Tarjinder Sahota (1), Brendan Johnson (1)
(1) Clinical Pharmacology Modeling and Simulation, GSK, UK.
Objectives: NONMEM has traditionally used gradient based algorithms (e.g. FO and FOCE) to estimate model parameters of nonlinear mixed effects models. These algorithms may fail to converge for numerical reasons. Newer expectation maximisation (EM) algorithms were introduced in version 7.0 of NONMEM which offered the prospect of increased numerical stability and reduced bias in model parameter estimation [1]. Unlike FO and FOCE estimation however, these algorithms come with many options and settings for the users to define which can be bewildering at first. The aim of this work is to derive efficient starting arguments using algorithm performance in real data case studies.
Methods: Three case studies are presented. 1) Mixture model PK with active metabolite 2) plasma lung model and 3) Hgb lifespan KPD model 4) Count data example. EM method performance was assessed relative to FOCE (LAPLACIAN for example 4) using the following criteria: 1) Likelihood of convergence, 2) Sensitivity to initial estimates 3) Run time and 4) MC noise evaluation. OFV vs ITERATION plots for assessing stationarity were used to select convergence criteria.
Results: Plotting parameters/OFV vs ITERATION is important component of assessing convergence. When initial estimates were close to final estimates, initial estimation steps (e.g. METHOD=ITS or METHOD=FO) can add to total run time and in some cases destabilise some models. Monte Carlo (MC) error in OFV evaluation was inappropriately high for some models. Use of RANMETHOD=3S1 and RANMETHOD=3S2 options resulted in drastic reductions in MC noise. Separate convergence criteria were defined for IMP/IMPMAP and SAEM methods.
Conclusions: The proposed argument settings provided a sensible starting point for efficient convergence with IMP and SAEM estimation methods.
References:
[1] Johansson, A. M., et al. Evaluation of bias, precision, robustness, and runtime for estimation methods in NONMEM 7. JPKPD (2014) 41:223-238
Reference: PAGE 24 (2015) Abstr 3630 [www.page-meeting.org/?abstract=3630]
Poster: Methodology - Estimation Methods