Kim Seongho, Hall Stephen D, Li Lang
Department of Medicine, Division of Biostatistics, School of Medicine, Indiana University, Indianapolis, Indiana 46032, USA.
J Biopharm Stat. 2009 Jul;19(4):700-20. doi: 10.1080/10543400902964159.
In this paper, various Bayesian Monte Carlo Markov chain (MCMC) methods and the proposed algorithm, the Gibbs maximum a posteriori (GMAP) algorithm, are compared for implementing the nonlinear mixed-effects model in pharmacokinetics (PK) studies. An intravenous two-compartmental PK model is adopted to fit the PK data from the midazolam (MDZ) studies, which recruited twenty-four individuals with nine different time points per subject. The three-stage hierarchical nonlinear mixed model is constructed. Data analysis and model performance comparisons show that GMAP converges the fastest and provides reliable results. At the mean time, data augmentation (DA) methods are used for the Random-walk Metropolis method. Data analysis shows that the speed of the convergence of Random-walk Metropolis can be improved by DA, but all of them are not as fast as GMAP. The performance of GMAP and various MCMC algorithms are compared through Midazolam data analysis and simulation.
在本文中，对各种贝叶斯蒙特卡罗马尔可夫链（MCMC）方法以及所提出的算法——吉布斯最大后验概率（GMAP）算法进行了比较，以用于在药代动力学（PK）研究中实现非线性混合效应模型。采用静脉注射二室PK模型来拟合来自咪达唑仑（MDZ）研究的PK数据，该研究招募了24名个体，每位受试者有9个不同的时间点。构建了三阶段分层非线性混合模型。数据分析和模型性能比较表明，GMAP收敛最快且能提供可靠结果。同时，对随机游走 metropolis 方法使用了数据增广（DA）方法。数据分析表明，DA可以提高随机游走 metropolis 的收敛速度，但它们都不如GMAP快。通过咪达唑仑数据分析和模拟比较了GMAP和各种MCMC算法的性能。
