--- Claude Francoeur <[email protected]> wrote:
> We are doing a polynomial OLS. An explanatory term is squared and
> cubed to account for non linearity of the association with the
> dependent variable. We also took the natural log of these variables
> to mitigate multicolinearity. We are getting significant
> coefficients. <snip> Is our approach valid?
It is unusual. If you care about multicollinearity of polynomial terms
you are better of using -orthpoly- (see -help orthpoly-). This results
in a model that is equivalent to a regular polynomial regression
whereas your model is not equivalent to a polynomial regression.
If you primarily care about whether or not the effect is non-linear
rather than polynomial, then you are probably better of with other more
flexible parameterizations. I am much more a fan of splines to
represent non-linear relationships (see: -help mkspline-), while others
like fractional polynomials (see: -help fracpoly- and -help mfp-)
Hope this helps,
Maarten
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting address:
Buitenveldertselaan 3 (Metropolitan), room N515
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
