Tutorial: Optimal design for population analysis

France Mentré

INSERM U194, Dpt de Biostatistiques et Informatique Médicale, CHU Pitié-Salpétrière, 91 Bd de 1'Hôpital, 75013 Paris, France

Introduction

  • General considerations on the accuracy of estimators in nonlinear regression, on the Fisher information matrix and D-optimality.
  • Extension to optimization of population designs or Bayesian designs
  • Optimal design for population analysis

Description of the problem: choice of the number of subjects, number of data points per subject, and individual designs for estimation of the distribution of the parameters.

First goal: increase the accuracy of the estimation, that is to say for parametric distributions, increase the accuracy of the estimates of means and variances

  • examples of simulation in population pharmacokinetics
  • some results in linear random effect model
  • a new approach (which used the NPML algorithm) for linearized models and gaussian distributions, with simulation for simple PK or PD models

Other goal: increase the predictive ability of the distribution as prior information for subsequent Bayesian analysis

  • example of some evaluation of predictive performance
  • a possible new approach: the evaluation of the predictive distributions for several designs

Bayesian design

Description of the problem: when the prior distribution is known and Bayesian estimation is scheduled, what individual design should be performed to increase the accuracy of the posterior estimates ?

  • limitations of D-optimality for that problem
  • Bayes-D optimality for linear models and gaussian distributions
  • extension of the Fisher information matrix for nonlinear random effect models
  • some simulated examples for simple PK, PD models
  • another approach: the Lindley information – related to the entropy of a distribution
  • illustration on a real example of the kinetics of radioiodine for treatment of Grave’s disease

Future: use utility functions based on the use of the posterior distribution for subsequent dosage optimization and not on the accuracy on the parameter’s estimates.

Conclusion

  • Some works for linear models and Gaussian distribution
  • Few applications in pop PK/PD
  • Most of the time no analytical solutions : specific software to be proposed

Reference: PAGE 2 (1993) Abstr 914 [www.page-meeting.org/?abstract=914]

Poster: oral presentation