Re: st: a cautionary tale re scaling

[Kit Baum <baum@bc.edu>]

<>
On Apr 16, 2010, at 2:33 AM, Maarten wrote:

> Because by just rescaling you estimate basically the same model, you 
> just solve any problems that might be the result of precision troubles
> caused by very very large or very very small numbers.
> 
> If you take the log of a variable you are estimating a different
> model, i.e. you are adding extra curvature to the effect.
> 
> It depends on your problem which one you want, however if you can 
> solve a problem by a linear transformation, then that must 
> certainly be preferable over solving that problem using a non-linear
> transformation.


Exactly. In my student's case his model explicitly indicates that gdp should enter in level form, not in log form. You do not want to change functional form in order to deal with problems of numerical accuracy or computability. Stata has no problem delivering sensible estimates of gdp in this case as long as the scale is altered.

We know that to be always an issue when fitting models with techniques such as nonlinear least squares or maximum likelihood, but it is not apparent that it may be an issue with linear models as well. Private correspondence with StataCorp points out that -regress- and -old ivreg- are much more bulletproof in this regard than -ivregress- due to the way that newer routine was implemented.

Kit Baum   |   Boston College Economics & DIW Berlin   |   http://ideas.repec.org/e/pba1.html
                              An Introduction to Stata Programming  |   http://www.stata-press.com/books/isp.html
   An Introduction to Modern Econometrics Using Stata  |   http://www.stata-press.com/books/imeus.html


*
*   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/