It depends on what your variables are. The first step is to transform your variables so that they could be used more efficiently (1st difference should be approximately normally distributed). Then you can use some model such as ARIMAX where you can find the best structure using information criteria.

this may help:

http://phd-inceif.blogspot.my/search/label/Structural%20VAR.%20SVAR

Option 1: Traditionally, you would still include all 9 variables in the underlying reduce form VAR, and merely apply the appropriate restriction in the shock identification (in the case of a Cholesky decomposition that would for example imply that the external variables are sorted first). 

Option 2: Theoretically, it is of course possible to additionally (!) impose zero restrictions on the reduced form as well, restricting all the lagged internal variables' coefficients to zero in the first three equations (that explain the external variables). However, I would suggest to be careful with the latter strategy. Technically, you would need to use a seemingly unrelated regression model rather than a standard OLS estimator to estimate your VAR if the right hand side variables are not the same in all equations.

(Of course neither of those voids the importance of using the appropriate stationary transformation of your variables, as pointed out by Andrey).

Best

Makram

Make sure you are using the correct transformation of your variables.

If you are using the Cholecky decomposition, place your exogenous variables first in the ordering, so they don´t get affected by the contemporaneous effects of the other variables. You might want to change the ordering of these tree variables to see if your results are sentitive to changes in the ordering.

However, if you are using buiding a SVAR, then, the appropiate way to do so is to impose restrictions and use SUR or another estimation method but OLS.

Best,

Jose

Thanks everyone.  Appreciate your responses.

The ordering of the variables in the standard reduced form VAR has nothing to do with exogeneity. It imposes an assumption to achieve identification of orthogonal shocks to the endogenous variables for computing impulse responses, etc.. In the realm of cointegration, weak exogeneity can be imposed or trested by restricting the error correction parameters for the relevant variables to zero in the VECM. Strongly exogenous variables are treated the same as constants and trends in a VAR. At least some econometric programs allow for such variables beyond constants and trends. It is also a possibility in cointegration VARs (VECMs), for example, in CATS in RATS.