s-news
[Top] [All Lists]

Re: glm(...,family=poisson) transformation : still trying

To: s-news@lists.biostat.wustl.edu
Subject: Re: glm(...,family=poisson) transformation : still trying
From: Joe Sexton <sharpernail@yahoo.com>
Date: Thu, 13 Feb 2003 14:02:16 -0800 (PST)
Thank you for your responses, Brad, John, and Jose.

As for specifying (...,type='response'), my problem is
that the larger dataset to which I am applying the
regression equation is pragmatically too large to
import into S Plus. (I'm regressing tree percent
canopy cover from about 150 Mb of satellite remotely
sensed imagery.) Therefore, I have to find a way to
manually predict my sample-calibrated equation in an
image-processing software that can't create GLM's on
its own.

As for using the exp() transformation, here are some
results:

> coefficients(dougfir100.x.glm)
                          Value Std. Error    t value 
         (Intercept)  -4.991452   1.176776 -4.2416314
   poly(fall.tc1, 1) -12.739892  15.805608 -0.8060362
  poly(fall.ndvi, 1)   7.341085   6.594062  1.1132872
  poly(sumr.tc1, 2)1   4.707552  20.419682  0.2305399
  poly(sumr.tc1, 2)2   9.380219  10.596375  0.8852290
  poly(sumr.tc2, 2)1   3.081442  20.380504  0.1511956
  poly(sumr.tc2, 2)2  -6.603397  12.827755 -0.5147742
poly(aspen.30.r, 2)1  -2.507069   6.497554 -0.3858481
poly(aspen.30.r, 2)2  -5.316410   4.269693 -1.2451505

Applying the coefficients above to the training
vectors (in S Plus),

>0-4.991452-12.739892*(fall.tc1)+7.341085*(fall.ndvi)+4.707552*(sumr.tc1)+9.380219*(sumr.tc1^2)+3.081442*(sumr.tc2)-6.603397*(sumr.tc2^2)-2.507069*(aspen.30.r)-5.316410*(aspen.30.r^2)

yields these results:

        1        2        3        4... 
 398728.1 350754.5 497222.4 317500.5...

But when the exp() function is applied to transform
the predictions--as Jose and I had thought
appropriate--to (...,family='poisson'), the values are
a bit off:

> exp(x)
   1   2   3   4... 
 Inf Inf Inf Inf...

This is especially interesting when compared to the
values yielded by predict(...,type='response'),
recommended by Brad and John:

> predict.glm(dougfir1.x.glm,contrain,type='response'
1           2           3         4...
0.003677987 0.008057072 0.0102533 0.008796681,

which are perfectly appropriate.

So, the question remains, is S Plus using some
function besides exp() as the transformation? Anyone?

Thanks again,
Joe


=====
Landscape Ecology: Modeling and Analysis Center  
& Department of Forest, Range, and Wildlife Science
Utah State University
     
"Go deeper."
-Zarathustra

__________________________________________________
Do you Yahoo!?
Yahoo! Shopping - Send Flowers for Valentine's Day
http://shopping.yahoo.com

<Prev in Thread] Current Thread [Next in Thread>
  • Re: glm(...,family=poisson) transformation : still trying, Joe Sexton <=