This question is about why I am getting strange results after running a
Hausman-McFadden test for the IIA assumption after multinomial logit
(-mlogit-) estimation.
I am estimating a McFadden choice model using -mlogit- with 8 choices and
about 1 million observations. The model is fully saturated (all my
right-hand-side variables are mutually-exclusive dummy variables). To test
the IIA assumption, I have run a Hausman-McFadden test:
1. I first conduct the estimation with all 8 choices, then save the
coefficient estimates (using -hausman, save-).
2. I re-estimate the model, but removing from the dataset all observations
that make a particular choice (it doesn't matter which one I remove). I
then test the equality of the coefficients versus the results of #1 above
(using -hausman, alleqs less-).
I am puzzled because all the coefficients from #2 are EXACTLY the same as
the coefficients from #1 above. (So of course, the chi-squared statistic on
the test is zero, with a p-value of 1.00.)
I am hesitant to interpret this result as saying that my model is PERFECTLY
specified. Rather, I suspect that some characteristic of my set-up (perhaps
the fact that the independent variables are mutually-exclusive dummies) is
generating this result, so that the test has no power in this situation.
Does anyone have any insight into what is going on?
Thanks in advance,
Dean
P.S. Sorry for sending this out twice. I originally sent it out over a
holiday weekend and wanted to make sure everyone on Statalist got a chance
to see the posting.
--------------------------------------------------------------------
Dean Yang
Ph.D. Candidate, Department of Economics
Littauer Center, Room 200
Harvard University
Cambridge, MA 02138
