Estimating Within-Subject Variability In Replicate Designs

Why Replication Is Needed

A formulation comparison concerns more than average exposure. Subjects can differ in their typical exposure, and the same subject can have different exposures on separate administrations of the same formulation. Replicate designs, in which a subject receives a formulation more than once, provide information to distinguish those sources of variation. Without replication, a single observation on a formulation cannot generally separate its occasion-to-occasion variability from other unexplained variation.

In model-based pharmacokinetic analysis, the observations may be concentrations used to fit a structural model, while the quantities of interest are exposure metrics such as AUC or Cmax derived from that model. Variability should therefore be defined on the scale of the quantity being compared. This tutorial uses a log-transformed exposure metric to make the variance components explicit. A concentration-level model can estimate occasion-level random effects on pharmacokinetic parameters and propagate them to exposure metrics, but concentration residual error is not automatically the same as within-subject exposure variability.

A Mixed-Effects Representation

Let \(y_{ifr}\) be the natural logarithm of an exposure metric for subject \(i\), formulation \(f\), and replicate occasion \(r\). A useful endpoint-level representation is

Here \(\mu_f\) is the population mean log exposure for formulation \(f\); \(b_i\) is a subject effect shared across formulations; \(u_{if}\) is the subject's stable formulation-specific deviation; and \(e_{ifr}\) is the occasion-specific deviation. A common model assumes \(b_i\) has mean zero and variance \(\sigma_b^2\), the vector \((u_{iR},u_{iT})\) has mean zero and covariance matrix \(\Sigma_U\), and the occasion errors are independent with \(e_{ifr}\sim N(0,\sigma^2_{Wf})\). The occasion errors are also assumed independent of the stable effects. These normal assumptions define a standard likelihood model; other distributions or covariance structures may be appropriate if the data warrant them.

The diagonal entries of \(\Sigma_U\) are the variances of stable subject-by-formulation deviations; its off-diagonal entry \(\sigma_{U,RT}\) is their covariance. These are distinct from \(\sigma^2_{WR}\) and \(\sigma^2_{WT}\), the formulation-specific occasion variances. Separating those quantities matters: repeated occasions within a formulation identify occasion variability, while the pattern of responses across subjects informs stable subject effects.

This decomposition also needs an identifiable parameterization. With only two formulations, a freely estimated common subject intercept and an unrestricted covariance matrix for two formulation-specific random effects can be redundant: several combinations can describe the same covariance in the observed responses. The model must constrain or otherwise define the shared and formulation-specific components, and the design must contain sufficient replication. A fitted model should not be treated as having estimated distinct components merely because software returned estimates.

Estimating Occasion Variances From Replicates

The simplest calculation applies when each subject receives a formulation twice, occasion effects have been adjusted for any systematic design effects, and the two occasion errors are independent, mean-zero, and have the same variance \(\sigma^2_{Wf}\). For each subject, form the difference between the two log exposures. The shared subject effects and the subject's stable formulation-specific effect cancel.

The variance of the subject-level differences, divided by two, is therefore a method-of-moments estimate of \(\sigma^2_{Wf}\). This derivation explains both the usefulness and the limits of the calculation: unequal occasion variances, correlated replicate errors, unadjusted systematic effects, or missing replicates change the relationship.

Estimating and Interpreting Covariance

The covariance of stable subject-specific formulation responses answers a different question from the occasion variance. Positive covariance means subjects who tend to have above-average stable responses for one formulation also tend to do so for the other, after the model's shared subject component is defined. Negative covariance means those deviations tend to move in opposite directions. Covariance is scale-dependent and should be interpreted alongside its corresponding variances; the correlation is \(\rho_U=\sigma_{U,RT}/(\sigma_{U,R}\sigma_{U,T})\), when both variances are positive.

Replication improves estimation of each subject's formulation-specific expected response by averaging over occasions, while a mixed-effects likelihood uses all observations and accounts for the greater uncertainty in a subject mean based on fewer replicates. In a simple balanced illustration, if adjusted formulation-specific stable deviations \(u_{if}\) were known, their sample covariance would estimate \(\sigma_{U,RT}\). In actual data those deviations are latent, not directly observed; restricted maximum likelihood (REML) or maximum likelihood estimates the variance components jointly rather than treating estimated subject effects as error-free observations.

Why the Components Matter for Comparisons

The covariance structure determines the variability of a formulation contrast. For a subject's stable deviations, the test-minus-reference contrast is \(u_{iT}-u_{iR}\), whose variance is

Thus, the covariance cannot be omitted when the target is the distribution of subject-specific formulation contrasts. A positive covariance reduces the variance of the difference; a negative covariance increases it. Occasion variability contributes separately to the uncertainty in observed or predicted exposures, according to the number of occasions and the estimand. For independent occasion errors, the variance of the difference between one T occasion and one R occasion includes \(\sigma^2_{WT}+\sigma^2_{WR}\); averaging replicate occasions reduces the contribution from each formulation's occasion variance.

Model Fitting, Checks, and Alternatives

A practical analysis can fit the mixed-effects model to log exposure metrics derived consistently from each subject's replicate profiles, or fit a concentration-time model with occasion-level random effects and derive exposure metrics from its predictions. In either case, the variance interpretation depends on where the random effects enter the model. For example, variability in an occasion-specific bioavailability parameter may induce variability in AUC and Cmax, while concentration residual error represents deviations of measured concentrations from the modeled profile. They should not be equated without a model-based justification.

REML is commonly used for variance-component estimation in a specified linear mixed model because it accounts for estimation of fixed effects; maximum likelihood is another option, particularly for comparing models with different fixed-effects structures. Both rely on the specified likelihood and covariance structure. Estimates near zero, unstable covariance estimates, or large changes under small modeling choices can signal weak information or a boundary problem. A covariance matrix must be positive semidefinite, so its variances cannot be negative and its correlations must lie between \(-1\) and \(1\).

Useful checks include confirming that the replicate schedule truly repeats formulations within subjects, inspecting replicate differences for outliers or systematic order patterns, checking whether variance estimates are sensitive to individual subjects, and verifying that the optimizer converged to a valid covariance matrix. A funnel-shaped pattern in replicate differences across fitted exposure levels may indicate that a constant variance on the log scale is unsuitable. These are warnings to revisit the model and data, not automatic grounds for selecting a more complex covariance structure.

Difference-based estimates are transparent and useful for a balanced two-replicate design, but they do not readily accommodate unequal replication, missing occasions, or several sources of random variation. Mixed-effects likelihood can use unbalanced observations and estimate components jointly, at the cost of stronger modeling assumptions and greater sensitivity to identifiability and convergence. A nonreplicate design may estimate an overall residual variation under additional assumptions, but it cannot provide the same direct within-subject replication information for each formulation.

The estimates matter when model-based comparisons depend on the distribution of individual formulation responses, not only on the population mean contrast. Their interpretation is conditional on the exposure scale, the random-effects parameterization, the replicate design, and the assumed error structure. A precisely reported component is not necessarily a well-identified or robust component, so report its scale and uncertainty and explain which sources of variation it represents.