On Sun, 01 Oct 2000 21:15:17 GMT, [EMAIL PROTECTED] wrote:
>
>
> My Rsquare is 0.99834 and Adj R square is 0.99377
- so your Error, which is now .00166 (instead of .00460 - what
happened to the 0.9954?), has increased by a factor of 3.75, to
.00623.
a) That is quite a bit bigger.
b) That is an underestimate of the increase: Your adjusted R-squared
computes as (1.0 - 22/8* Error), where 8 is your revised d.f. after
pulling 5 terms out of the original regression. Doing a computation
based on the original 3 d.f. is more conservative, and (usually) more
accurate.
c) I am guessing at your formula as using (n-k) in the numerator,
(n-k-1) in the denominator. I know that there is more than one
formula for shrinking the R-squared.
c) Cohen & Cohen's formula (3.6.4 in the 1983 edition) would make
that 30/8* error, assuming you were using the 8 d.f. Or it would be
30/3* error, for the original d.f.
If you are going to do "stepwise," you really can't rely on the
reports of significance of tests (or size of shrinkage) after
arbitrary stepping.
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
