s-news
[Top] [All Lists]

Choosing family link and variance in GAMs

To: <s-news@lists.biostat.wustl.edu>
Subject: Choosing family link and variance in GAMs
From: "Marc-Andre Villard" <villarm@UMoncton.CA>
Date: Tue, 11 Dec 2001 11:38:42 -0400
Dear S-Plus community,

I need some advice on the use of link functions in GAMs.  I am trying to model 
dispersal distances (in meters).  The data distribution has a high peak at low 
numbers and then tapers off at high numbers.

What are is the procedure to choose the best link and variance functions for 
such data?  Do I need to stick to canonical link functions and if not, then how 
can I know if certain combinations of link functions and variances actually 
mean something biologically?  For example, my data seem well modelled using 
family=quasi, link=log, variance="mu^2".  Does this estimate a certain 
distribution that has a name?  Do I need to try every combinations of link and 
variance?

I am using S-Plus 4.5 Professional Edition, release 2, on Windows 95.

Thanks for your help,

Marc-André Villard


<Prev in Thread] Current Thread [Next in Thread>
  • Choosing family link and variance in GAMs, Marc-Andre Villard <=