Near VAR

[AhmedSahlool]

Hi,

I hope this finds you well.

I am implementing a VAR model for 5 time series; 1st difference for Import price index, detrended external stock of debt, detrended GDP, Goods and services balance to GDP and inflation rate.

Given the different methodology of transforming the series to stationarity and their development, I included a constant to 1st difference of Import price index (it has a unit root trending upward), a constant and dummy variable to Goods and services balance to GDP to take the change in the level into consideration, and I didn’t include a constant for other variables.

Is this common in VAR models; including a constant, or a constant a dummy for some variables and not for others?

And If I would like to determine the lag length of this Near VAR, could I do it by estimating two models by SUR (because I need the option Robusterrors given the Heteroskedasticity) and Ratio instruction?

I am a little confused about the options of Ratio instruction in this case; "degree" will equal the number of restriction in the whole system of equations, and “Mcorr" will equal the equation with the constant and the dummy and the higher order of lags??

Could I use Sur instruction for block erogeneity in the same way?

Thank you very much.

Ahmed Sahloul

[TomDoan]

Hi,

I hope this finds you well.

I am implementing a VAR model for 5 time series; 1st difference for Import price index, detrended external stock of debt, detrended GDP, Goods and services balance to GDP and inflation rate.

Given the different methodology of transforming the series to stationarity and their development, I included a constant to 1st difference of Import price index (it has a unit root trending upward), a constant and dummy variable to Goods and services balance to GDP to take the change in the level into consideration, and I didn’t include a constant for other variables.

Is this common in VAR models; including a constant, or a constant a dummy for some variables and not for others?

No. VAR equations aren't structural -- they're reduced forms. If any of the variables needs a constant in its regression, then all do since the presence of regressors with non-zero mean would force a non-zero intercept even if the dependent variable has mean zero.

And If I would like to determine the lag length of this Near VAR, could I do it by estimating two models by SUR (because I need the option Robusterrors given the Heteroskedasticity) and Ratio instruction?

As I said, I wouldn't recommend using the different deterministics. However, if you want to go ahead, then you would need to do SUR anyway to get the maximum likelihood estimates.

I am a little confused about the options of Ratio instruction in this case; "degree" will equal the number of restriction in the whole system of equations, and “Mcorr" will equal the equation with the constant and the dummy and the higher order of lags??

MCORR is an approximate small-sample correction. I would use the average number of coefficients per equation.

Could I use Sur instruction for block erogeneity in the same way?

Yes.

[AhmedSahlool]

Thank you for your detailed reply.

However, I can include a dummy variable for only one variable to take into consideration the change in its level, non?

Thank you again, I do apprecaite your help

[TomDoan]

Thank you for your detailed reply.

However, I can include a dummy variable for only one variable to take into consideration the change in its level, non?

Thank you again, I do apprecaite your help

Again, a VAR is a reduced form. The effect on the mean of the process in throwing in a dummy isn't all that clear.