Re: [Gretl-users] ordered probit bug (was Re: Copy of Data File)
On Wed, 6 Aug 2008, Allin Cottrell wrote:
Anyway, you have exposed a bug: something has gone wrong with
gretl's ordered probit...
...there's a residual problem due to the fact that your
dependent variable has a gap in its sequence of values.
Sorry, I didn't understand this at first, but in fact the entire
problem was due to this gap. The ordered probit code in gretl is
predicated on the assumption that the values of the dependent
variable are drawn from, and exhaust, a continuous set of integer
levels or rankings, usually {1,2,3,...,n} or {0,1,2,3,..,n}.
The levels don't have to be zero-based, but there can be no
skipped values.
The problem is that we haven't actually enforced this requirement
up till now, and all hell breaks loose when it's not satisfied.
For the moment I've updated CVS and the Windows snapshot with a
check for that condition (along with a reasonably informative
error message if it's not met). When Jack gets back we can decide
if we want to go to the trouble of relaxing this condition.
Allin.