topic Re: Problem with my glimmix model in Statistical Procedures

Problem with my glimmix model

yahoo0806 — Thu, 23 Aug 2018 18:41:50 GMT

Hi all,

I recent run into a problem when I use proc glimmix, hope someone can help.

the data I have is a cross over data, each subject can have treatment sequence as AB or BA, my output is a binary responder: depending on if the change from baseline (CHG) value is >= 3 or not. For example, if the CHG value is >= 3 then responder variable = Y otherwise response = N.

The response status (responder varaible) will be analyzed using a generalized linear mixed model for binomial data (or mixed effects logistic regression model) with the logit of the probability of response as the dependent variable, treatment and period as fixed effects and subject as random effect. The Kenward-Rogers adjustment will be used to compute the denominator degrees of freedom for the test of the fixed effect. Adaptive Gauss-Hermite quadrature will be used to approximate the marginal log likelihood (SAS procedure GLIMMIX and METHOD=QUAD).

Below is the model I used:

Proc glimmix data=model METHOD=QUAD;
class subject period treatment;
model responder = period treatment/ddfm=KENWARDROGER solution link=logit;
random subject;
lsmeans treatment / cl ilink;
run;

The problem I run into is SAS gives me the following error message:

ERROR: For the multinomial distribution only the following links are possible: CUMLOGIT, CUMPROBIT, CUMLOGLOG, CUMCLL, GLOGIT.

Do you know why? and can you please let me know how I should fix my model?

Many thanks in advance!

Re: Problem with my glimmix model

Reeza — Thu, 23 Aug 2018 18:47:18 GMT

I think SAS see's the "responder" variable as 'multinomial', ie it has more than two responses for some reason. Run a PROC FREQ and check the results.

proc freq data=model;
table responder / missing;
run;

@yahoo0806 wrote:

Hi all,

I recent run into a problem when I use proc glimmix, hope someone can help.

the data I have is a cross over data, each subject can have treatment sequence as AB or BA, my output is a binary responder: depending on if the change from baseline (CHG) value is >= 3 or not. For example, if the CHG value is >= 3 then responder variable = Y otherwise response = N.

The response status (responder varaible) will be analyzed using a generalized linear mixed model for binomial data (or mixed effects logistic regression model) with the logit of the probability of response as the dependent variable, treatment and period as fixed effects and subject as random effect. The Kenward-Rogers adjustment will be used to compute the denominator degrees of freedom for the test of the fixed effect. Adaptive Gauss-Hermite quadrature will be used to approximate the marginal log likelihood (SAS procedure GLIMMIX and METHOD=QUAD).

Below is the model I used:

Proc glimmix data=model METHOD=QUAD;
class subject period treatment;
model responder = period treatment/ddfm=KENWARDROGER solution link=logit;
random subject;
lsmeans treatment / cl ilink;
run;

The problem I run into is SAS gives me the following error message:

ERROR: For the multinomial distribution only the following links are possible: CUMLOGIT, CUMPROBIT, CUMLOGLOG, CUMCLL, GLOGIT.

Do you know why? and can you please let me know how I should fix my model?

Many thanks in advance!

Re: Problem with my glimmix model

yahoo0806 — Thu, 23 Aug 2018 19:07:54 GMT

I did the proc freq as you suggested, but I do see only two results: Y and N.

Re: Problem with my glimmix model

Reeza — Thu, 23 Aug 2018 20:13:28 GMT

Do you have a format applied to the variable?

@yahoo0806 wrote:

I did the proc freq as you suggested, but I do see only two results: Y and N.

Re: Problem with my glimmix model

yahoo0806 — Thu, 23 Aug 2018 20:15:38 GMT

I don't think so, here is how I created the responder variable:

data model; set data;
if CHG >=3 then responder = "Y"; else responder = "N";
run;