[R] Box-Cox / data transformation question

Landini Massimiliano numero.primo at tele2.it
Mon Jan 31 18:30:24 CET 2005


On Sun, 30 Jan 2005 17:47:31 -0500, you wrote:

<<<<<-----------------SNIP
|=[:o)  >
|=[:o)  >  
|=[:o)  >
|=[:o)  Why are you using a double square root transformation? Is the 
|=[:o)  transformation for the response variable? Transfromation is one way to 
|=[:o)  help insure that the error distribution is at least approximately 
|=[:o)  normal. So if this is the reason, it certainly could make sense. 

Are you sure that (data^0.25) had sense??? Coud you explain me which is the
sense??
I know sense of boxcox exponents near zero when data are positively skewed and
log(data) make it  normally distributed, or all those case where variances grow
proportionally to means or when i know that there are interaction effects that
not follow additive model (AnOVa assumption);

I know 0.5 exponent (square root) [ as sqr(data) if all data differ from zero
else sqr(data+.5) else Asconbe propose sqr(data +3/8) else Tukey & Freeman
propose sqr(data)+sqr(data+1) particularly suitable when data domain is  (0,2) ]
for right skewed data, frequently  applied to count-data or
count-of-something-over -a -surface (bacteria, virus, nematode, lions) due to
n*p*q (variance)  is almost proportional to its mean (n*p)  so AnOVa fundamental
assumption is basically violated....

I know 1/3 exponent  applied to count-of-something-in-a -volume...and so on...

What is worth is that i'm trying to ask to Christoph to sit down and think: what
kind of number are these??
E.coli/mL?? ...so...i try cuberoot transformation and/or log transformation
Timing of a slug vs snail speed race?? ...so..i think that inverse
transformation it best.

BoxCox procedure have produced a fantastic implement that can help many people
but (IMHO) none procedure can be superior than Ripley + Bates + other gurus
experience. If you ask to those great statisticians how do you manage
electrophoresis velocity they could respond with "data^-1 why......blah blah
blah"
If you push data in BoxCox algorithm it will respond with "-0.97847164..."
Which answer had more sense???
I prefer -1

 
|=[:o) There  is no unique scale for making measurements. We choose a scale that helps 
|=[:o)  us analyze the data appropriately.
|=[:o)  
|=[:o)  Rick B.
|=[:o)  
|=[:o)  ______________________________________________
|=[:o)  R-help at stat.math.ethz.ch mailing list
|=[:o)  https://stat.ethz.ch/mailman/listinfo/r-help
|=[:o)  PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
|=[:o)  
|=[:o)  
|=[:o)  ---
|=[:o)  avast! Antivirus: In arrivo messaggio pulito.
|=[:o)  Virus Database (VPS): 0504-4, 28/01/2005
|=[:o)  Controllato il: 31/01/2005 16.23.36
|=[:o)  avast! - copyright (c) 1988-2004 ALWIL Software.
|=[:o)  http://www.avast.com
|=[:o)  
|=[:o)  



-------------------------------------------------------------------------------------------------------------------------
Landini dr. Massimiliano
Tel. mob. (+39) 347 140 11 94
Tel./Fax. (+39) 051 762 196
e-mail: numero (dot) primo (at) tele2 (dot) it
-------------------------------------------------------------------------------------------------------------------------
Legge di Hanggi: Più stupida è la tua ricerca, più verrà letta e approvata.
Corollario alla Legge di Hanggi: Più importante è la tua ricerca, meno verrà
capita.




More information about the R-help mailing list