s-news
[Top] [All Lists]

GAM modelling

To: <s-news@lists.biostat.wustl.edu>
Subject: GAM modelling
From: "Poh Siew Hua" <ESHPoh@ntu.edu.sg>
Date: Thu, 30 Oct 2003 17:12:51 +0800
Thread-index: AcOexf6ITSh82ujPRZOADg1uJXxgEQ==
Thread-topic: GAM modelling
Hi,
 
I have some queries on S-plus GAM.
 
If you have many covariates for GAM, eg: X1, X2, X3 etc, is it advisable to use 
step.gam to model the individual covariates and then put them together in the 
final model or incorporate more terms into the existing GAM model as you 
progress on?
 
For example, in a mortality analysis, if you model Seasonality first, then you 
go on to model Temperature.   When you are trying to find a suitable span for 
lo(Temperature), 
1. When you are trying to find a suitable span for lo(Temperature), do you 
include the term that you have already found-- lo(Seasonality, span a) into 
your step.gam function? Then if you model for a third term, you include the lo 
terms of Seasonality and Temperature?
2. Or do you find the spans required for each of the lo terms(X1, X2, X3,...) 
separately and then put them together in the final gam model? 

 

Is there any difference between these 2 ways?

 

Prof Ripley had pointed it to me earlier, "AIC is not even defined for a real 
GAM [ for AIC presumes maximum-likelihood fitting (amongst other things)], and 
gam() uses an approximation to a penalized fit."   Will someone tell me how 
useful/efficient is the 'AIC' produced in step.gam(), to determine the most 
suitable span based on AIC?    How do we calculate this 'AIC'?

Thanks.

Rgds,

Siewhua

 

 

<Prev in Thread] Current Thread [Next in Thread>
  • GAM modelling, Poh Siew Hua <=