Replicate Crossover Analysis of a Highly Variable Drug

Why Replication Matters

A highly variable drug can produce noticeably different exposure measurements when the same person receives the same formulation on different occasions. A conventional two-period crossover, in which each participant receives test once and reference once, estimates the average treatment difference but cannot separately estimate the within-subject variability of reference. Replicate crossover designs include repeated administrations of at least one formulation, making it possible to estimate that variability directly.

This tutorial uses a four-period, fully replicated design with the sequences TRTR and RTRT. Each participant receives test (T) twice and reference (R) twice. The analysis uses a log-transformed pharmacokinetic endpoint, such as AUC or \(C_{\max}\). The replicate data support an estimate of reference within-subject variability; they do not, by themselves, establish that test and reference are bioequivalent. The treatment comparison and its uncertainty still need to be assessed.

The coefficient of variation formula assumes a lognormal interpretation: a log-scale standard deviation \(\sigma_{wR}\) corresponds to a multiplicative spread on the original scale. It is a descriptive measure of repeatability, not a measure of the difference between test and reference means. In particular, high reference variability does not imply that the test formulation has a different mean exposure.

Model and Assumptions

Let \(Y_{ijk}\) be the log-transformed endpoint for subject \(i\), in period \(j\), under treatment \(k\), where \(k\) is T or R. One suitable mixed model for a balanced replicate design is

Sequence, period and treatment are fixed effects. The subject effect \(b_{i(\text{sequence})}\) is random and accounts for correlation among measurements from the same participant. Residual variability may be allowed to differ by treatment: \(\epsilon_{ijk}\) has mean zero and variance \(\sigma_{wT}^{2}\) under T and \(\sigma_{wR}^{2}\) under R. The participant-specific effect captures stable differences between people; the treatment-specific residual variances describe within-person variation not explained by the fixed effects.

This model assumes independent participants, a correctly recorded sequence and period for each administration, an adequate washout so that prior treatment does not affect later outcomes, and a reasonable fit of the log-scale distributional and variance assumptions. It also assumes that the fixed-effect structure adequately represents period and sequence patterns. Residual diagnostics and the design history matter: a variance estimate cannot repair carryover, a mislabeled treatment, or a systematic period effect omitted from the model.

The treatment effect of interest is the model-based difference \(\Delta=\mu_T-\mu_R\) on the log scale. Exponentiating it gives the test-to-reference geometric mean ratio, \(\exp(\Delta)\). For an average bioequivalence assessment, the conventional 90% confidence interval for this ratio is compared with 0.80 to 1.25. The variance estimate \(\sigma_{wR}^{2}\) is reported separately. Reference-scaled bioequivalence uses additional rules and is not inferred merely by observing that this variance is large.

Estimating Reference Within-Subject Variability

Conceptually, repeated reference observations from the same participant provide information about how much reference exposure varies within that participant. If two reference log observations have independent residual errors with variance \(\sigma_{wR}^{2}\), their difference has variance \(2\sigma_{wR}^{2}\): the shared subject effect cancels, while the two residual variances add. This motivates a simple moment estimate based on squared within-subject differences.

In a real four-period analysis, period effects and other fixed effects must be included. A mixed-model or other appropriate model-based estimate is preferable to treating raw reference differences as if periods were identical. The following calculation illustrates the variance scale using period-adjusted replicate differences; it is a transparent moment calculation, not a substitute for fitting the prespecified model to the full dataset.

Estimating the Treatment Effect and Assessing Average BE

The treatment comparison uses all observations in the fitted model, accounting for period, sequence and within-subject dependence. Let \(\hat{\Delta}\) be the estimated log-scale test-minus-reference effect, \(SE\) its standard error and \(df\) the model-based degrees of freedom. A two-sided 90% confidence interval is \(\hat{\Delta}\pm t_{0.95,df}SE\). Exponentiating both endpoints produces the interval for the geometric mean ratio. The interval, rather than the point estimate alone, determines whether the conventional average BE criterion is met.

The 90% interval has a frequentist interpretation: under the model and repeated use of this interval procedure, 90% of intervals constructed in this way would cover the true log-scale treatment effect. It is not a 90% probability statement about the fixed parameter after observing this particular interval. Likewise, a passing average BE interval does not show that every participant has similar exposures under the two formulations.

Putting the Results Together

A useful report keeps the treatment comparison and the repeatability estimate distinct. For each pharmacokinetic endpoint, report the sequence and period design, the analysis scale and model, the number of participants and usable observations, the test-to-reference geometric mean ratio, its 90% confidence interval, and the BE conclusion. For replicated reference data, also report the estimate of \(\sigma_{wR}^{2}\), its corresponding \(\sigma_{wR}\) or within-subject CV, and the estimation method. When available, include uncertainty for the variance estimate and state whether the model allowed treatment-specific residual variances.

A high estimated reference CV is relevant because it signals substantial within-person repeatability variation in this study. It does not change the arithmetic of the conventional average BE interval, nor does it make a confidence interval pass automatically. The later question of whether and how variability changes the decision criterion requires a separate reference-scaled assessment.

Diagnostics, Alternatives, and Limitations

Diagnostics should be connected to the model assumptions. Large or patterned residuals may indicate a poor distributional fit, an unmodeled period effect, data errors or an influential profile. A strong sequence or period pattern deserves investigation, not automatic removal: it may reflect chance, a design or conduct issue, or a failure of the assumed model. Examine whether the estimated variance is plausible relative to the observed replicate measurements and whether it is driven by one participant. Confirm that the result is stable under justified sensitivity analyses, especially when missing observations or protocol deviations affect the replicate structure.

A simple paired comparison of each participant's mean T and mean R can be useful as a descriptive check, but it does not fully replace the prespecified crossover model when period and sequence effects matter. Conversely, a model with one common residual variance can be more parsimonious, but it may not answer the specific question of reference within-subject variability if test and reference residual variability differ. Model choice should follow the study design and analysis plan rather than whichever method gives the most favorable result.

A replicate design still has limitations. Variance estimates can be imprecise, particularly with few participants; unequal or missing replications complicate estimation; and a variance estimate is sensitive to model assumptions and data quality. Replication also does not eliminate the need to evaluate carryover, period effects, endpoint definition or the appropriateness of the log transformation. Finally, the average BE confidence interval answers a question about average exposure, not individual interchangeability or clinical outcomes.