On Fri, 25 Jul 2003 18:03:52 +0200, Peter Frank
<[EMAIL PROTECTED]> wrote:

> Rich Ulrich <[EMAIL PROTECTED]> wrote:
> 
> >On Wed, 23 Jul 2003 17:50:24 +0200, Peter Frank
> ><[EMAIL PROTECTED]> wrote:
> >
> >> Which other study designs yield paired data other than before/after
> >> experiments?
> >
> >Left-right.
> OK. I can pair data from an individual's right hand with data from his
> left hand.
> 
> >Male-female.
> How can I specifically pair data from males and with data from
> females?

>From each litter of puppies, take one (at random) of each.
The fact of litter-membership defines the 'correlation'

> >Top-bottom.
> >Hot-cold.
> If these are factor to be examined but not the subjects, I understand
> it, otherwise I don't (see above).
> 

Is there something that defines a 'correlation'  between
the members? - if so,  then you might treat them as
paired  for an analysis.   This is a somewhat subtle point.
Correlation exists with-respect-to  something outside.
One hundred trials with my right hand might be 
considered uncorrelated -- if there was an experiment 
designed within those 100 trials.  However, they are
self-similar compared to trials with my left hand.  And
both my hands are self-similar (for me) compared to 
trials done with by someone else.  

The classic studies of 'case-control'  often started with a
limited number of 'cases'  and  matched each Case with the
candidate among 'controls'  who is the same (or closest)
with respect (say) to age and gender.  

These case-control studies are less popular today because  
(a) bad historical matches yielded some bad analyses; and 
(b) the statistical control by  covariates tends to be more
powerful; and 
(c) statistical control by covariates is more meaningful, since 
you can have tests on those covariates.

'Case-control'  is not a bad idea for collecting data, at times
when you don't want the samples to have disproportions of
various characteristics -- I am thinking especially of cases
where it is cheap and easy to screen for potential 'controls'
but expensive to actually carry out the exams to put them
in the sample.   Otherwise, you should add all your recruits
to the Control, to boost its N  and representativeness.

But the statistical analyses should be done as Case-control 
only when the pairs are *highly*  similar.  Most often that 
I think of, those would be pairs of family members.

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