I have a question with regard to stand error output when using glimmix.
I am trying to run a model on how the No. of veterinarians are determined by animal stocks. Since I have both male vets and female vets for each county, and the explanatory variables are the same for them, I think they are correlated, so I am trying to run a bivariate poisson model here.
I use the R-side covariance structure is used to model the correlations. I find in the output, General Chi-square/DF is 1 if I use this regresion.
My code is like:
First I combine "male vet" and "female vet" into one variable "gender":
length dist $7;
length class $7;
dist = "poisson";
gender = female;
class = "female";
dist = "poisson";
keep gender hog Cattle horse sheep dist class obs;
Then I run glimmix:
proc glimmix data=gender2000;
nloptions maxiter = 100;
class obs class;
model gender = class class*hog class*cattle class*horse class*sheep/
noint s dist=byobs(dist);
random _residual_ / subject=obs type=chol;
This is part of the output I get:
Solutions for Fixed Effects
Effect class Estimate Error DF t Value Pr > |t|
As you can see, several explanatory variables have standard error equal to 0. I have never encountered this problem before. As we always expect low standard error is a good sign, but with almost everything equal to 0 is weird here. Is it because my data? For example, I have the Mean for hog as 85603, with a standard deviation of 258881. I have total 1409 observations in the original data set.
I wonder whether anyone could provide an insight into my problem. Is there a way to correct it? Or is there some materials I can read to better understand the glimmix procedures, e.g, how this std. error is calculated? Thanks a lot!
It could be that you are in fact having numeric precision problems, as you suspect, given the large number of animals and what is probably a small number of vets. You could read up on numeric precision in SAS (you get lots of hits with Google), but it's even easier to just divide your animal count variables by 10,000 and see what happens.
Thanks a lot for your suggestion. I get the following result for the covariance test, yet the coefficient estimation result does not change much. Not sure whether the difference in variance is a problem or not.
Tests of Covariance Parameters
Based on the Residual Pseudo-Likelihood
Label DF -2 Res Log P-Like ChiSq Pr > ChiSq Note
Homogeneity 3 7913.87 0.00 1.0000 DF
DF: P-value based on a chi-square with DF degrees of freedom.
That result says that there is no evidence for a difference in variance between males and females, which to me means that a common estimate is sufficient for the analysis. You may not need the bivariate approach at all.