Covariate Model Building Method for Nonparametric Estimation Method in NONMEM VI: Application to Simulated Data
Paul G Baverel, Radojka M Savic, Mats O Karlsson
Department of Pharmaceutical Biosciences, Uppsala University, Sweden
Objective: To develop a method for covariate model building suitable for the nonparametric estimation implemented in the software NONMEM VI .
Methods: The method is based on the calculation of joint density parameter distributions for each individual from the population joint density distribution and the individual data. For each parameter, the marginal, individual, probability is used as the main weighting factor in a generalized additive model (GAM), implemented in the software R. The relative performance of the new method at detecting true covariate relationships was evaluated in comparison with parametric GAM analysis. A 1-compartment IV bolus model was used to simulate 10 datasets of 100 individuals following a rich sampling design. Ten different covariates were simulated of which 7 were continuous (4 with underlying log-normal distribution, one with underlying normal distribution and 2 following a uniform distribution) and 3 were binomial. Relationships between CL and a continuous covariate as well as V and a categorical covariate were included in the simulations. Re-estimation with the reduced model was then conducted for each set of data using either FOCE or FOCE-NONP method. For the parametric method, GAM analyses based on empirical Bayes estimates were performed. For the nonparametric method, a Perl script was used to automate and output the individual contributions (iOFV) from which the individual joint density parameter distributions were derived. Additional weighting by the parametric SEs of EBEs and individual variances of individual nonparametric distributions were also explored.
Results: Overall, preliminary results suggest that the new method for nonparametric estimation performed similarly to parametric GAM with respect to selecting true covariate relationships when applied to rich simulated data.
Conclusions: A covariate model building technique intended for the nonparametric method in NONMEM VI is proposed. When applied to rich simulated data sets, the performance of the nonparametric method in the stepwise search process performed similarly as the regular parametric GAM method.
 Savic RM, Kjellsson M, Karlsson MO. Evaluation of the nonparametric estimation method in NONMEM VI beta PAGE 15 (2006) Abstr 937 [www.page-meeting.org/?abstract=937]