I have run EFA on 50 items scale, EFA supported 35 items structure with 5 factors (eigen value >1, variance 68%, screeplot supported 5factors and PCA as well). Correlation among factors were quite low and some are negatives.
While running CFA, items reduced to 33 with the same model structure and this is supported only when I run each factor independenly in CFA.
Later I ran second order model where all five factors showed good model fit indices.
I am struggling in presenting my findings in a paper/thesis. Do i need to present tables/figures for each items separately as i did CFA ? or there are other options?
Any similar article/reference is appreciated.
It seems that you developed a new scale (50 items). In this case, you'd better run EFA and CFA with different samples. If your sample size is big enough, you can divide them into half: run EFA with one half, and run CFA with another half.
Then, report the EFA results with all the items, and report CFA results with the fit index (without items). you can refer to some articles published at top-tier journals. Meanwhile, you can also refer to the following article that developed a new scale and report EFA and CFA.
·Kim, T.-Y., David, E., Chen, T., & Liang, Y. (2023). Authenticity or self-enhancement? Effects of self-presentation and authentic leadership on trust and performance. Journal of Management, 49(3), 944–973. https://doi.org/10.1177/01492063211063807.
Hi,
as the EFA pointed to 5 factors, I would then approach the next step from the perspective of these 5 factors and select 3 items for each factor. These items would be the theoretically most salient given the interpreted meaning of the respective factor from the EFA.
The question whether to apply a CFA in the same data set is quite old but, against conventional wisdom, applying a cross-validation approach does not make much sense for two reasons:
1) CFA/SEM is a linear model and as any other linear model its variance across samples is the lowest of all models. Hence, there is not a big chance to overfit the data. As a consequence, no matter what you did in the first part, is what you will get in the second. That can be easily simulated. The more severe problem is misspecification of the structure which will be stable across samples (or parts of the sample)
2) you loose power to test the structure due to the lower N
Hence, my proposal (which is risky due to the potential diverging expectations of readers/supervisors/the commitee) would be:
1) Select the best indicators (see above)
2) Use the total sample
3) Now comes the key: Include variables into the model that should have some specific causal connections to your factors *conditional on your interpretations* of the factor. This will add new constraints that were not part of the EFA and these new constraints will be tested.
And: TEST your CFA model with the chisquare test.
https://www.researchgate.net/post/Model_fit_measures_in_CB_SEM
https://www.researchgate.net/post/Can_I_use_EFA_before_SEM_analysis
Hope that helps
Good luck
Holger
- Badges
- Science topic