1. What Is a Visual Predictive Check?
A visual predictive check (VPC) is a simulation-based diagnostic used to evaluate whether a population PK model can reproduce important features of the observed concentration-time data.
The central idea is straightforward: use the fitted population PK model to simulate many replicate datasets, summarize those simulated observations over time, and then compare the simulated distribution with the actual observations.
A VPC compares observed data with the range of observations expected under repeated simulations from the fitted population PK model.
2. Why Are VPCs Useful?
Population PK models contain multiple sources of variation. A model may have a reasonable typical concentration-time trajectory while failing to reproduce the observed variability, tails, or time-dependent distribution of concentrations.
A VPC provides a graphical way to examine several of these features simultaneously.
| Feature | What the VPC can reveal |
|---|---|
| Central tendency | Whether the model reproduces the typical location of the observed concentration distribution over time |
| Variability | Whether the model generates an appropriate spread of concentrations |
| Distributional tails | Whether unusually high or low observations occur at frequencies compatible with simulation |
| Time dependence | Whether model performance changes across absorption, distribution, and elimination phases |
| Systematic misspecification | Whether observations repeatedly fall outside or near the boundaries of the simulated distribution |
| Observation model | Whether residual error and other observation-level variability are represented adequately |
Unlike a simple observed-versus-predicted plot, a VPC focuses on the distribution of simulated observations. This distinction is particularly important for population models, where between-subject variability and residual unexplained variability are part of the model itself.
3. How Is a VPC Constructed?
A conventional VPC follows a simulation-and-comparison workflow.
- Fit the population PK model. Estimate structural parameters, between-subject variability, covariate effects, and residual variability.
- Simulate replicate datasets. Use the fitted model to generate many datasets under the study design.
- Calculate summary statistics. At selected time regions or bins, calculate statistics such as the 5th, 50th, and 95th percentiles.
- Construct simulation-based intervals. Across replicate datasets, determine the uncertainty around each simulated percentile.
- Plot the observed data. Overlay observed concentrations and observed summary statistics on the simulated prediction intervals.
- Interpret the comparison. Look for systematic differences between observed behavior and the behavior reproduced by the model.
The important point is that the model is being challenged to reproduce data generated under the same general design and variability structure, rather than being evaluated only through its fitted predictions.
4. What Gets Simulated?
For a population PK VPC, simulations generally include both the population model and the sources of variability specified in the model.
Conceptually, an individual simulated concentration can be written as:
where \(f(\cdot)\) represents the structural PK model, \(\theta\) represents population parameters, \(\eta_i\) represents individual-level random effects, \(\mathbf{x}_i\) represents covariates, and \(\epsilon_{ij}\) represents residual unexplained variability.
The exact implementation depends on the model and software, but the conceptual distinction is important:
| Component | Role in the simulation |
|---|---|
| Structural model | Determines the expected concentration-time behavior |
| Population parameters | Define typical PK characteristics |
| Between-subject variability | Generates differences among simulated individuals |
| Covariates | Allow simulated individuals to reflect the covariate structure of the study |
| Residual error | Creates observation-level variation around the model prediction |
5. What Are the Prediction Intervals?
A VPC commonly displays simulated percentiles such as the 5th, 50th, and 95th percentiles of concentration at different times.
For example, within a particular time bin, suppose 1,000 simulated datasets are generated. For each dataset, the 5th percentile of concentration is calculated. This produces 1,000 simulated values of the 5th percentile. The distribution of those values can then be summarized by an interval.
This interval describes the simulation uncertainty around the expected 5th percentile. The same process can be applied to the median and 95th percentile.
Conceptual VPC: simulated prediction intervals surround the range of behavior expected under the model, while observed concentrations are overlaid for comparison.
Thus, the shaded region in a VPC is not simply a confidence interval for an estimated model parameter. It is a simulation-based prediction interval for a feature of the data.
