Practical considerations for using the full random effects modeling (FREM) approach to covariate modeling
Gunnar Yngman (1,2), Joakim Nyberg (1), E. Niclas Jonsson (1) and Mats O. Karlsson (2)
(1) Pharmetheus AB, Uppsala, Sweden, (2) Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden
Objectives: The FREM approach to covariate modeling has been suggested to avoid issues with standard covariate model building approaches, e.g. selection bias [1][2]. FREM is implemented in PsN [3] and is similar to the full fixed effects modeling (FFEM) approach in that all parameter-covariate relationships are estimated simultaneously but has advantages over FFEM, e.g. with correlated covariates [4]. After estimation FREM also allows conditional interpretation of the results, i.e. only a subset of covariates (even a single covariate) without having to re-fit the final model. A disadvantage with FREM, though, is an unusual implementation of the covariate model. The aim of this work is to investigate practical usage considerations of FREM including the estimation method(s), specification of different parameterizations of the parameter-covariate relationships and the sensitivity to the additional distributional assumptions required by FREM.
Methods: A simulation study (n = 150) in NONMEM was performed, based on real data and the final parameters of a docetaxel model for neutrophil counts [5]. All 18 covariate-parameter relationships were re-estimated with FFEM and FREM using IMPMAP and FOCE. Bias, precision, termination and run-time were evaluated. Implementation details of parameterizations of covariate-parameter relations in FREM were investigated. Both transformation of covariate observations and the FREM model were tested.
Results: IMPMAP was found more stable than FOCE (FREM successful minimization: 100%, 62%). Bias and precision of the re-estimated covariate parameters were similar, as were mean run-times (FOCE 16 cores: 14 min, 30 min; IMPMAP 8 cores: 40 min, 90 min). As opposed to FFEM, FREM allowed both uni- and multivariate interpretation of the covariates and when the same relations as FFEM were selected the results were in close agreement (RMSE FFEM, FREM: 0.0272, 0.0271). While modeling covariates as random effects, FREM provided accurate covariate coefficients also for non-normal covariate distributions. Parameterization corresponding to different parameter-covariate relationships could be implemented as either data or model transformations with the same result.
Conclusions: The investigation gives information to implementation of FREM with respect to practical aspects such as estimation method, parameterizations, expected performance and robustness towards covariate distribution. Further, it supports FREM for unbiased covariate confirmatory analyses.
References:
[1] Karlsson, MO. A Full Model Approach Based on the Covariance Matrix of Parameters and Covariates, PAGE 21 (2012) Abstr 2455 [http://www.page-meeting.org/?abstract=2455]
[2] Ivaturi, VD, Hooker, AC, Karlsson, MO. Selection Bias in Pre-Specified Covariate Models, PAGE 20 (2011) Abstr 2228 [http://www.page-meeting.org/?abstract=2228]
[3] Harling, K, Nordgren, R et al. PsN: Perl-speaks-NONMEM (2017) [https://github.com/UUPharmacometrics/PsN]
[4] Gastonguay, MR. A full model estimation approach for covariate effects: inference based on clinical importance and estimation precision. AAPS J. 2004;6(S1): W4354.
[5] Friberg, LE, Henningsson, A et al. Model of Chemotherapy-Induced Myelosuppression With Parameter Consistency Across Drugs. J. Clin. Oncol. 2002;20(24):4713–21.
