Dear Colleagues:

I found a somewhat unusual way of using the confidence interval (CI) 
& prediction interval (PI) in multiple linear regression (MLR) in
practice and I'm wondering someone could help me

We have a multiple linear regression equation, let's say Y = b0 +
b1*X1+b2*X2+b3*X4+b4*X4. The derived equation is a very good one after
checking residual, p-values, multicollinearity, etc...  I understand
how to compute confidence interval estimation of the mean response and
prediction interval for a new observation in MLR at a particular point
(e.g., average values of independent variables) as usual from text
books.

However, I found that some people use the predicted-Y values
calculated by using the above equation as an independent variable and
the measured Y is regressed on the predicted-Y again (not independent
variables!).  Then CI or PI are computed at a nominal value of the
Y-predicted  (e.g. average of the predicted values).  By doing that,
they can draw CI and PI on a sheet where y-axis is actual Y values and
x-axis is predicted Y values.  I see their needs to show some errors
in prediction:  if we have more than 2 independent variables, it's
very difficult to draw CI or PI on a sheet. And they said that the
above equation has +/- 50 kg in errors.

I believe that the current practice is not correct.  If the above
approach is not the right way, how can I say in more elaborate way not
to use such an approach rather than say simply "It's wrong!"? Am I
asking a wrong question that is impossible to answer?  Or I have to
have them/myself in a training course.   Any thoughts or references
are welcome. I read through a few books on MLR and the "Statistical
Intervals" written by Gerald. Hahn but found no such practice.

Thank you very much.

Sangdon Lee, Ph.D., CQE, CRE 
Michigan
.
.
=================================================================
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