Prediction using Proc GLIMMIX

[Lijuu]

Dear Committee:

Greetings!

I have the following data:

My response variable is current status which has ordinal value (cured=0, improved=1, same=2, deteriorate=3 and death=4).

Proc glimmix data=lij.oncosort pconv=1e-4 method=laplace;
class Patient_Id cancer current_status BG anemia_at_dx Treatment Stage_grade;
model current_status= age anemia_at_dx Treatment Stage_grade sv / solution link=logit ;
random cancer;
covtest GLM;
run;

Then after I have two questions:

1. How can I make prediction probability for my response variable, current status?

2. How can I make the ROC curve plot to assess the number of correctly predicted current_status?

3. How can I match the predicted probability with the observed current status(0, 1, 2, 3 and 4)?

Thank you much in advance!

Lijuu

[Rick_SAS]

If your response variable is ordinal, you can't use the logit link.  use CUMLOGIT, CUMLOGLOG, or GLOGIT.

1. Use the OUTPUT statement.

2. The concept of an "ROC curve" for multinomial responses is not clear. Several researchers have proposed different methods. One method is Hand and Till (2001). SAS provides a macro for this method, but I haven't used it myself.

3. Not sure what you mean by "match." The predicted category is usually chosen to be the category that has the largest predicted probability.

[Lijuu]

Thank you Dear Rick!