Covariance Signatures of Subject, Occasion, and Residual Effects

Why covariance reveals where variability lives

Two concentration measurements can vary together for different reasons. Measurements from the same subject may be similar because that subject has persistently different pharmacokinetic parameters from other subjects. Measurements from the same dosing occasion may share an additional occasion-specific deviation. And each individual measurement can contain residual error. The key distinction is how long each source of variation is shared: across a subject’s occasions, within one occasion, or only for one observation.

Covariance summarizes this sharing. If two measurements tend to be high or low together across subjects, their covariance is positive. Comparing covariance for pairs with different sharing structures can therefore reveal which variability components are present. The interpretation requires care: concentration is a nonlinear function of time and pharmacokinetic parameters, so a simple additive covariance pattern is a useful local or simplified representation, not a universal description of raw concentrations.

Here \(Z\), \(B\), \(O\), and \(E\) are expressed on the concentration scale, so their variances and covariances have units of concentration squared. This representation is particularly transparent when comparing observations at the same sampling time and when a random contribution acts approximately like an additive shift. It is not a claim that all pharmacokinetic random effects are additive concentration shifts. In an NLME model, a subject-level effect such as an eta on a parameter can change the predicted concentration differently at different times.

Covariance patterns under the simple decomposition

For any observation, all three components contribute to its variance. For two distinct observations from the same subject and the same occasion, the subject and occasion terms are shared, but the residual terms are not. For observations from the same subject on different occasions, only the subject term is shared. Observations from different subjects share none of the random terms. These statements give the covariance signatures directly.

The third line is the persistent-subject signature: covariance remains between different occasions for the same subject. The difference between same-occasion covariance and cross-occasion covariance is the occasion variance. The difference between the observation variance and same-occasion covariance is the residual variance. In population terms,

where \(V\) is the variance of an observation, and the \(C\) terms refer to the corresponding covariances. These equalities depend on the stated independence and equal-variance assumptions. With finite data, estimated differences need not be nonnegative; a negative difference can signal sampling uncertainty, a poor covariance model, or both.

Worked covariance calculations

Why the covariance identities hold

The result follows from expanding covariance and using independence. For example, consider two different observations from the same occasion. Their covariance is \(\operatorname{Cov}(B_i+O_{ij}+E_{ijm},B_i+O_{ij}+E_{ijn})\). By bilinearity, this is the sum of the nine pairwise covariances between terms in the first and second arguments. Independence and zero means make every cross-component covariance zero. The shared subject term contributes \(\operatorname{Var}(B_i)=\sigma_B^2\), the shared occasion term contributes \(\operatorname{Var}(O_{ij})=\sigma_O^2\), and the distinct independent residuals contribute zero covariance. Therefore the result is \(\sigma_B^2+\sigma_O^2\).

For different occasions, replace the second occasion term with \(O_{ik}\), where \(k\ne j\). Occasion effects are independent across occasions, so their covariance is zero; the residual terms are also independent. Only \(B_i\) remains shared, yielding \(\operatorname{Cov}(Z_{ijm},Z_{ikn})=\sigma_B^2\). For one observation, variance adds because the three components are independent, giving the sum of all three variances. These derivations explain why the design must include the relevant kinds of pairs: without cross-occasion pairs, subject and occasion covariance are not separated by these identities.

Applying the signatures to concentration-time data

The decomposition gives a conceptual guide for interpreting repeated concentration measurements, but the expected pattern depends on the observation model. The additive form is most direct when measurements are compared at the same nominal time, after accounting for the population concentration-time curve, and when the random deviations can reasonably be treated as additive with similar loadings. If a subject effect changes clearance or volume, its effect on concentration can vary over time: it may be small at one sampling time and larger at another. Consequently, raw concentration covariance need not be constant across time, even when the underlying subject effect is persistent.

Residual error also need not have constant variance. For example, an observation model may imply residual variability that changes with the predicted concentration. Serially correlated assay or sampling deviations would likewise make the covariance of two residuals nonzero. In either case, a difference such as \(C_{\text{same}}-C_{\text{different}}\) no longer isolates occasion variance without adjustment. The earlier material on separating parameter variability from residual error motivates this distinction: the covariance comparison is about the pattern of shared contributions, not a shortcut around specifying the observation model.

In practice, examine covariance patterns among appropriately matched observations and inspect whether the assumed mean curve and residual structure are plausible. Important warning signs include:

• Cross-occasion covariance changes markedly with sampling time, suggesting that a constant additive subject contribution is inadequate or that the mean curve has not been properly accounted for.
• Same-occasion covariance does not exceed cross-occasion covariance, beyond what sampling uncertainty could explain, despite a proposed positive occasion variance.
• Residual plots show changing spread with predicted concentration, or paired residuals show serial dependence, contradicting independent equal-variance residual assumptions.
• Estimated variance differences are negative or highly unstable. Variance components cannot be negative in the model, so such raw differences signal estimation noise or a mismatch between the summaries and assumed structure—not a negative physical variance.

Design, alternatives, and limitations

Replication determines which signatures can be observed. Multiple occasions per subject permit cross-occasion covariance to inform persistent subject variability. Multiple distinct observations within an occasion make it possible, in principle, to compare same-occasion covariance with cross-occasion covariance. If each occasion contributes only one observation, there is no within-occasion pair of distinct samples; residual and occasion contributions may then be difficult or impossible to separate using this simple covariance comparison alone.

Covariance contrasts are an explanatory tool, not the only estimation strategy. A nonlinear mixed-effects model can estimate variance components by fitting the concentration-time predictions and observation model jointly, using the full sampling schedule rather than only matched-time covariance summaries. A simpler analysis of occasion averages may be easier to communicate, but averaging changes the residual contribution and can discard information. Neither alternative removes the need for an appropriate mean model, a defensible residual model, and enough replication to distinguish the components.

The additive signatures also rely on independence, mean-zero effects, and a consistent comparison scale. Correlated occasion effects, correlated residuals, unequal residual variance, or random effects that alter the curve shape can all change the pattern. In such settings, covariance remains informative, but its terms must be derived from the specified model rather than assigned by the simple formulas above. Covariance patterns can indicate which effects a design may distinguish; they do not, by themselves, guarantee that a complex NLME model is identifiable or that its estimates are precise.

A common interpretation mistake

The useful question is not simply whether repeated measurements are correlated. Ask which components each pair shares. A pair on different occasions from one subject shares the persistent subject contribution; a pair within one occasion shares that contribution plus the occasion contribution; an individual observation additionally contains its own residual error. This sharing map is the covariance signature.