Royal Statistical Society
Statistical Computing Section Meetings
The Potential of High Performance Computing (HPC) for Statistical Modelling
in the Social Sciences (Half day meeting)
This meeting will review the current and potential role of HPC and the GRID
(a virtual supercomputer) for statistics in the social science community.
The vast majority of quantitative researchers in the social sciences use
statistical packages, such as SPSS and SAS as a tool for performing their
analyses. This in itself can have a limiting effect on the type of analysis
performed. Speakers from a range of backgrounds will discuss some of the
various options currently available in modern computing environments that
enable a more realistic treatment of the scientific issues that arise in
the analysis of observational data on complex socio-economic processes.
R. Crouchley and R. Fligelstone ( University of Lancaster
)
"The relevance of HPC for the analysis of observational data"
Synopsis: Observational social science data sets are relatively small, but the intricacies of human behaviour create a complex set of interdependencies between the variables of a data set. We illustrate how the complex and comprehensive nature of the models that reflect these intricacies take our analysis beyond what can be usefully estimated on a PC.
R. Ecochard ( University of Lyon
)
D.G. Clayton ( University of Cambridge
)
"Fitting complex random effect models with standard software using data
augmentation"
Synopsis: Clayton and Rasbash (1999) have proposed an alternating imputation-posterior (AIP) algorithm that provides an easy way of dealing with large and complicated crossed-hierarchy designs. The AIP algorithm may be implemented in two distinct ways: first by running the steps sequentially and secondly as parallel tasks with multi-processor architecture or on separate computers in network. We illustrate the approaches on an example of artificial insemination by donor.
J. A. Doornik ( Nuffield College, Oxford
)
"HPC in Econometrics using Distributed Ox"
Synopsis: In a recent paper presented to the Royal Society, Doornick, Hendry and Shephard argued the need for HPC in econometrics and statistics. We identified a shortage of easy-to-use statistical software for HPC, and provided a solution based on the Ox matrix programming language. We now extend this work and illustrate Ox for the simulation of distributions for outlier detection procedures in generalized autoregressive conditional heteroscedasticity (GARCH) models, and simulation-based estimation of stochastic volatility models.
J. Rasbash ( Institute of Education, London
)
M. Bull ( Edinburgh Parallel Computing Centre
)
"Parallel implementation of a multilevel modelling software package"
Synopsis: This talk describes a portable parallel implementation of the IGLS algorithm used by the software package MLwiN for fitting random effects models. Particular attention is paid to crossed and multiple membership random effects models, which pose significant computational challenges for the IGLS algorithm which is optimised for nested random effects. Comparisons of performance between the sequential and parallel approaches across a range of shared memory parallel architectures are given.
The slides for this talk are available in pdf format.
The talks will be followed by a period of general discussion.
Date & Time
Wednesday 27 March at 1:30 - 5:30 pm
Place
Errol Street
(nearest UG stations, Barbican, Moorgate & Old Street)
Refreshments
There will be tea and biscuits served at 3:00 - 3:30 pm.
