npde is a metrics designed to evaluate non-linear mixed effect models. Model evaluation is an important part of model building. Since the development of statistical software allowing parameter estimation in non-linear mixed effect models in the early 70's, model evaluation and diagnostic graphs have mostly used standardised prediction errors (a.k.a. weighted residuals or WRES in the software NONMEM), but this metric is computed using a first-order linearisation of the model function and as such, its shortcomings have been amply characterised, most notably an increase in the type I error when comparing models or building covariate models [1].
Prediction discrepancies (pd) were recently proposed as an improvement over this previous metric: they take into account the full predictive distribution of each observation [1]. Normalised prediction distribution errors (npde) further improved over pd by taking into account the correlation between multiple observations in each subject [2].

We designed the software npde to easily compute npde and pd.

npde was programmed as an add-on package for the R software. Documentation is available in the relevant section and includes detailed explanations and examples as to how to use npde.

[1] Mentré F, Escolano S. Prediction discrepancies for the evaluation of nonlinear mixed-effects models, Journal of Pharmacokinetics and Biopharmaceutics, 2006, 33: 345-­67. Link to PDF of paper
[2] Brendel K, Comets E, Laffont C, Laveille C, Mentré F. Metrics for external model evaluation with an application to the population pharmacokinetics of gliclazide, Pharmaceutical Research, 2006, 23: 2036-­49. Link to pdf of paper
See also references in the reference page.

Brendel K, Comets E, Laffont C, Laveille C, Mentré F. Metrics for external model evaluation with an application to the population pharmacokinetics of gliclazide, Pharmaceutical Research, 2006, 23: 2036-­49

©2007 INSERM U738