An example I used to assign in an intermediate course, though not for
this purpose:  data were from a textbook (Cochran, I think), being dry
weights of chick embryos at ages 6 to 16 days, at 1-day intervals.
Very nicely fitted an exponential growth curve (R^2 = 97% or so).
But to fit the curve  y = a*e^(b*x) one took logarithms of the response
variable:  log(y) = log(a) + b*x.  Here the R^2 based on the obvious
ANOVA is the square of the correlation between <log(y)> and <predicted
value of log(y)>.  If you take the regression results and back-transform
(using antilogs = exponentials) to get actual predicted values of y, you
can then correlate those with the original (or observed) y.  But these
are now two different variables.  Related to those in the ANOVA R^2, but
related nonlinearly;  so the square of that correlation would not be
expected to be identical to (SSpred/SStotal).  (In this case the two
values of R^2 are necessarily quite close, because they're so close to
1.  Haven't investigated data with a less compelling fit, so don't know
what to expect in general.)

[If you want the data to play with, ask.  I've got it somewhere around
here, but it will take a little time to find it.]

Cheers!   -- Don.

On Fri, 26 Mar 2004, jim clark wrote in part:

> I would really like to see an example where the correlation between
> the original y and the predicted y does NOT equal the multiple R
> defined as sqrt(SSpred/SStotal).

 ------------------------------------------------------------
 Donald F. Burrill                              [EMAIL PROTECTED]
 56 Sebbins Pond Drive, Bedford, NH 03110      (603) 626-0816
.
.
=================================================================
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