Classification table for ordinal logistic regression

[ammarhm]

Hi everyone, 

I have a dataset similar to the one below:

DATA HAVE;
   input ID	Age	Sex	DM	IHD	Cholesterol	Outcome;
   DATALINES;
1	49	1	0	0	0	0
2	20	1	0	1	1	0
3	23	1	0	0	0	1
4	56	1	1	1	1	2
5	39	0	1	1	1	0
6	57	1	1	1	1	2
7	28	0	0	0	0	0
9	34	0	1	0	0	0
10	45	0	0	0	0	0
11	48	1	0	0	0	0
14	28	1	1	1	1	2
15	41	1	0	0	0	0
16	26	1	0	1	1	0
18	59	0	1	0	1	0
19	51	0	0	0	0	0
20	49	1	1	0	0	0
21	50	0	0	0	0	0
24	56	1	0	1	0	0
25	51	1	1	1	1	2
26	57	0	1	0	0	0
27	32	1	0	0	0	0
28	33	0	0	0	0	0
29	36	0	0	0	0	0
30	54	1	0	0	0	0
32	41	0	0	0	0	0
33	65	1	0	0	0	0
;

I am running a logistic regression with the dependant variable 'outcome'. As this variable has three ordered levels, this is essentially an ordinal logistic regression. I am also trying to get the classification table using ctable as below:

proc logistic data=have;
class Sex	DM	IHD	Cholesterol	Outcome/param=glm;
model Outcome=age  Sex	DM	IHD	Cholesterol	Outcome/CLPARM=wald ctable;

run;

The attached data is only a small dataset so the estimates are nothing to worry about, but the classification table is not generated and a note is generated:

NOTE: Since there are more than 2 response levels, the following options have no effect --
      CTABLE.

Not sure to solve this issue. I would appreciate very much any suggestion on how to obtain the classification table.

Best regards. 

[SteveDenham]

Since CTABLE is dependent on the cutpoint, things get difficult for a multinomial response variable.  I suppose you could get the predicted probability for a case to be in each of the levels for a vector of cutpoints, and then present the results, but the post-processing code may not be simple.  Perhaps someone out there has written a macro or even a datastep to do this.  I think it would need the PREDPROBS= option in the OUTPUT statement, and a variety of cutpoint values for the various levels.  You would have to then look at a table of correctly classified/incorrectly classified vs level.  Seems that getting out sensitivity and specificity from a 2xK table would be difficult if not impossible as they may not be defined, but at least you would have a table.

SteveDenham

[StatDave · Accepted solution]

If you just want to obtain a predicted by actual classification table (sometimes called a "confusion" matrix), that is discussed in this note. If you want the area under the ROC curve (AUC) as a measure of model performance, a multinomial version of the usual AUC for binary response models is available with the MultAUC macro.