On Tue, 16 Sep 2003 16:24:55 -0400, Rajarshi Guha
<[EMAIL PROTECTED]> wrote:

> Hi,
>   I'm working on a problem in which I'd like to determine whether two
> sets of data come from the same distribution. It seems that the Kolmogorv
> Smirnov test will give me the information I need.

KS  gives you a distance as the  *maximum*  of linear distances,
in one direction, or in both directions.  Would you want a maximum, or
some sort of sum?  Or, you could use the area;  or the average
squared distance.  Et cetera.  Check a book on Goodness of fit tests.

> 
> However I'd like to go further than just accept or reject the H0 for the
> test. Is there any way (using this test or someother test) to determine
> *how* similar the distributions of two sets are?
> 
> Am I correct in thinking that  a P value for the KS test would provide this
> information? I looked up Conovers book on non parametric statistics for
> the algorithm of the KS test. However it does not mention any way of
> calculating a P for the test. Is it possible?

If you go to the table in the back, which has p-values for one 
end of the test (serving both one-tailed and two-tailed 
hypotheses for the location parameter), you will find citations 
to the literature.  I assume that one of those says something 
about How.

"Moderately similar"  or "not greatly different"  is usually all
that folks care about for the shape.   Do you want to see a 
Confidence interval?  - A CI  computed on a difference?

I could mention, that it can require a large N if you want to 
put a  *narrow*  confidence limit on any computed
"difference"  between two distributions.

> 
> Are there any other tests that would be able to tell me how similar the
> distributions of two sets of observations are
> 
(a) see the books, and (b) some versions of that 'similarity' 
are    really   not    interesting.
  ... for instance ...  whether something comes out looking 
similar-versus-HIGHLY-similar,  may turn out to be the fault 
of various TEMPORARY and artificial conditions, that 
were poorly measured, and that aren't really in control ...

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
"Taxes are the price we pay for civilization." 
.
.
=================================================================
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