Nicky Best, Andrew Thomas, David Lunn, Jon Wakefield and David Spiegelhalter
Imperial College, London and *MRC Biostatistics Unit, Cambridge
BUGS is a program that carries out Bayesian inference on complex statistical models for which there may be no exact analytic solution, and for which even standard approximation techniques have difficulties. A Markov chain Monte Carlo (MCMC) approach to numerical integration is used: BUGS version 0.5 (currently freely available by internet) implements the Gibbs sampling algortihm, and permits sampling from
distributions which are either discrete, of standard form or log-concave. This enables BUGS v0.5 to handle a wide range of problems, but it has so far prevented its application to many non-linear models such as those encountered in population pharamcokinetic (PK) and pharmacodynamic (PD) studies, due to the non-log-concave sampling distributions involved.
We are currently extending the BUGS software to handle more general sampling distributions by implementing a Hastings-Metropolis algorithm. In addition, a PC-based user-interface to BUGS is being developed. The general interface – WinBUGS – allows graphical description of the model, includes menu-options for sampling and monitoring quantites of interest, and produces on-line graphics such as trace and kernel density plots. A second interface – PharmaBUGS – is being developed specifically for use in PK/PD modelling. This will provide menu-driven options for model building, covariate analysis and residual diagnostics, and facilitate prediction and analysis of integrated PK/PD systems.
Demonstrations of protype versions of WinBUGS and PharmaBUGS will be given.
Reference: PAGE 6 () Abstr 597 [www.page-meeting.org/?abstract=597]
Poster: oral presentation