1) The current implementation is buggy in dsge_likelihood.m and in non_linear_dsge_likelihood.m. Writing
likelihood = penalty;
DOESN'T penalize the likelihood because penalty is a base value to which it is necessary to ADD a positive amount proportional to the violation.
2) I don't understand the initialization penalty = xparam1 in non_linear_dsge_likelihood.m
3) for linear models, we make sure that the solution (dr.ghx and dr.ghu) are real. Complex likelihood can appear only if matrix F in the Kalman filter is not positive definite. I'm not sure that this problem would arise only for some value of the estimated parameters that we want to eliminate. I rather think it is the result of a problem with the model or the priors that we should reject with an error. But others may disagree.
4) For nonlinear models, maybe it makes sense to filter out parameters generating complex likelihood. Stéphane and Frédéric should tell their opinion.
Best
Michel
On 06/06/2012 12:09 PM, Johannes Pfeifer wrote:
Dear all,
To close ticket 59, I would require feedback on the treatment of complex likelihoods, see http://www.dynare.org/pipermail/dev/attachments/20120425/8000f828/attachment...
To recap: currently complex likelihoods are assigned a penalty and continue estimation. They do not result in an error.
Hence, I would suggest to delete the text part of the error message referring to complex likelihoods.
On a related issue, the treatment of NaN-Likelihoods in linear and non-linear estimation is not consistent. In the former they results in an error, while in the latter they result in a penalty. Is this distinction on purpose? And if not, how should we treat this case in the future.
Best,
Johannes
Johannes Pfeifer Haußerstr. 29 72076 Tübingen Tel.: +49-(0)7071-6396184 Mobil: +49-(0)170-6936820 jpfeifer@gmx.de
-----Ursprüngliche Nachricht----- Von: dev-bounces@dynare.org [mailto:dev-bounces@dynare.org] Im Auftrag von Sébastien Villemot Gesendet: Mittwoch, 6. Juni 2012 11:35 An: Marco Ratto Cc: List for Dynare developers Betreff: Re: [DynareDev] estimation with analytic scores and hessian
Marco Rattomarco.ratto@jrc.ec.europa.eu writes:
would it be possible to add an option to estimation
analytic_derivation
default = 0 (already set in global_initialization)
=1, triggers estimation with analytic gradient and the final hessian is also computed analytically.
Done.
-- Sébastien Villemot Researcher in Economics& Debian Maintainer http://www.dynare.org/sebastien Phone: +33-1-40-77-84-04 - GPG Key: 4096R/381A7594
Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev