Influence Of Estimation Of Inter-Occasion Variability On Detection Of Time-Varying Covariates

Authors: Wu H, Hirankarn S, Fiedler-Kelly J
Conference: AAPS
Keywords: auto induction, covariates, inter occasion, IOV, methods, NONMEM, PE%, pharmacokinetic, PK, PKPD methodology, simulated data, simulation, stochastic, time varying, time-dependent
Division: Cognigen

Abstract

Purpose: To explore the influence of the estimation of inter-occasion variability (IOV) on the ability to detect time-varying covariates influencing PK parameters from a population PK (PPK) analysis using NONMEM.

Methods: Based on a published PPK model developed using data from a phase III clinical trial of a drug which exhibits enzyme auto-induction, 10 replicated clinical trial data sets, each consisting of 430 patients and 18 measurements per patient following repeated oral dosing over 9 weeks were generated using stochastic simulation. The PPK model is a one-compartment model with first-order absorption and elimination, with clearance expressed as a function of time (to characterize the induction process) and body weight, and volume expressed as a function of body weight. IOV was introduced on clearance at a moderate level relative to IIV (IOV CV%/IIV CV%: 0.83). The simulation model was fitted to the simulated data sets using two different approaches: estimating or ignoring IOV in clearance. Backward elimination of covariate effects was then performed on each dataset. The final models achieved from the different approaches were compared with respect to the bias and precision of the parameter estimates. The detection of the time-varying covariate was judged by its statistical significance during backward elimination.

Results: Except parameters associated with volume, bias of fixed effect parameters from the analyses estimating IOV was generally improved over the analyses where IOV was ignored by up to 1 % decrease in mean percentage error (-0.7 ~ 6.1% versus -1.6 ~7.2%). As expected, precision of random effect parameters in IOV models was improved compared to models ignoring IOV by up to 7% decrease in mean absolute percentage error (4.8~11.3% versus 9.6~17.9%). The time-varying covariate, time on therapy, was a statistically significant predictor (P < 0.0001) in all datasets regardless of whether IOV was included.

Conclusions: As expected, incorporation of IOV generally warrants better estimation of parameters in the cases studied here. In addition, in the presence of IOV with a moderate magnitude relative to IIV, the estimation of IOV does not preclude the ability to detect a time-varying covariate.

American Association of Pharmaceutical Scientists (AAPS); San Diego, California; November 2007

By Huali Wu, Sarapee Hirankarn, Jill Fiedler-Kelly