Page 7 of 16
suspicious. Indeed, it makes more biological sense to assume that weight is always a covariate
for clearance and other PK model parameters such as volume of distribution (7).
The example of weight vs. creatinine clearance with respect to avoiding correlations between
covariates is suboptimal: although there is undeniably a correlation between them in most
subjects, they are separate physical descriptors, which are useful for different things. We propose
making this argument with a less ambiguous example, such as weight vs. body surface area.
2.7 Model evaluation
We believe that “model validation” is not an appropriate term (8), and that the new guidance
should instead use “model evaluation”. One can never fully “validate” a nonlinear mixed-effects
model, only evaluate it given its purpose.
An ISoP working group composed of internationally recognized experts in pharmacometrics
published a tutorial covering model evaluation for longitudinal models with continuous data (9),
but this appears to have escaped attention when preparing the new guidance. We recommend that
due consideration be made of its contents, which we believe represent the current best practice in
model evaluation.
Additional observations and suggestions regarding model evaluation in the updated guidance are
below:
• Residual-based diagnostics are useful under most circumstances, but appropriate and
relevant simulation-based diagnostics such as visual predictive checks (VPCs) are equally
essential components of any PPK analysis, and are often more informative than standard
goodness-of-fit (GOF) plots. We suggest that their use be strongly recommended by this
guidance.
• There are different ways to compute a population prediction (PRED) and it is not always
the prediction with the typical parameters that should be used for GOF plotting purposes.
This has been extensively described in the ISoP tutorial quoted above. We used xPRED
where x could be nothing, C, or P (PPRED is called EPRED in NONMEM). Particularly,
CWRES should be plotted versus CPRED not PRED.
• There is no universally accepted definition of acceptable shrinkage (20% versus 30%) or
of what is excessive. In addition, it should be noted that models that have high shrinkage
might still be valid. This is also discussed in the ISoP guidance: “A shrinkage value of
30% or 50%, if calculated from SD or variance, respectively, has been suggested as a
threshold for high shrinkage, but whether this threshold should be applied for all models
and population parameter values remains to be evaluated.” We agree that shrinkage
should always be reported since it can result in loss of power (10,11).
• Using normalized prediction distribution errors (NPDE) as a simulation-based model
evaluation tool is not mentioned, although it is now available in most PPK software. In
the ISoP tutorial, some graphs based on NPDE vs. time and xPRED are part of the core
set of recommended diagnostics.
• Not all estimation methods provide eigenvalues to determine condition number. In
addition, it should be noted that models with high condition number might still be
appropriate and useful.