turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- PROC GLIMMIX error message

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

01-12-2016 01:55 PM

**I am trying to estimate a pooled random effects proportion. I get an error message I have never encountered before: **

NOTE: At the starting values the G matrix is the zero matrix. Adaptively determining the number of quadrature points is not meaningful. The QMIN=1 value is chosen. You can provide starting values for the covariance parameters with

the PARMS statement.

**It also says: (which I understand better)**

NOTE: Convergence criterion (ABSGCONV=0.00001) satisfied.

NOTE: Estimated G matrix is not positive definite.

**The confidence interval obtained is really, really, really wide. **

**The code I used is: **

**PROC** **GLIMMIX** DATA=IPDALL2 METHOD=QUAD;

CLASS COUNTRY;

MODEL NUM/DENOM= /DIST=BIN SOLUTION CL;

RANDOM INT /SUBJECT=COUNTRY;

**RUN**;

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

01-12-2016 04:11 PM

That is not an error message but a statement about how the procedure behaved. It is likely indicating that method=quad was not appropriate for the specific data used in this case.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

01-13-2016 08:41 AM - edited 01-13-2016 08:44 AM

@ballardw's response is right on. Method=quad isn't appropriate--I would guess that the number of countries being fit or at least with very different proportions is small. You could try METHOD=LAPLACE (which is what QUAD with one point is doing anyway), but the whole thing looks as if there is insufficient variability due to country for that to work well (that's the meaning of NOTE: Estimated G matrix is not positive definite.), and it is why the confidence interval is so large

If that doesn't help, you'll probably have to go to a marginal model, using the default pseudo-likelihood method.

Steve Denham