Proceedings of the 3rd International Workshop on Distributed Statistical Computing (DSC 2003) March 20���22, Vienna, Austria ISSN 1609-395X Kurt Hornik, Friedrich Leisch & Achim Zeileis (eds.) http://www.ci.tuwien.ac.at/Conferences/DSC-2003/ JAGS: A Program for Analysis of Bayesian Graphical Models Using Gibbs Sampling Martyn Plummer Abstract JAGS is a program for Bayesian Graphical modelling which aims for com- patibility with classic BUGS. The program could eventually be developed as an R package. This article explains the motivations for this program, briefly describes the architecture and then discusses some ideas for a vectorized form of the BUGS language. 1 BUGS BUGS (Bayesian inference Using Gibbs Sampling) is a program for analyzing Bayesian graphical models via Markov Chain Monte Carlo (MCMC) simulation (Spiegelhal- ter, Thomas, Best, and Lunn (2002b)). It is aimed at applied statisticians who have a problem that does not fit into a standard model class, but who do not wish to write their own custom software. BUGS fills an important role by making the analysis of problems with various sources of ���complexity��� such as random effects, measurement error and missing data accessible to a wide audience. The BUGS project has promoted the Bayesian approach not only through the BUGS program, but also through the accompanying documentation, which includes a large collection of worked examples taken from the literature, and practical advice on model construction and criticism. The development team have also used BUGS as a platform for exploring new ideas in Bayesian modelling. Recently Spiegelhalter, Best, Carlin, and van der Linde (2002a) proposed the ���Deviance information cri- terion���, a Bayesian analogue of the Akaike Information Criterion, which has been implemented in WinBUGS and will most likely become very popular as a result. More speculative ideas, such as ���cuts��� in the graph, which allow only a one-way flow of information in the graph, have also been implemented.

Proceedings of DSC 2003 2 1.1 History of BUGS The original form of BUGS (now referred to as ���classic��� BUGS) had a simple com- mand line interface that could used interactively or from a script. Classic BUGS had a very limited capacity to analyze the output of the Markov chain. The cus- tomary usage was to dump selected output from the chain to a file and then analyze the contents with an output processor, such as the S-PLUS library CODA (Best, Cowles, and Vines (1995)), which is also available as an R package. In 1997, classic BUGS was superseded by WinBUGS. WinBUGS includes a pow- erful GUI which allows a model to be defined by literally drawing the graph on a ���doodle���. The GUI also includes a compound document interface which allows the model, the data and the output to be stored together in a single document. Win- BUGS has built in capabilities for summarizing the output of the chain in graphical or tabular form, and these summaries can be incorporated into a compound docu- ment. WinBUGS is therefore a much more comprehensive package, and has a much reduced need for an output processor. Two capabilities were lost in the transition from classic BUGS to WinBUGS. The first was the ability to run on platforms other than Windows. Classic BUGS was written in MODULA-2 and could be compiled on some versions of Unix using the Gardens Point Modula compiler (http://www.citi.qut.edu.au/projects/). WinBUGS is written in Component Pascal and depends on the Black Box component framework (http://www.oberon.ch) which is only available for Windows. A second capability that was lost in the transition to WinBUGS was the ability to run a model from a script. However, a new scripting language has been introduced in the latest version (WinBUGS 1.4) and solutions have been developed to run WinBUGS remotely from S-PLUS, R, SAS, Matlab and Python (http://www.mrc-bsu.cam.ac.uk/bugs/ winbugs/remote14.shtml). 2 JAGS: Just Another Gibbs Sampler 2.1 Motivation There are two distinct motivations for wanting to clone BUGS. The first is to have the ability to modify the program. This is a typical motivation for free software. As an example, I have been exploring, with Vincenzo Bagnardi of the University of Milan, multi-state Markov models for analyzing panel data. Such models can be analyzed with the R package msm (Jackson, Sharples, Thompson, Duffy, and Couto (in press)), but the attraction of the Bayesian graphical modelling approach is the ability to adapt the analysis to complex study designs. Bayesian analysis of multi-state Markov models has been considered, in an epidemiological context, by Sharples (1993), and Guihenneuc-Jouyaux, Richardson, and Longini Jr (2000), and in the context of actuarial statistics by Bladt, Gonzalez, and Lauritzen (in press). In order to represent such models in the BUGS language, two new distributions are required: ��� A discrete distribution describing the probability of being in state j at time t, given that the subject was in state i at time 0. ��� A survival distribution describing the time to an absorbing state (e.g. death), which is known exactly, or may be right censored.