I'm using AMOS 20 to analyse some data. I have four humour styles (latent variables), three types of peer-victimisation (PV), gender, and a measure of depression (latent variable).
The model is one where all three types of PV have direct effects on depression. They also have direct effects on the four humour styles, and the four humour styles have direct effect on depression as well.
When I run this model using observed indicators for the three PV variables (using a simple mean of the relevant items for each type of PV) I get a good fitting model, and sensible standardised estimates. However, when I replace the PV indicators with full measurement models the estimates change radically. In the second model, with PV measurement models, the standardised values are totally different to the standardised values in the first model. Not only that, but some of the values are very high and one is 1.23. However, in model two the unstandardised estimates are very similar to the standardised estimates in model one.
Any ideas why this is happening? The only explanation I can think of is that in the second model all paths between latent variables are paths between varaibles which are already standardised, and therefore the unstandardised estimates should be used. But I don't want to rush ahead if this explanation is wrong.
My view is similar to that of Levente Kriston. If a path analysis model also includes measurement models then it is a good practice to test the measurement models first and then to include them in the structural association model. Thus, checking the goodness of fit of measurement model for peer-victimization, humour style, and depression before testing the structural association among the latent and observed variables may be helpful.
Thank you all. I'm glad to say that I had already tested the measurement models before putting them together as a full SEM, so the problem should not be with a poorly fitting measurement model. sample size may, I guess, be a problem although we have a sample of 1242 so I'd not considered that it would be problematic. I'll check that though.
Thanks for those other suggestions (a-d) Levente, they are helpful to consider.
Go back and check the distribution of your data first. Look for outliers and check distributions. The full measurement model is the correct individual model, discard the mean centred model. Usually standardised measures are reported, I would not worry about the unstandardised measures. Also think about restricting one of the measurement paths , or item loadings to 1, when you originally run the model. You also need to tell use what the fit indcies are. SEM is all about finding the best model that fits the data, rather than any model that fits the data. If you are unsure about all this compare your SEM measurement model to a simple Factor analysis, are the structures similar? Lastly, have you tested competing SEM factor models of one, two and three PVs? You need to do this order to show evidence for your accepted factor structure. Think also about whether you have a formative or a reflective model for the measures. This requires a close look at the theory and the way the items have been worded.
Hope this helps!
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics