|Date: ||Tue, 3 Apr 2007 07:06:06 -0700|
|Reply-To: ||Shawn Haskell <shawn.haskell@TTU.EDU>|
|Sender: ||"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>|
|From: ||Shawn Haskell <shawn.haskell@TTU.EDU>|
|Subject: ||Re: an easy one: overfitting in proc mixed|
|Content-Type: ||text/plain; charset="iso-8859-1"|
On Apr 2, 1:22 pm, val_har...@hotmail.com wrote:
> Just to be sure of myself...
> There are 1190 observation in my dataset but those observations are
> repeated measure on 17 individuals which was follow using satellite
> telemetry during 7 months. I consider individual as my experimental
> unit (instead of the number of observation) and I used proc mixed (and
> Akaike information criterion) determine the factors (month, sex and
> age class of individual) that best explained the variation in the
> data. To account for repeated measures on the same individual,
> inindividual' was considered as repeat and random factors in the
> How many varaible may I fite, at maximum, in my model to avoid
> overfitting? 3 (which including the interaction)??
> thank you
It is good you are using a mixed model for clustered data. In theory,
the AICc stats tell you when you have overfit - the model with lowest
AICc is the best fit parsimonious model. If the fully reduced
intercept-only model has the lowest AICc then none of your covariates
are much good. Partial p-values should tell about the same story.
Check out Burnham and Anderson (2002. Model selection and multiomodel
inference) or just start with David Anderson's website and pdf files
there. Note: there is a school of thought that says you may use all
useful stats available and not just AIC all the time. good luck.