Prediction Correction Under Dose Variation

Why Dose Variation Matters in a Prediction-Corrected VPC

A visual predictive check (VPC) compares observed data summaries with the corresponding summaries from data simulated under a population pharmacokinetic model. When subjects receive different doses, their expected concentrations may differ systematically even at the same nominal sampling time. A pooled VPC can then mix dose-related differences in typical concentration with differences in model fit. Prediction correction attempts to put those observations on a common prediction scale before calculating the summaries.

That adjustment is a rescaling, not a new source of information about the dose-response relationship. It can make a pooled display easier to interpret when the model’s predictions adequately represent dose effects and the remaining variation behaves comparably across doses. But when those conditions fail, correction can itself change the spread or shape of the data. The important question is therefore not simply whether doses vary, but whether the correction is appropriate for the way dose affects both expected concentrations and residual variability.

The predictions in this definition are model-based predictions for the observed design; they are not individual posterior predictions inferred from each observed concentration. The reference value is a common scale for the bin, not a dose that every subject received. Both \(Y_{ij}\) and \(Y^{\mathrm{pc}}_{ij}\) have concentration units because the ratio of predictions is dimensionless. Predictions must be positive for this ratio to be usable.

When the Rescaling Has a Clear Interpretation

The simplest justification is multiplicative. Suppose, within a bin, the concentration can be represented as \(Y_{ij}=m_{ij}R_{ij}\), where \(R_{ij}\) is a dimensionless relative deviation from the model prediction. Then

The dose-specific prediction cancels, while each observation retains its relative deviation \(R_{ij}\). If the model captures the typical dose effect and the distribution of relative deviations is comparable across dose groups, corrected observations can be pooled on the reference-prediction scale. The same calculation for simulated data lets the observed and simulated summaries be compared on that scale.

This argument does not require dose to have a linear effect on concentration. If dose effects are nonlinear but the model predicts them appropriately, the correction uses each occasion’s predicted concentration, rather than assuming that concentration is directly proportional to dose. What matters is whether those predictions describe the expected concentrations well enough and whether rescaling preserves a meaningful comparison of the remaining variation.

Correction is most naturally interpreted when departures from the prediction are approximately relative—for example, when variability is roughly constant on a proportional scale. With additive error, the absolute error is instead multiplied by the correction factor. In that setting, the expected value may align while the variance does not. This distinction is central to interpreting corrected quantiles.

Nonlinear Dose Effects and the Model’s Role

Prediction correction uses model predictions, so it can accommodate a nonlinear dose-response relationship only insofar as that relationship is represented by the model. It does not independently validate the dose effect. If the model predicts the wrong relative concentrations across doses, the correction factor can amplify or suppress observed differences. The resulting pooled summary may appear more compatible with simulation—or less compatible—even though the underlying issue is the dose prediction.

How Error Structure Can Distort Corrected Summaries

A common failure mode arises when residual variation is additive rather than proportional. Suppose \(Y=m+\epsilon\), where \(\epsilon\) has mean zero and variance \(\sigma^2\) that is the same across dose levels. After correction,

Although the expected corrected value is \(m_{\mathrm{ref}}\), the corrected variance depends on \(m\). Lower predicted concentrations receive larger multipliers and therefore larger corrected additive errors. Higher predicted concentrations receive smaller multipliers. Pooling the corrected data can create or alter spread across doses, so the corrected quantiles need not represent a common residual distribution.

Checks Before Interpreting a Pooled Corrected VPC

Prediction correction is useful only when its consequences are checked. Examine the predictions, correction factors, and dose composition alongside the corrected summaries. The following questions help identify when the display may be misleading:

• Are predictions positive and well away from zero? A small denominator produces a large correction factor, making corrected values sensitive to small prediction errors. Nonpositive predictions make the ratio undefined or uninterpretable.
• Does the model describe dose-related typical concentrations? Compare predictions across doses and time with the model’s intended dose-response behavior. A correction based on a misspecified dose effect can obscure the discrepancy it is meant to control.
• Is residual variation plausibly comparable after rescaling? Consider whether residual error is chiefly proportional or additive and whether variability changes across doses. The correction factor multiplies absolute residual deviations.
• Do dose groups have overlapping sampling times and adequate representation? If a bin is dominated by one dose, its reference prediction and corrected summaries may mainly reflect that group. Sparse representation at another dose limits what a pooled display can show.
• Do the observed and simulated data use matching correction rules? Apply the same defined factors and reference-prediction rule to observed and simulated data at the same design points. A mismatch can introduce an artificial difference.

A helpful comparison is to inspect the ordinary VPC and the prediction-corrected VPC together. The ordinary display retains the dose-related concentration scale; the corrected display aims to remove the predicted typical dose effect. If their conclusions differ, ask whether this is the expected consequence of removing that effect or a sign that the model or error structure needs closer examination.

Dose-specific summaries are an alternative when dose groups have distinct residual behavior or when the dose prediction is uncertain. They retain within-dose comparisons rather than combining all doses on one rescaled axis. A pooled correction is more compact, but it relies more heavily on the prediction and residual-structure assumptions. Neither display eliminates the need to examine whether the model can reproduce the features relevant to the diagnostic target.

Common Mistake and Practical Limits

Wrong reasoning: “Prediction correction removes dose effects, so any remaining difference between corrected observations and simulations must be unrelated to dose.” This is not justified. The correction removes only the dose-related differences represented by the predictions, and its multiplication changes absolute residual errors. Dose-related residual variance, misspecified dose effects, or unstable correction factors can still influence corrected summaries.

Correction: Interpret a pooled corrected VPC as a comparison on a model-defined reference scale. Check dose-specific predictions and the effect of the correction factors, and consider whether the residual variation is comparable after rescaling. If it is not, examine dose groups separately rather than relying on a single pooled summary.

Prediction correction is also limited by the information in the data. It cannot recover dose-response behavior where some dose groups have little or no representation at relevant times, and it cannot show which model component caused a discrepancy. It is a visualization tool for a specified predictive check, not a replacement for model evaluation or for choosing a diagnostic target.