Re: st: Not Quite Quadratic Regression

[David Hoaglin <dchoaglin@gmail.com>]

Maybe.

The Box-Cox approach is likely to be more appropriate for y than for x
here, though it would be interesting to see what transformation
emerges for x.

The data on y (or x) must be positive, and -boxcox- assumes that the
errors have a normal distribution.

The estimated power for the transformation should be used as a
starting point in choosing a power that is familiar and likely to be
interpretable.  For example, 0.28 would lead me to consider the square
root and the logarithm (the limiting case as the exponent goes to 0).

David Hoaglin

>
> Sounds like a job for Box & Cox... help boxcox
>
> Kit
*
*   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/