Hi all,
For the declaration of calibrated parameters we may use the
estimated_params block by specifying priors with zero variance... Or we
may add dirac delta function as a new prior shape.
For the choice of the prior density I agree with Michel. The user has to
choose the prior variance knowing that the weight of the prior (relative
to the likelihood) will depend on the sub-sample size. Obviously we
could imagine a way to automatically reweight the priors, but I don't
think it would be a good idea.
Best, Stéphane.
Le mardi 26 janvier 2010 à 13:39 +0100, Michel Juillard a écrit :
> Hi George,
> >
> > There are however few outstanding issues to be resolved.
> >
> > 1.a) Is Detrend data called only once as it looks like to me now?
> No. in EstimationModule, function data_filtering is called a first time
> and then several time in Loop on periods 1 to NP
>
> > I.e. can trend depend on different draws of parmaters and different
> > steady states?
> yes
> >
> > 1.b) Also, Can trend and the constant be different for different periods?
> yes
> >
> > 2) Can there be fixed (calibrated) parameters' sets for different
> > regimes/periods?
> >
> Good point. This should be permitted, but we don't have provision for
> it. Neither an interface for declararing changing calibrated parameters,
> nor an internal representation in the preprocessor or in Matlab.
> We need to think about it some more, but the resolution of this problem
> will only affect the design of parameter updates
>
> > 3) Passing SteadyState:
> > Do we need to return SS from estimation?
> yes
> > If so, which one since each estimation period may eventually have its
> > own different SS?
> >
> all of them. SS is then a matrix with as many columns as there are
> regimes in the model
> > 4) Should we not weight prior density for the length of the period the
> > parameter and its prior are valid for?
> >
> I don't think so. The prior density describes believes about the
> possible values of a parameter irrespective of the number of periods for
> which this parameter matters. If the subsample where a given paramter
> matters is long, the prior will have less influence on estimation as in
> a subsample that is short. But it is what we want.
>
> All the best,
>
> Michel
>
>