New Bayesian Methods For Model Comparison
A one-day ESRC Masterclass. Professor Murray Aitkin, Department of Mathematics and Statistics, University of Melbourne.
The course will provide a description and discussion of new methods for model comparison, of nested or non-nested models, based on the posterior distributions of the likelihood for each model and the likelihood ratio between them.
The methods resolve the difficulties of Bayes factor definition and computation, by replacing one-point summaries of the likelihood (integrated, maximized or penalized likelihoods) by the full posterior distribution of the likelihood under each model. The preference for one model over another is then expressed in terms of the stochastic ordering of their likelihood distributions.
Computationally, this is done through the comparison of the deviance distributions, where the model deviance is -2 log likelihood; these are much better behaved in their small- and large-sample posterior distributions. The new approach provides a general solution to model comparisons of all kinds. It gives new solutions to nested model comparisons, providing simple Bayesian equivalents of the t-test and other standard frequentist tests.
Computations are straightforward, requiring only standard posterior distributions for the model parameters under each model being compared. Standard MCMC output for complex models can be used directly to compute the posterior distributions for the model deviances and compare them.
Cost :
Postgraduate Students - £30
Academic Staff from HE Institutions - £60
Participants from non-HE Institutions - £220
The cost of the course includes course materials, lunch, tea and coffee.
Further details and registration:
http://shortcourses.maths.lancs.ac.uk/new_bayesian_methods_for_model_comparison
The course is sponsored by two nodes of the ESRC National Centre for Research Methods – the Lancaster-Warwick-Stirling node in Statistical Modelling and the BIAS node in Bayesian Methods at Imperial College.
