08-22-2016 02:10 PM
Of the three polytomous item IRT models PROC IRT claims to estimate, it appears that only one (the Graded Response Model) actually works. Recognizing that the procedure is still experimental, there seems no way to even specify the Generalized Partial Credit Model, contrary to information in the documentation. The Nominal Response Model provides some basic preliminary output, including a notice that the convergence criterion has been satisfied, and occasionally the following warning:
WARNING: Standard errors for some parameter estimates are close to zero.
However, fit statistics and parameter estimates are never produced.
Has this been your experience, too?
08-22-2016 02:35 PM
The IRT procedure is production in SAS/STAT 13.2 and 14.1. It was experimental in 13.1. What version of SAS are you running? Is it possible that you are viewing documentation that is for a later release than you are running?
08-23-2016 10:48 AM
Yes indeed, I must apologize. There was a mismatch between the software I had installed (SAS 9.4, SAS/STAT 13.1) and the newer (SAS 9.4, SAS/STAT 14.1) documentation. I have since updated the software to 14.1. After re-checking the 14.1 documentation, I still don't see a MODEL RESFUNC option for the Generalized Partial Credit Model, although this model is mentioned several other places in the documentation. Though not specifically documented, RESFUNC=GPC works to obtain the Generalized Partial Credit Model in PROC IRT 14.1. As far as the Nominal Response Model is concerned, specifying RESFUNC=NOMINAL in 14.1 (correctly) returns an error in the log, rather than truncated output as in 13.1.