Hello,
I intend to develop a social scale that can be used in conjunction with other dimensions of a modular framework.
Reading the literature, I found three potential social dimensions. I conducted a focus group and came up with 7, 8, and 12 items for them.
Similar to other scales developed for the aforementioned framework, I ran three separate EFAs and retained one factor for the first two dimensions each with 4 items.
The third dimension (with 12 items) retained two factors. However, there was one item with cross-loading of 0.387 and 0.390 for two factors (CLS, BRZ). I know I should have removed it, but I kept this item for the second factor because I wanted to have the 4-item format for all the dimensions, and theoretically, it makes sense based on the items.
I have collected another sample to conduct CFA. It does not give great goodness of fit indices but my main problem is the high correlation between the two factors I mentioned above (0.91). My understanding of another thread in researchgate is that one can use second-order CFA to solve discriminant validity of highly correlated factors. Is it correct? Can I do it?
You can see the AMOS results attached.
One more question: in case I delete one factor altogether should I go back and repeat the EFA or I can continue with the 3 factors.
Thank you
P.S: I am not very knowledgeable in statistics so I would be grateful if you could explain it a little. Thank you
A second-order factor does not "solve" the issue of two or more factors being highly correlated. It merely allows you to "model" this correlation. The more important thing for you to think about is why the two factors are so highly correlated. For example: Could the high correlation be due to shared item content? Are those really two distinct dimensions or could they be summarized into a single dimension?
Also, you mention that the model(s) showed a suboptimal fit. This should be addressed first because a non-fitting model may result in biased (incorrect) parameter estimates that could be misleading. For example, the high factor correlation of .91 may in part be a result of model misspecification (e.g., omitted cross-loadings).
Hello Eshan,
That the two factors (CLS, BRZ) in the first model have such a high correlation and correlated virtually identically with the other two factors argues that Christian Geiser 's reply is on target: The CLS/BRZ pair appears to be more a single factor than two distinguishable factors. You'd probably fare better by going back and reconceptualizing the merger of the two as one factor rather than trying to paste in a second order factor to "fix" the problem.
Good luck with your work.
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts