MERLÉ Y. (1) and TOD M. (2)
(1) INSERM U436, 91 BD DE L'HOPITAL, 75634 PARIS CEDEX 13, FRANCE
Various methods for designing population experiments have been recently developed and rely on the optimization of an appropriate design criterion. These procedures, as well as the population analysis itself, implicitly assume that the (PK/PD) model is known and adequately describes the PK/PD of a given drug in every individual. Unfortunately, for various reasons, a non negligible uncertainty often remains about model adequacy. Here, our aim is to assess the influence of model misspecifications on optimal design and hyperparameter estimation in population pharmacokinetics. This assessment is conducted from simulated data sets. First, two PK nested models are chosen, the full one being assumed to be true. A population design optimization settlement is then considered and an experiment is optimized for each model considered by two design criteria. Replicated data sets are generated from the full model for every optimal experiment and analyzed by two widespread parametric methods on the basis of each model. Optimal designs obtained for each criterion and model are compared to each other as well as corresponding hyperparameter bias and precisions. Powers of approximate likelihood ratio tests used for discriminating between models are also evaluated for each model based design. Our results show that optimal designs obtained on the basis of each model largely differ. The impact magnitude of a model misspecification on the estimates depends on the step at which it occurs (design optimization and / or analysis), on the hyperparameter considered and on the estimation method. Bias and precision might be affected differently. Model misspecifications during the design optimization can lead to a decrease of the power of the above mentioned likelihood ratio tests. Some practical guidelines which might be useful when a model misspecification is suspected at the end of the data collection or at the end of the analysis are suggested.
Reference: PAGE 10 () Abstr 206 [www.page-meeting.org/?abstract=206]
Poster: poster