MCMC in StRD

Published

October 16, 2008

Author(s)

Hung-Kung Liu, William F. Guthrie, D Malec, Grace L. Yang

Abstract

The numerical inaccuracies caused by floating point arithmetic, although often not important, can change the conclusions of an analysis. Computational accuracy is of increasing concern because the number of software packages has exploded as computers have evolved and statistical software is increasingly written and used by non-statisticians who may not be aware of potential computational problems.To address this problem, SED developed the Statistical Reference Datasets (StRD) web site (http://www.itl.nist.gov/div898/strd/index.html) which provides datasets with certified values for assessing the numerical accuracy of software. Four areas of statistical computation were originally addressed, univariate statistics, linear regression, nonlinear regression, and analysis of variance. Recently Markov chain Monte Carlo (MCMC) has become popular and is a new area in which intensive statistical computations are used. Despite its importance, the numerical accuracy of the software for MCMC is largely unknown. By way of specific datasets, we demonstrate in this paper some of the anomalies in MCMC computations.

Citation

Proceedings of the Joint Statistical Meetings

Pub Type

Journals

Keywords

Bayesian analysis, floating point arithmetic, Markov chain Monte Carlo (MCMC), numerical accuracy, Statistical Reference Datasets (StRD)

Citation

Liu, H. , Guthrie, W. , Malec, D. and Yang, G. (2008), MCMC in StRD, Proceedings of the Joint Statistical Meetings (Accessed September 12, 2024)