This takes me back to my original example of a bunch of separate
probits, but this time contrasted with, e.g., the default behavior of
-gologit2- (with link(pr) to make it a generalized ordered probit). The
latter imposes only the constraint that the cutoffs are ordered,
corresponding to order-ability of the outcome variable. How would you
test this constraint?
Actually, an unconstrained gologit doesn't even impose that
constraint. Indeed, Clogg argued that the model isn't ordinal (or at
least need not be ordinal). You can even rearrange the categories of
Y and still get about the same fit! Since unconstrained gologit has
as many parameters as mlogit, this actually isn't that surprising.