Re: st: RE: nonsense category in ordered logit

[Richard Williams <Richard.A.Williams.5@nd.edu>]

At 04:52 PM 2/21/2005 -0500, you wrote:
> That was my intuition, too, but I was a little fuzzy on the test to use.
> I couldn't convince myself that the models were nested
> thanks,
> michael
Long and Freese (2003, p. 203) describe how to test whether categories in a 
multinomial logistic regression can be combined.  For example, if y has 
categories 1, 2 and 3 you can do something like

mlogit y x1 x2 x3, basecategory(1)
test [2 = 3]

If the test is insignificant, categories 2 and 3 can be combined.

Better yet, if you have Long and Freese's -spostado- installed (and 
everyone should!) give the command

mlogtest, combine lrcomb

and it will do wald and likelihood ratio tests of all possible combinations 
of categories.

Of course, you are doing ordinal regression, so I don't know if these tests 
would be considered legit for that, but it seems like a place to start.

As for Nick's suggestion:


> By showing that -logit- with two and -ologit-
> with three give essentially equivalent results?
That was my first impulse too - but how exactly would you do that?  You are 
comparing results from a 2 category DV with a 3 category DV, so what 
exactly would you compare?  Is it legit to compare model fit, BICs, or 
whatever?  I am not sure.  Perhaps you can compute the predicted scores 
from each model and correlate them?  When you talk about nested models you 
are generally talking about testing constraints on the Xs, are you not?  I 
don't believe these models are nested so I'm not sure what the formal 
statistical test would be.  I suspect Nick is right but I am not sure how 
to do it.

*
*   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/