On Thu, 16 May 2002 15:48:37 -0500, "Abram Fridman"
<[EMAIL PROTECTED]> wrote:

> Hello,
[ snip, first questions...  where local statistical advisor
is communicating poorly, or erroneously. ]
 
> If I should not, then I have an additional question:
> SPSS has an option for computing multiple regression without
> coefficient C. It makes a lot of sense for my model: I need
> force equal to zero to "break" specimen with zero thickness.
> 
> When I turn this option ON R increases. Why it happens?

I thought that SPSS gave a warning and explanation 
these days.  This has been discussed in the stat-groups
more than once, and my (old)  FAQ includes comments.

One value that can be called R-squared  will increase when
regression is forced through zero  because the model has been
changed:  The "sum-of-squares accounted-for"  is no longer
the sum around the mean, but has become the sum around zero.

The uniform opinion of previous discussions, as I recall it:
Forcing regression through zero is only rarely justifiable,
and even if you can justify it, you probably are better off, NOT.
(WHY?  Models are often weird near zero, so the intercept could
be an improvement, even though you do not justify it.
Or, the fitted value is either essentially zero, in which case
it does not hurt the fit; or it is not zero, which warns you that
something odd is going on.)

Using the definition of R-squared where it is computed around
the mean, regression forced through zero can give you a 
negative R-squared -- when 'prediction'  gives more error
than using the mean to predict.  That bothered me at first, but
now I feel happy to recommend that people stick with that one.
Though, I could allow special cases.

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
.
.
=================================================================
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