I am using R to perform an EFA on ordinal data using DWLS estimation, and used fa.parallel(data, cor="poly") to produce a scree plot. It gives two curves - one for PC and one for FA. The PC curve makes more sense theoretically as it produces 5 components, whereas the FA indicates only 3 factors, but I thought it is incorrect to use PCA for reflective variables such as mine. (I am looking for factors, not data reduction). When I continue with 5 factors, my cfa and sem work beautifully aligned with theory. Are there any problems using the PC curve instead of the FA curve to inform how many factors I extract?
I am curious: why did you perform PCA or EFA if you have a theoretical model? If you have strong theoretical ideas, it is always better to test this instead of PCA or EFA. If you have an article with information on the structure of the factors, you can test this. When you do a PCA or EFA, it might pick up on sample dependent idiosyncrasies you might not want.
Or do you mean something else if you say that is makes 'theoretically more sense'. If you even have theoretical ideas, they might be better then using PCA or EFA.
Why: because you are in fact testing yourself. Do I really understand what I am working on?
I prefer the betting model of Wim Hofstee to advance science. If you think you know something, you could bet on it. If you lose, you didn't understand it well enough. If you win, you knew. [ HOFSTEE, W. (1984). Methodological decision rules as research policies - a betting reconstruction of empirical-research. Acta Psychologica, 56(1-3), 93-109. ]. He also has a book in Dutch: de empirische discussie. (the emperical discussion].
Thanks Peter. I performed a cfa using a common model, the discriminant matrix indicated a 2nd order construct had emerged/was present in the data which needed explaining, especially given the fit was so-so. My research aim is not to fit a model but to see what our community thinks about a new technology, so I ran an eta to see what emerged naturally. The same 2nd order construct that I saw in the cfa emerged, but included another construct as well. When I ran a sem using this structure the fit was very good. I understand the theory and why it emerged and will report it in my findings.
My question is more about the validity of PC vs FA in an efa with reflective variables. I was under the impression they are different. My research assistant says 'they're the same thing just presented differently'. I want to write a robust method and results, and so I would like to understand more fully whether the efa technique in R runs a PCA and FA at the same time and puts them on the same scree plot. I would like to use the 5 components indicated by the PCA, but wonder if this is indeed valid theoretically instead of going with the three factors indicated by the FA curve in the attached image.
Hi Andrew Kemp ;
fa.parallel gives PC and FA results. They are a little bit different. Components means that there are umidimensional construct but it has some components. Factors means that there are non-correlated factors (but they may be correlate result of EFA).
fa.parallel results may not be correct. You can see our article about it.
Article Comparison of Factor Retention Methods on Binary Data: A Sim...
If you want to use parallel analysis for factor retention you can use Factor 10.09 software. ( http://psico.fcep.urv.es/utilitats/factor/soft/factor.10.9.02.WIN64.zip )
Because you can select optimal implementation for parallel analysis for more accurate results in Factor 10.09 software. I added ss for this selection.
Hope helps.
Best,
Faruk
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics