Evaluation of FREM and FFEM including use of model linearization
Hwi-yeol (Thomas) Yun, Ronald Niebecker, Elin M. Svensson, Mats O. Karlsson
Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden
Objectives: In full model approaches, covariate relations are predefined . Attaching covariate relations selectively to only some of the model parameters can lead to selection bias [2,3]. By allowing all covariates of interest to affect all parameters, this risk of selection bias is mitigated. In the present work we evaluate and compare two full model approaches that both allow estimation of all parameter-covariate relations: a full random effects model (FREM [3,4]) and a full fixed effects model (FFEM) saturated with respect to parameter-covariate relations.
Methods: A semi-mechanistic myelosuppression model with four structural parameters and a dataset containing 636 individuals and 3549 observations was used . In addition, two dummy covariates having correlations of 0.5 and 0.75 respectively with a clinically relevant covariate were generated to investigate the performances of correlated covariates in both models. Linearization to decrease run times during model development and evaluation was assessed for both methods [6,7]. The performance was evaluated in terms of model run times, estimates and precision of parameters and ability to identify clinically relevant covariates. Precision was derived from variance-covariance matrix and bootstraps. FOCE-I with NONMEM 7.2 assisted by PsN was used.
Results: Both FREM and FFEM were successfully implemented, also as linearized models with good agreement and several magnitudes shorter run times. FFEM and FREM were found to be similarly precise. Run times for FREM and FFEM were similar. Both methods identified the same parameter-covariate relationships to be clinically relevant. However, in the case of correlated covariates, only FREM was able to identify all clinically relevant parameter-covariate relations. Furthermore, the coefficients were more precisely estimated compared to FFEM.
Conclusions: Although FREM and FFEM performed equally well in this case with an informative dataset and predominantly uncorrelated covariates, FREM has advantages in comparison with FFEM when investigating correlated covariates. This first combination of linearization and FREM/saturated FFEM appears to be promising and should be further evaluated.
Acknowledgement: This work was supported by the DDMoRe (www.ddmore.eu) project.
 Gastonguay, The AAPS Journal 2004 (6), S1, W4354.
 Ivaturi et al, PAGE 20 (2011) Abstr 2228 [www.page-meeting.org/?abstract=2228]
 Karlsson, PAGE 21 (2012) Abstr 2455[www.page-meeting.org/?abstract=2455]
 Ivaturi et al, WCoP (2012) Abstr WCoP-152
 Kloft et al, Clin Cancer Res. 2006;12(18):5481-90
 Khandelwal et al, The AAPS Journal 2011; 13(3): 464-72
 Svensson et al, PAGE 21 (2012) Abstr 2404 [www.page-meeting.org/?abstract=2404]