Help with understanding CFA...

#1
I hope someone can give me some pointers here... Or just point out to my muddled mind the porobably obvious things I am missing!
I am running an online study where I have several statements that are worded to measure either A(say, attitude) or B(e.g. common sense), and these statements are rated on a 4 point likert scale. Most of these statements are based on a pre-existing questionnaire, so I know that it has at some point measured A and B satisfactorily. I feel like I should also mention that the original questionnaire has been vetted in the same way I am trying to do currently(new version is being vetted because I have changed some aspects and questions to try and make it more accessible for other languages, original q was not written or tested me). Also, my demographics this time are a different cultural group to the original set, which throws in cultural variation to consider as well. In addition to that I am also running this study on another set of participants of the original cultural group. I am hoping that the results will show a two-factor(A&B) model fit for both of these groups(i.e. questionnaire measures same thing regardless of culture). So this is the background.

Now I am trying to run a CFA using AMOS to see whether the data we have collected fits our original model, with the statements being the observed variables and A and B being the latent variables (previous q fitted the model satisfactorily). The current version does not fit the model at all, with fit indices as far removed from a good fit as possible. I have still meddled with it and taken out variables etc. just to see what may happen, but didn't get any change or ideas from that. In my prelim analysis I had noticed that there were some statements that had very high mean values and was wondering whether that may be making the difference..? But from my (very new)understanding of CFA, distributions like that should not really matter, right? As I would want to use the questionnaire to reliably measure A and B I am feeling a bit on the spot now... I know there was no real question in this, but does anyone have any thoughts on this?
 

Lazar

Phineas Packard
#2
My initial guess is that if your fit is "as far removed from a good fit as possible" then you have incorrectly specified your model. If you are new to CFA this might be likely.

However assuming this is hyperbole and you specified your model correctly, with a four point scale you might consider changing you link function to either probit or logit (you should be able to declare your items as ordered categorical in AMOS somehow and it should take care of this for you; why does everyone on this site use AMOS or Lisrel!). After that you will want to think about changin your estimator the default ML is most likely fine in this case (all be it a little slow) but if you are having problems see it AMOS has a diagonally weighted least squares estimator. If AMOS cannot deal with ordered categorical data for some reason switch to lavaan in R (its free).