David Radwin <radwin@berkeley.edu>:
Any technique may lead you astray, but the technique you describe is,
I suspect, demonstrably inferior to others. The reference you cite
applies only to estimating the mean or median of an open-ended
category with a Pareto distribution, and its abstract says that "the
choice of method and the selection of a mean or median estimator for
the open-ended category midpoint have substantial effects on analyses
in which income is the dependent variable."
The -intreg- approach proposed by Maarten relies on some shaky
assumptions, but I expect it is far preferable to imputing the
midpoint of intervals and testing for a difference in means as if you
have real data.
I hate ordinal variables like this - and I think it is even worse
when they are independent rather than dependent variables. You don't
like to treat them as continuous, nor do you like to create a bunch
of dummies. Mid-point scoring is popular and easy but still problematic.