Re: [Gretl-users] Using a generated dummy in following estimations
On Tue, 8 Apr 2008, Antonello Lobianco wrote:
Hello all.. following the reference manual I generated a dummy
variable for year 1992 (1992=1, 0 elsewhere) with command:
genr tpD92 = (obs=1992)
The dummy is created without problems and I can see it.. however when
I try to use the new serie in estimations (hsk) I receive an error
"Not numerical value in a calculation".. why??
There's nothing wrong with the dummy variable itself, but I see
what's happening.
When we do hsk we start by taking the residuals from OLS
estimation, squaring them, and taking logs. If the model contains
an observation-specific dummy variable, the residual at that
observation will be exactly zero, yielding a zero square and a log
of minus infinity.
I'm not sure what to do about that. The idea behind hsk is that
we do WLS with weights based on the inverse of the fitted error
variance, but this breaks down if some of the residuals are
by construction equal to zero. In particular, if a dummy variable
is included for a single observation, this is equivalent to
dropping that observation, not giving it a large weight.
Anyway, expect some sort of fix soon.
Allin Cottrell.