[Limdep Nlogit List] MNL Dummy Variable issues

[Thomas C. Eagle]

You cannot have a constant attribute or variable across all alternatives in a
choice set.  That is causing your error.

Tom Eagle

-----Original Message-----
From: limdep-bounces at limdep.itls.usyd.edu.au
[mailto:limdep-bounces at limdep.itls.usyd.edu.au] On Behalf Of contactemt
Sent: Thursday, October 12, 2006 9:36 AM
To: limdep at limdep.itls.usyd.edu.au
Subject: [Limdep Nlogit List] MNL Dummy Variable issues

Hi,
 
I have a discrete choice model with no universal choice set and a variable
number of choices.

I code it thus:

NLOGIT ; Lhs = CHOICE, SETSIZE
    ; Rhs = v
    ; Prob = probs $

Where v is the set of attributes common across all choices ( may be several
dozen).

This seems to work fine, with each attribute returned as a variable parameter.

Now I add a further "category type" variable coded as a dummy (1,0) or (1,-1)
The variable is constant across an individual set  of alternatives (it describes
a category)
but does vary across choices as one would expect. 

I thought the above model could handle this type of attribute without
adjustment.

However, when running the enhanced model I get:

"
Hessian is not positive definite at start values.
  Error   803: Hessian is not positive definite at start values.
B0 is too far from solution for Newton method.
Switching to BFGS as a better solution method.
"

And several of the variables are now returned as fixed parameters.
Also, if I change the value of just one element in a choice set from 1 to -1 (or
vice versa) the model runs 
as before, so clearly the fact that the dummy variable does not change within a
choice set causes the 
issue.

Is this to be expected and what substantive effects does this have on my model?

I may have, in the full dataset, several dummy variables which describe
categories for the choice set and 
be constant across a choice set as above. I wanted to produce a "global model"
and combine these 
categories into one set of data. By doing so, using a larger dataset, I hoped to
increase the accuracy of 
the parameter esitimates and so the model when classifying new data. 

So my questions are:

Can I just plough ahead and accept that some/many of the variables will be fixed
parameters in my general 
model or should I run each category as a separate model? - the danger here would
be the datasets, given 
the number of attributes may not be large enough.
Is there a transformation I could/should apply to the category(s) which will
allow the "global model" to 
run "better".

Thanks
_______________________________________________
Limdep site list
Limdep at limdep.itls.usyd.edu.au
http://limdep.itls.usyd.edu.au