[R] AIC and model selection; not a R question

[Ben Bolker]

Lynnette Dagenais wrote:
> 
> Hi,
>  
> I was wondering if someone could help me answer a question that is bound
> to come up in my Master's defense. I'm using AIC to select models and my
> question is how do I know that the models I developed a priori contain the
> 'best' models in the system. How do I not know that some models which I
> didn't include aren't actually the 'best' model??
>  
> 

Answer: you can't (or probably not).

While there are some goodness-of-fit tests against generic alternatives,
most model selection approaches 
test only among a set of specified alternatives.

  On the other hand, if you have discrete data and a standard sampling
distribution
(binomial, Poisson, etc.), and the residuals from your most complex model
are _not_
significantly overdispersed, then you have some evidence that everything not
explained by your most complex model is due to (irreducible) sampling
variation.
(If you do have overdispersion that doesn't mean that your model is a bad
fit,
though.  Most ecological data, at least, is overdispersed.)

  good luck,
    Ben Bolker

-- 
View this message in context: http://www.nabble.com/AIC-and-model-selection--not-a-R-question-tf4867273.html#a13929620
Sent from the R help mailing list archive at Nabble.com.