Skip to content

On model selection and model evaluation

2 messages · Manuel Spínola, David Hewitt

Original
Manuel Spínola
# 
Dear List,

Thank you very much for your responses.
After looking at the responses I have some comments/questions.
It is my understanding that model selection is one step in data modeling 
and model evaluation need to follow after model selection.  After a 
model has been selected is necessary to evaluate the model and look at 
the parameter estimates, is that correct?  I see that many people think 
that the AIC value and model averaging is the last step in data modeling 
but I am not sure if that is appropriate.  It is also my understanding 
that using IT methods you can select the best worst model of a set of 
bad models, so model evaluation is needed for your selected model.
I wrote "it is my understanding" because I am far for being an expert on 
this.
Any comments on that?
Best,

Manuel
David Hewitt
# 
Of course. You model stuff because you want parameter estimates (effect
sizes). The model selection table goes hand-in-hand with the parameter
estimates and SEs in the final "analysis".

Ben et al.'s recent thread about assessing model fit covers the other part
of your question: evaluating the model.

http://www.nabble.com/glm-model-evaluation-to17525503.html
Model averaging is used to get model-averaged parameter estimates, so folks
are interested in the estimates if they're doing model-averaging.
Ditto on this. David Anderson's recent book (2007) covers all of this in
simpler terms than B&A (2002), so is a good introduction.

-----
David Hewitt
Research Fishery Biologist
USGS Klamath Falls Field Station (USA)