6. Why Are Time Bins Often Used?
PK datasets usually contain observations at many different times, and the density of observations may vary substantially across the study. Directly estimating a percentile at every exact time can therefore be unstable or difficult to interpret.
VPCs commonly divide time into bins. Within each bin, observed and simulated concentrations are summarized.
| Time bin | Purpose |
|---|---|
| Early sampling | Evaluate absorption and early concentration behavior |
| Distribution phase | Assess whether the model captures changing concentration variability during distribution |
| Later sampling | Evaluate elimination and terminal behavior |
| Sparse regions | Provide broader bins when there are insufficient observations for stable summaries |
The choice of bins can influence the appearance of a VPC. Bins that are too narrow may contain very few observations, while bins that are too broad can obscure important time-dependent model deficiencies.
7. How Do You Read a VPC?
The first question is whether the observed data generally fall within the simulation-based prediction intervals. But that is only the beginning of the interpretation.
Central tendency
Compare the observed median with the simulated median region. A systematic separation can indicate that the model's typical concentration-time behavior is inconsistent with the observations.
Variability
Compare the observed lower and upper percentiles with the corresponding simulated prediction intervals. If the observed spread is consistently wider than simulated, the model may be underestimating variability.
Time-dependent patterns
A model can perform adequately during one part of the concentration-time profile and poorly during another. For example, systematic discrepancies during early times may suggest problems with absorption, while discrepancies during the terminal phase may point toward the disposition model.
Extreme observations
Individual observations outside a prediction interval are not automatically evidence of model failure. A prediction interval describes a distribution, so some observations are expected to fall outside it. The important question is whether the pattern of observations is inconsistent with what the model produces.
8. Worked Example: Interpreting a Population PK VPC
Suppose a population PK model is developed for an orally administered drug. The model contains first-order absorption, a two-compartment disposition model, between-subject variability on clearance and volumes, and a proportional residual error model.
After fitting the model, 1,000 replicate datasets are simulated using the original study design.
Step 1: Examine the observed median
Suppose the observed median concentration lies close to the simulated median across most of the profile.
This suggests that the model's typical concentration-time trajectory is broadly consistent with the observed data.
Step 2: Examine the observed spread
Suppose the observed 5th and 95th percentiles also remain mostly within the corresponding simulation-based prediction intervals.
This provides evidence that the model reproduces the overall magnitude of variability reasonably well.
Step 3: Examine early concentrations
Now suppose the observed concentrations during the first two hours are systematically higher than the simulated upper prediction region.
This pattern is more informative than a handful of isolated observations. It suggests that the model may not adequately reproduce the early concentration behavior. Possible explanations could include an inappropriate absorption model, an incorrect absorption rate, dose-history assumptions, or other structural features that affect early concentrations.
Step 4: Examine the terminal phase
Suppose the later observations are well reproduced.
The discrepancy therefore appears concentrated in the early part of the profile rather than being a general failure of the model.
9. What Does a VPC Say About Variability?
Population PK models typically contain multiple levels of variability. A useful conceptual decomposition is:
The exact statistical representation depends on the model, but both components can influence the width and shape of the simulated concentration distribution.
For example, if between-subject variability is too small, simulated individuals may be too similar to one another. The VPC may then show a simulated distribution that is narrower than the observed distribution.
Conversely, excessive variability can produce prediction intervals that are substantially wider than the observed data.
| Observed pattern | Possible modeling question |
|---|---|
| Observed spread wider than simulation | Is between-subject variability adequately specified? Is residual error too small? |
| Observed spread narrower than simulation | Is variability overestimated? |
| Median reproduced but tails poorly reproduced | Does the variability distribution or residual error model adequately represent the tails? |
| Variability mismatch only at certain times | Does residual error or structural model behavior change appropriately over the concentration range? |
These are diagnostic questions rather than automatic conclusions. A VPC identifies patterns; additional diagnostics and model investigation are needed to determine their cause.
