The following is the coding for a model I have run:
proc glimmix data=data2b ;
class ndens tank arch;
model out/in = arch ndens days /s dist=binomial link=logit e3
output out=binom_pout predicted(ilink noblup) =p resid(ilink noblup)=r;
estimate "S125" intercept 1 ndens 1 0 0 arch 1 0/ilink cl;
Some items to notice:
* ndens, and arch are fixed categorical variables
* tank is a random categorical variable
* days is a fixed continuous variable
* the random residual statement is included to help capture
overdispersion in the binomial response.
What I am most interested in is this: Is the estimate statement giving
me the predicted value for my first treatment of density and
architecture at the average of the variable "days", and averaged across
all "tanks"? Is that a correct assumption, and if not, how do I change
the code to reflect the "average" conditions.
The reason I ask is, if I use the estimate statement, I get one estimate
of the predicted outcome. If instead, I use the output statement, then
create the summary statistics of the predicted values, I get a
substantially different answer. The value from the estimate statement
seems to reflect the estimate when days=0, not at the mean of days.
There is some unbalancedness within the design, but not so severe I
would expect to see the differences I see.
HOH Fisheries Science Center
5103 Junction Hwy
Mt. Home, TX 78058
Phone 830.866.3356 x214
Learn how you can help Texas State Parks <http://bit.ly/sVdilb>