Takayuki Katsube (1, 2), Akash Khandelwal (1), Andrew C Hooker (1), E. Niclas Jonsson (1), Mats O Karlsson (1)
(1) Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden, (2) Clinical Research Department, Shionogi & Co., Ltd., Japan
Objectives: We reported on a stepwise covariate model building combined with cross-validation (XV SCM) [1]. This method is useful to determine suitable model size based on prediction error using multiple data subsets.
Methods: The objective of this study is to further characterize XV SCM and make comparisons of estimated model sizes and predictive performance of developed models with other covariate modeling methods, e.g. standard SCM and lasso.
Results: All model building strategies improved the prospective OFV compared to the base model regardless of data set size and number of covariate relations used in the simulation. The prospective performance, in percent of the true-base model OFV difference, was for the small, medium and large data set: 41%, 92% and 99% (non-linear XV SCM), 52%, 91% and 99% (linearized XV SCM), 55%, 92% and 99% (SCM) and 71%, 91% and 97% (lasso) on average. The number of covariates in the final model relative to the true model was on average -0.5 (non-linear XV SCM), -0.04 (linearized XV SCM), -1.3 (SCM) and 2.4 (lasso). The predictive performance of the XV methods was similar regardless of number of splits.
Conclusions: These results suggested XV SCM provides a suitable model size with good predictive performance except for extremely small data and XV SCM using the fewer splits might be a good compromise.
References:
[1] Katsube T, Khandelwal A, Harling K, Hooker AC, Karlsson MO. Evaluation of Stepwise Covariate Model Building Combined with Cross-Validation. PAGE 2011.
[2] Jonsson EN, Karlsson MO. Automated covariate model building within NONMEM. Pharm Res. 1998. 15:1463-1468.
[3] Khandelwal A, Harling K, Jonsson EN, Hooker AC, Karlsson MO. A fast method for testing covariates in population PK/PD Models. AAPS J. 2011. 13:464-72.
[4] Ribbing J, Nyberg J, Caster O, Jonsson EN. The lasso–a novel method for predictive covariate model building in nonlinear mixed effects models. J Pharmacokinet Pharmacodyn. 2007. 34:485-517.
Reference: PAGE 21 () Abstr 2482 [www.page-meeting.org/?abstract=2482]
Poster: Covariate/Variability Model Building