Proc Glimmix multinomial model zero standard error

FabioMC — Wed, 21 Sep 2016 08:25:19 GMT

Dear SAS support,

I tried to fit the following multinomial model:

proc glimmix data=data;
freq count;
class gender class_var ;
model class_var = gender age var1 var2

/link=glogit ddfm=kr;
random int / group=class_var;
run;

The model converges but parameter estimates are near to zero (0, without decimals) and their standard errors are 0 for var1 and var2. Moreover degrees of freedom of parameter estimates are extremely high or are 0. Please note that freq variable is very high (from 10000 to 500000), (because class_var are blood white cells populations and count is their count in blood analysis).

Number of observations is 35 subjects (in the "long" form, data are 35*5, given 5 levels of class_var).

I read on the internet some advices about rescaling the freq variable but rescaling (dividing counts by 10,100,1000...) has not effect on the problem, moreover for the other continuous variable of the model (age) there is not this problem (altough standard error is very low).

I tried some model modification, varying degrees of freedom and random statement syntax but this did not solve the problem. I tried and analogous model using genmod but I got zero both parameter estimates and standard errors; I found no collinearity in regressors.

I performed model with only one regressor using one of variables var1, var2 but this did dot change the "zero" situation.

Thank You for help,

Fabio

Re: Proc Glimmix multinomial model zero standard error

howarder — Wed, 21 Sep 2016 15:17:35 GMT

One thing I noticed about your code is that you do not have a 'dist=multinomial' option on the model statement. Not sure if you just missed it when including your code or not. I hope that helps.

Re: Proc Glimmix multinomial model zero standard error

FabioMC — Thu, 22 Sep 2016 10:32:20 GMT

Thank You for the answer but now I have found the solution (also after inserting dist=multi the problem remained without rescaling regressor).

It was a matter of "mental confusion" (and I am sorry for this) :my mind keep concentrating on response variable frequencies and I was not aware that I had a regressor with high numbers .

Regards,

Fabio

topic Re: Proc Glimmix multinomial model zero standard error in Statistical Procedures

Proc Glimmix multinomial model zero standard error

Re: Proc Glimmix multinomial model zero standard error

Re: Proc Glimmix multinomial model zero standard error