Referenced Visual Predictive Check (rVPC)
Heiner Speth, Martin Burschka, Ruediger Port
Objectives: The prediction-corrected VPC (pcVPC) [1] reduces the variance in the observed data and in the model distributions by normalizing to an "identically distributed" situation for all individuals. The computation of the prediction correction factor for each bin separately, limits this method to situations of balanced data within each bin. The goal is to extend the prediction-corrected VPC to situations of sparse and unbalanced data by avoiding binning on the independent variable (idv) scale and to show:
1. The equivalence of the new rVPC method to the pcVPC in cases of balanced data.
2. The robustness of the rVPC in cases of sparse and unbalanced data.
In addition to model evaluation, the rVPC is to provide a base for discussion with clinicians on dosing adequacy in untypical individuals.
Methods: For every observation, the model is requested to give a normalized prediction, based on a data tuple of normalized covariate values (reference) and the same random effects from the fit for this observation. The implementation is particularly simple in NONMEM.
Results: The rVPC is independent from binning and calculates the correct reference factor in any observed/ simulated data-point.
Conclusions: The rVPC implements the idea of covariate adjusting to any kind of data and allows an arbitrary choice of referenced covariate values.
References:
[1] Bergstrand, M., Hooker, A.C., Wallin, J.E., Karlsson, M.O.: Prediction-corrected visual predictive checks for diagnosing nonlinear mixed-effects models. The AAPS Journal pp. 1-9 (2011)
