Why Decompose Between-Subject Variability?
A population PTA is a proportion: it summarizes how many simulated individuals meet an exposure target. It does not, by itself, explain why some individuals attain the target and others do not. Variability in clearance, volume, absorption, or other model parameters can change exposure, but the size and direction of the resulting PTA change depend on the regimen, endpoint, and target threshold. Decomposition uses controlled simulation comparisons to investigate these contributions while keeping the comparison interpretable.
The goal is not to claim that a model parameter causes an individual outcome. It is to quantify how a specified simulation result changes when a specified source of interindividual variability is included or removed. These are model-based contrasts: their meaning depends on the model, the population, and the comparison design.
Specify the Variability and the Contrast
A typical population model writes an individual parameter as a typical value, adjusted for covariates, multiplied by an individual random effect. For a positive parameter such as clearance, a common specification is
Here \(CL_{\mathrm{typ}}\) is a typical clearance, \(g(x_i;\beta)\) represents the modeled covariate effect for individual \(i\), and \(\eta_{CL,i}\) is the individual clearance random effect. The covariance matrix \(\Omega\) describes the modeled distribution of random effects, including correlations where specified. For a variability ablation, setting \(\eta_{CL,i}=0\) removes clearance random-effect variation; it does not remove the covariate effect or change the typical clearance.
A useful comparison holds the dose regimen, target definition, covariates, structural model, and other random-effect components fixed. Reusing the same simulated individuals and the same draws for the components that remain in both scenarios makes the contrast paired. This common-random-number design reduces Monte Carlo noise in the difference, because the two PTA estimates are not needlessly exposed to different random variation.
The contrast must also state what happens to correlated random effects. If clearance and volume random effects are correlated, independently setting one component to zero while leaving the other unchanged may define a mathematical intervention that does not preserve the original joint distribution. That may still be a deliberate sensitivity analysis, but it should not be described as a uniquely determined share of naturally occurring variability. One option is to examine a correlated group together; another is to define an explicit conditional or latent-factor simulation scheme.
Controlled Comparisons and Their Interpretation
For each scenario \(s\), calculate the exposure \(E_{i,s}\) using the same exposure definition and time horizon, then apply the same target rule. For a minimum exposure target \(T\), for example, \(I_{i,s}=1\) if \(E_{i,s}\geq T\), and \(I_{i,s}=0\) otherwise. The paired individual difference \(D_i=I_{i,\mathrm{full}}-I_{i,\mathrm{ablated}}\) is 1 for an individual who attains only under the full-variability scenario, -1 for one who attains only under the ablated scenario, and 0 if attainment status is unchanged.
A negative difference means that PTA is lower with the specified variability included than in the ablated comparison. It does not mean variability always lowers exposure: some individuals may have higher and others lower exposure. The result reflects the chosen target and the full distribution of outcomes. Because PTA is a thresholded outcome, a variability change can substantially affect the exposure distribution yet have little impact on PTA if few individuals cross the target boundary.
Worked Example: Clearance Variability and AUC Attainment
In a hypothetical simulation, ten individuals receive the same regimen. The target is AUC at least 80 mg·h/L. The full model includes interindividual clearance variability; the paired ablation sets the clearance random effect to zero while retaining the same typical clearance and regimen. Suppose the full-model AUC values are 55, 65, 72, 78, 84, 91, 100, 112, 130, and 155 mg·h/L. In the ablation, all ten have AUC 100 mg·h/L.
Under full variability, six values meet the target: 84, 91, 100, 112, 130, and 155 mg·h/L. Thus \(\widehat{\mathrm{PTA}}_{\mathrm{full}}=6/10=0.600\), or 60.0%. With clearance variability removed, all ten meet the target, so \(\widehat{\mathrm{PTA}}_{\mathrm{ablated}}=10/10=1.000\), or 100%.
The paired contrast is \(0.600-1.000=-0.400\), a decrease of 40.0 percentage points in this particular hypothetical simulation. The simulated AUCs are consistent with the direction expected from \(AUC \propto 1/CL\) for a fixed dose under linear pharmacokinetics: individuals with higher clearance have lower AUC. The contrast describes this regimen, this AUC target, and this model; it is not a general claim that clearance variability reduces PTA at every target.
