item removal

[Sara]

Hi everyone,
I'm working on a CFA model using the MHI-5, which includes 5 items: nervous, depressed, calm, down, and happy. I reversed the negatively worded items so that higher scores consistently indicate better well-being across all items.

Based on prior EFA results , a one-factor solution seemed appropriate. I ran a CFA using a single latent factor, but the model fit indices are poor 

I then tried removing Item D (calm), and the fit improved considerably. However, I’ve read that high modification indices alone are not sufficient justification for removing an item. I'm hesitant to drop it unless there is a stronger theoretical basis. I suspect there may be content redundancy between calm, down and depressed. Do you think removing calm can be justified in this case? Thanks in advance

[Matt Diemer]

Hi Sara,

Thanks for sharing this info with all of us. I have a few questions/ideas:

1) What were the factor loadings in the CFA for Calm and the other items?

2) Looking at the ModIndices makes me wonder if shared error covariances might be another solution to this problem, given the chi square decrements associated with those. Do those error covariances make sense, theoretically?

3) Sometimes the model fit estimates can change from EFA to CFA bc the estimation method changes. Is your estimator (ML vs MLR vs WLSMV) the same across EFA and CFA?  

4) Calm and happy seem to be the only naturally positively valenced (i.e., before recording) items. I'm surprised they don't share a stronger relationship - more of an observation.

Hope this helps...

matt