R help - Jul 2005 - Boxcox transformation / homogeneity of variances

Dear r-helpers,

Prior to analysis of variance, I ran the Boxcox function (MASS library) to 
find the best power transformation of my data. However, reading the Boxcox 
help file, I cannot figure out if this function (through its associated 
log-likelihood function) corrects for * normality only * or if it also 
induces * homogeneity of variances *. I found in Biometry (Sokal and Rohlf, 
p. 419) that the box-cox transformation can be extended to induce 
homogenity of variances in conjunction with Bartlett's test of homogeneity 
of variances. Does the Boxcox function implemented in R refer to this 
extension ?
Thanks a lot,


- - - - - - - - - - - - - - - - - - - - - - -
Arnaud DOWKIW
INRA
Forest Research
Avenue de la Pomme de Pin
BP 20619 ARDON
45166 OLIVET CEDEX
FRANCE
Tel. + 33 2 38 41 78 00
Fax. + 33 2 38 41 48 09
- - - - - - - - - - - - - - - - - - - - - - -

Please consult the reference on the help page of that function: it _is_ 
support software for a book.  It implements the Box-Cox procedure (as it 
says).  The original Box-Cox paper has three aims, two of which you have 
mentioned (but perhaps the most inportant one is the one you 
have not mentioned, additivity).

It is probably worth stressing that the Box-Cox procedure is about finding 
the best transformation within a specific family for fitting a particular 
_model_ to a set of data, not for the data per se.  There is a long 
history of people using an inappropriate model and finding an 
uninterpretable transformation.

On Wed, 13 Jul 2005, Arnaud Dowkiw wrote:
> Prior to analysis of variance, I ran the Boxcox function (MASS library) to
> find the best power transformation of my data. However, reading the Boxcox
> help file, I cannot figure out if this function (through its associated
> log-likelihood function) corrects for * normality only * or if it also
> induces * homogeneity of variances *. I found in Biometry (Sokal and Rohlf,
> p. 419) that the box-cox transformation can be extended to induce
> homogenity of variances in conjunction with Bartlett's test of
homogeneity
> of variances. Does the Boxcox function implemented in R refer to this
> extension ?
-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595