GLIMMIX SAS CODE with different slopes for ordinal outcome

[Gilles-Protais]

Please I wish to know whether there is SAS code, on GLIMMIX to  handle when the proportional odds assumptions is not verified. So I wish to model the ordinal outcome with different slopes, taking into consideration the random effects. 

Thanks

[Gilles-Protais]

Please I wish to know whether there is SAS code, on GLIMMIX to  handle when the proportional odds assumptions is not verified. So I wish to model the ordinal outcome with different slopes, taking into consideration the random effects. 

Thanks

[Ksharp]

It looks like you are doing survival analysis . Do you check PROC PHREG 's RANDOM statement ?

[Gilles-Protais]

Greetings,

Thanks very much for your reply,

Actually am working on a project concerning discrete longitudinal analysis.

And actually there is evidence of different slopes, so I want to have initial values for me run the NLMIXED procedure. 

when there is evidence of equal slopes, e usually do the GLIMMIX procedure to get the initial values.

[sbxkoenk]

You need a "Random Coefficient Model" with random intercepts and random slopes for each "class"?

See this paper:

Paper SAS026-2014
Analyzing Multilevel Models with the GLIMMIX Procedure
Min Zhu, SAS Institute Inc.
https://support.sas.com/resources/papers/proceedings14/SAS026-2014.pdf

This PROC GLIMMIX is on page 11.

proc glimmix data=mctrial;
class patient center treat;
model dbp = dbp1 treat visit / solution;
random intercept visit / subject = center type=chol;
random intercept       / subject = patient(center) ;
covtest 'Diagonal G' DIAGG;
estimate 'Carvedil vs. Atenolol' treat 1 -1 0;
estimate 'Carvedil vs. Nifedipi' treat 0 1 -1;
run;

Koen

[Gilles-Protais]

No , the code below shows a cumulative logit model for  equal slopes across the category of the outcome, But I want a was procedure that can produce me different slopes across different categories; if it is possible

proc glimmix data=longil_cat  method=RMPL;
title 'PROC GLIMMIX analysis, ordinal, RMPL';
class id sex(ref="0") smoking(ref="0");
model bmi_cat = time smoking **bleep**e sex time***bleep**e time*sex/ dist=multinomial link=cumlogit solution;
/*random intercept time/ subject=id type=un;*/
random intercept/ subject=id type=un;
/*nloptions maxiter=5000;*/
run;

[Ksharp]

PROC LOGISTIC has UNEQUALSLOPES option in the MODEL statement. Check the example Example 79.18: Partial Proportional Odds Model in its documentation.

[Gilles-Protais]

But it does not take into consideration the random effect

[sbxkoenk]

Hello,

Not sure if UNEQUALSLOPES can be done in a multinomial logit model that is fit with PROC GLIMMIX (because there are mixed effects).

I cannot see that covered in:

Usage Note 22871: Types of logistic (or logit) models that can be fit using SAS®
(Date Modified: 2018-09-18 11:31:00)
https://support.sas.com/kb/22/871.html

6 times the string "unequal" is there on that page, but not when GLIMMIX is talked about.

Last update of above usage note was done in September 2018, so maybe @StatDave has more info??

BR, Koen

[StatDave · Accepted solution]

See the Addendum at the bottom of this note which shows how you can fit a nonproportional odds model to an ordinal response in PROC NLMIXED. You could add a random effect in the model if needed.

[Gilles-Protais]

Thanks for your reply,  I have tried the code, below and I have very strange results like very high standard errors, what might be the case? because I sude glimmiw for equal slopes and the results look attractive

proc nlmixed data=longil_cat noad  qpoints=30;
title 'BMICAT, PROC NLMIXED, ordinal, adaptive, q=30';
parms int1=2.5080 int2=4.5586 d= 0.07258;

eta1 =int1+ b+beta11*time + beta12***bleep**e + beta13*smoking + beta14*sex + beta15*time***bleep**e + beta16*time*sex;
eta2 = int2 + b + beta11*time + beta22***bleep**e + beta13*smoking + beta14*sex + beta25*time***bleep**e + beta26*time*sex;

if bmi_cat=1 then z = (1/(1+exp(-eta1)));
else if bmi_cat=2  then z = (1/(1+exp(-eta1))) - (1/(1+exp(-eta2)));
else z = 1 - (1/(1+exp(-eta2)));
if z > 1e-20 then ll = log(z);
else ll = -1e10;
model bmi_cat ~ general(ll);
random b ~normal(0,d**2) subject= id out=EB;
estimate 'var(d)' d*d;

run;