- I'm responding to statistical questions about my post -
On 16 Dec 2002 01:53:33 -0600, Brian Sandle
<[EMAIL PROTECTED]> wrote:

> In talk.politics.drugs Rich Ulrich <[EMAIL PROTECTED]> wrote:
[ snip, his and mine]
BS > 
> Are you saying you get exactly the same correlations when you put ranks in 
> Pearson as you do when you put the same ones in Spearman?

Absolutely.

BS > 
> Isn't it better to be `regardful' of non-linear stuff and use the proper 
> formula?

What?  What "proper formula"?  The raw data may have
poor scaling *for our purpose*, where the measured, equal
intervals are not constant, *for our purpose*.  Ranking works
a distortion of whatever scaling is there; sometimes, that is 
preferable.  Where it works pretty well is for *simple*  tests,
such as the correlation between two variables, or the difference
between two groups.  Where it works badly, almost assuredly, 
is for complicated tests -- two-way ANOVAs, partial 
correlations  -- whenever the R-squared is large.

RU > 
>   and what we sacrifice 
> > might be (a) the test where we can rely on p-levels, and 

BS > 
> In SPSS, if you were to do rank correlations, and use the Spearman, you 
> would still have p-levels, i.e. significance. And I think you would still 
> have p-levels if you did Spearman partials, that is with linear stuff &c.

I was not clear.  (And you are not reading carefully well.)  
I intended to draw the distinction, for complicated statistics using, 
ranks, between computing a p-level, which you can do; and  
"relying on p-levels", which you can't do,

[ snip ]
RU >
> >   Instead,  I talk about doing a
> > 'rank-transformation' to the data;  and whether that is 
> > desirable or useful or damaging.
BS > 
> By that you are turning scores into ranks? You *have* to do that when the 
> data is non-linear is my guess. Then you put it into a formula designed to 
> work with ranks. Otherwise what you get out has no meaning.

You can turn scores into ranks if you want robustness
for the simple test, and you are willing to pass up the chance
for complicated tests.  When data need a different transformation,
you can do a different transformation, and you might preserve
the chance for fancier modeling.

"Your guesses"  are particularly uninformed on statistics,
and mostly seem to be wrong.  Computers can ignore 
the formulas designed to work for ranks when
general formulas give *precisely*  the same thing -- for
example,  Pearson for Spearman.  Oh, the programs 
often do have some exact formulas for the p-values for 
small N (under 10, at least, for Spearman's).  Those will 
be an improvement if there are no ties in the scores.

It is a matter of indifference, to any users, just how the 
computation is performed.  Textbooks have special 
formulas for ANOVA, too, starting with equal N, and
moving up to multi-way analyses with unequal N.
The basic books will omit the matrix notation that makes
it easier for professionals manipulate.  The basic 
formulas are useful for purposes of teaching, even though
the computer's internal logic can be implemented as 
regression.

[ snip]
> 
> Perhaps you should give some refs.

On what area?  Try my stats-FAQ.  Look at the tags
of other folks who post to the stats-groups, and look
at their FAQs, too.   And look at the FAQs and pages
(or texts) that they recommend.... 

-- 
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