Balke(2000) Threshold VAR

[TomDoan]

Yes. Use the correct value for SSTART. Your working sample starts one period later.

[pbrigante]

Hi Tom,

Can you say me where I´m wronging in these irf´s ? Why are they have such irregulars moviments ?

Greatful !

[TomDoan]

Isn't your output measure in growth rates? That's what responses of growth rates look like.

[pbrigante]

Hi Tom,

please ! I have some doubts about how to interpreat the results of the Nlirf´s.

1. How Can I know if the variables to react significantly on shocks ?

2. In the file in attachment, for example, is it correct to conclude that the responses of the variables: "produto" (product) and "inflação" (inflation) not react significantly on monetary shock ?

Greatful

[TomDoan]

Hi Tom,

please ! I have some doubts about how to interpreat the results of the Nlirf´s.

1. How Can I know if the variables to react significantly on shocks ?

There's no simple way to do that. The NLIRF's require a double level bootstrap just to get the point estimates. Confidence bands would require a completely new level of simulations to vary the model coefficients as well.

2. In the file in attachment, for example, is it correct to conclude that the responses of the variables: "produto" (product) and "inflação" (inflation) not react significantly on monetary shock ?

No. For instance, the values on the Produto IRF depend upon the relative scales of the two variables.

[Jules89]

Dear Tom,

again I have a follow up question on the near-T-VAR. You said that I have to use SUR on different subsamples and then add the %logls.

1) How would I define the subsamples. For SUR there does not exist the convenient GROUP function. How can I use the SMPL option on SUR to get estimation for all the different subsamples?

2) What do you precisely mean by adding the %logls for the subsamples? Do you mean that when I have a threshold and estimate a lower and upper SUR, that I have to add the two %logls resulting from the two models defined by the sample split?

Best

Jules

[TomDoan]

Dear Tom,

again I have a follow up question on the near-T-VAR. You said that I have to use SUR on different subsamples and then add the %logls.

1) How would I define the subsamples. For SUR there does not exist the convenient GROUP function. How can I use the SMPL option on SUR to get estimation for all the different subsamples?

SUR with SMPL=credthr{d}<=thresh and SUR with SMPL=credthr{d}>thresh

2) What do you precisely mean by adding the %logls for the subsamples? Do you mean that when I have a threshold and estimate a lower and upper SUR, that I have to add the two %logls resulting from the two models defined by the sample split?

Correct. Note that that will only give you the estimation with heterogeneous covariance matrices---there is no simple way to get a common covariance matrix when you are using SUR. (With identical regressors, the coefficients in each subsample are independent of the choice of covariance matrix, which is not true with a near-VAR).

[Jules89]

Thanks,

that means that running the bootstrap tests as in Balke (2000) is going to be difficult. Do you thing the distortion from using var=hetero in the bootstrap is large?
Another question: why using sur rather thsn estimate?

Best

Jules

[TomDoan]

Thanks,

that means that running the bootstrap tests as in Balke (2000) is going to be difficult. Do you thing the distortion from using var=hetero in the bootstrap is large?

That's hard to say, since the fixed regressor bootstrap won't apply otherwise.

Another question: why using sur rather thsn estimate?

ESTIMATE does least squares equation by equation (whether that's maximum likelihood or not). SUR does joint estimation (whether it's needed to not).