10. Why Might a Stratified VPC Be Needed?
Population PK models frequently include covariates such as body weight, renal function, age, or treatment group. A model may reproduce the overall population distribution while failing for a particular subgroup.
A stratified VPC separates the data into clinically or scientifically meaningful groups and evaluates the model within each group.
| Potential stratification | Question |
|---|---|
| Dose group | Does the model reproduce concentration distributions across dose levels? |
| Body-size category | Does the model adequately describe concentration differences associated with body size? |
| Renal-function category | Does the model reproduce differences in exposure associated with renal function? |
| Study or formulation | Does the model reproduce differences associated with study design or formulation? |
| Age group | Does the model adequately represent age-related differences? |
Stratification can reveal problems hidden by an aggregate VPC. However, excessive stratification can leave too few observations in individual groups and make the diagnostic difficult to interpret.
11. What Is a Prediction-Corrected VPC?
When different subjects have substantially different dosing regimens or covariate values, their expected concentrations can differ even when the model is correct. A standard VPC may then be difficult to interpret because the observed concentration distribution reflects differences in design as well as model performance.
A prediction-corrected VPC (pcVPC) adjusts observations and simulations relative to a typical prediction so that distributions can be compared more effectively when study design or covariates produce systematic differences in expected concentrations.
Conceptually, the correction asks:
The exact prediction-correction procedure depends on the implementation, but the goal is to reduce the influence of known differences in typical predictions while retaining information about unexplained distributional behavior.
12. Why Do VPCs Have Shaded Prediction Regions?
It is useful to distinguish two different ideas:
- Variability: the range of concentrations generated among individuals or observations in a simulated dataset.
- Simulation uncertainty: the variation in a summary statistic across repeated simulated datasets.
Suppose 1,000 datasets are simulated. Each dataset has its own median concentration within a particular time bin. Those 1,000 medians will not be identical.
The VPC can display the resulting distribution of medians as a prediction interval. Thus, the shaded band around the simulated median reflects uncertainty in what the model would produce for that summary under repeated simulation.
This distinction prevents a common interpretation error: the shaded VPC region is not simply the range containing 95% of individual simulated concentrations.
13. Standard VPC or Prediction-Corrected VPC?
| Situation | Diagnostic consideration |
|---|---|
| Similar doses and sampling schedules | A conventional VPC may provide an intuitive comparison. |
| Wide range of doses | Prediction correction or dose stratification may improve interpretability. |
| Strong covariate effects | Stratification or prediction correction may help separate model performance from expected covariate-driven differences. |
| Highly heterogeneous study design | A standard aggregate VPC may hide important patterns. |
| Multiple studies combined | Study-stratified diagnostics may be useful. |
The choice should follow the scientific question and the design of the data. A more sophisticated VPC is not automatically more informative.
14. What a VPC Cannot Tell You by Itself
A VPC is an important population-level diagnostic, but it should not be treated as a complete model evaluation.
- A VPC does not prove that the structural model is correct. Different models can sometimes reproduce similar observed distributions.
- A VPC does not diagnose the exact source of a discrepancy. Additional diagnostics are needed to determine why a pattern occurs.
- A few observations outside the prediction interval do not automatically indicate model failure.
- A visually attractive VPC does not guarantee reliable parameter estimates.
- A VPC can be influenced by simulation settings, binning, stratification, and prediction correction.
- A population-level VPC does not establish individual-level predictive accuracy.
15. Common VPC Interpretation Mistakes
Mistake 1: Treating every point outside the band as a failure
Prediction intervals describe a distribution. Some observations can legitimately fall outside the simulated region. Look for systematic patterns rather than isolated points.
Mistake 2: Looking only at the median
A model can reproduce the median while substantially underestimating variability. Examine both central tendency and distributional spread.
