Accurate vs. approximate likelihood methods in population analysis of sparse data sets

Abstract

Objective Compare the bias, statistical consistency, and statistical efficiency of conventional NONMEM and similar parametric population pharmacokinetic methods using F.O. and F.O.C.E. likelihood approximations to a novel parametric method PEM and to the nonparametric NPAG method, which both use accurate likelihoods, in analyzing sparse data sets often encountered in population modeling. Methods Controlled simulations with a one-compartment, two-parameter (volume of distribution and elimination rate constant) model with Gaussian parameter population distributions were performed with two data points (sparse data) and with one data point (very sparse data) per subject with sample populations ranging from 25 to 1000 subjects. Bias, statistical consistency, and statistical efficiency were evaluated for each method. Results NONMEM and similar methods using F.O. and F.O.C.E approximations showed poor statistical efficiencies (0.9% and 29.0%, respectively) relative to PEM (75.4%) and NPAG (61.4%) for sparse data sets. The approximate methods were significantly biased and the F.O. method was clearly statistically inconsistent, whereas PEM and NPAG were both consistent and unbiased. For very sparse data sets, the approximate methods failed to produce acceptable results, while PEM and NPAG remained viable. Conclusion PEM and NPAG avoid the severe statistical performance degradation observed for NONMEM and similar approximate likelihood methods applied to sparse data sets. Clinical Pharmacology & Therapeutics (2004) 75, P30–P30; doi: 10.1016/j.clpt.2003.11.113