Problem fitting an SEM with latent variable

#1
Hello,

I am having trouble with a Structural Equation Model.

As you can see in the attached picture, my main outcome variable (F4) is a latent variable. When I run the model, all of the paths are significant in the hypothesized direction, but the Chi-square is very high (almost 10,000) and other fit indexes are low.

However, if I create a satisfaction 'observed' variable first in SPSS (by adding the scores together of x1, x2, and x3 - the cronbach's alpha is .921 (answers on a Likert scale)), then run the model again, the model is fit much better - chi-square is 16.754 at a .005 probability level, NFI at .996, CFI .997, etc.

Is there something I can do to improve the full model which uses the latent variable? I can provide more information.

Thank you.
 

Lazar

Phineas Packard
#2
From the picture it looks like you have constrained the error terms to be 1. Hence there is not much left for the latent variable to explain. Free the error terms and then see what the fit looks like.

EDIT: To put this more plainly, your model is basically saying items x1 through x3 have variance but no covariance. I am pretty sure this is NOT the model you hypothesised.
 
#3
Why would you constrain the error term of your variable to be one (signifying essentially no explanatory value for that variable)?
 
#4
Hmm - thank you for the answers. Actually, I was just using the Amos defaults. When I release the error term for F2 and F4, I have some trouble running the model at all. However, I keeping playing with it to see if I can get it to run properly.
 
#6
Hmm - thank you for the answers. Actually, I was just using the Amos defaults. When I release the error term for F2 and F4, I have some trouble running the model at all. However, I keeping playing with it to see if I can get it to run properly.
You might also try Mplus over Amos. I think you would find the former much better than the later.