On Wed, 10 Dec 2003 16:06:27 -0500, Rich Ulrich <[EMAIL PROTECTED]>
wrote:

> On 10 Dec 2003 07:30:40 -0800, [EMAIL PROTECTED] (Donald Burrill) wrote:
> 
> > So far all responses have appeared to assume that the data represent
> > left and right ovaries on the SAME women.  But this cannot be the case,
> > because the total number of left ovaries in the data is 132, and of
> > right ovaries 103.  So:  before we can offer useful advice, we need to
> > know whether the pathology being documented occurs, when it occurs, only
> > in one of a woman's ovaries (for these data, anyway), or whether some of
> > the 235 instances of pathology represent both ovaries for some of the
> > women.  If there are any of these last, those are the only cases for
> > which one has paired data.

 
[ snip, to what I posted before ... ]
>If the L and R  instances are disjunct, then the paired t-test
> would take place with a negative correlation, and the 
> absolute correlation would be *larger*  for a smaller total  N.

[ snip, rest ]

Correlated?  Or negatively correlated?  Or assumed to
be not-correlated?

Ignore the scores for a moment, except Some vs None.

Here are the cells, for Left versus Right,  if the table is 
assumed to be disjunct, where 103 and 132  represent
the cells with R-not-L  and L-not-R.

   0 + 103
132  + [the rest]

This can be tested with McNemar's, and a good approximation
is the difference D=(132-103) squared, divided by their sum S:
And that gives  a chisquared test of  3.58, which is not enough
to leave  anyone heavily impressed.  
If there were *correlation*  which would mean that L&R  tend to
coincide, the table would have the same D, but a smaller S:

 50 + 53
 82 + [the rest]

Here the test is 6.23 -- enough to be a bit impressive.
Keep in mind:  The proper test uses the proper overlap.
The "independent"  t-test  would require a proper N  for the
cell marked [the rest].   I believe that if we stick in any 
number that is rather large, this test will come out intermediate
between the two versions of McNemar that I just demonstrated.
At least, that is what I get when I test the numbers as
Poisson counts, using the square-root transform
(so the results for each sample have SD= 0.5).


Those are results while ignoring the rated severities
of disease.  When you put in those scores again, 
you can write similar models, with similar results.  
You cannot do an "independent t test"  unless you have
the extra cases for the diseaseless-cell.   - Note that the
question is *not*  whether the 132  cases have a higher
'average score'  than the 103 -- the profiles tend to match,
and that is not a problem.  In fact, the question might be 
construed as whether the 132 have a higher *total*  score.

-- 
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