Worked Random-Effect Structure For Repeated-Dose PK

Why repeated-dose random effects need structure

With repeated dosing, a subject can differ from the population in a way that persists across all occasions, and can also differ from their own typical value on a particular occasion. These are distinct sources of variation. A subject whose clearance is consistently low contributes to interindividual variability; a temporary change in that subject’s clearance on one occasion contributes to interoccasion variability. Combining the two without a clear structure can obscure what the model says and what the data can estimate.

This tutorial builds a compact random-effects structure for clearance and volume. The structural PK model determines how those parameters generate concentrations over time. The random-effects structure describes how the parameters vary across subjects and occasions; it does not replace the structural model or the residual-error model.

A two-parameter repeated-dose model

We model positive clearance and volume using their natural logarithms. The log scale is convenient because normally distributed random effects then give positive parameter values after exponentiation. Let \(CL_{ij}\) and \(V_{ij}\) be the clearance and volume applicable to subject \(i\) on occasion \(j\), and let \(\theta_{CL}\) and \(\theta_V\) be their population typical values. Specify

The random-effects vectors, in the order clearance then volume, are

Here \(N(0,\Omega)\) denotes a multivariate normal distribution with zero mean vector and covariance matrix \(\Omega\). A common starting model assumes that subject effects are independent across subjects, occasion effects are independent across subjects and occasions, and \(\eta_i\) is independent of every \(\kappa_{ij}\). These are modeling assumptions, not consequences of repeated dosing. The PK structural model must also specify how each occasion’s parameters apply to the dose history and concentration-time predictions, including any drug remaining from earlier doses.

A full two-parameter covariance structure at each level is

The diagonal entries are variances on the log-parameter scale. The off-diagonal entries are covariances on that same scale. Each matrix must be positive semidefinite; for a two-by-two matrix, this requires nonnegative diagonal entries and an off-diagonal entry whose magnitude does not exceed the product of the corresponding standard deviations. Correlations are often easier to compare across scales: for example, the IIV correlation is \(\rho_{\mathrm{IIV}}=\omega_{CL,V,\mathrm{IIV}}/(\omega_{CL,\mathrm{IIV}}\omega_{V,\mathrm{IIV}})\), when both standard deviations are nonzero.

Worked calculations for the omega structure

The assumed independence of \(\eta_i\) and \(\kappa_{ij}\) makes the occasion-level covariance pattern especially clear. Adding independent random effects adds their variances. For two different occasions \(j\ne k\), the shared subject effect remains in both log parameters, but the independent occasion effects do not contribute to their cross-occasion covariance.

Why the covariance result holds

The result follows directly from covariance rules and the independence assumptions. For example, write the log clearance and log volume for one occasion as their fixed log typical values plus subject and occasion effects. Fixed constants do not affect covariance. Expanding the covariance gives four terms:

The two mixed terms are zero because subject and occasion effects are assumed independent. The first term is the IIV covariance, and the last is the IOV covariance. Thus the same-occasion covariance is their sum. For \(j\ne k\), the last term becomes \(\operatorname{Cov}(\kappa_{CL,ij},\kappa_{V,ik})=0\) under independence across occasions. The remaining covariance is the IIV covariance. This derivation also shows why the conclusion would change if the independence assumptions changed.

Choosing terms that the data can support

A full pair of two-parameter matrices contains six variance and covariance parameters: two variances and one covariance at each of the IIV and IOV levels. That may be a reasonable candidate, but it is not automatically justified. A term can be mathematically defined yet weakly informed by the sampling design, the number of subjects, or the way the structural model maps parameters to concentrations.

For example, a clearance-volume covariance at the IOV level requires information about how the two occasion-specific parameter deviations co-vary. If each occasion has sparse samples concentrated in a time region that is weakly informative for one parameter, that covariance may be poorly estimated. Similarly, estimating an IOV variance requires repeated occasions within subjects; estimating a stable IIV variance separately from it requires the repeated-occasion covariance information described above. Merely having many concentration records is not sufficient if they do not inform the relevant parameter contrasts.

Common alternatives impose more structure:

A diagonal matrix does not prove that the true effects are independent; it fixes their covariance to zero. Likewise, removing an IOV term is a modeling restriction, not proof that occasions are biologically identical. Candidate structures should be assessed for estimability, fit, plausibility, and stability rather than selected solely because a larger matrix is possible. The earlier material on covariance structures and information criteria provides a framework for comparing supported candidate models.

Diagnostics, limitations, and a common pitfall

Warning signs include variance estimates near zero when the model contains many terms, correlations close to \(1\) or \(-1\), large uncertainty in covariance estimates, unstable estimates under small changes to the model, or numerical difficulties in fitting. These can indicate that the data do not distinguish the terms well, that the chosen parameterization is too ambitious, or that the assumed distribution and independence structure are inadequate. They are not, by themselves, proof of a particular biological alternative. Check whether subjects truly have repeated informative occasions, whether occasions are defined consistently, and whether the structural model can distinguish clearance from volume in the observed sampling windows.

The log-scale model also has an interpretation worth keeping precise. A zero-mean normal random effect on the log scale makes the typical parameter a median under the random-effects distribution, not generally its arithmetic mean. Exponentiating a normal random effect yields a lognormal distribution; for a single normally distributed log deviation with variance \(v\), the multiplicative factor has mean \(\exp(v/2)\), not 1. This distinction matters when translating model parameters into natural-scale summaries.

• Common mistake: “The total variance on one occasion is IIV variance, so any extra variability must be residual error.” Correction: Under this structure, occasion variance also contributes: the log-parameter variance is IIV plus IOV. Separating these components requires repeated occasions and a design that informs occasion-level PK parameters; residual error is a separate model component.
• Common mistake: “A positive clearance-volume covariance means the parameters cause one another to increase.” Correction: It describes a modeled association between random deviations on the specified scale. It does not establish causation.
• Common mistake: “The covariance matrix entries are the same on the log and natural scales.” Correction: The omega matrices here describe log-parameter random effects. Exponentiation changes the distribution and its covariance.

The structure is a useful working model, not a guarantee that every random effect is identifiable or that the Gaussian assumptions are correct. Its practical value is that it separates persistent subject differences from occasion-specific deviations and makes the implied dependence transparent enough to evaluate against the study design.