[Gretl-users] Re: MLE estimation N/A std errors even when based on Outer product matrix
On Thu, 26 Mar 2020, Alecos Papadopoulos wrote:
Gretl 2020a windows 64bit.
Consider the following output from mle estimation (pay attention to the
Gradient values and the NA std errors):
[...]
I have a question, and a suspicion about what the answer might be.
My question is: since the Std errors are based on the Outer Products matrix,
why are they not computed/presented ? Never mind whether this is a successful
estimation attempt -it is not. My question is "mechanical". I would
understand "NA" if I was asking for std errors based on the inverted Hessian.
But where is the problem in computing std errors based on the gradient?
Your suspicion, I suspect, is just slightly off the mark. The OPG matrix
is in fact calculated as G'G, banking on the idea that sumc(G) should be 0
at convergence (note: differently from your message, here I'm using G for
the n x k matrix of gradients per observation). The name you gave to the
coefficients seem to suggest that you have a variance parameter whose
estimate is numerically 0. In those cases, it wouldn't surprise me if the
G'G matrix (which is positive semidefinite by construction) was
numerically singular; upon inversion, God knows what could happen. For all
we know, it could be full of negative elements.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------