On Tue, 23 Apr 2002 09:17:08 +0100, "Robin Hayman"
<[EMAIL PROTECTED]> wrote:

> Hi all,
> 
> I have two columns of numbers that I wish to compare to see how
> similar/ dis-similar they are.  The columns of numbers are a series of
> Pearson Product Moment Correlation Coefficients (PPMCC).  In order to
> compare the two columns of numbers in the past I have employed a
> standard independant samples t-test.  Recently I have plotted the
> distribution of the data and it is NOT normally distributed (the range

Several considerations do come to mind.
There are columns:  are the numbers 'paired'  in a meaningful way?

When it comes to data collection and tests in general,
a correlation is a lousy thing to compare,  since it *assumes*
equality of  variances.  That is, a whole lot of people do 
ask to compare two correlations when they ought to be
asking to compare two regression coefficients or regression
lines.  

The worst problem with non-normality, when it comes to doing
ANOVA (including t-tests)  is that there are far outliers that
thoroughly distort the means.   If you are comfortable with 
comparing the sets of numbers as *means*, then you should
be comfortable with the t-tests.   

> in actual fact is more like -0.1 to +0.8).  Therefore the assumptions
> underlying the t-test are violated and I shouldn't be conducting a
> t-test on the data. Right??

 - with correlations over .5, you do get into the 
range where Fisher's z  is ordinarily used to 
stretch out the larger  values.   That is usually 
considered a stabilizing transformation for correlations, 
so it is acceptable without much argument.

Of course, any time you are taking the average of
correlations, someone might argue that you ought
to consider the N's  that underlie the r's  -- and 
provide a test to show that the r's  are basically
from a homogeneous set.   That does assume that
the r's  are intended to represent a central tendency
instead of being a composite value.


> 
> I have thought about using a non-parametric equivalent of the t-test
> and also about using some kind of normalisation on the data set so
> that I can use a t-test (determined to use a t-test!!!).  What is the
> correct way to go forward?

Hope this helps.

-- 
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