[R] question about AIC
Prof Brian Ripley
ripley at stats.ox.ac.uk
Tue Nov 16 10:33:54 CET 2004
On Tue, 16 Nov 2004 dlavecchia at tiscali.it wrote:
> we are a pool of phd students and we have a question about AIC. We are interested
> in calculating the AIC for a mixture model on galaxies data. So far we have
> found AIC works only for regression models, whereas we need AIC for a mixture
> of Normal with mean, sd and weights given by our EM algorithm.
What is the problem? The concept AIC or the function AIC()? If the
concept, it was originally developed in the area of AR(p) time series
models, using asymptotic theory for MLEs, and you need to check that the
concepts apply to your situation (the original papers do not).
Note there are some fundamental problems here: the likelihood for a
mixture of normals is unbounded, the MLE corresponding to degenerate
(zero-variance) components. And AIC is about maximized likelihoods, with
maxima in the interior of the space. So you need to take some
professional statistical advice about the concepts.
Brian D. Ripley, ripley at stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595
More information about the R-help