Dang, Jeff wrote:
> Edstat,
> 
> I have personally found that a lot of health researchers like to aggregate
> normally distributed, continuous outcomes into dichotmous outcomes.  In some
> cases, this is done because the researcher is more familiar with dicohotmous
> outcomes (disease/no disease) and seeks to interpret their results in terms
> of odds ratios within a logistic regression.  
> 
> In some cases, this can be problematic because you lose information.  For
> instance those near the cut-off point are forced into one group or another.
> Thus, you exaggerate the differences for some individuals.  
> 
> If anybody on the listserv can refer me to articles related to this problem
> I would be most appreciative.  Thank you.
> 
> Jeff Dang  


The following is from a thread that appeared on the same 
topic last year.  I don't know if the links given are still 
active.

Cheers,
Bruce



On 2/4/03 2:25 PM, "Wuensch, Karl L"
<[ mailto:[EMAIL PROTECTED] [EMAIL PROTECTED]> wrote:

 > Thanks to all who have participated in the discussion on
 > dichotomization. I have logged selected parts of the 
discussion into a
 > document for both my students and my colleagues to read. 
It is available at
 > [ 
http://core.ecu.edu/psyc/wuenschk/StatHelp/Dichot-Not.doc 
]http://core.ecu.edu/psyc/wuenschk/StatHelp/Dichot-Not.doc
 > <[ 
http://core.ecu.edu/psyc/wuenschk/StatHelp/Dichot-Not.doc 
]http://core.ecu.edu/psyc/wuenschk/StatHelp/Dichot-Not.doc> 
. If any of
 > you would rather not have your comments included in that 
document, please
 > let me know.
 >

---- Gary McClelland replied ------

I've put together a Java applet that illustrates the issues.

http://psych.colorado.edu/~mcclella/MedianSplit/

There is a slider that allows you to move the values of X 
smoothly from the continuous analysis to the dichotomized 
analysis. Doing so you can see that as the variance of X is 
thereby reduced, the correlation is reduced, but the
estimate of the slope remains unbiased. The effect of 
dichotomization is simply to reduce statistical power. If X 
is normally distributed, the effect of dichotomization is 
equivalent to discarding a random 50% (approximately) of 
one's observations. I can't imagine why anyone would
want to do that.

Note: to use the above applet, PC users will need to have 
downloaded Sun's Java plug-in to replace the crippled Java 
that Microsoft distributes. Mac users will need to be using 
OS X.

gary


-- 
Bruce Weaver
[EMAIL PROTECTED]
www.angelfire.com/wv/bwhomedir/

.
.
=================================================================
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