1. For my one instrument I added items - therefore I have run an EFA - do I then need to run a CFA before using the scales for my other stats?
2. After running a CFA on a different instrument from question 1, a pre-established instrument with just slight tweaks, the model is just shy of meeting the accepted RMSEA and CFI. How do I determine which items to delete, or how to tweak the model to increase model fit?
Dear Daniel de Freitas
I will go by your questions one by one.
1. The explanatory factor analysis is generally used once there is no theory and the goal is bring a sort of model to the table or the researcher is developing a new scale to measure a specific dimension using a set of newly designed (items/questions) . However, some scholars still use in a favor of ensuring the unidimensionality of a newly developed (construct/factor). A main emphasis about EFA, that it allows the data to play the role of specifying which (items/questions) go to which (construct/factor). In the other hand, the confirmatory factor analysis is often applied as a step of the implementation of the latent variables modeling using the structural equation modeling. The main aim of CFA is to ensure the (construct/factor) reliability and validatly is fulfilled to further proceed with the structural model. Therefore, if you are aiming to use the structural equation modeling in later steps, I would suggest that you rely on CFA solely, while if your aim to develop a new sacle for a (construct/factor) or a new theory is to use the EFA.
2. Well, at first let's pay an attention that those fit indices are trying to measure the difference between the initial correlation matrix and the caclduated correlation matrix with the spefication of the model. I would assess the average variance extracted of each (construct/factor) and check what then minimum value is, which can be a starting point if the (construct/factor) has relatively many (items/questions) then dropping the ones with the lowest can an approach to try. In addition, if you are using Lavvan package in R, you can use a fit. Indices function where it would useful information for what might change if you want to drop any of the items.
You must take into account the previous literature and if you have adaptation of the instrument in your context, because not always EFA and CFA obtain the same results in the same sample, the adjustment values of the model in CFA can be the elimination of items that do not meet the psychometric parameters (e.g. factorial load, item-test correlation or communality), or even these models allow to correlate the errors of items similar or belonging to the same factor (theoretical justification) in order to improve the proposed model.
Hii Daniel
I agree with Belal and Cristian advice. In addition, I have a few suggestions that may help. First, you need to justify the reasons for adding items to a pre-existing scale. Actually, this is very fundamental and reviews may raise questions (I'm telling you from my personal experience). If your justifications are convincing and contributing to a theory, then perform the EFA to determine the structural dimensionality, and CFA to determine the psychometric properties. However, they must be performed in different samples. Or divide the sample data (e.g. n=500) into two parts, perform EFA (nefa=200)in one and CFA (ncfa=300) in rest (such an approach could produce biased results, but still it is common practice).
Now about, fit indices, it can be achieved by examining the modification indices (MIs) and correlating error terms of items (belongs to the same factor) that shows high MIs. Though modification indices provide statistical support for such changes, these changes must be driven by psychological reasoning as well. Read the work by Furr and Bacharach (2013), for better insights.
Thanks
Sayoni
Reference:
Furr, R.M & Bacharach, V.R. (2013). Psychometrics: An Introduction (2nd ed.).
Thousand Oaks, CA: Sage.
Yes after doing an efa you just come to know the items are getting loaded in the right factor.It is CFA which tells you that the model has run or the path of how variables effect each other based on beta coefficient values
EFA and CFA are steps in construct validation. An adapted instrument should go through these 2 steps before proceed to the next step.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis