Date: Fri, 29 Aug 2008 06:40:02 -0700
Reply-To: Shawn Haskell <shawn.haskell@STATE.VT.US>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Shawn Haskell <shawn.haskell@STATE.VT.US>
Organization: http://groups.google.com
Subject: Re: Random coefficients and GLIMMIX code
Content-Type: text/plain; charset=windows-1252
On Aug 28, 1:45 pm, geraldine.mabi...@UQAR.QC.CA (Geraldine Mabille)
wrote:
> Dear all,
> I’m a biologist working on GPS habitat use data collected on mooses
> (Animal) during three consecutive years (yr). I want to compare attributes
> of the points used by animals (1) with attributes of random available
> points (0). My response variable is a binary variable (U_A) and I use proc
> GLIMMIX to determine which factors of the habitat (slope, altitude,
> habitat type : 4 types : food, shelter, food_shelter, others, coded as 3
> dummy variables) explain whether a point is a used or a random point. One
> question I specifically want to adress is whether different animals show
> different selection for food habitats as the availability of the food
> habitat increase in their home range. I use random coefficient models of
> that type :
>
> proc glimmix data=mydata;
> class animal yr;
> model u_a= slope altitude food shelter food_shelter/ dist=binary solution;
> random intercept food/subject=animal(yr) type=un solution;
> run;
>
> I then use the random coefficients obtained for food and plot them against
> the availability of the food habitat for each animal(yr).
> Here is for the context! Now, I have several questions regarding my SAS
> code
> Q) Is it correct to code my 4 habitat types with three dummy variables and
> include those three dummy variables into the model without including them
> into the class statement?
> Q) From what I understood, the most logical variance structure to use in
> random coefficients models is the type=un structure. Can anyone confirm
> and may be justify that? Is it worth trying other covariance structures
> for my question?
> Q) For each animal(yr), I have thousands of points collected and my data
> are therefore repeated by animal(yr). I wanted to take that into account
> by adding a « repeated » statement in my model
>
> proc glimmix data=mydata;
> class animal yr;
> model u_a= slope altitude food shelter food_shelter/ dist=binary solution;
> random intercept food/subject=animal(yr) type=un solution;
> random _residual_/subject=animal(yr) type=cs;
> run;
>
> I wanted to use a cs covariance structure because I expect all points to
> be equally correlated for a given animal(yr). When I run this model with
> type=vc, the processing time is very short and I obtain an estimate for
> residual(vc). However, if I try to run this model with type=cs, the
> processing time is very long (it has been already running for 6 hours and
> is still running right now!). Is my code correct? Is the type=cs
> appropriate or should I try other covariance structures?
>
> Thanks heaps if anyone can help. I’m quite a beginner with statistics and
> will really appreciate any input.
> Geraldine Mabille
Geraldine, If you haven't already, I would recommend you read Manly et
al.'s Resource Selection by Animals: Statistical Design and Analysis
for Field Studies - i think you'll find some SAS code in it (PROC
LOGISTIC perhaps), and you might check the WEST Inc website too. Also
see special section on resource selection in Journal of Wildlife
Management 2006 vol 70 no. 2. You may not get all the specifics you
are looking for but you'll find a great foundation. Otherwise, try
Googling "SAS resource selection function" - many folks have studied
habitat selection of various critters with GPS, satellite, and GIS
data - surely some have used SAS - maybe you can post your findings
for the benfit of others. good luck. SH
| 
