On 1 Feb 2004 13:19:12 -0800, [EMAIL PROTECTED] (Niko
Tiliopoulos) wrote:

> Dear all,
> 
> I hope someone can help me out with the following issue, because it is
> giving me a headache.
> 
> For practical reasons I will simplify the problem dramatically.

 (the answer, in detail, might hide in the simplification.)
> 
> I have a number of manifest variables, some acting as predictors and
> some as outcomes. One of those predictors has a behaviour I do not
> understand:
> 
> 1. In bivariate correlations it has a positive association with all
> outcomes
> 
> 2. In regression models (all predictors) against each of the outcomes,
> it still shows positive (standardised) betas
> 
> BUT
> 
> when in a structural equation model (with all predictors making a
> single exogenous, and all outcomes a single endogenous variable) its
> regression weight sign is minus!
> 
> In addition:
> 
> This predictor does not have high correlations with the rest of the
> predictors (approx. max r = .30)
> 
> Frankly I do not understand this behaviour.
> 
> Any ideas will be greatly appreciated

The first thing I would do:   Be 100% sure about the
behavior.  Can you confirm that the N  is exactly the same 
in all analyses?  Is there ever any problem or question 
about Missing codes?

The second thing:  Make sure about the collinearity.
Okay, the maximum r is 0.30  with predictors.  
What about the multiple r?   (No dummy variables
involved?  nothing that makes part-scores, when
the total is also around?)  

Third item:  What is the total 'fit'?  What is your
r-squared, and what *should*  it be?  What warnings 
do you have for overfit, or for bad fit?
 - I don't do structural models, but my understanding 
is that 'fit'  is problematic.  I guess that I would expect
your problem to arise somewhat-legitimately from a
overfitting,  maybe as a symptom of odd scaling.  

I hope this will help -- your point (2)  does tell me that
you have considered - already ruled out -  the most 
obvious aspects of what I have in mind.

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