🚀 go-pugleaf

  If you are having problems like those described, where one groups looks
like it needs one transformation and another group needs another, I see no
way that using both transformations respectively is anything other than
garbage.  I suppose you could normalize them so that the means stay in the
same place, but that is really a fudge.  On the other hand are this is still
an interesting finding.  It tells you a *difference* between the distributions
of the two groups, even if it does not tell you a difference between the
means.
  But, let's consider the transformation vs non-parametrics.  Both do not
make reference to the mean of the original variable anymore.  The np method
may be making inferences on the median or the ordinal structure; the trans
method looks at a specific type of ordinal structure.  If the transformed
scale means something (many people argue that money, for example, naturally
lives on a log scale) then testing difference of means of this meaningful
scale seems to be saying something.  The np/trans argument is very context
dependent, but in my experience seldom makes any difference anyway.

Clark K. Gaylord                  ****************************
Dept. of Statistics               *** Thelonious Monk -    ***
Virginia Tech                     ***   a melodious thunk  ***
703/552-2868                      ****************************
e-mail: gaylord@vtvm2 or gaylord@vtvm2.cc.vt.edu