Determining Threshold Value in the VAR model

[nazif]

Dear all,
I estimate the following simple VAR model to compare the effectiveness of credit channel in the high and the low inflation regime. In case of a linear regression i can easily use ThreshTest procedure to find a threshold break, however it is not applicable in VARs. Is there any procedure to find a threshold value in the VAR model similar to one used in the linear regression or could you please help me to modify this program?
Thanks in advance.

[TomDoan]

See the Tsay JASA 1998 example files on the RATS distribution, or here:

http://www.estima.com/forum/viewtopic.php?f=8&t=1004

[nazif]

Dear Tom,
Thanks. I tried to modify Tsay's code. I write the following. but I'm afraid it is not running properly. Furthermore when I set the first lag of intbrate (interbank interest rate) as a threshold variable, the best break is founded as -0.30000 but It doesn't make sense since I dont have such an observation in the threshold variable.
I would be very happy if you check what is wrong with this code.
Thanks again
Nazif

[TomDoan]

You copied the original definition of the search grid series from the Tsay example, which (as you note) makes no sense for your data. In fact, none of the values are in range for your threshold series, so all the values fit the same. Try

@gridseries(from=.08,to=.90,n=300,pts=ngrid) rgrid

instead.

[nazif]

Many Thanks ,
It is working very well right now. but I have another question. Does the range of grid search should have to cover all values of the threshold variable?
Nazif

[TomDoan]

Many Thanks ,
It is working very well right now. but I have another question. Does the range of grid search should have to cover all values of the threshold variable?
Nazif

You can't cover the whole range because there won't be enough data points on either end to estimate the model. Excluding about 10% on each end would be common, but if the number of regressors in the VAR is large relative to the number of data points, you might need to leave out more than that.

[nazif]

Dear Tom,
Thanks for your help again. Now my model is seem to be working properply. I used montevar.src procedure to create impulse responses with error bands. Is it possible to modify this procedure to create only responses for specific variables such as intbrate or lm1 or to print the results similar the one obtained from impulse instruction.
Regards
Nazif
ps I'm using RATS version 7.1

system(model=creditmodel2)
variables LINDTR_SA LCPI_SA LM1 INTBRATE LCREUS
lags 1 to p
det constant lpetpri{1 to 4} ffr{1 to 4} indus{1 to 4}
end(system)
estimate(smpl=low)
impulse(model=creditmodel2,window="Low Inflation Regime", steps=nsteps)
@MonteVar(model=creditmodel2, step=24, draws=10000, header= "Low Inflation Regime")

[TomDoan]

If you need a more flexible set of graphs, see the discussion on

http://www.estima.com/forum/viewtopic.php?f=4&t=992

about using a different set of procedures rather than montevar.

[iloverats]

dear
if i determining Threshold Value in the VAR model

how to calculate the impulse response function of threshold VAR?

[TomDoan]

dear
if i determining Threshold Value in the VAR model

how to calculate the impulse response function of threshold VAR?

http://www.estima.com/forum/viewtopic.php?f=30&t=948

is for a threshold VECM, which is a bit more complicated, but the idea is the same.