[Limdep Nlogit List] Best practice for ordered models

[Bruce Rayton]

Hello fellow Limdep users,

I wonder if any of you would be willing to share your views about what
constitutes "best practice" for dealing with ordered models relative to
continuous models. Specifically:

"How many categories must a dependent variable have before you are willing
to regard the variable as approximately continuous?"

This may seem trivial, but we do (and I think should do) this all the time
with monetary values (e.g., really measured discretely in cents), but life
is not always so clear. Imagine you have a survey question which is measured
with five ordered responses (say, strongly disagree to strongly agree). This
sounds like a classic place for application of an ordered estimator to me.
What if there are seven categories? Twelve? Twenty? Is there a line to be
drawn? If so, how should it be set? 

I'd particularly welcome academic references that address this question.

Many thanks,  Bruce