[Limdep Nlogit List] MNL Dummy Variable issues

contactemt contactemt at bigfoot.com
Sat Oct 14 21:37:36 EST 2006




> You have to have one alternative have a parameter of zero for the 
> constant.  If
> you do not then you have a vector of ones across all alternatives in the 
> choice
> set which leads to a linear dependency.  This is critical in choice 
> modeling.  I
> say choose on carefully only to make you think about the ramifications of
> setting the alternative you choose to have a constant parameter equal to 
> zero.

I really don't understand this. The alternatives are generic. What 
difference does it make whichever parameter I choose?

I think I am missing something conceptually here. I know nothing substantive
about the alternatives just that one is chosen- and remember there are
variable numbers of alternatives for each choice set.
The only distinguishing thing about the alternatives in a choice set is
which one was chosen.
The alternatives are not adjustable they are hard coded and unique to an 
individual choice set, the order they are presented in any choice set is 
arbitrary as they are generic, and the attributes are a "black box" - at 
least at this stage.

I am trying to do a sort of unsupervised learning I suppose.

> If you do not think predictability will be improved with the pRF1*gender
> interaction then do not do it.....

Why not try it and see?

 I am pretty sure that gender (or whichever category) is relevent, can help 
increase the accuracy
(predictability) so I want to include it.


> I really think you better get a handle on what the terms of you model mean 
> in
> terms of interpreting the utility function.  Personally explanation is 
> more
> important to me than predictability.  Why build the model if you are not 
> trying
> to explain something.  Why add gender if you do not think there is a 
> difference
> between genders.  Build you model based upon theory!
>

Because I am trying to predict something :)
I understand that having a grasp of the utility function may be beneficial
in producing a better, more accurate model through some sort of iterative 
process.
But I was hoping that I can include categorical variables naively without
having to make arbitrary decisions about the model in the first place.

The original "simplistic" model I gave in my initial post allows me to do
this. I need to know nothing about the type or number of alternatives in any
choice set or make decisions regarding the attribute relationships.(Of
course I have had to make some decisions in data preparation).
But it doesn't allow me to include dummy variables constant across a choice
set.

If you could give me a model that can predict 100%  accurately and "explain"
0 then I would be very happy. The whole point of the model is prediction.

If I can include a (several) dummy variables and the predictive capacity is 
increased I am very happy. If it isn't I just disregard them - or redefine 
them and have another go.



More information about the Limdep mailing list