SEM or path analysis using AMOS, issues with sample size and model complexity

Question

I am new to SEM / AMOS and as such I have done a lot of reading around my problem but find I am getting lost, any help would be appreciated and I am sorry if this sounds rather basic;

1. I have a model that looks at a number of factors (distal, indirect factors) which include 5 factors of personality and a couple of 'other factors' shown to be related to 2 distinct accident behaviours (outcome variables, DVs etc)

2. There are 2 mediating factors between the distal and outcome variables

3. the mediating variables and outcome variables are bi directional so I have an non recursive model.

4. I have relatively small sample of 200 for the purposes of SEM depending on the perspective you take on this.

I understand the principles of SEM and here lies my problem, to include a measurement model with observed variables that underlie these factors is not possible (the 5 individual factors are measured by a 120 item questionnaire), this puts unrealistic demands on sample size and simply drawing a model like this seems impossible.

I was looking at the use of Path Analysis which I understand is a form of SEM. Here I treat the factors as observed variables so lose the ability to account for measurement error. There are also issues with testing a non-recursive model here.

I should say that I conducted a series of multiple regressions to establsh the best predictors of the 2 mediating and 2 outcome variables to form I suppose a 'conceptual' model that I would like to test and test the mediating variables. I hope this makes sense to someone..The more I read, the more I am getting a little overwhelmed and was hoping someone could offer some guidance or comments?

Many thanks in advance

Answer 1

Wow. 120 items is a lot. Are they yes/no (0/1)? Do you just use a simple sum?

You could try approaching this with fixed reliabilities of your factors. Compute your scores and store them as say f1score through f5score. In AMOS, specify latent variables as circles with f1 through f5 in them, and connect them to the rest of the model. Now, connect each of the latent with the corresponding scores, f1 -> f1score, etc. Set the path coefficient to be 1; set the measurement error for f1score to be 1 minus the (generalized) reliability of the score. If you don't have the latter from previous research on the psychometric characteristic of your scales, you would have to compute that, too; see Raykov's work on the topic.

Be aware that regression when regressors are measured with error attenuates coefficients (biases them towards zero). See also Skrondal and Laake 2001.