On Tue, 01 Oct 2002 12:49:51 +0100, Thom Baguley
<[EMAIL PROTECTED]> wrote:

> Donald Burrill wrote:
> >  And, to address your last question, a Mann-Whitney test (aka a Wilcoxon
> > rank-sum test) is exactly equivalent to a t-test on the mean ranks;  so
> > yes, you can, but use alpha/2 for your tabled significance level.
> 
> Pedantry alert! Is it _exactly_ equivalent - don't the results
> differ very slightly (though not to a degree that would make a
> difference in practice)?

I think I remember the details from what DB cites -
  Conover, W.J. & Iman, R.L. (1981).  Rank transformations as a 
bridge between parametric and nonparametric statistics.  
_The American Statistician, 35,_ 124-129.

When there are no ties, the  t-test computed on ranks gives 
exactly the same  *ordering of outcomes*   as you would tabulate
by hand.  So you can *compute*  the t-test.   If you look up the
tabled value for t (so-many d.f.),  you get a good approximation
to the p-value.  If you compute a t-value for all the permutations 
of the samples,  and assign the p-value according to the extreme,
then you get the same p-value, exactly.

The large-sample tests on ranks often use normal-approximation
instead of t-test approximation, since "the variances are known."  
However, variance estimates are not precise if there are ties.  
So the alternate tests are no longer precise or exact.

Simple t-tests or F-tests  on the ranked data gave answers 
that were generally as accurate, and their errors tended to be 
on the conservative side (not-significant)(which is good).

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to