What Replicate Observations Identify

What Repetition Adds to a Crossover Study

Reference-scaled bioequivalence depends on two different kinds of information: a comparison of the test and reference treatment means, and an estimate of how much the reference response varies within a subject across administrations. The mean comparison concerns the relative average responses to T and R. The reference-variability quantity concerns repeated responses to R in the same subject. These are not interchangeable features of the data.

The treatment sequences determine which comparisons the observations can support. A study in which every subject receives T and R once can compare treatments within subjects, but it cannot observe that subject’s response to R on a second occasion. Replicate sequences add those repeated administrations. Their value is not simply that they produce more measurements: they create within-subject comparisons that are unavailable in a conventional two-period crossover.

Reading the Design Before Analyzing the Data

In a crossover study, a sequence lists the treatment each subject receives in each period. For example, TRR means that the subject receives T in the first period and R in each of the next two periods. For a specified endpoint, the recorded response is commonly analyzed on the log scale. The administrations must be separate treatment periods, not merely multiple instrument readings from one sample: the latter are technical replicates and do not measure variation across treatment occasions.

A full replicate design gives each subject two administrations of T and two of R. Example sequences include TRTR and RTRT. Each subject can contribute repeated observations under both treatments. A partial replicate design gives each subject two administrations of R but only one of T; example sequences include TRR, RTR, and RRT. Across suitable sequences, both designs provide T and R observations for comparing treatment means, and repeated R observations for assessing within-subject reference variability. Only the full replicate also repeats T within each subject.

The design must be considered together with its assumptions. Treatment assignment to sequences should be randomized or otherwise appropriately controlled, and the period structure should permit treatment effects to be separated from period effects. Washout must be sufficient for the intended crossover interpretation; meaningful carryover could make a later-period response depend on an earlier treatment. Repeated reference administrations must represent comparable occasions for the reference formulation. If the occasions have systematic differences, or if within-subject variability changes with period, a simple repeated-measure contrast may not represent one common \(\sigma_{WR}^2\).

The observations do not identify these quantities by sequence labels alone. The endpoint, period, sequence, missingness, and analysis model matter. The basic mapping below describes what information a design supplies, not a guarantee that any particular estimator is unbiased or precise.

Why a Repeated Reference Contrast Contains Variability Information

Consider a subject with two log responses to R, \(Y_{R1}\) and \(Y_{R2}\), from separate administrations. A useful conceptual representation is \(Y_{Rj}=\theta+\delta_j\), where \(\theta\) represents the subject’s stable level and \(\delta_j\) represents occasion-specific variation. The subject’s stable level is shared by both observations. Subtracting the responses removes that stable component.

If the occasion deviations have mean zero, equal variance \(\sigma_{WR}^2\), and are independent across administrations, then the difference has mean zero and variance

Thus, the spread of within-subject R differences contains information about reference variability. In an actual analysis, period and sequence effects must also be handled as appropriate, and the exact estimator depends on the design and model. If the two occasions are correlated, the difference variance is instead \(2\sigma_{WR}^2-2\operatorname{Cov}(\delta_1,\delta_2)\); the simple division-by-two interpretation then requires the covariance to be zero or otherwise accounted for. This is why the repeated administrations and their assumptions matter.

The T–R mean contrast answers a different question. It is estimated from observations under both treatments, with the crossover structure used to control stable subject differences and the period structure handled in the analysis. Repeating R does not itself define the mean contrast, and observing T and R once each does not itself reveal the variation between repeated R administrations.

Worked Reference Contrasts

What to Check, and What Replication Cannot Fix

Before interpreting replicate observations, check whether the intended repetitions are actually present and comparable. Verify that each relevant subject has the planned number of administrations, confirm treatment and period assignments, and examine whether missing periods or deviations from sequence have left particular treatments or periods poorly represented. A repeated R response recorded in a different endpoint, on a different scale, or under a materially different protocol is not automatically a valid within-subject reference contrast.

A pattern in which most R differences are positive or negative may indicate a period shift, an order effect, or another systematic feature rather than purely random occasion variation. Large differences concentrated in a few subjects may make a variance estimate sensitive to individual observations. These patterns are reasons to inspect the data and the design assumptions; they do not, by themselves, identify the cause. A later model-based analysis can address structured effects, but cannot make an unobserved repeat or an unrepresented period appear in the data.

Three design configurations are especially useful to distinguish:

• Conventional two-period crossover: each subject receives T and R once. It supports a within-subject treatment comparison, but has no repeated R administration to identify within-subject reference variance.
• Partial replicate: each subject receives R twice and T once. It supports the mean comparison and repeated-reference information, but does not provide a within-subject T repeat.
• Full replicate: each subject receives T twice and R twice. It supports both the mean comparison and repeated within-subject observations under each treatment, at the cost of additional periods and operational complexity.

Replication has limitations. It does not remove the need for suitable randomization, adequate washout, period control, or a defensible variance model. It also does not guarantee a precise estimate: a small sample may produce an unstable estimate even when the design is capable of identifying the relevant variance. Technical replicates, duplicate readings from a single administration, cannot substitute for separate administrations when the target is within-subject occasion variability.