Variable Specific Prior (Mean) Restrictions on BVARs

[rangangupta]

Hi,
Suppose, I have a VAR model with two variables, one stationary and the other non-stationary.
I was wondering if, while estimating a BVAR model with these two variables, it is possible to impose a prior mean of 0 on the first own lag of the stationary variable, while continuing to impose a prior mean of 1 on the first-own lag of the non-stationary variable.

I know the from the drop down menu (Statistics -> VAR (Setup/Estimate)) allows me to choose 0 or 1 as the prior mean on the first own lag of a variable in a VAR, but not really sure how I can make it variable specific depending on whether it is stationary or non-stationary.

Any help would be highly appreciated.



Regards
Rangan

[TomDoan]

The wizards don't include some of the less-used options. (With too many choices, they can get confusing).

SPECIFY has an MVECTOR option, which in your case would be MVECTOR=||1.0,0.0|| if the 1st variable is the non-stationary and 2nd is stationary (and ||0.0,1.0|| if they're reversed). You can just take the instructions generated by the wizard and edit the SPECIFY to put the MVECTOR option in. (Take the MEAN option out if it's in your SPECIFY).

[Gary]

Hi there,

I have a similar question to Rangan's below. I am running a 5 variable BVAR with 3 foreign and 2 domestic variables. I am treating the foreign variables as block exogenous, but not imposing exogeneity directly using the near-VAR approach. Instead, I want to shrink the parameters on the domestic variables to zero in the equations for the foreign variables.

This is less restrictive than the near-VAR approach, as it allows for a non-zero posterior in the event the data strongly disagrees with the prior. What I have done is to use the specify(type=general) option with a declared 5 x 5 prior matrix, with zero's placed on the 04th and 05th elements in the first, second, and third equations ( the three foreign variables are ordered first in the VAR). Is this the correct way to do it, or do I need to use a "dummy observation" prior? Are there any examples of the dummy observation prior approach that I can use for reference?

Many thanks in advance for your help

Best wishes,
Gary

[TomDoan]

That's the right idea, but you can't do a hard zero because it needs to divide by the value. Use a value number like .0001 instead. If you look at the CANMODEL.PRG example, the final model run there nearly zeros out several of the lags.

[Gary]

Hi Tom,

Many thanks, much appreciated. The procedure seems to work all right - the impulse responses of the foreign variables are fairly flat around zero as far as domestic shocks are concerned.

Best wishes,

Gary.

[rangangupta]

Thanks a lot. Really appreciate this.
Regards
Rangan