This seemed basic enough, but I'm having trouble coming up with the
right search terms to do my own research...
I've been given a project that wants a logistic regression using
classification in the upper or lower quartile for a standardized
continous effector as the explanatory
a la a proc logistic model like
   model finish = score;
where finish is a 0/1 event and score is 'low' or 'high' (i.e. only
data from the upper and lower quartiles used)

This seemed popular and intuitive enough, but isn't an approach I've
used so i went looking for the rationale or the 'here's a better
modern way'. Why would this be advantagous? I am apparently searching
on the wrong line.

Can anyone point me in the right direction?

In educational research (at least when I took the courses about 100
years ago or so) a similar argument was made for comparing the upper
and lower 27th percentiles which appears to be very similar.  You can
find some references for that approach at:

 http://www.rasch.org/rmt/rmt121r.htm

HTH,
Art
--------------
On Aug 24, 12:30=A0pm, A Sanders <asand...@ufl.edu> wrote:
> This seemed basic enough, but I'm having trouble coming up with the
> right search terms to do my own research...
> I've been given a project that wants a logistic regression using
> classification in the upper or lower quartile for a standardized
> continous effector as the explanatory
> a la a proc logistic model like
> =A0 =A0model finish =3D score;
> where finish is a 0/1 event and score is 'low' or 'high' (i.e. only
> data from the upper and lower quartiles used)
>
> This seemed popular and intuitive enough, but isn't an approach I've
> used so i went looking for the rationale or the 'here's a better
> modern way'. Why would this be advantagous? I am apparently searching
> on the wrong line.
>
> Can anyone point me in the right direction?

When you run the logistic regression with just the two quartilies, you
distort the statistical relationsips that exist. I recall my Proffesor
used to call that "throwing away statistical information". The same
argument goes for any other regression analysis.

If you really what to compare the lower and upper deciles you can
always call on the CONTRAST statement in PROC LOGISTIC.

T!