Estimating Reference Variability in Partial Replicate Designs

Why Partial Replication Can Estimate Reference Variability

Reference-scaled analyses require information about the variability of repeated reference administrations within the same subject. A partial replicate design supplies that information without repeating every treatment: each subject receives the reference product at least twice, while the test product is administered once. In commonly used three-period patterns, the sequences are TRR, RTR, and RRT. The repeated reference observations provide a within-subject contrast; the single test observation does not provide a corresponding test-versus-test contrast.

The key is to separate a subject’s general response level from the variation between that subject’s repeated reference measurements. A difference between two reference observations from one subject removes the subject’s shared level. If period effects are also accounted for, the remaining contrast measures the difference between two reference-specific residual errors.

Model and Identifiability

For an endpoint analyzed on the log scale, a useful starting model for an observation from subject \(i\) in period \(j\) is

Here, \(Y_{ij}\) is the log endpoint, \(\mu\) is an overall level, \(\tau_{T_{ij}}\) is the treatment effect, \(\pi_j\) is the period effect, \(b_i\) is the subject’s shared deviation, and \(\varepsilon_{ij}\) is the within-subject residual. For estimating reference variability, assume that reference residuals are independent across periods, normally distributed with mean zero and common variance \(\sigma_{WR}^2\), and independent of subject effects and treatment assignment. The design must include subjects with two observed reference administrations. Period effects must be absent, negligible, or properly adjusted; otherwise period differences can be mistaken for replicate variability.

For two reference periods \(a\) and \(b\) in the same subject, subtract the period-adjusted observations. The common subject effect and the common reference treatment effect cancel. With period effects removed, the resulting contrast is \(d_i=\varepsilon_{ia}-\varepsilon_{ib}\). Since the two residuals are independent and each has variance \(\sigma_{WR}^2\), the contrast has variance \(2\sigma_{WR}^2\). This identity is the basis for the simple estimator below.

If \(m\) independent, period-adjusted reference differences are available, their sample variance is \(s_d^2=\sum_{i=1}^{m}(d_i-\bar d)^2/(m-1)\). Under the stated assumptions, the method-of-moments estimate is \(s_d^2/2\). In a real crossover analysis, period adjustment and the exact estimator depend on the design and model. The simple calculation is most direct when the differences are independent and period effects are known or have been removed without adding relevant uncertainty.

Worked Reference-Variance Calculations

The log-scale variance can also be expressed as a within-subject coefficient of variation when a lognormal interpretation is suitable. The conversion is \(\mathrm{CV}_{WR}=\sqrt{\exp(\sigma_{WR}^2)-1}\). It is a transformation of the log-scale standard deviation, not a second variance estimator.

Which Treatment Contrasts Remain Estimable?

The repeated reference is not the only useful feature of the design. When both test and reference occur in a connected, appropriately balanced crossover design, the average test-versus-reference treatment contrast can be estimated. For example, each subject in a TRR, RTR, or RRT design receives both treatments, so treatment effects can be compared using the observed treatment data while accounting for period and subject structure. The exact estimate requires the model appropriate to the design; simply subtracting unadjusted group averages is not generally a substitute.

What the design does not identify is a within-subject test variance from repeated test administrations: each subject has only one test observation, so there is no test-versus-test within-subject contrast. Nor does this design alone provide a within-subject test replicate contrast or a subject-specific covariance between repeated test and reference responses. Estimability of an average treatment contrast should not be confused with identifiability of all treatment-specific variance components.

Assumptions, Diagnostics, and Failure Modes

The contrast estimator has a clear interpretation only under its assumptions. Reference observations must be comparable after adjustment; residual variation should be reasonably stable across the relevant periods and subjects; and the paired reference errors should not have an unmodeled correlation that changes the contrast variance. If the reference residuals have variance \(\sigma_{WR}^2\) in both periods but correlation \(\rho\), then

Thus, dividing the difference variance by two estimates \(\sigma_{WR}^2\) only when the paired residuals are uncorrelated. Positive correlation would make that simple estimate too small for the marginal within-subject variance.

Diagnostics should focus on whether the contrasts are constructed appropriately and whether a few observations dominate their spread. Check that each included subject truly has two reference observations, that period and sequence are represented as intended, and that the analysis uses the correct endpoint scale. Inspect period-adjusted contrasts by sequence and period pair; markedly different spreads may indicate unequal variance, an adjustment problem, or a design feature requiring a more appropriate model. Extreme contrasts can have a large effect on a sample variance because deviations are squared. Such observations warrant data checks and sensitivity analysis, not automatic deletion.

Missing one of a subject’s reference administrations removes that subject’s direct reference replicate contrast. A complete-pair calculation then uses fewer subjects and may no longer represent the planned design if missingness is selective. More complex model-based methods can use incomplete records under additional assumptions, but they do not create reference replication where none was observed. In particular, a subject with a single reference observation contributes information to treatment means but not a direct within-subject reference difference.

Alternatives and Limits of the Simple Estimator

The simple difference-variance calculation is transparent and useful for understanding the source of information. Its strength is that the cancellation of shared subject effects is visible. Its limitation is that it relies on a clean set of independent, appropriately adjusted paired differences. A model-based analysis can account for treatment, period, sequence, and the covariance structure together, especially when the allocation is unbalanced or observations are incomplete. Such an analysis must still be supported by a design that repeats reference within subjects.

A full replicate design repeats both test and reference within subjects. It can therefore supply replicate contrasts for both treatments and information about their within-subject variabilities, subject to the model and design. A partial replicate design is more limited: it can identify within-subject reference variability and, in a connected crossover, the average test-reference treatment contrast, but not a within-subject test variance. A parallel design, in which subjects receive only one treatment, lacks the within-subject reference contrast altogether.

The reference variance estimate is an input to reference-scaled procedures, not a standalone conclusion about bioequivalence. Its uncertainty matters, particularly when few subjects have usable reference pairs. The point estimate alone does not show how precisely the variance is known, and a nominally precise-looking number should not obscure limited replication. The inferential procedure must account for the variance-estimation method and its degrees of freedom; those are not determined solely by the identity \(s_d^2/2\).

Common Pitfall