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 noint & random effect

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

01-11-2012 05:14 AM

Dear all

I'm using this code

**(1)PROC** **GLIMMIX** DATA=input plots= all;

class hosp mese;

model y= x_1 x_2 x_3 mese /dist=poisson link=log s ;

output out=gmxout predicted=pred resid=res;

random intercept /subject=hosp ;

**run**;

I've a client request to run the same model without intercept.

How can I do it? I would like a model with no intercept but with the random effect of hosp variable.

I tried with

**(2) PROC** **GLIMMIX** DATA=input plots= all;

class hosp mese;

model y= x_1 x_2 x_3 mese /dist=poisson link=log s noint;

output out=gmxout predicted=pred resid=res;

random intercept /subject=hosp ;

**run**;

but I think the option noint and statement

random intercept /subject=hosp ;

could be in contrast (however model (1) and (2) give the same output)

I tried also with this code but I'm not sure that the model is what I want.

**(3) PROC** **GLIMMIX** DATA=input plots= all;

class hosp mese;

model y= x_1 x_2 x_3 mese /dist=poisson link=log s noint;

output out=gmxout predicted=pred resid=res;

random _residual_ /subject=hosp ;

**run**;

Thank in advance for any help

Kind Regards

Accepted Solutions

Solution

01-11-2012
09:07 AM

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

01-11-2012 09:07 AM

Many thanks Steve for your help.

And what about model in the third code?

**(3) PROC** **GLIMMIX** DATA=input plots= all;

class hosp mese;

model y= x_1 x_2 x_3 mese /dist=poisson link=log s noint;

output out=gmxout predicted=pred resid=res;

random _residual_ /subject=hosp ;

**run**;

What kind of model I'm running?

Thanks in advance

Have a nice time

All Replies

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

01-11-2012 08:54 AM

I am somewhat surprised by the request to remove "intercept" but still have the random effect of "hosp." So far as I have ever been able to tell, the following RANDOM statements are equivalent:

RANDOM intercept/subject=hosp;

RANDOM hosp;

The only advantage is that the first is somewhat more stable algorithmically, and is absolutely needed in PROC GLIMMIX for METHOD= LAPLACE or METHOD=QUAD.

I would ask client some questions regarding why they want to fit things differently, and in particular, do they wish to change the inference space to the narrower space of only those hosp observed?

Steve Denham

Solution

01-11-2012
09:07 AM

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

01-11-2012 09:07 AM

Many thanks Steve for your help.

And what about model in the third code?

**(3) PROC** **GLIMMIX** DATA=input plots= all;

class hosp mese;

model y= x_1 x_2 x_3 mese /dist=poisson link=log s noint;

output out=gmxout predicted=pred resid=res;

random _residual_ /subject=hosp ;

**run**;

What kind of model I'm running?

Thanks in advance

Have a nice time

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

01-11-2012 10:28 AM

Your model 2 is the one to use, if you really do not want a fixed effects intercept. As stated by Steve, there are multiple equivalent ways of writing out the random effect terms using GLIMMIX or MIXED. Model 3 is different. You are basically fitting a marginal effects model with overdispersion, not a conditional effects model (as in model 2). That is, fitting model 3 adjusts the standard errors in a multiplicative sense for variability greater than expected for a Poisson distribution.

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

01-11-2012 10:51 AM

Many thanks

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

01-11-2012 10:42 AM

That is a repeated measures model of sorts. I would expect it to differ slightly, as the denominator degrees of freedom may change to reflect the clustering in the subject factor. Looking at some data I had on hand, the repeated measures approach looked way underdispersed overdispersed (as per lvm's note above) and had a substantially larger -2 Res , Log Pseudo-Likelihood. That indicates to me that the model with RANDOM intercept/subject=animal (NB. for MY data) is a better model.

Steve Denham

Message was edited by: Steve Denham