Hi Mike.

There are a number of things that can result in an unidentified model. Some are simply operational in that mistakes are made in instructing AMOS how to define the model. These are easily avoided. Others are due to mis-specifications of the model, limitations in the data, or complications in the nature of the model. These are less easy to detect, but some of them are identified by the AMOS program and warnings given.

Looking at your model, having only one indicator for every latent construct is definitely causing this problem. you need to have at least 3 indicators for every construct.

Good luck

Michael, at first glance it appears to be because your latent variable only has two indicators. With more than one latent variable, this is acceptable, but not with a single one. Check out http://faculty.ucr.edu/~hanneman/soc203b/lectures/identify.html#obrien.

Also, why is the path from A to your latent fixed to '1'? 

http://faculty.ucr.edu/~hanneman/soc203b/lectures/identify.html#obrien

Hi Mike.

There are a number of things that can result in an unidentified model. Some are simply operational in that mistakes are made in instructing AMOS how to define the model. These are easily avoided. Others are due to mis-specifications of the model, limitations in the data, or complications in the nature of the model. These are less easy to detect, but some of them are identified by the AMOS program and warnings given.

Looking at your model, having only one indicator for every latent construct is definitely causing this problem. you need to have at least 3 indicators for every construct.

Good luck

You might try equating the loadings of the two items on the latent variable. 

Hi again. Thank you all for some useful insights.

@Jeffrey Williams - the path from A to latent was fixed to '1', as I though that the unit of measurement of each unobserved variable by suitable constraints on the parameters has to established. It turned out to be a problem (probably, as I'll explain below)

*

I 'experimented' with the model a bit and found the state when it is identifiable.

First, look to the attachment and see model 1. I removed a constrain from the A-->X regression and then the model turned out to be identified. Unfortunately, the strange values for regression from X to Alpha occurred (1,28), as well as the absurd value of R square (1,65) - see Model 1: estimation. Nevertheless model fit data perfectly (?) as Chi-square was ,972 df 9 p ,99.

In the second attempt, I constrained an error variance e3 to the specific value and then the model was identified and all the values seemed reasonable (see attached Model 2: estimation). However, it didn't fit the data...

Any ideas what happened? Is it right to manipulate the variance of error (e3), to set it up at the certain level (as a constrain), which make the model identified and providing reasonable values of parameters?

Okay, let's take a step back. Which variables are the indicators of the latent variable?  Alpha and Beta or A-D? Your model indicates that it's alpha and beta, which is why Steven suggested that you constrain the loadings to equality because the measurement model isn't identified as is. However, this statement makes me wonder if it's the other way around:

"the path from A to latent was fixed to '1', as I though that the unit of measurement of each unobserved variable by suitable constraints on the parameters has to established". 

If A-D are the indicators of the latent variable, then your model is backwards (the arrow from the latent should point toward A-D, not alpha and beta). 

@Jeffrey - In fact I wanted to model the following relationships:

The central latent construct (X) is dependent on observed variables A-D in a particular way (beta A, beta B.. etc). They explain altogether some - hopefully considerable - fraction of the X general variance.

In the same time the latent construct (X) loads Alpha and Beta in some particular way (loading X-Alpha and loading X->Beta).

I wanted to show these all on the one path model.

Now I suppose, that the fixation of the path from the predictor (A) to latent (X) probably made no sense.

Okay, then it seems you have it correct, but I agree with Steve in that you should try constraining your two factor loadings to equality (loading X-Alpha=loading X-Beta) because your measurement model is not identified. 

Thank You for good contribution.