Bug in BVAR Add-in?

[donihue]

Further to my previous post in this Bug Report forum, when using the Diffuse (Flat-Flat) prior, the parameter estimates are (as they should be) identical to those produced by the standard VAR command (although the standard errors differ).

IRFs of course depend upon the definition of the shock. Below is an example using the non-factorised 1 sd definition under VAR. Here, the first column is the same for response of pi to pi, but everything else is different.. (Note also that the name of the response variable is incorrectly shown by the BVAR Add-in.)

A second example uses the df-corrected Cholesky definition under VAR. Although the IRFs are identical for the first period, they are different for subsequent periods

These odd results may well be caused by definitional differences rather than a bug, so it would be helpful to know how the BVAR Add-In IRFs are calculated.

Regards
Donihue

[EViews Esther]

Further to my previous post in this Bug Report forum, when using the Diffuse (Flat-Flat) prior, the parameter estimates are (as they should be) identical to those produced by the standard VAR command (although the standard errors differ).

Yes. Since the Diffuse (flat-flat) prior is a non-informative prior (which has no distributional information), the parameter estimates should be identical to those in the standard VAR.

IRFs of course depend upon the definition of the shock. Below is an example using the non-factorised 1 sd definition under VAR. Here, the first column is the same for response of pi to pi, but everything else is different.. (Note also that the name of the response variable is incorrectly shown by the BVAR Add-in.) A second example uses the df-corrected Cholesky definition under VAR. Although the IRFs are identical for the first period, they are different for subsequent periods. These odd results may well be caused by definitional differences rather than a bug, so it would be helpful to know how the BVAR Add-In IRFs are calculated.

As mentioned, IRFs depend on the *shock* matrix, which is a square matrix of initial *shock* vector. One way to get such a matrix is to set shock = @cholesky(sigma) where error_t ~ N(0, sigma).
The odd results *could* be a bug because of the odd Cholesky ordering.

In the updated BVAR add-in, the Cholesky ordering is fixed. To get the shock matrix corresponding to *different* ordering, you can use shock = @cholesky(P*sigma*@transpose(P))*P where P is a permutation matrix.

Thank you for your attention! It was a big help.

Best,
Esther

[donihue]

Many thanks, Esther. It is marvellous to be able to dialogue with programme developers. What a (positive) change from the old "black box" days of EViews!

Regards
Donihue