No, I would not recommend that, because there is no sense of convergence
with this matrix. In the standard implementation, the matrix starts as
an identity matrix, and is then updated as the iterations progress. So
it's final value depends on the starting values, while the converged
parameters do not depend on the starting values (in a well behaved
problem). In an extreme case, start at the optimum: no iterations are
required, and the quasi Hessian stays at the identity matrix, which does
not provide a reasonable estimate of the variance.
On the other hand, the BFGS matrix does tend to lead to a more robust
maximization process than the OPG version. The reason is that it does
provide a better measure of the curvature, as the iterations proceed,
than the OPG matrix does. However, referring back to the first point,
there is no control (if there would be, I think you could use it).
Jurgen.
Mogens Fosgerau wrote:
> I wonder if the quasi-Hessian output from the MaxBFGS routine can be used
> as the variance of scores/outer product of the gradient in the
> sandwich/Huber/White variance matrix estimator?
>
> Thanks
>
--
oxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxox
oxoxoxox 3rd OxMetrics user conference August 2005
oxoxoxox Cass Business School, London
oxoxoxox Dates still to be announced
oxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxox
Dr Jurgen A Doornik
Nuffield College, Oxford OX1 1NF, UK
tel. UK: +44-1865-278610 fax +44-1865-278621
http://www.doornik.com
http://www.oxmetrics.net
oxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxoxox
|