Three Different Sources of Variation
A dose-justification simulation can produce a wide range of exposures for three fundamentally different reasons. Virtual subjects may differ in their pharmacokinetic parameters; measured concentrations may fluctuate around each subject’s model-predicted profile; and the population model’s estimated parameters may themselves be uncertain. These sources are not interchangeable. Including the wrong one can change the simulated quantity, the spread of results, and the interpretation of a dose comparison.
The key decision is whether the target is variation among subjects under a specified model, variation in observed measurements, or sensitivity to uncertainty in the fitted model. Exposure metrics calculated from an individual’s model-predicted profile, such as AUC or Cmax, describe the profile before residual observation error is added. A simulated concentration assay result, in contrast, may require residual error.
Write the model-predicted concentration for subject \(i\) at time \(t\) as \(f(t;\theta,\eta_i)\), where \(\theta\) denotes population model parameters and \(\eta_i\) denotes that subject’s random effects. For an observation, a residual model may be written \(Y_{ij}=f(t_{ij};\theta,\eta_i)+\epsilon_{ij}\) for an additive error model. Other residual structures, including proportional or combined error, use different expressions. The important distinction is that \(\epsilon_{ij}\) is attached to an observation, whereas \(\eta_i\) determines subject-specific PK parameters across that subject’s profile.
In a typical dose-justification simulation, the model parameters are estimated quantities and are held at specified values for a conditional simulation. A population of new virtual subjects is then generated using the estimated distribution of random effects and the relevant covariate distributions. If the question also asks how conclusions depend on uncertainty in the fitted model, parameter uncertainty is added as a separate layer.
What Each Component Belongs To
| Component | What varies | Use it when | Usually omit it when |
|---|---|---|---|
| IIV | Subject-level PK parameters across virtual subjects | The target concerns a population distribution of true or model-predicted exposures | The target is only a typical-subject profile |
| Residual error | Observed values around a subject’s predicted profile | Simulating measured concentrations or evaluating a measurement-based procedure | Calculating latent exposure from the model-predicted profile |
| Parameter uncertainty | The population model parameters across plausible fitted models | Assessing sensitivity of dose conclusions to estimation uncertainty | Reporting a conditional simulation at fixed parameter estimates |
These components can coexist, but their simulation roles differ. IIV is usually sampled once per virtual subject and retained across that subject’s entire time course. Residual errors are sampled for individual observations under the chosen residual model. Parameter uncertainty is represented by drawing a plausible parameter vector, then using that same vector for all virtual subjects in the corresponding simulation scenario. Drawing a new population parameter vector independently for every subject would incorrectly turn uncertainty about the population model into apparent between-subject variability.
Parameter vectors should preserve estimated dependence. For example, if typical clearance and volume estimates are correlated, sample them jointly from an uncertainty distribution that reflects that covariance, rather than independently sampling each marginal distribution. The same principle applies to uncertainty in random-effect variances and correlations. A fitted covariance matrix of random effects describes IIV; uncertainty about the entries of that matrix is a separate parameter-uncertainty layer.
Worked Example: IIV in Clearance and AUC
Suppose a hypothetical one-compartment model has typical clearance \(\theta_{CL}=5.00\) L/h and \(\log(CL_i)=\log(\theta_{CL})+\eta_i\), with \(\eta_i\sim N(0,0.04)\). A 100 mg dose is given, and the model is linear with complete exposure represented by \(AUC_i=Dose/CL_i\). Find the typical-subject AUC and the AUC for subjects with \(\eta_i=-0.20\) and \(+0.20\). Treat the stated model parameters as fixed.
At \(\eta_i=0\), clearance is \(5.00\) L/h, so the typical-subject AUC is \(100/5.00=20.0\) mg·h/L. For \(\eta_i=-0.20\), \(CL_i=5.00\exp(-0.20)=4.094\) L/h and \(AUC_i=100/4.094=24.43\) mg·h/L. For \(\eta_i=+0.20\), \(CL_i=5.00\exp(0.20)=6.107\) L/h and \(AUC_i=100/6.107=16.38\) mg·h/L.
