SEM Problem with AMOS:

[Bastia...@gmail.com]

Hello,

I got a problem with my structural equation model which consists of 4
latent constructs. Three of these are factors which are responsible
for a change of the 4th construct (unobserved endogenous variable).
The model is identified and a minimum is achieved. My problem is that
AMOS returns the following messages in the text output:

Model Fit: The following variances are negative. (error term of the
4th construct (endogenous variable))
Model/Group Fit: This solution is not admissible.

If I omit the error term for the latent endogenous variable these
messages are of course no longer in the output.

Now I am a little bit confused:
- Should I omit an error term for the construct, also it is
theoretically possible that there are other input factors than the
three constructs I identified?
- How could I care about the negative variance?

I hope someone can help me! Thank you very much...

Bastian

[Ryan]

On Apr 26, 5:18 pm, Bastian.P...@gmail.com wrote:
> Hello,
>
> I got a problem with my structural equation model which consists of 4
> latent constructs. Three of these are factors which are responsible
> for a change of the 4th construct (unobserved endogenous variable).
> The model is identified and a minimum is achieved. My problem is that
> AMOS returns the following messages in the text output:
>
> Model Fit: The following variances are negative. (error term of the
> 4th construct (endogenous variable))
> Model/Group Fit: This solution is not admissible.
>
> If I omit the error term for the latent endogenous variable these
> messages are of course no longer in the output.
>
> Now I am a little bit confused:
> - Should I omit an error term for the construct,

There are several reasons for a negative variance. It would be helpful
if you posted more information about your model, such as

1. # of manifest variables you have per latent variable
2. level of measurement for your manifest variables, including the
range
3. if by "justified" do you mean "overjustified?"
4. your total sample size
5. To get a better sense of the model, it would be helpful if you
provided the sets of equations.

I would not omit an error term. There are other ways to deal with this
issue. Moreover, a negative variance can reflect a very serious
problem with your model, and removing the symptom does not necessarily
cure the disease.

> also it is
> theoretically possible that there are other input factors than the
> three constructs I identified?

Yes, it is possible to have more than three constructs. Your decision
to have three constructs should be based on theory/previous research.
Perhaps I'm not understanding this question.

> - How could I care about the negative variance?

Again, I'd need more info before I can make a suggestion.

>
> I hope someone can help me! Thank you very much...
>
> Bastian

Sorry if my response wasn't very helpful. Here's a nice primer on SEM,
and you'll notice a discussion on possible reasons for negative
variances in the FAQ section:

http://www2.chass.ncsu.edu/garson/pa765/structur.htm

Ryan

[Bastia...@gmail.com]

On 28 Apr., 16:32, Ryan <Ryan.Andrew.Bl...@gmail.com> wrote:

> On Apr 26, 5:18 pm,Bastian.P...@gmail.com wrote:
>
>
>
> > Hello,
>
> > I got a problem with my structural equation model which consists of 4
> > latent constructs. Three of these are factors which are responsible
> > for a change of the 4th construct (unobserved endogenous variable).
> > The model is identified and a minimum is achieved. My problem is that

> >AMOSreturns the following messages in the text output:

Dear Ryan,
thanks a lot for your help. I tried to read something about my problem
and looked for literature on Heywood cases.

I've got 4535 cases and my model looks like this: http://www.baspo.de/SEM_1.gif
23 indicators load on 3 factors which forms 1 latent construct
(measured by two items).
I'm not sure whether the model is correctly identified, but a problem
could be a poor discrimination of my items. A pretty poor value is my
CMIN/DF = 66,314.
I'll try to read more and get a better idea of my problem, but if you
could give me some more help I was really grateful!

Best regards,
Bastian

[Ryan]

> Bastian- Hide quoted text -
>
> - Show quoted text -

Bastian,

From the graphical model, it looks like you've set it up correctly. As
you pointed out, a chi-square of 66 thousand indicates an extremely
poor fit. I cannot read all of the path coefficients, but I can see
that several of them are very low.

To be clear, the graphical model you have shown is NOT testing a
second order confirmatory factor analysis model, it is testing the
validity of a causal model. If you are proposing a second order CFA
model, then the current model is NOT correct.

So what are the typical steps to checking a causal model?

The first step would be to test the validity of the measurement model
BEFORE looking at the structural model (running the 4 CFAs separately
before looking at all latent variables together). Something tells me
that if you were to look at them separately, you'd find issues that
need to be addressed (e.g. very low path coefficients, cross loadings
etc.). You should address/fix these issues before moving on to the
whole model. The measurement of each latent variable MUST be
psychometrically sound, before you look at the relationships between
these latent variables. I am going to assume that you are familiar
with running and interpreting CFAs, as the causal model is more
complex. There are so many issues that you could potentially find that
it wouldn't make sense for me to speculate.

I cannot stress enough that your model needs to be grounded in
empirical research/theory.

I have used Barbara Byrne's book, entitled "Structural Equation
Modeling with AMOS," as a basic guide when "testing the validity of a
causal model." Another great book is "Principles and Practice of
Structural Equation Modeling" by Rex B. Kline.

HTH

Ryan