Skip to content

compare observed and fitted GAM values

3 messages · Lucía Rueda, David Winsemius, Simon Wood

Original
David Winsemius
#
On Aug 21, 2009, at 5:03 PM, Luc?a Rueda wrote:

            

      
      
      
    

        
So then,  you are complaining because it "works" but _unexpectedly_  
well.

            

      
      
      
    

        
If you have a balanced design in Year and it is the only covariate,  
then that is precisely what should happen with any regression method:

 > newdat <- data.frame(val=rnorm(40), cat=factor(1:4))
 > aggregate(newdat$val, list(newdat$cat), mean)
   Group.1          x
1       1  0.4972092
2       2 -0.1042936
3       3  0.1549305
4       4 -0.1500513
 > lm(val~cat, newdat)

Call:
lm(formula = val ~ cat, data = newdat)

Coefficients:
(Intercept)         cat2         cat3         cat4
      0.4972      -0.6015      -0.3423      -0.6473

 > 0.4972092 + coef(lm(val~cat, newdat))[2:4]
       cat2       cat3       cat4
-0.1042937  0.1549304 -0.1500514

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    
2 days later

    

      
      

        


  

        

      Simon Wood
    

    

      
      #
    

  


  

      
You haven't given quite enough information to be sure, but I would guess that 
this is not really a problem, but rather the interesting proporty of GLMs 
fitted with a canonical link described in e.g. section 2.1.8 of Wood (2006) 
Generalized additive models: and introduction with R, or at the beginning of  
the GLM chapter in Venables and Ripley MASS.

        
On Friday 21 August 2009 22:03, Luc??a Rueda wrote: