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
- /
- model spatial correlation using GLIMMIX in poisson...

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

01-14-2013 05:18 PM

Hi!

I want to use a poisson regression for my data with adjusting for spatial autocorrelation.

I have tried to do it by using GLIMMIX

proc glimmix data=sasdata.all;

nloptions maxiter=500;

model ABUND=logP logN/ dist=poisson link=log solution;

random _residual_/type=sp(exp)(lat_dd lon_dd) subject=intercept;

run;

There is a warning message: "Obtaining minimum variance quadratic unbiased estimates as starting values for the covariance parameters failed."

What should I try to make this work?

Thank you!

Feng Zhang

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

01-15-2013 07:22 AM

Without seeing the data, I can't be sure if this will even be helpful. I note that you have subject=intercept in the random statement. Unless there is a variable called "intercept" in sasdata.all, this error could occur. Of course, if the variable weren't there, then there should be other error messages, and the proc would error out before the "Obtaining..." message ever showed.

So, I am very curious as to what the subject really is. Suppose it is called locate (not that it is, but for examples sake only). You might try:

random intercept/residual type=sp(exp)(lat_dd lon_dd) subject=locate;

which would give a random intercept model with a spatial correlation between the subjects. You might also consider using different techniques for optimization. The default is QUANEW, so you might examine NRRIDG or NEWRAP. Another possibility is to include the INITGLM option in the proc glimmix statement.

But what I think is happening is that your data is such that there just isn't a feasible solution. You may have to "creep up on" the proper covariance structure, starting with something remotely applicable that is simpler and describes things in an approximate manner.

Steve Denham

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

01-15-2013 09:20 AM

The subject=intercept option is OK. It is the MIXED and GLIMMIX convention to indicate that all the data are in one big subject, meaning that all observations can be correlated. You probably want to experiment with the PARMS statement, to explicitly give initial guess for the parameters.The procedure will use these instead of the ones it is trying, and failing, to obtain.

By using the RANDOM _RESIDUAL_, you are doing so-called R-side analyses, which means that you no longer have a true Poisson conditional distribution. The residual variance is being adjusted for overdispersion. You could try so-called G-side analysis:

proc glimmix data=sasdata.all;

nloptions maxiter=500;

class ID;

model ABUND=logP logN/ dist=poisson link=log solution;

random ID / type=sp(exp)(lat_dd lon_dd) subject=intercept;

run;

Note: You need an ID for each observation (a unique value for each), and make this a class variable. This model and code comes directly out of the great new book by Walter Stroup (Generalized Linear Mixed Models [CRC Press]). Interpretation is slightly different than R-side analyses (here, you have a true conditional Poisson distribution).

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

01-15-2013 10:30 AM

Thank you for you your help. Actually I want to include over-dispersion factor into my model because the variance of the data is much bigger than the mean.

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

01-15-2013 10:32 AM

If the variance is greater than the mean, you may wish to consider type=negbin, to get a negative binomial distribution.

Steve Denham

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

01-16-2013 07:29 AM

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

01-16-2013 10:53 AM

Thank you for your help. I will try The G-side analysis with negative binomial.