Mistake 3: Ignoring the study design
Dose, sampling schedule, formulation, and covariate distributions can strongly influence the concentration distribution. The simulations should reflect the relevant design.
Mistake 4: Using arbitrary bins
Very narrow bins can create unstable percentile estimates, while overly broad bins can conceal time-dependent discrepancies.
Mistake 5: Assuming the VPC identifies the problem
A VPC can show where a model disagrees with the observed distribution. Determining why requires additional model diagnostics and scientific reasoning.
Mistake 6: Confusing prediction intervals with confidence intervals
A VPC prediction region concerns the distribution of simulated data or simulated summary statistics. It is not simply a confidence interval for an estimated PK parameter.
16. A Practical VPC Workflow
- Finalize the candidate model. Ensure the structural model, covariates, between-subject variability, and residual error model are appropriately specified.
- Define the simulation design. Decide whether to replicate the original dosing and sampling design or use another design appropriate to the diagnostic question.
- Generate many replicate datasets. Use the fitted population model to simulate observations.
- Choose useful summaries. Common summaries include the 5th, 50th, and 95th percentiles.
- Choose sensible time bins. Ensure sufficient observations are available for stable summaries while retaining useful temporal resolution.
- Calculate simulation-based prediction intervals.
- Overlay the observed data. Plot observations and observed percentile summaries against the simulated regions.
- Consider stratification or prediction correction. Use these when dose, covariate, or design heterogeneity makes the aggregate VPC difficult to interpret.
- Investigate discrepancies. Use additional diagnostics to determine whether patterns suggest structural, covariate, variability, or residual-error misspecification.
- Document the simulation settings. Record the number of simulations, binning approach, stratification, prediction correction, and other choices that affect interpretation.
17. VPCs in the Population PK Diagnostic Toolbox
A population PK model should generally be evaluated from several complementary perspectives.
| Diagnostic | Primary question |
|---|---|
| Observed vs. predicted plots | Are predictions systematically different from observations? |
| Residual diagnostics | Are residuals consistent with the assumed observation model? |
| Individual prediction plots | How well does the model describe individual concentration-time profiles? |
| VPC | Can the model reproduce the population distribution of observations over time? |
| Parameter uncertainty | How precisely are the model parameters estimated? |
| Bootstrap or related simulation methods | How stable are the parameter estimates under repeated sampling? |
| Covariate diagnostics | Are systematic relationships inadequately represented? |
The VPC therefore occupies a specific role: it is especially useful for evaluating whether the model reproduces the distributional behavior of the observed population over time.
18. Key Takeaways
- A visual predictive check is a simulation-based diagnostic for population PK models.
- The basic VPC workflow compares observed concentration-time data with data simulated from the fitted model.
- Common VPC summaries include the 5th, 50th, and 95th percentiles of concentration.
- Simulation-based prediction intervals describe how those summaries vary across replicate simulations.
- A VPC evaluates both central tendency and variability rather than only the typical prediction.
- Time-dependent discrepancies can reveal problems that are hidden by aggregate measures of model fit.
- Between-subject variability and residual unexplained variability both influence the simulated concentration distribution.
- Stratified VPCs can reveal model deficiencies that are hidden when all subjects are analyzed together.
- Prediction-corrected VPCs can be useful when dose, covariates, or study design produce substantial differences in expected predictions.
- Individual observations outside a prediction interval do not automatically imply model failure; systematic patterns are more informative.
- A VPC does not identify the exact cause of a model discrepancy and should be interpreted alongside other diagnostics.
- The most useful VPC is one that is designed around the scientific question, study design, and characteristics of the population PK model.
Where to Go Next
A natural progression after learning VPCs is to study prediction-corrected VPCs, followed by goodness-of-fit diagnostics, conditional weighted residuals, normalized prediction distribution errors (NPDEs), bootstrap validation, and external model evaluation.
These methods answer related but different questions. Understanding how they complement one another is essential for a rigorous population PK model evaluation strategy.