Glimmix Covariance Matrix Confidence Levels

[j***@gmail.com]

I'm trying to predict a binary outcome (participation in interracial
relationships) using clustered data, so it seemed like proc glimmix
would be the way to go. However, it lacks the 'covtest' option, so I
was told that using the 'asycov' option to get the asymptotic
covariance matrix for the covariance parameter estimates, and that I
could use that to construct confidence intervals and confidence levels
for the covariance parameter estimates. Can someone point me to a way
to do that, or explain how? Thanks in advance for the help. SAS code
follows:

proc glimmix data=WORK.sfnomiss abspconv = 0.0000001 asycov;
class region;
model irw1 (event='1') = female immigrant education /s dist=binomial
link=logit;
random intercept /subject=region;
random _residual_;
weight gswgt3;

J

[O***@medizin.uni-halle.de]

Post by j***@gmail.com
I'm trying to predict a binary outcome (participation in interracial
relationships) using clustered data, so it seemed like proc glimmix
would be the way to go.  However, it lacks the 'covtest' option, so I
was told that using the 'asycov' option to get the asymptotic
covariance matrix for the covariance parameter estimates, and that I
could use that to construct confidence intervals and confidence levels
for the covariance parameter estimates.  Can someone point me to a way
to do that, or explain how?  Thanks in advance for the help.  SAS code
proc glimmix data=WORK.sfnomiss abspconv = 0.0000001 asycov;
        class region;
        model irw1 (event='1') = female immigrant education /s dist=binomial
link=logit;
        random intercept /subject=region;
        random _residual_;
        weight gswgt3;
J

Dear J,
the new version of PROC GLIMMIX (Version 9.2) has a COVTEST statement.

If you are still using an older version then the following might help:
To calculate confidence intervals for your covariance parameters you
can use the standard formula for a 95%-Wald confidence interval which
is

Estimate +/- 1.96*SE(Estimate).

"SE(Estimate)", the standard error of the estimate, is just the square
root of the variance of the estimate which is given on the main
diagonal of the asymptotic covariance matrix.
Please be careful not to mix up the different variance terms. What you
want is a confidence interval for the random effects variance and what
you need for it (and what is given in the covariance matrix) is the
variance of its estimation. These two variances are very different
things.
Just aside, "1.96" is the 0.975(=(1-0.05)/2)-quantile from the
standard normal distribution. If you want a 99%-CI, you would use the
0.995-quantile which is 2.576.

Hope that helps,
Oliver

[jjantzer]

Post by O***@medizin.uni-halle.de

Post by j***@gmail.com
I'm trying to predict a binary outcome (participation in interracial
relationships) using clustered data, so it seemed like procglimmix
would be the way to go.  However, it lacks the 'covtest' option, so I
was told that using the 'asycov' option to get the asymptotic
covariancematrix for thecovarianceparameter estimates, and that I
could use that to construct confidence intervals and confidence levels
for thecovarianceparameter estimates.  Can someone point me to a way
to do that, or explain how?  Thanks in advance for the help.  SAS code
procglimmixdata=WORK.sfnomiss abspconv = 0.0000001 asycov;
        class region;
        model irw1 (event='1') = female immigrant education /s dist=binomial
link=logit;
        random intercept /subject=region;
        random _residual_;
        weight gswgt3;
J

Dear J,
the new version of PROCGLIMMIX(Version 9.2) has a COVTEST statement.
To calculate confidence intervals for yourcovarianceparameters you
can use the standard formula for a 95%-Wald confidence interval which
is
Estimate +/- 1.96*SE(Estimate).
"SE(Estimate)", the standard error of the estimate, is just the square
root of the variance of the estimate which is given on the main
diagonal of the asymptoticcovariancematrix.
Please be careful not to mix up the different variance terms. What you
want is a confidence interval for the random effects variance and what
you need for it (and what is given in thecovariancematrix) is the
variance of its estimation. These two variances are very different
things.
Just aside, "1.96" is the 0.975(=(1-0.05)/2)-quantile from the
standard normal distribution. If you want a 99%-CI, you would use the
0.995-quantile which is 2.576.
Hope that helps,
Oliver

That is exactly what I was looking for. Thanks for the help!

J