Different Results of A0 using SBVAR


I am currently trying to estimate a (recursive) constant-coefficient SBVAR using Sims and Zha (1998)'s prior.

I am running into an issue when using contemp_reduced_form option. If the option is selected, the code with crash just before the Gibbs sampling step.

If I instead run the code without the contemp_reduced_form option, the code finishes. However, the point estimates of the A0 matrix have different signs on some of the columns compared to the one produced with the contemp_reduced_form option on.

How is this possible?

Attached are the code and data.
constantRecursiveBVAR.mod (462 Bytes)

MacroData.mat (18.4 KB)

The results of both cases are also included.
Results with contemp_reduced_form
CRFoutdata_a0dp_const.mat (2.4 KB)

Results without contemp_reduced_form
outdata_a0dp_const.mat (2.4 KB)

Thank you in advance!

This is almost surely a bug. Thanks for reporting it. I opened a ticket at https://git.dynare.org/Dynare/dynare/issues/1716

Hello again,

I was wondering if there has been any progress on this issue?

From what I can tell, I think the problem is how the code is constructing linear restrictions.

We forwarded the issue to Zha/Waggoner. They are working on it, but we haven’t heard back with a solution yet.

Thank you for the update.