Dropping Insignificant Variables

[trainee]

Hi guys,

I'm a massive rookie just starting to use EViews so please bear with me. I am trying to write a code that will drop insignificant variables from a system of equations (in a more technical language, I am trying to create an unbalanced Seemingly Unrelated Regression from a balanced one because I see some variables are not significant). Any ideas on how to proceed?

Thanks..

[startz]

Out of curiosity, why would you want to drop variables just because they are insignificant?

[trainee]

I heard somewhere that the balanced SUR is really just OLS but I could be wrong..

[startz]

Yes, balanced SUR is just OLS. Is that bad?

[trainee]

Not if the errors are still assumed to be correlated across equations. Is this still true?

[startz]

In the case where the RHS variables are identical, SUR and OLS are numerically identical without regard to the correlation across equations. It's sort of an algebraic oddity.

[trainee]

Would it then not make sense to start dropping insignificant variables so that you end up with an unbalanced SUR that does acknowledge the correlation across equations? (if theory suggests your LHS variables must be somewhat correlated)

[startz]

No.
(1) The fact the LHS variables are correlated doesn't require the errors to be correlated. Correlated RHS variables will do it.
(2) The fact that a variable is insignificant does not mean it's coefficient equals zero. It means that the data can't tell you whether the coefficient equals zero or not.

[trainee]

(1) Bit nervous about the second part of your statement as errors by defintion are outside of the RHS variables

(2) True, but there are costs to leaving the coefficiens in

Anyway would really appreciate if someone could comment on my original question. Thanks!