Hi dear collegues in Bayesian methodology,
I am currently working on several applications involving BUGS, and I am
repeatedly faced with the problem of model choice.
I have adopted the following methods:
1) for non embedded model, I use a kind of Akaike criterion,
2) for embedded models, I use the maximum likelihood ratio test
In both cases, I just replace the usual maximum likelihood by its means
computed by BUGS.
I suspect this method is not quite valid, since the result depends of
the prior choice (I use "vague" priors), but I do not know any "provably
better" method. I do not believe in "Bayesian factors", which have also
drawbacks.
Is there any method you think better ?
Thanks for any comments or suggestions.
Jean-Louis Golmard
INSERM U 436
Dépt de Biomathématiques
CHU Pitié-Salpêtrière
91, bd de l'hôpital,
7563 Paris cedex 13
Phone: + 33 1 40 77 98 47
Fax: + 33 1 45 85 15 29
Email: [log in to unmask]
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|