[Abnms] Open question: justifying Bayesian methods over other techniques

Ban, Stephen stephen.ban at jcu.edu.au
Thu Mar 10 16:56:30 AEDT 2011


The list seems to be rather quiet (either that, or I've been missing the e-mails), so I'd like to pose a question that should spark some discussion.

In the context of a relatively straightforward model (e.g., predicting a discrete outcome from several putative predictor variables), what reason(s) would you give someone sceptical or unfamiliar with Bayesian methods to use them over other statistical techniques, particularly logistic regression? I've looked at a few papers that have actually done head-to-head comparisons with logistic regression, but thus far I've basically just found that the main drawback of logistic regression is the requirement that you assume linearity in the logit and additivity, which makes logistic regression poorly suited to large datasets. Any reviews or other papers on this subject would be much appreciated.

Secondly, there's the issue of programs like Netica/Genie and discretization - it would seem that the requirement to discretize variables would put a BBN at a disadvantage compared to a logistic regression where you can preserve the continuous nature of your predictor variables. Obviously, this is more of a software/interface limitation than a problem with BBNs generally, but how would you address this limitation (if it is indeed a limitation)?

Cheers,
Stephen


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://abnms.org/pipermail/abnms_abnms.org/attachments/20110309/1a339d6d/attachment-0003.html>


More information about the ABNMS mailing list