The RATS Software Forum

Yes. Though I would caution you (as I tell anyone who tries to adapt this), that the EEV program is rather specifically aimed at splitting the sample based upon the variance of the oil price.

compute sigmav(2)=%diag(%xdiag(sigmav(1)).*||.25,.25,4.0||)

is for the EEV data, as it guesses a 2nd regime where the variance of the 3rd equation (oil in the EEV model) is higher than in the 1st. Whether that's appropriate to your model is your decision.

TomDoan wrote:Yes. Though I would caution you (as I tell anyone who tries to adapt this), that the EEV program is rather specifically aimed at splitting the sample based upon the variance of the oil price.

compute sigmav(2)=%diag(%xdiag(sigmav(1)).*||.25,.25,4.0||)

is for the EEV data, as it guesses a 2nd regime where the variance of the 3rd equation (oil in the EEV model) is higher than in the 1st. Whether that's appropriate to your model is your decision.

Thank you for the warning, Tom. It fits to my model. However, I wonder if ||.25,.25,4.0|| can have some other values as long as the third one is kept as the highest.

Those are just guesses so it doesn't really matter other than for trying to force a separation in the regimes.

atarca wrote:

TomDoan wrote:You can't test anything with the MCMC estimates---any testing has to be done with the maximum likelihood. You can do a LR test for covariance and coefficients switching vs just covariances by estimating the models with and without the coefficient switching. That will have standard asymptotics.

Is the attached code consistent with what you recommend, Tom?

Tom, I would be thankful if you do not mind my two more questions about testing.
(1) I want to do the tests individually, e.g., BETASYS(1)(1,2)=BETASYS(2)(1,2).
(2) I want to test a Granger non-causality-like hypothesis, e.g., BETASYS(1)(3,1)=BETASYS(1)(4,1)=0 in regime 1 and BETASYS(2)(3,1)=BETASYS(2)(4,1)=0 in regime 2 separately to compare the outcomes.
How can I do these tests, Tom?

Those are Wald tests on the results of the full switching model. You can use the Regression Tests wizard to set them up.

I'm not sure what testing a difference between individual coefficients is going to tell you in a model like that.

TomDoan wrote:Those are Wald tests on the results of the full switching model. You can use the Regression Tests wizard to set them up.

I'm not sure what testing a difference between individual coefficients is going to tell you in a model like that.

I just thought to make a table of results showing only the significantly switching parameters like in Ehrmann et. al (2001, p.15) . Thanks for all Tom, that's very kind of you.

Not everything that's published is a good idea. Individual coefficients in a multi-lag dynamic model really don't have useful interpretations.

TomDoan wrote:Not everything that's published is a good idea. Individual coefficients in a multi-lag dynamic model really don't have useful interpretations.

Certainly, I will take your word for it, Tom. Do you think that testing a Granger non-causality-like hypothesis is also useless?

They are reasonable as long as they are properly structured for a 3 variable system. Zero lags on one variable in one equation aren't really interesting, just like they aren't interesting in a regular VAR.

Hi Tom,
Would you please explain what "-12.0*3" means in

compute sigmav(2)=%diag(%xdiag(sigmav(1)).*||.25,.25,4.0||)
compute logdetlimit=%MSSysRegInitVariances()-12.0*3

in the eev_ml.rpf? In what conditions should we change these two figures?
Thanks

As it says on the first page of this, this is a very technically demanding model. The Structural Breaks and Switching Models course has an entire chapter which explains the details. I would very strongly recommend that you get that.

That test is designed to prevent problems with one covariance matrix going singular when they are allowed to switch.

is it possible to get the regime dependent response of the variables to shocks in other variables?

Sure. They're already being computed, just not graphed. (The ONLYSHOCKS option on the @MCGRAPHIRF at the end controls what shocks are graphed).

I have one more question is it possible to get the impulse response in a tabular format.....

Use @MCPROCESSIRF to put the IRF information into SERIES.