Simulation studies on the estimation of total area under the curve in the presence of right-tailed censoring

Abstract

The effect of extrapolated area (%AUCextrap) on estimating mean AUCinf in a simulated single-dose clinical trial was examined. Concentration-time (C-t) profiles from 12 to 36 subjects for 1- and 2-compartment models after single dose administration were simulated with increasing right-tailed censoring. Each subject's %AUCextrap and AUCinf was calculated using eight different methods, including noncompartmental analysis (NCA), population-based methods, and maximum likelihood (ML) accounting for censoring. Each method's geometric mean AUCinf and percent relative error (PRE) from the true AUCinf was calculated. This was repeated 100 times and the mean PRE (MPRE) was calculated. Mean %AUCextrap ranged from 1 to ~30 % for the 1-compartment and 2 to 32 % for the 2-compartment model at the lowest and highest degree of censoring, respectively. NCA methods using all subjects to estimate the population mean AUCinf had similar or less bias (within ± 20 %) than when those subjects with >20 % %AUCextrap were removed. Using Cpred compared to Clast in the calculation of individual AUCinf resulted in no performance improvement. Linear mixed effects models to estimate λz and ML methods accounting for censoring resulted in either no improvement or increased bias when censoring was high. Population pharmacokinetic method bias was dependent on the nature of the C-t profile. When the C-t profile declined biphasically, population models had higher bias than NCA methods but were superior when the C-t profile decline in a log-linear manner. It is recommended that subjects with high %AUCextrap should not be removed from the estimation of mean AUCinf in NCA analyses.