The two subjects differ because they received different draws of the IIV random effect, not because the model parameters were re-estimated for each subject. Under this lognormal model, the coefficient of variation of clearance is \(\sqrt{\exp(0.04)-1}=0.2020\), or about 20.2%. That describes the model-implied between-subject spread in clearance, not uncertainty in the estimate of the typical clearance.
Residual Error Depends on the Simulation Target
If the simulated outcome is an exposure metric calculated from the model-predicted profile, residual error generally does not belong in the calculation. Adding independent noise to sampled concentrations and then calculating an AUC produces a measurement-based, noisy approximation to exposure. It is not the same as calculating AUC from the underlying individual profile. Whether measurement noise is appropriate depends on the question: assay precision or a sampling-based estimation method may make it relevant, while the population distribution of model-predicted AUC usually does not require it.
Residual error should be generated at the observation level using the fitted residual model and its assumptions. Under an additive model, for example, the error is added on the concentration scale. Under a proportional model, its effect depends on predicted concentration. The choice matters: the earlier material on concentration-dependent residuals explains why a residual model that fails to match the data can distort simulated observations. Residual errors also need not be independent or identically distributed if the model specifies otherwise.
Worked Example: Deciding Whether to Add Residual Error
A hypothetical simulation compares the distribution of true model-predicted AUC after a dose with the distribution of measured concentrations at 1, 4, and 8 hours. The PK model predicts concentrations of 8.0, 3.0, and 1.0 mg/L for one virtual subject. The fitted observation model is proportional error, but no residual-error magnitude is needed for the requested AUC comparison.
For the AUC comparison, generate the subject’s random effects, calculate the model-predicted concentration-time profile, and calculate the exposure metric from that profile. Do not add proportional residual error to the three concentrations: doing so would introduce observation noise into a target defined as model-predicted exposure. The three listed predictions alone are not enough to calculate AUC without a specified integration method and profile, so no numeric AUC is inferred here.
For a separate simulation of measured assay results at those three times, residual error is needed. Under a proportional model, each observed value is generated from its prediction using the model’s specified proportional-error distribution. The resulting values represent noisy measurements, not a changed set of individual PK parameters. The distinction is the estimand: latent exposure for the first task, observed concentrations for the second.
Adding Parameter Uncertainty as an Outer Layer
A conditional simulation uses a fixed parameter vector, often the fitted estimate \(\hat{\theta}\). This answers: “What population outcomes does this model imply at these parameter values?” It does not account for uncertainty in those values. To address that uncertainty, specify a distribution of plausible parameter vectors, obtained by an appropriate method such as a parametric approximation based on the parameter-estimate covariance matrix, bootstrap refitting, or a Bayesian posterior. The chosen distribution and its assumptions should be stated.
The simulation is nested. First draw a parameter vector \(\theta^{(m)}\) for outer scenario \(m\). Then, conditional on that vector, generate the virtual subjects’ covariates and IIV random effects, and calculate their model-predicted profiles. Add residual error only if the outcome is an observed measurement. Use the same \(\theta^{(m)}\) throughout that outer scenario; otherwise each subject is implicitly governed by a different assumed population model.
A useful justification follows from the law of total variance. Let \(G\) be a simulated outcome, such as an individual’s model-predicted AUC, and suppose the generation order is parameter vector, subject random effects, then residual error. Repeatedly applying the conditional-variance identity gives:
The first term is variation across plausible population parameter vectors; the second is average variation across subjects conditional on each vector; and the third is average observation-level variation. For a latent exposure calculated from the predicted profile, \(G\) does not depend on residual error, so the third term is zero. This decomposition does not imply that every practical summary can be obtained by simply adding three reported standard deviations: the terms are variances under the specified hierarchy, and the outcome and conditional distributions must be defined correctly.
