[Limdep Nlogit List] Question on socio-demographic characteristics and the MNL model

[Marcelo Baena Moreno]

Dear Oliver,

If your sample is large such as mine (2000 passengers in an airport choice
experiment), you may use "age" as market segmentation criterium. I believe
you can use the mean age to make two segments of same size. Then you compare
the results of all variable coefficients for "older customers" - seniors,
and "younger customers" - juniors.

Marcelo.

2009/10/1 Oliver Ehrlich <oliver.ehrlich at hhl.de>

> Dear all,
>
> I have a question concerning the inclusion of socio-demographic
> characteristics (SDCs) in discrete choice MNL models:
>
> Let's say I want to model 1) the influence of perceived sales channel
> quality (service quality, convenience, relative product price) as stated
> by consumers and 2) the influence of SDC "age" of consumers on sales
> channel choice (store, internet, hotline) with a classical MNL model.
>
> In NLOGIT 4.0, I would use the NLOGIT command with
> lhs=choice;
> choices=store,internet,hotline;
> model:
>
> U(store)=asc_s+par_qual*quality+par_conv*convenience+par_price*price+par_age*age/
>
> U(internet)=asc_i+par_qual*quality+par_conv*convenience+par_price*price+par_age*age/
> U(hotline)=par_qual*quality+par_conv*convenience+par_price*price $
>
> The question for me now is: is there any meaningful interpretation of the
> parameter "par_age" for the SDC "age"? For the channel attributes, the
> interpretation is of course straightforward (e.g., positive parameter
> "par_qual" = the higher service quality in a channel, the higher its
> channel choice probability); but as age is by definition invariant across
> the alternative channels, the interpretation of a positive/negative
> parameter par_age would be "the older/younger a consumer, the higher his
> overall channel choice probability" - which doesn't make too much sense...
>
> I have also tried alternative-specific parameters for the SDC; however,
> the sign of these parameters turned out to be either always positive or
> always negative across alternatives, which leads to the same meaningless
> interpretation as in the generic parameter case.
>
> Against this background, I can't currently think of a setup where a SDC
> would lead to a meaningful interpretation in my model setting... Maybe
> some of you have an idea how SDCs could be included in the model in a
> meaningful way - or if they in general should be left out completely in
> such a model setting.
>
> Thanks in advance for your help,
> regards
>
> Oliver
>
>
> _______________________________________________
> Limdep site list
> Limdep at limdep.itls.usyd.edu.au
> http://limdep.itls.usyd.edu.au
>