[Limdep Nlogit List] HEV model results

[William Greene]

M. Huybers.
In the output you provided, the estimated scale parameters for the first
group (all equal) are found not to be significantly different from 1.0,
not 0.0. It sounds like you did not not note this in your comment below.
the reason for the "different significances" is that the first is a test
that the parameter estimated is different from 1.0 while the second is
a test that pi/(theta*sqr(6)) is different from zero. They are different.
On your second question, given your specific model specification, if you 
divide the first set of utility functions by the appropriate scale 
parameter, what remains will be an MNL model.
/Bill Greene

----- Original Message -----
From: "Twan Huybers" <T.Huybers at adfa.edu.au>
To: limdep at limdep.itls.usyd.edu.au
Sent: Wednesday, June 10, 2009 12:04:51 AM GMT -05:00 Columbia
Subject: [Limdep Nlogit List] HEV model results

Dear All,

 

I have two queries regarding the estimation results of an HEV model.

 

1) I have run an HEV model which contains eight generic alternatives
with various attributes.  The dataset is a stacked one with the first
four alternatives coming from one sample and the second four from
another sample.  The IVs are set to be equal for the first four, and
equal (and normalised) for the second four.  The estimated scale
parameters themselves are not significant but the implied standard
deviations of the random components are significant - see output below.
In assessing the departure from MNL, which ones should prevail, the
scale parameters or the implied standard deviations?  And how come these
two sets of parameters have different significances in the first place?

2) My second query concerns the calculation of HEV choice probabilities.
My understanding has always been that these are not in a closed form
and, therefore, have to be obtained using simulated maximum likelihood.
However, I have now come across a ''parameterised heteroskedastic MNL
model' whose choice probabilities seem, simply, to be equivalent to the
MNL ones after duly multiplying by each alternative's estimated scale
parameter.  Does this mean that I can use my HEV scale parameters to
"adjust" each alternatives utility value and, hence, probability?

 

Thanks for any comments/suggestions.

 

Regards,

Twan Huybers