st: Multicollinearity and SUEST

["Bowe Hansen" <bhansen1@vt.edu>]

I have a model which includes several interaction variables, including three
way interactions of a few variables with most of the interactions including
one of two related indicator variables. (Those variables are: one for firms
which pay a small dividend, and one firm firms that pay large dividends,
non-dividend paying firms have a zero for both indicators). Not surprisingly
this results in some very high VIFs and a condition index above 30. It
occurred to me that one way to remove the effect of the interaction
variables on collinearity would be to run separate regressions for groups of
firms designated by the indicator variables. (Run separately on non dividend
paying firms, small dividend paying firms, and large dividend paying firms).
Then I have three separate regressions with one two way interaction (among
other variables) remaining and no three way interactions. I looked at the
VIFs and condition indexes for each of the 3 separate regressions and there
is no evidence of multicollinearity for them. Then, I need to test whether
or not particular coefficients are statistically different in each of the 3
regressions. I believe the SUEST command is ideal for this, but haven't been
able to find any documentation suggesting that this would be an appropriate
way to deal with multicollinearity. 

My question is whether or not this is a valid use of the SUEST command. I
would greatly appreciate thoughts on this. 

Thanks.

P.S. I understand that this approach implicitly includes an interaction
between every variable in the regression and the variable on which I split
the sample, but there is no theoretical reason why the coefficients on all
other variables couldn't be different for the three groups, so I am not
concerned that that would invalidate any of my results.


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/