Didier Concordet
National Veterinary School. 23, Chemin des Capelles, 31076 Toulouse Cedex France.
When a kinetic study is performed, concentrations and covariates are measured on different individuals. Then, using a population pharmacokinetic model, mean (or individual) drug parameters are estimated in order to predict the appropriate dosage regimen for the individuals. Since theses means (or individuals) predictions are based on a model, it is necessary to check the validity of the assumptions which make it “usable”. Particularly, individual parameters are often assumed to be gaussian (at least in a given scale) when the covariates are fixed to a given value. Even if the population mean parameters have been shown to be robust to slight departure from normality, the variance of the random effect and the individual predictions can be biased. Since both (and their graphical counterparts) are used to select covariates, it is prudent to identify large departure from this assumption.
We propose a graphical method permitting to check the appropriateness of the assumed distribution of the individual parameters (the random effects) based on a bayesian predictor such as the Maximum A Posteriori (MAP) or the a posteriori mean. When the number of observations per individual is small the distribution of the predictor is different from the true random effects distribution whatever the number of individuals. The random effect is usually a multidimensional random variable from which we construct an unidimensional “omnibus” predictor to check the normality without too a large loss of power.
We construct the distribution of the omnibus predictor that would have been obtained if the random effect was gaussian. The direct comparison of the empirical (coming from the data) distribution of the random effect to the expected one (from the gaussian assumption), allows to identify departure from normality.
Reference: PAGE 9 () Abstr 80 [www.page-meeting.org/?abstract=80]
Poster: oral presentation