Talk:Box-Cox transformation

Perhaps this page should be expanded to include information on when performance of the test is appropriate or finding the appropriate transformation.

Dubious information
I have removed this material:


 * In regression analysis, one sometimes carries out a series of Box-Cox transformations of the response variable with a range of values of &lambda; and &alpha;, and one then compares the residual sum of squares at these values in order to choose the transformation which gives the best results. Because the residual sum of squares is proportional to the log-likelihood, this procedure amounts to approximate maximum likelihood estimation.

from the Background section. Reducing RSS is not the goal&mdash;otherwise often using the stronger transformations without limit in either direction, depending whether the data are less than or greater than one, will yield better and better RSSs. I think what is meant here is the shape of the residual distribution is observed, e.g., for symmetry, equal widths across data space, etc. Baccyak4H (Yak!) 15:16, 27 April 2007 (UTC)

Merge proposal feedback
As currently written, the power transform article mentions just the Box-Cox, so a merge/redirect would make sense. However, I think that other article could be improved by generalizing beyond just the Box-Cox (e.g., taking the square root of Poisson data before performing linear regression). Baccyak4H (Yak!) 17:16, 19 July 2007 (UTC)
 * Also related but broader is Data transformation (statistics). That might be a better place for things such as square-root transform (and Anscombe transform) for Poisson data? It has a short section Data transformation (statistics) to which I've just added Main article: power transform. Qwfp (talk) 14:59, 12 March 2008 (UTC)