A question on reducing dimension of VAR system
Posted: Sun Dec 19, 2010 5:08 am
Dear all, here I have many variables total in 12, let us name them as "v1, v2, ..., v12" all are integrated of order 1, and I want to analyze them through VAR/VEC framework. For this analysis I found that, I need to reduce the number of variables as I do not have enough number of times series observation to analysis all those variables simultaneously. So have thought of following 2 approaches:
1. I found that there is very high correlation estimate (simple correlation on the 1st difference of them) between v1 and v2. Does it make sense to omit either v1 or v2 from my system of variables, and do analysis on remaining 11 variables?
2. Similarly I found that there is fundamentally zero correlation between v11 and v12. Therefore I have constructed 2 systems (v1, v2, ..., v11) & (v1, v2, ..., v12) and do analysis separately for these 2 systems.
Therefore I am seeking expert advice on whether above approaches of reducing dimension make sense? What I may loose or gain if I go with above approach?
Thanks for your time.
1. I found that there is very high correlation estimate (simple correlation on the 1st difference of them) between v1 and v2. Does it make sense to omit either v1 or v2 from my system of variables, and do analysis on remaining 11 variables?
2. Similarly I found that there is fundamentally zero correlation between v11 and v12. Therefore I have constructed 2 systems (v1, v2, ..., v11) & (v1, v2, ..., v12) and do analysis separately for these 2 systems.
Therefore I am seeking expert advice on whether above approaches of reducing dimension make sense? What I may loose or gain if I go with above approach?
Thanks for your time.