)s_D^2/\operatorname{Var}(D_i)\) has a chi-square distribution with \(m-1\) degrees of freedom. Substituting \(\operatorname{Var}(D_i)=2\sigma_{wR}^2\) gives a confidence interval for the variance. This interval is useful for seeing how imprecise an estimate from a small number of contrasts may be; it does not by itself decide bioequivalence.

When Contrasts Mislead

The calculation is only as meaningful as its assumptions. A contrast removes a subject’s additive level, but it does not automatically remove every design or measurement feature that affects repeated observations. Important checks include:

• Period patterns: Plot or tabulate contrasts by the periods being compared. A common period shift changes the contrast mean, while a period effect that varies across subjects can also affect the spread.
• Outlying contrasts: Because squared deviations grow quickly, one unusually large contrast can dominate the estimate. Check the underlying measurements and protocol records rather than deleting a value solely because it is large.
• Equal residual variances: If the two reference occasions have different residual variances, the contrast variance is their sum, not twice either one variance. The simple estimate then does not isolate a single common variance without additional assumptions.
• Independence within a subject: If the two residuals have covariance \(c\), then \(\operatorname{Var}(D_i)=2\sigma_{wR}^2-2c\) when their variances are equal. Positive correlation makes the contrast-based estimate smaller than \(\sigma_{wR}^2\) if one still divides by two.
• Independence across subjects: Dependence between subjects can make the usual uncertainty calculation unreliable, even if the point estimate can still be computed.

A useful diagnostic is to inspect both the signed contrasts and their squares. The signed values can reveal an overall shift or an ordering pattern; the squares show which subjects contribute most to the variability estimate. Neither plot proves the assumptions, but unusual patterns can signal that the simple contrast summary is inadequate.

How This Approach Compares With Alternatives

Reference-replicate contrasts are transparent: the subject-level subtraction is visible, the cancellation of a stable subject effect is explicit, and each squared deviation has a direct contribution to the estimate. They are especially helpful for understanding where information about within-subject reference variability comes from.

A raw standard deviation calculated across all reference observations is not an equivalent shortcut. It mixes between-subject differences with within-subject repeatability, so it can be large even when each subject’s repeated measurements are close together. Conversely, reporting only the average absolute difference does not estimate the same variance quantity.

A model-based analysis can use the design structure and accommodate more complicated patterns, such as multiple periods, incomplete replication, or additional fixed effects. Such an analysis requires its own model specification and assumptions. A simple paired-contrast estimate is a useful direct summary when each subject has the required pair of reference observations and the assumptions are plausible; it is not a substitute for a model that the design requires.

Common Mistake and Practical Limits

Another tempting shortcut is to divide the uncentered sum of squared contrasts by \(2m\) even when their mean is not known to be zero. Since the average squared contrast includes both its variance and the square of its mean, a nonzero mean can inflate that calculation. Centering estimates the contrast variance, but it does not explain why the mean differs from zero: period effects or other design features still deserve attention.

Finally, a contrast estimate from few subjects can be unstable, particularly when a single squared difference is large. Its role is to summarize information about reference variability under stated assumptions. It does not establish that the assumptions hold, and it should not be interpreted as a direct measurement of every subject’s individual variability.