Main Page

From MCMCpack

MCMCpack is a software package designed to allow users to perform Bayesian inference via Markov chain Monte Carlo (MCMC). More specifically, MCMCpack is a formal R package. R is an extremely powerful language and environment for statistical computation and graphics. The lead developers of the MCMCpack project are Andrew D. Martin (admartin@wustl.edu), Washington University, and Kevin M. Quinn (kevin_quinn@harvard.edu), Harvard University. We assume that users of this package will be familiar with R.

Currently MCMCpack allows the user to perform Bayesian inference via simulation from the posterior distributions of the following models: linear regression (with Gaussian errors), Quinn's dynamic ecological inference model, Wakefield's hierarchial ecological inference model, a probit model, a logistic regression model, a one-dimensional item response theory model, a K-dimensional item response theory model, a robust k-dimensional item response theory model, a Normal theory factor analysis model, a mixed response factor analysis model, an ordinal item response theory model, a Poisson regression, a Poisson changepoint model, a tobit regression, a multinomial logit model, an SVD regression model, and an ordered probit model. The package also contains densities and random number generators for commonly used distributions that are not part of the standard R distribution, a general purpose Metropolis sampling algorithm, functions to compute Bayes factors for some models, a handful of teaching models, and some data visualization tools for ecological inference.

To maximize computational efficiency, the actual sampling for each model is done in compiled C++ using the Scythe Statistical Library. The posterior samples returned by each function are returned as mcmc objects, which can easily be summarized and manipulated by the coda package. Please consult the coda documentation for a comprehensive list of functions that can be used to analyze the posterior sample.

[edit] MCMCpack Development Page

For information about current MCMCpack development go to the MCMCpack Development Page.

[edit] Current Release (September 7, 2009)

We are pleased to release Version 1.0-4 of the MCMCpack package. This is a major release of MCMCpack that contains a dynamic item response theory model as well, some new models for Bayesian time series analysis, and a Bayesian quantile regression function. All models are now coded using Scythe 1.0.2, which whould make estimation substantially faster. It also contains our first Bayesian time series model coded in C++. We are happy to add Jong Hee Park as a full collaborator on the MCMCpack project. This release (mostly) conforms to the developer environment described in the specification and uses the newest version of the Scythe Statistical Library. It also supports parallel computing using suitable random number generators, which are now licensed under the GNU GPL. This site contains both the source code for Linux, Unix, and MacOS X installations. A binary file for Windows will become available on Comprehensive R Archive Network. This version of MCMCpack depends on version 0.11-3 or greater of the coda package, which is available from the Comprehensive R Archive Network, R version 2.8.0 or greater, and gcc 4.0 or greater.

[edit] Installation

Installing MCMCpack from source on a Linux, Unix, or MacOS X workstation is easy. First, download the package source MCMCpack_1.0-4.tar.gz. Then, login as superuser and type:

  R CMD INSTALL MCMCpack_1.0-4.tar.gz

To install MCMCpack on Windows, we recommend downloading the binary file from CRAN.

One can also very easily install the latest version of MCMCpack available at CRAN on any machine (regardless of platform) with a live internet connection. To do this, open R, and then at the R command prompt issue the following command:

  > install.packages("MCMCpack")

[edit] Documentation

Documentation for the package is available as an Adobe PDF file [1].

Slides from a presentation about MCMCpack at useR! 2006 are available [2].

Slides from older presentations about MCMCpack at useR! 2004 are available [3]. Slides from a presentation about MCMCpack at the 21st Annual Meeting of the Political Methodology Society are also available [4], along with demonstration files [5].

[edit] Acknowledgements

We gratefully acknowledge support from the United States National Science Foundation (Grants SES-0350646 and SES-0350613), the Department of Political Science and the Weidenbaum Center at Washington University, and the Department of Government and the The Institute for Quantitative Social Science at Harvard University. Neither the Foundation, Washington University, nor Harvard University bear any responsibility for this software.

The MCMCpack project is now hosting its own CRAN mirror at http://cran.wustl.edu.

We are interested in adding additional models to the library. If you have R, C++, FORTRAN, GAUSS, or other code that you would like to distribute as part of this package, please contact us. Please email Andrew D. Martin] (admartin@wustl.edu) or Kevin M. Quinn (kevin_quinn@harvard.edu) with any comments or questions.