I have faced a problem in a couple of models while conducting CFA. I have a big model say 6 or 7 variables and around 40 items. When I run CFA, global fit indices don’t show an acceptable fit. Usually RMSEA and RMR and CMIN are in acceptable ranges but CFI and GFI are below 0.90 or even 0.70. it doesn’t get much better after deleting insignificant items etc. On the suggestion of a colleague, I ran CFA separately for each construct one by one. CFA for single constructs shows good fit. My question is: can we report CFA for separate constructs one by one in a research paper? Or it must be reported for all the factors altogether?
In the case, you can mention CFA for each construct but at last, it is necessary to go for full CFA. Moreover, it's very rare that your individual construct CFA are significant and final measurement model is not. Must check the model and minimise modification indices. This may help you out in increasing CFI, GFI and other indices.
This paper may help in your research: Cheng, C. C., & Shiu, E. C. (2012). Validation of a proposed instrument for measuring eco-innovation: An implementation perspective. Technovation, 32(6), 329-344.
Morning, Sharjeel,
I think you should run the complete mod3l since by the CFA convergent and discriminant validity issues are of the main importance. If you cannot show that you are measuring the right variable with the right items, then it seems to be problematic.
Some of researchers in cases similar to yours report only some model fit statistics (simply ignoring RMSEA or CFI, or others). I think this is not good at all.
Try to work on your model. May be you can delete a factor from your model? Maybe some items are interdependent, so modification indices might help? May be you can split your sample and work with a part of it? Sometimes adding an additional factor helps?
Sincerely,
Eugene
what is your sample size? small sample size can also cause model fit problem. For each manifest variable you must have 5-10 respondents. If you can not increase the sample size you can reduce the manifest variables by parceling the variables.
sample size is really important and may I please know if you have translated some scale? Faced this issue many times especially with scales having more than 30 items..Some constructs may not emerge in the same way in our culture as we expect them to be as per existing theory. We usually go for an EFA later. because if modification indices are suggesting too much change...better to go for an EFA then..thats what we practice in cross cultural validation of psychological instruments
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers