Why Replication Matters for Individual Bioequivalence
Planning individual bioequivalence (IBE) requires information about more than the average difference between formulations. The relevant question includes whether individuals vary in their response to the formulation change. A crossover in which each subject receives each formulation only once cannot cleanly distinguish that subject-by-formulation variation from ordinary within-subject measurement variability. Replicating both formulations supplies the repeated observations needed to estimate these components separately.
This tutorial focuses on design information for estimating subject-by-formulation variability. It does not replace the IBE power target or criterion discussed in the earlier material on individual bioequivalence and population and individual bioequivalence power targets. The central planning decisions are how many subjects to enroll and how many evaluable administrations of each formulation each subject needs.
A common starting point is a balanced replicate crossover: each subject receives test (T) twice and reference (R) twice, with treatment order arranged across sequences to balance period and carryover concerns. For subject \(i\), let \(d_i\) be the subject’s test-minus-reference contrast, calculated from that subject’s mean response on T minus mean response on R. A useful conceptual model is \(d_i=\Delta+b_i+e_i\), where \(\Delta\) is the population mean contrast, \(b_i\) is the subject-specific departure in the contrast, and \(e_i\) is sampling variation from the finite number of administrations. The variance of \(b_i\) represents between-subject heterogeneity in formulation response; the variance of \(e_i\) depends on within-subject variability and the number of repeats.
For positive exposure measures, such as AUC or Cmax, the analysis is generally conducted on the log scale. In this tutorial, all response values and variability parameters are therefore assumed to be on that scale unless stated otherwise. The estimated contrast is then in log units, and variance components are in log-unit-squared.
Assumptions and the Information Supplied by Repeats
A planning calculation for variance-component precision relies on assumptions about the design and data-generating process. The model should be appropriate for the endpoint and should account for systematic period, sequence, and formulation effects. After those effects are accounted for, subject effects and subject-by-formulation effects are assumed to have the specified variance structure; residual errors are independent across administrations, approximately normally distributed, and have formulation-specific variances that are stable across periods. Washout must be adequate so that residual drug from an earlier period does not contaminate a later observation.
These assumptions are not guaranteed by the word “replicate” in a protocol. A poorly balanced schedule can confound treatment with period, and missed administrations can leave some subjects with incomplete contrasts. The model and estimand must be specified before using a sample-size calculation. The earlier material on evaluating variance estimates for sample-size planning is relevant when selecting plausible variance inputs.
With \(r_T\) test administrations and \(r_R\) reference administrations, each additional repeat helps estimate that formulation’s within-subject variance and makes the subject’s mean response more precise. In contrast, recruiting another subject adds another independent subject-level contrast, which directly informs the distribution of subject-by-formulation responses. More periods are not always an efficient substitute for more subjects: repetitions within one individual cannot provide as many independent subject contrasts as recruiting additional individuals.
For a balanced design with \(r\) administrations of each formulation, the measurement contribution to the variance of a subject’s mean contrast is \( (\sigma_{WT}^{2}+\sigma_{WR}^{2})/r \), where \(\sigma_{WT}\) and \(\sigma_{WR}\) are the within-subject standard deviations for T and R. This shows why the number of repeats matters. It does not, by itself, give the precision of a variance-component estimate: that precision also depends on the number of independent subjects and on the method used to estimate the components.
Worked Example: Why One Observation per Formulation Is Not Enough
Hypothetical design. Suppose a two-period crossover gives each subject one T observation and one R observation. The observed contrast for subject \(i\) is \(d_i=Y_{iT}-Y_{iR}\). Under the model above, its variability includes both subject-by-formulation heterogeneity and the within-subject errors from the two observations.
If the within-subject errors have variances \(\sigma_{WT}^{2}\) and \(\sigma_{WR}^{2}\), then \( \operatorname{Var}(d_i)=\operatorname{Var}(b_i)+\sigma_{WT}^{2}+\sigma_{WR}^{2} \). There is only one observation per formulation for that subject, so the data do not separately estimate the two within-subject variances from that subject’s repeated values.
With two observations per formulation, the subject’s contrast uses formulation means: \(d_i=\bar{Y}_{iT}-\bar{Y}_{iR}\). Under independent within-subject errors, \( \operatorname{Var}(d_i)=\operatorname{Var}(b_i)+\sigma_{WT}^{2}/2+\sigma_{WR}^{2}/2 \). The replicate observations also provide within-subject differences for estimating each formulation’s residual variance. Replication therefore adds the information needed to distinguish noisy measurement from genuine heterogeneity in subject response.
This distinction depends on the model and the balanced design; replication does not automatically make every variance component estimable. A suitable schedule must ensure the treatment contrasts and residual variation are identifiable after accounting for period and sequence.
Planning Precision for a Variance Component
A useful planning objective is a target standard error (SE) for the estimate of subject-by-formulation variability. This is a precision target, not a substitute for the IBE power target. It can identify a design with too few independent subjects to estimate the variance component stably, even when an enrollment number based only on a mean contrast appears adequate.
For a balanced two-replicate design, consider the estimator \(\hat{\tau}^{2}=S_d^{2}-q\), where \(S_d^{2}\) is the sample variance of the independent subject contrasts and \(q=(\sigma_{WT}^{2}+\sigma_{WR}^{2})/2\) is their within-subject measurement variance. If the contrasts are normal with variance \(\tau^{2}+q\), and the within-subject variances are treated as known, then \( \operatorname{Var}(\hat{\tau}^{2})=2(\tau^{2}+q)^{2}/(n-1) \). The square root gives the SE for this estimator. This calculation does not include uncertainty from estimating the within-subject variances.
The result follows because subtracting the known constant \(q\) does not change the sampling variance of \(S_d^{2}\), and a sample variance from \(n\) independent normal observations with variance \(v\) has variance \(2v^{2}/(n-1)\). Thus the subject-contrast contribution depends on the total contrast variance, not just \(\tau^{2}\). If within-subject variances are estimated, their uncertainty and covariance with the contrast variance must also be accounted for using the planned estimator.
Worked Example: Finding the Minimum Subject Count for a Precision Target
Hypothetical planning inputs. Suppose the estimator-specific calculation uses \(\tau=0.0688\), \(\sigma_{WT}=\sigma_{WR}=0.20\), and a target SE no greater than \(0.0060\) log-unit-squared. The number \(n\) is the number of evaluable subjects. Here \(q=(0.20^{2}+0.20^{2})/2=0.04\), so use \( SE(n)=\sqrt{2(0.0688^{2}+0.04)^{2}/(n-1)} \).
At \(n=112\), the estimated variance of the variance-component estimate is \(2(0.00473344+0.04)^{2}/111=0.0040021613/111\approx0.0000360555\). Thus \(SE(112)\approx0.00600462\) log-unit-squared, which exceeds the \(0.0060\) target.
At \(n=113\), the corresponding calculation is \(0.0040021613/112\approx0.0000357336\). Therefore \(SE(113)\approx0.00597776\) log-unit-squared, which is below the target. Since this SE decreases as \(n\) increases, 113 evaluable subjects are the minimum that meets the precision criterion under this calculation. Enrollment would need to be higher if some participants are expected to be nonevaluable.
Choosing Replicate Counts and Checking the Design
The number of replicates should be chosen in light of both the estimand and the practical study burden. A four-period design with two administrations of each formulation is a common way to obtain balanced repeated observations. A design with additional repeats can improve within-subject variance estimation and reduce the measurement noise in individual contrasts, but it requires more periods per subject and may increase withdrawals or missed visits. If the main information need is the population distribution of individual contrasts, adding subjects can be more valuable than adding further periods to the same small group.
Before finalizing a plan, check the proposed treatment sequences, period balance, washout, and expected completion pattern. Check that each subject has enough evaluable administrations of both formulations for the planned estimator. Use plausible values for each variance component rather than assuming that test and reference within-subject variability must be equal. Evaluate sensitivity to uncertainty in those inputs, as emphasized in the earlier material on sample-size planning under uncertain variance.
Worked Example: Comparing Two and Three Repeats
Hypothetical design comparison. Suppose the within-subject standard deviations are both \(0.15\) log units. With two administrations per formulation, the measurement contribution to the variance of a subject’s mean contrast is \( (0.15^{2}+0.15^{2})/2=(0.0225+0.0225)/2=0.0225 \) log-unit-squared.
With three administrations per formulation, it is \( (0.15^{2}+0.15^{2})/3=0.0450/3=0.0150 \) log-unit-squared. The difference is \(0.0225-0.0150=0.0075\) log-unit-squared. In this model, the subject-by-formulation component is unchanged; the extra administrations reduce the contribution from within-subject error to each individual’s mean contrast.
This comparison does not establish that three repeats are the preferred design. They cost an additional period per formulation and may affect completion. The next decision is whether the reduction in contrast noise is worth that burden compared with enrolling more subjects. That decision should use the planned analysis, the intended precision or power target, and realistic completion assumptions—not the variance reduction alone.
Diagnostics, Alternatives, and Limitations
After data collection, inspect the estimates and their uncertainty before interpreting the estimated subject-by-formulation variability. Warning signs include a variance-component estimate at or near zero despite substantial observed contrast variation, highly imbalanced numbers of usable repeats, extreme subject contrasts, and marked differences in residual spread across periods or formulations. A negative unconstrained variance-component estimate can occur with finite samples; it signals that the data provide weak evidence for a positive component under that estimator, not that a true variance is negative. Constrained estimation may report zero, but that boundary estimate still warrants attention to uncertainty and model fit.
A replicate design is not the only possible approach. A nonreplicate crossover can often inform the average test-reference contrast, but it has limited ability to separate subject-by-formulation heterogeneity from within-subject error. Replication of only one formulation can estimate that formulation’s within-subject variability, but does not provide the same direct repeated information for both formulations. A balanced full replicate design is more informative for the two within-subject variances and individual contrasts, at the cost of extra treatment periods.
No simple SE target fully describes operating characteristics for a nonlinear or multi-component IBE criterion. The approximation depends on normality, variance stability, independence, balance, and the accuracy of the planning inputs. It can be a useful design screen, but it may be unreliable with small samples, substantial missingness, strong heterogeneity, or a complex covariance structure. Where these features matter, a more complete model-based power assessment is needed; complex power calculations are beyond this tutorial’s scope.
State which subject-by-formulation variance is needed and how period, sequence, and residual variation will be handled.
Ensure both formulations have enough repeated administrations to estimate their within-subject variation and form subject-level contrasts.
Choose an SE target for the variance-component estimate, alongside—not instead of—the relevant IBE power objective.
Use a formula matched to the design, check neighboring values of \(n\), examine plausible variance inputs, and account for nonevaluable subjects.
Check Your Understanding
Answer these questions using the definitions and planning logic above. Solutions are provided after each question.
- Why does a replicate design need repeated observations of both formulations to estimate subject-by-formulation variability? Solution: Repeats provide within-subject information for each formulation, helping separate residual measurement variability from variation among subjects in their test-reference response contrast.
- In a balanced design with three observations per formulation and within-subject standard deviations \(0.10\) and \(0.15\) log units, what is the measurement contribution to the variance of a subject’s mean contrast? Solution: \( (0.10^2+0.15^2)/3=(0.0100+0.0225)/3=0.01083 \) log-unit-squared, to four significant figures.
- What does the count \(n\) represent in a subject-level variance-component planning calculation? Solution: It is the number of independent evaluable subjects, not the total number of administrations.
- For the inputs in the sample-size example, does \(n=77\) meet an SE target of \(0.0060\) log-unit-squared? Solution: No. Under the estimator-specific calculation, the SE is approximately \(0.007257\) log-unit-squared, which is greater than \(0.0060\).
- Name one assumption whose failure could undermine the planning calculation. Solution: Examples include residual dependence across administrations, inadequate washout, changing residual variance across periods, or a schedule that confounds treatment with period. Each can invalidate the assumed variance decomposition.