s-news
[Top] [All Lists]

magnitude of response variable and AIC

To: <s-news@lists.biostat.wustl.edu>
Subject: magnitude of response variable and AIC
From: "Hughie Broders" <n127g@unb.ca>
Date: Sun, 25 Aug 2002 15:06:20 -0300
Hi all,
I am trying to model activity data against weather variables in S-PLUS 2000 using AIC as a model selection criterion....
 
The range of the response variable is 0 - 850,000 and is poisson-like in distribution.  I am using a poisson family GLM to do this.  The AIC values of these models are huge ( in the millions).  Therefore what tends to happen when you are ranking different models via the Akaike weights is that the best model is ranked as one (of course), and because the actual magnitude of the difference between the first and second best is in the thousands it gets an Akaike weight of essentially 0 and therefore all other candidate models become ranked as tied for second best.  This seemed odd..
I experimented with transforming the response variable by dividing by 100,000.  What I am finding is that the values of the AIC are now in the order to 100 or so and the ranking the models via Akaike weights appears to be ok.
HOWEVER, when I rank the models using both approaches from lowest to highest AIC values they are not the same-- The AIC selected best models were different, WHY??.
 
Can anyone provide insight into why this occurs.
Thanks in advance 
<Prev in Thread] Current Thread [Next in Thread>
  • magnitude of response variable and AIC, Hughie Broders <=