On 20 May 2004 03:37:10 -0700, [EMAIL PROTECTED] (G.S.Clarke) wrote:

[snip, about educational level and marital status > 

> OK � two questions.
> 
> First � is there any good reason why kruskal-wallis (which I would have 
> taken as being the more powerful technique) doesn�t show significance 
> when the chi-squared tests do (in other words, am I correct in my 
> assumptions as to which is the more powerful technique).
> 
I will go further than Thom's reply did.

The contingency table is more powerful against the so-called
non-ordered alternatives.  The KW is more powerful against ordered 
alternatives.

If two tests are of equal size (honest alpha) and obtain different 
p-values, then they are testing different hypotheses - e.g., here - or
they are weighting the errors differently.  An example of the latter
would be the Pearson contingency chi-squared versus the likelihood
chi-squared:  Pearson is more powerful against large deviations by 
cell, the (O-E) term, and the other is more powerful against multiple 
small deviations.  This might be considered a more subtle form of
"testing a different hypothesis".


> Secondly chi-squared provides a �symmetrical� analysis � it says (at 
> least I think that it does) that having information about one of the 
> variables provides information about the other variable � and vice 
> versa.  However is Kruskal-Wallis �symmetrical� in this sense.  Does not 
> KW say (had it been significant) that the marital status of the subjects 
> (the factor) informs us somewhat on their educational status (the 
> dependent variable). Now logically if this is true then the educational 
> status should inform us somewhat on the marital status, however is this 
> also the case �statistically� i.e. will any significance level generated 
> by the KW using education (dependent) v marital status (independent) 
> also apply to phrasing the question the �other way around�?

Yes, these tests give us a report on "relation", not causation.  That
is one reason why I don't like those terms very much, Dependent and 
Independent.  The causation may run opposite to what we expect,
or both things might be caused by something else.

A regression gives us coefficients that are asymmetrical, but
the ordinary regression of A on B  has the same test as B on A,
which is the test on the correlation.

There may be something more to be said about symmetry, but 
nothing comes to mind.

-- 
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