Separating Parameter, Individual, and Trial Uncertainty

Why Separate the Sources of Uncertainty?

A PBPK prediction of a drug-drug interaction may be expressed as a distribution of exposure ratios rather than a single number. That distribution can reflect different things: uncertainty about shared model inputs, genuine differences among individuals, or the fact that a clinical trial observes only a finite sample. These sources have different meanings. More virtual subjects can improve a simulated estimate of a population mean, for example, but they do not resolve uncertainty about a shared enzyme parameter or eliminate variability among patients.

The distinction matters when relating a simulated population to a clinical study. A distribution across virtual individuals is not the same as a sampling distribution for a trial estimate. Nor is either one, by itself, a distribution describing uncertainty about a shared PBPK parameter. The earlier material on propagating PBPK uncertainty to exposure ratios provides the context for ratio predictions; here the task is to identify what each part of the resulting spread represents.

A Nested Model for Exposure Ratios

It is often convenient to work on the log scale. Let \(R_i\) be an individual's exposure ratio for a specified endpoint, such as AUC with the perpetrator drug divided by AUC without it, and let \(B_i=\log(R_i)\). The log scale turns multiplicative ratios into additive outcomes. The target endpoint and the direction of the ratio must remain fixed throughout the analysis.

Let \(\theta\) denote shared PBPK parameters or inputs that are uncertain, such as a parameter governing enzyme activity. For a given \(\theta\), let \(Z_i\) represent the physiology or other individual-level characteristics of virtual subject \(i\). Write the subject's modeled log ratio as \(B_i=g(\theta,Z_i)\). Conditional on \(\theta\), suppose subjects are sampled independently from the target population, with \(E(B_i\mid\theta)=m_\theta\) and \(\operatorname{Var}(B_i\mid\theta)=v_\theta\).

To represent possible observation or endpoint-measurement error in a trial, let the observed log endpoint be \(Y_i=B_i+\varepsilon_i\). This term is optional when the analysis concerns a noiseless model output. If included, the variance decomposition below requires that \(E(\varepsilon_i\mid\theta,B_1,\ldots,B_n)=0\), and that the errors are conditionally uncorrelated across individuals and with the subject outcomes. For a simple equal-variance case, assume \(\operatorname{Var}(\varepsilon_i\mid\theta)=\tau_\theta^2\). These conditions are stronger than merely requiring \(E(\varepsilon_i)=0\); a marginally mean-zero error could still have a nonzero conditional mean at a given \(\theta\).

The quantities \(m_\theta\) and \(v_\theta\) can change with \(\theta\). The uncertainty distribution over \(\theta\) describes uncertainty about the shared model inputs, not the biological claim that one common parameter is literally redrawn for every person. Conversely, a subject-specific physiological characteristic belongs in \(Z_i\) when it represents between-individual variation. A characteristic can be treated differently in different models, but its role must be explicit to avoid counting the same source twice.

Deriving the Trial-Mean Decomposition

For a trial of \(n\) individuals, define the sample mean log endpoint as \(M_n=n^{-1}\sum_{i=1}^nY_i\). Conditional on a fixed \(\theta\), the subject outcomes have mean \(m_\theta\), and the stated error condition gives each observed outcome the same conditional mean. Conditional independence and zero covariance give the variance of the mean by adding the individual variances and dividing by \(n^2\).

The unconditional variance follows from the law of total variance: the variance of a random quantity equals the variance of its conditional expectation plus the expected conditional variance. Substituting the two conditional results gives the decomposition.

The first term is the contribution from uncertainty in the shared parameter because different plausible values of \(\theta\) can imply different population mean log ratios. The second term is finite-sample uncertainty from drawing individuals and, if present, measuring their endpoints. If \(\theta\) is fixed rather than uncertain, the first term is zero in this representation. If a trial has unequal weights, clustering, or repeated observations, the second term is not generally the simple variance divided by \(n\).

Worked Calculations

Diagnostics, Alternatives, and Limits

A useful diagnostic is to examine conditional predictions before pooling them. For each plausible shared parameter setting, inspect the conditional mean and spread of virtual-subject log ratios. Then compare the variation of those conditional means with the within-setting variation. If pooled spread is large but conditional spreads are modest, shared-parameter uncertainty may be an important contributor. If conditional spreads are large even at fixed parameters, individual variability is important. Replicate trial simulations at a fixed parameter setting can reveal finite-sampling variability separately.

These checks depend on the model structure. If the same individuals contribute repeated measurements, or if participants are sampled in correlated clusters, individual outcomes are not conditionally independent. For example, with a common positive within-cluster correlation, the variance of the trial mean can exceed the independent-sampling formula. Use the study's actual covariance structure rather than dividing a marginal variance by \(n\). Likewise, time-varying interactions or paired treatment periods can require a model for within-person contrasts rather than independent individual ratios.

A pooled predictive distribution is a useful alternative when the goal is to describe outcomes integrating over parameter uncertainty and subject variability. However, by itself it may conceal which source creates its width. A conditional analysis fixes \(\theta\) and emphasizes population heterogeneity; a parameter-sensitivity analysis varies \(\theta\) and compares resulting predictions; nested simulation can represent both at once and then simulate finite trials. The appropriate choice depends on whether the question concerns virtual individuals, uncertainty in a population prediction, or the behavior of a specific trial estimate.

A further limit is that the decomposition applies to the defined scale and summary. The variance of a mean log ratio is not the variance of the arithmetic mean ratio, because exponentiation is nonlinear. If the target is a ratio of arithmetic mean exposures, it must be calculated from the exposure-scale quantities under the relevant model rather than obtained by simply exponentiating the mean log ratio. Finally, Monte Carlo error from using a finite number of simulation draws is computational error, not one of the three biological or clinical sources defined above; it should be checked by increasing simulation size or using simulation diagnostics.