New Estimation Methods in NONMEM 7: Evaluation of Robustness and Runtimes
Sebastian Ueckert, Åsa M. Johansson, Elodie L. Plan, Andrew Hooker, Mats O. Karlsson
Department of Pharmaceutical Biosciences, Uppsala University, Sweden
Background: NONMEM is the most widely used software for population PKPD analyses. The latest version, NONMEM 7 (NM7), includes several new sampling-based estimation algorithms in addition to the classical methods. Besides the recently evaluated accuracy and precision inherent in these methods [1], time to complete estimation and sensitivity with respect to initial values might be critical in practice.
Objectives: To investigate the runtimes and robustness of the FOCE, LAPLACE, ITS, IMP, IMPMAP, SAEM and BAYES methods in NM7.
Methods: Five models representing different types of PKPD data handling, continuous, binary, count, ordered categorical (OC), repeated time-to-event (RTTE), were used to simulate 100 datasets that were subsequently reestimated using NM7 through PsN 3.1.8 (http://psn.sf.net). All datasets were analyzed twice, (A) starting with initial estimates set to the simulation values and (B) starting at values randomly generated using the CHAIN option. For the latter, fixed effects were sampled from a uniform distribution to [0;2θTRUE] (IACCEPT=1); for the random effects, a Wishart density of variance ωTRUE with 20 degrees of freedom was used. All estimation methods were used with their default settings and a test for convergence if available.
Average estimation time for each method was calculated from runtimes reported by NM7 (running on similar, dedicated machines). Median absolute deviation in final estimates between approach A and B was computed.
Results: Across models LAPLACE (FOCE for the continuous model) had the shortest runtimes, followed by ITS with runtimes equal to LAPLACE for the OC model and approx. half as fast for the remaining models. Runtimes for SAEM were between 20 and 40 times longer than for LAPLACE, but always faster than IMP and IMPMAP (40-60 times slower than LAPLACE). For all models BAYES was slowest (up to 250 times LAPLACE).
In general, sensitivity w.r.t. initial values was higher for random than for fixed effects and differed considerably between methods, models and parameters. However, the BAYES method showed the highest sensitivity in all scenarios. Furthermore, FOCE/LAPLACE had the lowest sensitivity across parameters for all but the OC model, where the IMP method performed best. For the continuous, binary, count and RTTE models, ordering of methods by increasing sensitivity yielded in: FOCE/LAPLACE, ITS, IMP, IMPMAP, SAEM, BAYES. For the OC the same ranking gave: IMP, IMPMAP, SAEM, ITS, LAPLACE, BAYES.
References:
[1] Johansson ÅM, Ueckert S, Plan EL, Hooker AC & Karlsson MO: New Estimation Methods in NONMEM 7: Evaluation of Bias and Precision. PAGE 19. 2010.
