Jamin Day posted on Friday, June 12, 2015 - 5:05 am
The model fit for my EFA is okay (but not ideal), and on closer inspection of the rotated factor loadings there are some items cross-loading (>0.4) on multiple factors, and others that load poorly (<0.3) across all factors.
I have 79 items and am using a 7-factor solution so there is some room to drop items to improve model fit but I'm unsure where to begin RE deciding which items to remove. I realize any action needs to make sense substantively, but my question is around whether the Mplus output offers any guidance or statistical indicators that might help guide this process. I tried MODINDICES (ALL) however for each n-factor model I get two (very) large theta matrices, labelled modification indices and expected parameter change respectively. I'm not sure how to interpret these - for example should I be looking for items with large values...? Or am I missing something?
A large modindex for a Theta element tells you that two item correlate beyond what their factors can explain. That can have big or small reasons. It can call for adding a factor or call for deleting one of the items due to common phrasing.
I think it is very hard to rely on statistical procedures to drop items. We discuss substantive reasoning in our Topic 1 handout and video on EFA.
Jamin Day posted on Tuesday, June 16, 2015 - 1:02 am
Thanks Bengt. I have watched the video, very helpful.