[Gretl-users] Re: Hessian with negative/missing diagonal elements ?
Am 23.09.2024 um 23:13 schrieb Alecos Papadopoulos:
I was discussing with a colleague an empirical application, and he
told me that the only problem he had was that some elements of the
Hessian was "nan" because they came up with the wrong sign. But the R
code that he used did print out the Hessian, with these elements as
"nan".
Can we do that in gretl (natively)? I ask because in standard
mle-implementation with the option --hessian we may get "Hessian is
not negative-definite, dropping back to OPG".
If you're in a middle of an
algorithm that uses the Hessian as a
numerical input, I don't see how you can (or should) ignore that this
input is not what you think it is. Just like you wouldn't want to
standardize your data with a standard error estimate that somehow came
out as negative, or would you?
What I ask is whether we can instruct gretl to stick with the Hessian,
returning some "nan" values in the diagonal / std errors.
Not that I'm aware of. Regarding the cited "R code" of your
colleague,
maybe it has a clever solution, maybe it just prints out interim stuff
and stops there, or maybe it does very unwise things; we have no way of
knowing.
cheers
sven