Statistics seminar
Liverpool University Statistics Division
Time: 2pm, Wednesday February 24th
Venue: Room 1.05, Maths & Oceanography Building, University of Liverpool
Speaker: Simon Godsill, University of Cambridge
Title: Simulation methods for inference in processes with symmetric stable
innovations
Abstract:
This talk will describe some recent work for parameter estimation and signal
extraction in models where one or more component is from a symmetric stable
distribution. Likelihood and posterior probability-based methods for inference in
the presence of stable components are typically hampered by the unavailability of
analytic expressions for the stable density function. Here I demonstrate a simple
Markov chain Monte Carlo method for performing inference when such components are
present. The method uses a scale mixture of normals (SMiN) representation of
symmetric stable distributions, in which the original model is augmented to
include scale parameters at each data point, in order to facilitate simulation
from parameters and hidden states from their conditional posterior probability
distributions. A rejection sampling method is proposed for sampling
from the augmented scale parameters, and simple schemes based upon asymptotic
tail approximations are described for cases where rejection rates are too high.
Examples are presented from the field of conditionally Gaussian time series
analysis, where these methods are ideal since they retain conditionally the
simple Gaussian structure of the model, but application of the methods to more
general scenarios is stressed.
Other forthcoming seminars:
March 24th (Joint seminar with Applied Mathematics)
David Greenhalgh (University of Strathclyde)
Mathematical modelling of immunization programs for Hepatitis A in Bulgaria
26 May
Helen Wilson (University of Liverpool)
A comparison of a parametric and a non-parametric approach to the assessment
of replicated spatial point patterns
Full seminar programme available via http://www.liv.ac.uk/maths/SOR/
Damian Clancy.
----------------------
[log in to unmask]
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|