New Estimation Methods in NONMEM 7: Evaluation of Bias and Precision
Åsa M. Johansson, Sebastian Ueckert, Elodie L. Plan, Andrew C. Hooker, Mats O. Karlsson
Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden.
Objectives: The aim was to investigate the performance with respect to bias and precision of all estimation methods available in NONMEM 7 for a diverse set of PKPD models.
Methods: Five PKPD models with different types of PD data: continuous (1), binary (2), ordered categorical (3), repeated time-to-event (4) and count (5) were used in the study. The estimation methods investigated were: BAYES, IMP, IMPMAP, ITS, SAEM, FOCE (for model 1) and LAPLACE (for models 2-5). The options for the estimation methods were kept as default except that a convergence test was applied if available. Stochastic Simulations and Estimations (SSE) using PsN (http://psn.sf.net) were utilized to compare the methods: for each model, 500 datasets were simulated and then reanalyzed with the different methods. Relative root mean squared error (RMSE) (100*sqrt(mean[(θEst-θTrue)2]) /θTrue) in estimates were evaluated for each parameter. A score based on ranks according to RMSE was attributed to each method for each model.
Results: The average fraction of runs for which the minimization completed successfully differed between the methods; 100% with SAEM and FOCE, 43% with ITS and between 81% and 95% for the other methods. The average rank scores resulted in the following ranking of the methods: SAEM (2.2), IMP (2.6), IMPMAP (2.8), ITS (3.2), FOCE/LAPLACE (4.2) and BAYES (5.8). SAEM performed best for models 1, 2 and 3, whereas IMPMAP was the best method for model 4 and IMP for model 5, followed by ITS. The BAYES method had the lowest ranking for all models, except model 3 where it was ranked number 5 and for which ITS performed worst. The average ranking scores for the fixed effects were similar to the total ranking scores, whereas for the random effects best ranking score was shared with FOCE/LAPLACE and IMPMAP. The FOCE/LAPLACE methods performed well with model 1, 2 and 5, whereas IMPMAP performed well with model 3, 4 and 5. The BAYES method performed worst regarding the random effects and had the lowest rank with all models.