It is a established practice in psychology and management that in a test development process exploratory factor analysis is essential to determine/define the factors/construct of test. However, in my opinion where the construct are well defined through established markers, EFA will not serve a useful purpose and we may consider CFA to validate the tool. Please share your view with suitable justifications. Thank you in advance.
I agree with you. In many (if not most) social science applications, EFA is not needed because we don’t need to explore the number of factors or loading pattern. Our theories or prior studies already predict a specific factor model. We can just go ahead and test it with CFA.
CFA is better suited to test a hypothesized factor model and/or to statistically compare competing factor models (e.g., a 1-factor solution vs. a 2-factor solution). Therefore, in most situations, my recommendation would be to only conduct CFA.
Nonetheless, EFA can be useful, for example, when you are truly exploring a set of variables without any prior theory about which dimensions may be reflected by those variables. Also, EFA can be helpful when a CFA shows a very bad fit and a researcher has trouble figuring out why the solution does not fit. Also, in some cases, a hybrid approach (ESEM) can be helpful, for example, when there are many cross-loadings that are different from zero.
Factor analysis is dependent on the assumptions you make, for instance about whether the factors are correlated or not. And staring at a table of factor loadings and uniquenesses is more like a Rorschach test than a statistical procedure.
May I commend Mokken scaling, which is nonparametric, iterative, based on a strong definition of a univariate scale and capable of detecting items that just don's scale. Where the item pool contains correlated subscales, you can detect this situation by increasing the threshold value of Loevinger's scaling coefficient.
But the main advantage is that at the end of the exercise you have a definable, defensible scale.
I am thankful to both the learned faculty for their valuable and insightful comments, in fact I was expecting more comments opposite of my view to have more deliberation and it will beworthwhile to mention that even when there is a multidimensional scale and subject to parametric test.
My warm regards and good wishes on the eve of new year 2024 to Professor Ronán Michael Conroy and Professor Christian Geiser.
Theory may not be sound. I'd always begin with a parallel analysis to detemine that I at least assumed tthe correct number of factors.
Again, it's nice to know in advance whether your theory is correct on even the number of factors. It would enable an earlier start to further theoretical fishing.
Mr. Kohn, I was waiting for some more comments before sharing my experience. I do must share that I did a study using open-ended semi structured questionnaire using critical incident techniques and analyzing data by phenomenoligical and qualitative content analyses on a sample N=2000 and could identified a number of behavioral makers for 24 behavoral competencies and large pool of items were finalized for each competencies with help of 12 SME and after performing items analysis, I preferred not to go for EFA as my variables were well established and performed CFA and results were quite satisfactory. We have over 3400 user feedback who used the tool for self assessment and reported that the test accuracy level is 80% or more on every behavoral competencies.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python