Per Olsson Gisleskog1 & Mats O Karlsson2
1Clinical Pharmacology, GlaxoWellcome R&D, Greenford, UK.; 2Division of Biopharmaceutics and Pharmacokinetics, Department of Pharmacy, Uppsala University, Uppsala, Sweden.
Modelling of drug behavior during drug development is usually an on-going process, where the models are refined, validated and expanded as new studies provide more data. Prior information is often used to help define the structural model, and sometimes as initial estimates for modelling of new data. In a more formal context, prior information can be used both to stabilise complex models when fitted to sparse data, and to propagate information about parameters and associated variabilities throughout the drug development program (1).
With these two objectives in mind, we explored different ways of incorporating prior information when performing mixed effects modelling using NONMEM. Two approaches were examined using simulations based on one small rich dataset, typical of phase I data, and one larger sparse dataset, typical of phase II data:
- Rich and sparse data were combined into one dataset and were fitted together.
- Parameters derived from the rich dataset were used as priors for the fitting of the sparse dataset
Scenarios were simulated both were the parameters of the two populations were identical and where parameters and their associated variabilities were different between populations. Diagnostic tools for assessing the relative amount of information on a parameter in the two datasets and for determining whether a parameter was different between the two populations were examined and different model building methods explored.
The two methods performed in a predictable way and gave broadly similar results. The fits of the sparse dataset were stabilised, decreasing the standard error in the estimations substantially. For parameters where adequate information was present in the sparse dataset, any changes from the rich dataset were reflected in the estimates, while for parameters where little information was present in the sparse dataset, the parameters remained close to those of the rich dataset. Aspects regarding methods for detection of parameter differences and nominal significance levels will also be presented.
1. Wakefield J, The use of predictive distributions in drug development. In: Aarons, L et al (Eds.), The population approach: measuring and managing variability in response, concentration and dose. Brussels: The European Commission. p.353-362.
Reference: PAGE 8 (1999) Abstr 137 [www.page-meeting.org/?abstract=137]
Poster: oral presentation