The Chicken and the Egg in Interoccasion Variability
Tarjinder Sahota (1), Nuria Buil (2), Katsutoshi Hara (2), Oscar Della Pasqua (1,2)
(1) Division of Pharmacology, LACDR, Leiden University, Leiden, The Netherlands; (2) Clinical Pharmacology Modelling & Simulation, GlaxoSmithKline, Stockley Park, UK; (3) Clinical Pharmacology, GlaxoSmithKline, Tokyo, Japan
Objectives: The inclusion of interoccasion variability (IOV) on model parameters has been justified previously [1]. However, it is unclear at which step to include it during the model building process when covariates are being evaluated.
One school of thought emphasises the inclusion of fixed effects before introducing random effects. This view favours exploration of covariates before inclusion of IOV. This ignores the possibility of model misspecification with fixed effects. The opposing view is that random effects should come first and covariate explored only after IOV has been identified. We aim to use pharmacokinetic (PK) modelling of an oncology drug as an example to assess the difference between the two approaches in terms of model performance.
Methods: An integrated dataset on PK, demographic and treatment covariates was used from a variety of phase I and II studies. The basic model was obtained by including between-subject variability only. Stepwise covariate selection (SCM) was performed in PSN v.2.3.2 using forward addition (p<0.05) and backward deletion (p<0.01). IOV was explored on all parameters with occasion as varying between a) steady-state/non-steady-state, b) 5 day intervals and c) daily sampling. Model building was performed in NONMEM 6.2 using FOCEI estimation. Two different approaches were implemented. First a model was built by applying the SCM to the basic model and then incorporating IOV (Model A). Secondly, a model was built by applying IOV to the basic model and then performing the SCM (Model B). Model performance was assessed by goodness-of-fit plots, VPCs, NPCs, and PPCs of AUC and Cmax statistics.
Results: The use of IOV for each sampling day on bioavailability gave the largest drop in objective function for model A and B. However model A showed the large bias in population predictions. Model A identified co-administration of lapatanib and drugs affecting PH balance in the stomach as significant covariates, Model B did not. Posterior predictive checks showed that model A did not predict average exposure in individuals receiving co-medications. Model B correctly predicted these statistics.
Conclusions: Incorporation of IOV after exploration of significant covariate relationships may produce greater bias in final model predictions. Non-significant covariates may also be selected due to a biased fixed effects structure. Conventional wisdom should favour IOV being included before optional covariate relationships.
References:
[1] Karlsson MO, Sheiner LB. The importance of modeling interoccasion variability in population pharmacokinetic analyses. J Pharmacokinet Biopharm 1993; 21(6):735-50.