A. Lemenuel-Diot(1), N.frey(2), C. Laveille(2), R. Jochemsen(2), A. Mallet(1)
(1)INSERM U436, Dept Biomathematics, CHU Pitie Salpetriere, 91 bd de l’Hôpital, 75013 Paris, France. (2)Institut de Recherches Internationales Servier, 6 place des Pléiades, 92415 Courbevoie Cedex, France
Introduction and Objectives: The Gauss Hermite quadrature method allows to get an accurate approximation of the likelihood function by expressing the integral-based individual likelihood according to an adjustable and an easy way to compute expression: weighted values of the individual likelihood function calculated for tabulated nodes. In order to determine the advantage of an accurate approximation of the likelihood function, the Gauss Hermite quadrature method was compared to NONNEM for different features.The aim of the present work was to study the ability of Gauss Hermite quadrature and NONMEM for one specific feature : the detection of subpopulations.
Methods: Simulations/estimations were performed to compare Gauss Hermite quadrature method to NONMEM. Populations made up of two components were simulated with different distribution characteristics. The criteria of comparison between the two approaches were the power of detection of these two simulated subpopulations and the precision of estimations of the parameters. Since the two approaches do not use the same statistical test for the detection of the mixture model, it was necessary to adjust these tests before doing any comparison. The selection between two nested models (i.e. one component versus two components) is made using the log likelihood ratio test for NONMEM and the Kullback Leibler test for the Gauss Hermite quadrature method. A specific simulation/estimation approach was used to adjust each method in order to obtain different type I errors.
Results: With these two methods, the power of detection is quite linked to the characteristics of the mixture distributions : the differences between the mean of the components, their variances and the weights of the mixture. However, the ability to detect subpopulations is not the same for the two methods and these differences will be presented at the PAGE meeting.
Reference: PAGE 12 () Abstr 431 [www.page-meeting.org/?abstract=431]
Poster: poster