07-14-2014 08:14 PM
Hi, I have normally distributed data that is spatially correlated. When I put in correction code for spatial correlation the chi square/DF value becomes very large showing that it is over dispersed. Is there a correction I can do for this? Also, how do I refer to this code for spatially correlated data in my methods? The code I have is:
PROC GLIMMIX data=WORK.Ch_2_tree_inverts;
CLASS Site Treatment Block;
MODEL pests = Treatment
/ dist=gaussian solution oddsratio;
random residual / type=sp(exp) (Lat Long) subject=Block;
07-15-2014 09:24 AM
I think that if you change the random statement just slightly, it should accommodate overdispersion. Try:
random _residual_ / type=sp(exp)(lat long) subject=block;
Note the use of the underscores on residual when placed before the slash.
Now on to other questions. You list Site in the CLASS statement. This leads me to suspect that there are blocks at multiple sites, and that you have not yet included the random effect of site. To make sure the overdispersion statement works, you must have unique values for blocks. By this I mean that if you have block=1 at site=1 and also have block=1 at site=2, you are going to run into problems. I would suggest, just to be on the safe side, that you change the subject= option to:
Another question, and this one just reflects my not knowing the field very well, has to do with the oddsratio option. What does this represent in a model with gaussian errors? Since you have an identity link, the best that will happen if you include this is a NOTE or WARNING in the log. At worst, it will stop the procedure. The oddsratio option only applies if there is a logit link (could be cumulative or generalized logit).