Y.Merlé(1) and M. Tod(2)
1) Inserm U436, 91 Bd De L'hopital, 75634 Paris Cedex 13, France; 2) Departement De Pharmacie Et Toxicologie, Hopital Avicenne, 125 Route De Stalingrad, 93009, Bobigny, France
Suitable scalar functions of the population information matrix can be used as design criteria when optimizing an experiment for estimating population characteristics. Unfortunately, no closed form of this matrix does exist for nonlinear models. An approach, relying on the model linearization has been proposed and considerably simplifies the population information matrix computation. However, the quality of this approximation has not been really assessed and its impact when evaluating a design or optimizing an experiment is not known. To address that point we considered a sample of 100 subjects constituting a single group, each individual receiving two or three dose levels, the latter constituting the design variable. The drug effect was supposed to be described by an Emax model. Pharmacodynamic parameters were assumed to be
normally distributed in the population with known mean and variance. Population information matrix was computed for the mean population parameters and various designs (i.e. dose levels) by two distinct approaches: i) the above cited method leading to an approximation; ii) a procedure using numerical techniques, not requiring any model linearization, and considered as the reference method. D-optimal population designs were also determined using a Powell’s algorithm, the population information matrix being evaluated by the two procedures. Lower bounds of hyperparameter estimates accuracy were derived from the population information matrix for each design. In order to assess the impact of non linearity on the approximation
quality, the RMS intrinsic and parameter effects curvatures of Bates and Watts were computed for each design. Our results show that the two calculation procedures lead to the same D-optimal designs. Accuracies of mean parameter estimates are similar with both methods. In contrast substantial differences can be observed between the accuracies of parameter variability estimates obtained by the two approaches, the magnitude of these differences increasing with the RMS parameter effects curvature only. Hence, the latter needs be determined to assess the adequacy of the linear approximation on the computation of the population information matrix.
Reference: PAGE 9 () Abstr 105 [www.page-meeting.org/?abstract=105]
Poster: poster