Re: Analysing performance data (somewhat OT stat question; sorry)

[Henry F. Mollet]

Sorry about mis-spelling of "Abramowitz" is correct.

More than rusty myself on stats questions. How about non-parametric test for
association if two variable are known *not* to be bivariate normally
distributed. E.g. Kendall's coefficient of rank correlation (tau) and
Spearman's rank correlation (r_sub_s). See Sokal and Rohlf "Biometry" for
details. Mike below suggests a Wilcoxon Rank-Sum test, which I surmise might
be something similar.
Henry 


on 9/22/04 9:28 PM, Mike Miller at address@hidden wrote:

> On Wed, 22 Sep 2004, James Knowles wrote:
> 
>> I've been collecting performance data for different software
>> configurations, and examining it in Octave.
>> 
>> It's been about 15 year since college stats class, so I'm rusty. Some
>> data distributions are normal, and a simple t-test works great for
>> testing whether there's a significant difference between configurations.
>> Many are visually "obvious," but some are not. They all need to be
>> documented, however.
>> 
>> Some are heavily skewed, scrunched up near zero. I don't remember what
>> this kind of distribution is called. I do not remember what to use here
>> to test a null-hypothesis of the two data populations being the same.
>> (Yes, inaccurate terminology; sorry.)
>> 
>> Octave has alot built in, and I don't want to just plug data in
>> randomly.
> 
> 
> I don't know what your question is.  The chi-square distribution on 1
> degree of freedom has infinite density at zero.  Maybe you're thinking fo
> that one.
> 
> Anyway, the "Mann-Whitney U test," also called "Wilcoxon Rank-Sum test" is
> nice for comparing means when the distributional assumptions of the t test
> are not met.  There is not great loss of power when the distributions are
> normal, and there can be very substantial gains in power when the
> distributions are non-normal.
> 
> Mike



-------------------------------------------------------------
Octave is freely available under the terms of the GNU GPL.

Octave's home on the web:  http://www.octave.org
How to fund new projects:  http://www.octave.org/funding.html
Subscription information:  http://www.octave.org/archive.html
-------------------------------------------------------------