Page 1 of 1

Acceptange rate when mode_compute=6 and when mode_compute=0

PostPosted: Fri Jan 20, 2017 1:52 pm
by pepito_bm
Dear all,

I am currently estimating a small open economy model with a housing sector and I encountered the following issue. When I haven't found the posterior mode yet, I use mode_compute=6 (since it seems to be the least problematic procedure towards finding the posterior mode). My mh_jscale=0.15 otherwise but still the acceptance rate of the MH algorithm is not satisfactory (i.e. below 0.2). Now, since I'm not happy with it, I re-run the estimation whti mode_compute=0 since the posterior mode has already been found. But then I need a much larger value for the scale parameter, say, mh_jscale=0.34 otherwise the acceptance rate skyrockets. So I've got two questions:

1. Shouldn't the acceptance rate be the same irrespective of what posterior mode finder I use?

2. Does the mh_jscale value have any impact on finding the posterior mode when mode_compute=6 is selected?

Kind regards and thanks a lot in advance,
Peter

Re: Acceptange rate when mode_compute=6 and when mode_comput

PostPosted: Mon Jan 23, 2017 10:59 am
by jpfeifer
In Dynare 4.4.3 there are interaction effects between the user-provided
Code: Select all
mh_jscale

and the tuning of the scale-factor by
Code: Select all
mode_compute=6
. Those issues should be resolved in the unstable version/Dynare 4.5.

Sidenote: if only mode_compute=6 works, there are usually other issues in the model. Have you looked at the
Code: Select all
mode_check
plots?

Re: Acceptange rate when mode_compute=6 and when mode_comput

PostPosted: Thu Jan 26, 2017 8:40 am
by pepito_bm
Hi Johannes,

Thanks for your reply. In the first stages of the model, when I was simply playing around with a basic NK model there were ocassions when Chris Sims' posterior mode finder (so the default in dynare) would do, but that was in a very limited number of cases. So I've just been sticking to mode_compute=6 which works fine. Haven't tried any others. The posterior mode looks fine, well, depending on the observables and the model, I get good convergence, but basically I've decided to draw 2 million times from the posterior during the MH state for reliable convergence. And I just noticed that sometimes, when I use mode_compute=6 and mode_compute=0, the j_scale parameter has to be different to get an acceptance rate between 25% - 33%. But things are working fine now more or less.

Kind regards,
Peter