Using Replicate Concentrations To Identify Residual Variance

Why Replicates Help Separate Sources of Variation

A concentration measurement can vary because subjects differ in their pharmacokinetic parameters, because a subject’s concentration differs from the structural model prediction, or because repeated measurements of the same sample are not identical. Replicate measurements at the same time point are useful because they share the subject and the underlying concentration at that time. Under suitable assumptions, their differences therefore reveal within-sample residual variation without including between-subject variation.

The key is to distinguish a replicate measurement from a new subject or a new time point. Two concentrations from different subjects can differ because the subjects have different parameter values. Two measurements of the same sample at the same time share those subject-specific parameters and the same underlying concentration, so a difference between them can be attributed to measurement-level variation—provided the replicate errors are independent and the sample itself is genuinely shared.

This interpretation depends on what “replicate” means in the data. Repeated assays of aliquots from one collected sample primarily inform assay-level error. Separate blood draws at nearly the same time may also include collection and short-term biological variation. The earlier material on Separating Assay Error From Structural Model Error is relevant here: replicate measurements do not automatically identify every source of model discrepancy.

Model and Estimator

Let \(Y_{ijr}\) be the \(r\)th replicate concentration for subject \(i\) at time \(t_{ij}\). Write the observation as a shared subject-and-time concentration plus a replicate-specific residual:

Here \(f\) is the structural pharmacokinetic model, \(D_i\) denotes the subject’s dose information, \(\eta_i\) represents subject-level random effects, and \(\theta\) denotes population parameters. At a particular subject and time, all replicates have the same \(\mu_{ij}\). For the simplest variance calculation, assume that conditional on \(\mu_{ij}\), replicate residuals have mean zero, common variance \(\sigma^2\), and are independent:

For \(m_{ij}\) replicates in one subject-time cell, calculate their sample mean and sample variance. The mean describes the cell’s observed concentration level; the variance describes the spread around that mean.

The sample variance uses deviations from the replicate mean, not from the population model prediction. That choice removes the unknown shared concentration from the calculation. If residual variance is constant across the cells being pooled, a pooled estimate is:

Why the Shared Concentration Cancels

The estimator’s justification follows directly from the replicate model. For one cell, \(Y_r=\mu+\epsilon_r\), so \(Y_r-\bar{Y}=\epsilon_r-\bar{\epsilon}\). The shared concentration \(\mu\), which includes the subject’s parameter effects, disappears. The expected sum of squared deviations is then:

The last equality uses independent, zero-mean residuals with variance \(\sigma^2\), which imply \(E(\bar{\epsilon}^{\,2}\mid\mu)=\sigma^2/m\). Dividing by \(m-1\) therefore gives an unbiased estimate of the conditional residual variance. Summing these within-cell sums of squares and dividing by their total degrees of freedom gives the pooled estimator when the same residual variance applies to every included cell.

Replicate differences give the same cancellation in another form. For two replicates, \(Y_r-Y_s=\epsilon_r-\epsilon_s\), so under independence its variance is \(2\sigma^2\). A large difference is evidence of substantial within-sample residual variation; it is not evidence that the subject’s parameters changed between those two measurements.

Replicate Variation Versus Between-Subject Variation

The within-cell estimator removes the shared subject-time level. To see how a between-subject component can be distinguished in a simple balanced setting, consider a model at one fixed time:

Here \(b_i\) is a subject-level deviation and \(\epsilon_{ir}\) is replicate residual error. Assume subjects are independent, residuals are independent of subject effects, and each subject has the same number \(m\) of independent replicates. A subject’s replicate mean then has variance \(\omega^2+\sigma^2/m\): averaging reduces residual variance, but it does not remove the subject effect. This gives a simple method-of-moments estimate of \(\omega^2\), the variance among subjects’ concentration levels at that time:

where \(S_{\bar{Y}}^2\) is the sample variance of the subject replicate means. This is a simple decomposition, not a general replacement for fitting the population model. In particular, \(\omega^2\) at one time point is not itself the variance of a specific pharmacokinetic parameter; the mapping from parameter variation to concentration variation depends on the structural model and time.

Assumptions, Diagnostics, and Alternatives

Before interpreting a replicate-based estimate, check what was replicated and whether the assumptions fit the design. In particular, the measurements should share the same underlying concentration, and replicate errors should be independent conditional on it. If replicates are correlated, then:

Positive correlation makes replicate differences smaller than they would be under independence. Treating those differences as independent evidence can therefore underestimate the residual variance. Shared batch effects, use of the same preparation step, or other common measurement processes can create this problem.

Useful checks include examining within-cell spreads across subjects and concentration levels, confirming the number of replicates contributing to each estimate, and checking for data-entry or assay anomalies. A single extreme replicate can strongly affect a sample variance, especially when there are only two or three measurements. A negative method-of-moments estimate of \(\omega^2\) can also occur when subject-mean variation is small relative to the estimated residual contribution; it signals weak evidence for a positive component under this calculation, not a physically negative variance.

Replicates are most informative when they are genuine repeated measurements of a shared sample and the residual model is appropriate. A constant-variance estimate should not be pooled indiscriminately across a range of concentrations if residual spread changes with concentration. In that case, estimate or model variability in a way that respects the scale and variance assumptions of the data. Replicate information can clarify the residual component, but it does not by itself resolve every ambiguity between structural error and other sources of variability.