tag:blogger.com,1999:blog-7959705296201073323.post9152516529042974685..comments2023-02-24T11:14:00.053+01:00Comments on The Bayesian kitchen: Model averaging versus model selectionAnonymoushttp://www.blogger.com/profile/09710797049914216414noreply@blogger.comBlogger3125tag:blogger.com,1999:blog-7959705296201073323.post-57134555979814382042014-10-06T22:22:55.658+02:002014-10-06T22:22:55.658+02:00We often do model selection in a frequentist frame...We often do model selection in a frequentist framework, using a greedy step-up procedure with AIC as a stopping criterion. This means we never have to run the most general (complex) versions of the model.<br /><br />An example would be where one is using an n-category general discrete distribution to approximate a continuous reality, with computation time depending on n. We start with n=1 and increment it until model comparison shows no further improvement.<br />Konradhttps://www.blogger.com/profile/06867375994008638278noreply@blogger.comtag:blogger.com,1999:blog-7959705296201073323.post-72897146189311580562014-10-05T08:28:39.906+02:002014-10-05T08:28:39.906+02:00I can see some contexts where model selection is p...I can see some contexts where model selection is perhaps cheaper than model averaging. <br /><br />But in many, if not most, cases, what I see in fact is the opposite: people go through a laborious series of Bayes factors, whereas they could in fact use the most general model, thus implicitly and rapidly averaging over all submodels.<br /><br />one could argue that the most general model is also the one that is computationally the most challenging. But doing the Bayes factor analysis implies that this model will have to be conditioned on the data anyway.<br /><br />So, I am not sure that the cost-saving approximation is really the main reason.<br />Anonymoushttps://www.blogger.com/profile/09710797049914216414noreply@blogger.comtag:blogger.com,1999:blog-7959705296201073323.post-73680161802778732052014-10-04T20:45:59.858+02:002014-10-04T20:45:59.858+02:00Nice post. I would add that, in practice, computat...Nice post. I would add that, in practice, computational considerations matter when making these decisions. In many contexts we can think of model selection as a cost-saving approximation to model averaging, and the question of interest becomes where we can cut corners without sacrificing statistical performance.<br />Konradhttps://www.blogger.com/profile/06867375994008638278noreply@blogger.com