st: Some questions on IRT using GLLAMM

[Marcello Pagano <pagano@hsph.harvard.edu>]

Prathiba Natesan wrote:

Dear Statalist:

   I have been working on GLLAMM for  some research projects . I find 
the program very handy and useful  particularly when it comes to Item 
Response Theory.

However, I have some specific questions about applying GLLAMM to IRT. I 
would appreciate it if someone could get back to me with answers to 
these questions. I am trying to estimate the difficulty, discrimination 
and threshold values for a polytomous dataset. I have run the gllamm 
with item specific thresholds and now have the parameters. I need some 
assistance in identifying the parameters. I have _cut11 for d2-d6 and 
cons,..., _cut14 for d2-d6 and cons and the level 2 coefficients for 
random effect 1.  Here are my specific questions:

1. There should be only 3 thresholds (for 4 categories). So how do I 
tell which of the parameters are the threshold, difficulty and 
discrimination parameters? 
2. I constrained the standard deviation of the latent variable using the 
constraint "def 1 [id1_1]d1=1" syntax and then estimated the oprobit 
model  again with the constraint and the frload option. Are the level 2 
coefficients I have now, the discrimination parameters?

On a different note:

1. For a polytomous dataset, what causes the estimates to not converge? 
Can something be done about it? Does it say something about the data?

2. When the program says, flat or discontinuous region encountered, I 
understand that there is trouble in evaluating the integral due to this. 
But why does this happen? Or rather, what do we do in such situations?

3. This is a very basic question. When I have certain factors (from the 
factor analysis), should I perform GLLAMM for each factor (GLLAMM on 
items belonging to this factor) individually? I am doing this because I 
think that it would be
easier to tackle unidimensional models than multidimensional models. 
Will there be any problems in doing this (I am thinking analogous to 
performing many ANOVAs instead of one MANOVA which inflates the type-I 
error rate)?


Any advice would be of great help. Thanks

Prathiba
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/