Why the Comparison Is Paired
The paired design estimates a difference for matched simulation draws rather than subtracting two unrelated Monte Carlo estimates. For each individual, keep the covariates and all random effects not being ablated the same in both scenarios. Then calculate \(D_i\) and average these differences. The simulation standard error for the estimated mean paired difference can be estimated as \(s_D/\sqrt{N}\), where \(s_D\) is the sample standard deviation of the \(D_i\) values. This quantifies Monte Carlo error for the contrast under the simulated model; it is not a measure of uncertainty in the model parameters or proof that the model is correct.
This also gives a short justification for common random numbers. If the same latent draws are used in both scenarios, the variance of the difference of the two estimated means is
When matching draws makes the two estimates positively correlated, the covariance term reduces the variance of their difference compared with independent simulations. Pairing does not alter the expected contrast; it makes its Monte Carlo estimate more precise when the matching induces positive correlation.
Worked Example: Ablating Clearance or Volume Variability
A hypothetical eight-person simulation evaluates whether steady-state \(C_{\min}\) is at least 2.5 mg/L. All scenarios use the same dose, covariates, target, and matched individual draws for retained random effects. The exposure outputs are:
| Scenario | Simulated \(C_{\min}\) values (mg/L) | Attaining individuals | PTA |
|---|---|---|---|
| Full variability | 1.8, 2.2, 2.4, 2.6, 2.8, 3.1, 3.5, 4.0 | 4, 5, 6, 7, 8 | 5/8 = 62.5% |
| Clearance random effect set to zero | 2.3, 2.4, 2.5, 2.6, 2.7, 2.8, 2.9, 3.0 | 3, 4, 5, 6, 7, 8 | 6/8 = 75.0% |
| Volume random effect set to zero | 1.9, 2.1, 2.3, 2.5, 2.8, 3.2, 3.6, 4.2 | 4, 5, 6, 7, 8 | 5/8 = 62.5% |
For clearance, the full-minus-ablated contrast is \(5/8-6/8=-1/8=-0.125\), or -12.5 percentage points. For volume, it is \(5/8-5/8=0\) percentage points. In this hypothetical setup, removing clearance variability changes attainment for the comparison, whereas removing volume variability leaves the count unchanged. That does not establish that volume has no effect on exposure: the displayed values differ, but none of the changes alters whether an individual crosses 2.5 mg/L.
The controlled comparison is informative only if the model represents the relevant dependence correctly. If clearance and volume random effects are correlated, these separate ablations should be reported as specified scenario contrasts, and the correlation-handling choice should be stated. The two contrasts should not be added and presented as an overall decomposition without further justification.
Interactions, Attribution, and a Shapley Example
A parameter-specific ablation is order-dependent when variability drivers interact. The reason is that PTA is a nonlinear function of exposure: each individual's indicator changes only when that individual's exposure crosses the threshold. Therefore, the PTA change from adding clearance variability can depend on whether volume variability is already present. Controlled ablations answer clear questions, but they do not automatically assign unique, additive shares of a total PTA difference.
One allocation for multiple drivers is the Shapley approach. For two drivers \(A\) and \(B\), define \(v(S)\) as PTA when the drivers in set \(S\) have variability included and the other drivers are set to their prespecified reference condition. The Shapley contribution for \(A\) averages its marginal change when added first and when added after \(B\). With \(v(\varnothing)\) denoting the reference PTA, the contributions sum to \(v(\{A,B\})-v(\varnothing)\). This allocation is meaningful only after the scenarios and treatment of dependence have been defined.
Worked Example: Allocating an Interaction Between Two Drivers
Suppose a hypothetical simulation with independent variability drivers \(A\) and \(B\) gives \(v(\varnothing)=0.90\), \(v(\{A\})=0.82\), \(v(\{B\})=0.78\), and \(v(\{A,B\})=0.70\). These are PTA proportions, not percentages.
For driver \(A\), the marginal change when it is added first is \(0.82-0.90=-0.08\). When added after \(B\), its marginal change is \(0.70-0.78=-0.08\). Its Shapley contribution is their average, \((-0.08-0.08)/2=-0.08\), or -8 percentage points.
For driver \(B\), the marginal changes are \(0.78-0.90=-0.12\) and \(0.70-0.82=-0.12\). Its Shapley contribution is -0.12, or -12 percentage points. The contributions add to \(-0.08+(-0.12)=-0.20\), which equals the total change \(0.70-0.90=-0.20\). Here the effects happen to be additive. With other scenario results, the averaged allocation can divide an interaction between drivers while still summing to the overall change.
If \(A\) and \(B\) are correlated random effects, independent scenario generation may change their joint distribution and make these values answer a different question. The Shapley arithmetic remains valid for the stated scenario values, but the scenario construction must preserve or explicitly redefine dependence. When that cannot be justified, report joint-group contrasts or the individual ablations without claiming unique shares.
Diagnostics, Limitations, and Common Pitfalls
A sound decomposition starts by checking that each scenario differs only as intended. Confirm that the same population, dosing regimen, endpoint definition, target, simulation settings, and retained random-effect draws are used. Inspect exposure summaries as well as PTA: a change in exposure spread or location may not change attainment at a particular threshold. Repeat simulations with enough replicates to make Monte Carlo error small relative to the differences being interpreted, as discussed in the earlier material on choosing Monte Carlo replicate counts and quantifying Monte Carlo error in PTA.
Interpret results alongside the target definition. A driver can have a clear effect on exposure but a small PTA contrast if most simulated individuals remain on the same side of the target. Conversely, a modest exposure change can move many individuals across a threshold and produce a substantial PTA change. Results can also differ across dose levels, endpoints, or targets; one contrast is not a universal ranking of parameter importance.
- Pitfall: treating ablation differences as causal effects. The incorrect reasoning is, “Removing clearance variability raised PTA, so clearance variability causes treatment failure.” The correction is that the result is a model-based scenario contrast for the specified population and target; it does not identify a clinical causal effect.
- Pitfall: adding one-at-a-time changes. The incorrect reasoning is, “The clearance ablation changed PTA by 10 points and the volume ablation by 5, so their combined effect must be 15 points.” Thresholding, dependence, and interactions can make the combined contrast differ. Simulate the combined scenario or use a clearly defined allocation method.
- Pitfall: changing all random draws between scenarios. The incorrect reasoning is to subtract two independently simulated PTA values and interpret the difference as a parameter-specific effect. Unmatched Monte Carlo noise can obscure the contrast. Use paired draws for components held in common.
- Pitfall: assuming unchanged PTA means unchanged exposure. Equal attainment counts do not imply equal exposure distributions. Compare exposure summaries or distributions in addition to the target indicator.
Finally, a decomposition inherits limitations of the structural and covariate models, the estimated variability distribution, and the chosen reference scenario. It describes sensitivity within those assumptions. It does not establish that a random-effect component is biologically independent, that it is the only relevant source of variability, or that the resulting dose decision is robust to uncertainty in the model.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- In a paired simulation, full-model PTA is 0.68 and PTA with one random effect ablated is 0.74. State the contrast and interpret its direction.
- Why can removing a random effect change exposure summaries but leave PTA unchanged?
- What must be considered before separately ablating two correlated random effects?
- In a two-driver Shapley analysis, \(v(\varnothing)=0.80\), \(v(A)=0.76\), \(v(B)=0.74\), and \(v(A,B)=0.68\). Calculate the Shapley contributions.
- What does a small Monte Carlo standard error for a PTA contrast establish, and what does it not establish?
Solutions. (1) The full-minus-ablated contrast is \(0.68-0.74=-0.06\), a decrease of 6 percentage points with the random effect included in this specified comparison. (2) PTA changes only when attainment status crosses the target; exposure values can move without changing any individual's status. (3) The analyst must specify how the joint distribution and dependence are handled; setting one effect to zero while retaining the other may alter the natural joint structure. (4) Driver \(A\) contributes \(\{(0.76-0.80)+(0.68-0.74)\}/2=(-0.04-0.06)/2=-0.05\). Driver \(B\) contributes \(\{(0.74-0.80)+(0.68-0.76)\}/2=(-0.06-0.08)/2=-0.07\). Together they sum to -0.12, equal to \(0.68-0.80\). (5) It indicates Monte Carlo precision for the estimated contrast under the simulation setup; it does not validate the model or quantify uncertainty in its estimated parameters.