Worked Example: A Shared Parameter Draw Changes a Cohort Together
Consider a simplified hypothetical one-parameter model in which typical clearance is uncertain. Two plausible outer draws are \(CL_{typ}=4.50\) L/h and \(CL_{typ}=5.50\) L/h. Each scenario simulates 100 subjects with their own IIV draws. For illustration, focus on the typical-subject AUC for a 100 mg dose, \(AUC=100/CL_{typ}\).
Under the first parameter draw, the typical-subject AUC is \(100/4.50=22.22\) mg·h/L. Under the second, it is \(100/5.50=18.18\) mg·h/L. Each value applies to its entire corresponding simulated cohort: the parameter draw changes the population model scenario, while IIV generates differences among individuals within that scenario.
It would be incorrect to assign the 4.50 L/h draw to some subjects and the 5.50 L/h draw independently to other subjects and label that spread IIV. That procedure mixes model uncertainty into the subject distribution. If both sources are of interest, retain the outer-scenario labels and describe between-scenario uncertainty separately from within-scenario subject variability.
Diagnostics, Alternatives, and Limitations
Before running a simulation, state the outcome and the target population explicitly. Then check that each random quantity has the right role and sampling level. A practical audit asks whether each virtual subject has a persistent subject-level random effect, whether residual errors occur only at simulated observations, and whether parameter uncertainty is shared across subjects within an outer scenario.
- Residual error inflates exposure variability. If the target is latent model-predicted exposure, check whether noise was added to concentrations before calculating the metric. Remove it unless measurement-based exposure is the target.
- Parameter uncertainty is mistaken for IIV. Check whether a new population parameter vector was drawn for each subject. Parameter vectors should be drawn by scenario, then held fixed across that scenario’s cohort.
- IIV is accidentally resampled over time. Check that each subject’s random effects are held constant over that subject’s profile, unless the specified model includes a time-varying component.
- Uncertainty draws are implausible. Inspect parameter correlations, boundary behavior, and whether draws produce invalid parameter values or implausible profiles. A convenient multivariate approximation is not automatically adequate near boundaries or under strong nonlinearity.
- Individual estimates replace population sampling. Empirical Bayes estimates are shrunken subject-specific estimates from observed data; they are not generally a substitute for drawing new subjects from the population random-effects distribution when simulating a new population.
The main alternatives answer different questions. A fixed-estimate simulation is direct and transparent when the goal is conditional on the fitted model. A nested uncertainty simulation adds uncertainty about the fitted model but requires a defensible parameter uncertainty distribution and more computation. A residual-error simulation is appropriate for measurement-level questions but can obscure the distribution of latent exposure if treated as though it were IIV. None is universally best: choose according to the quantity to be justified.
All three components depend on model adequacy. A simulation cannot repair misspecified random-effects distributions, an unsuitable residual model, or uncertainty estimates that omit important parameter dependence. Nor does a wide uncertainty interval automatically identify which source drives the spread; report conditional subject variability and between-model uncertainty in a way that keeps their meanings distinct.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A simulated subject has one clearance random effect used for every time point. Is this IIV or residual error, and why?
- A team calculates model-predicted AUC for each virtual subject. Should it automatically add residual error to predicted concentrations before calculating AUC? Explain.
- In a nested simulation, a parameter vector is drawn once for a cohort of 500 subjects. What source of variation does that draw represent, and why is it shared?
- For \(AUC=100/CL\), calculate the typical-subject AUC when clearance is 4.00 L/h. Give units if the dose is 100 mg.
- Name one diagnostic for detecting parameter uncertainty that has been incorrectly treated as IIV.
Solutions. (1) It is IIV: the random effect changes the subject’s PK parameter and persists across the subject’s profile; residual error varies at the observation level. (2) No. For model-predicted AUC, use the predicted profile without observation error unless the target is specifically measurement-based. (3) It represents uncertainty about the population model parameters. It is shared because all subjects in that scenario are generated under the same assumed population model. (4) \(AUC=100/4.00=25.0\) mg·h/L. (5) Check whether the parameter vector is being redrawn independently for each subject; that would mix between-model uncertainty into the apparent subject-to-subject spread.