On Thu, 31 Jul 2003 16:59:36 +0200, agentsmart <[EMAIL PROTECTED]>
wrote:
> [ snip...  down to, univariate tests ]
> 
>       Y-X1    Y-X2    Y-X3    Y-X4
> r     ,50     -,35    ,01     -,27
> p     ,005    ,061    ,951    ,156
> 
[ ... to multiple regression ]
> 
> dependent var: Y
> independent: X1, X2, X3, X4
> 
> Adj. R2 = 0,37
> p=0,003
> 
> intercept: B=97,5 ; p=0,000
> 
>       X1      X2      X3      X4      
> beta  ,29     -,36    ,83     -,70    
> p     ,09     ,051    ,011    ,034
> 
> 
> my conclusions:
> - there is a significant dependance between Y and X3 and X4
> - there is no dependance between Y and X1, X2
> 

Gus offered some corrections to language. 

Here is some more stuff that is not said quite-right.
The multiple regression shows that  --
'The near-significant relation of X2  is practically unchanged; 
the apparent effect for X1  is reduced by half, 
so that its partial contribution no longer meets 
an arbitrary test criterion of p < 0.05. 

'The beta coefficients for X3 and X4  are newly significant.
Also, they are large and in the opposite directions -- that 
is a pretty robust  indication that what counts for prediction,
coming from them, is some measure of the *difference*
between the two.'

Gus gave a logical example.  I am emphasizing that
the large standardized betas are a diagnostic clue,
always, that a difference probably matters most.
One proper step in model building, if it makes sense, 
is to compute the 'difference'  (left leg minus right;  
or longer leg minus shorter?)  and use that in the 
prediction instead.  - When I see two variables that
are *very*  highly correlated, I automatically consider
whether I should use just one variable;  or replace 
the two  with a sum-and-difference to represent 
the 2 degrees of freedom.

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
"Taxes are the price we pay for civilization."  Justice Holmes.
.
.
=================================================================
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