Results from Exploratory Factor analysis on my data (N=546) suggest 10 variables whereas Confirmatory Factor Analysis suggests combing them into 4 directions for better fit (No results with the 10 factor model).
Is it fine to have 10 variables in EFA and combing them into 4 in CFA?
Dear Mr. Butt,
10 are too many factors....do you have ten Eigen values more than 10? How much variance is being explained by those factors? Did you use the screen plots too to confirm the number of factors to be ten? If not just go ahead and get these facts varified first. And if, still you have so many factors then you should proceed for CFA as per the theoretical base of the relationship rather than to just improve the fit.
if you can share the results of EFA and CFA, we can be in better position to suggest.
Best wishes
Hi,
from a scientific viewit is not. Confirmation is used if there is a theory to be tested. If a confirmatory FA is used for exploration, it is an exploratory FA.
Please also reconsider the number of factors (10 variables I assume), which seems high to me. Also take into account, that Axis and Component-Analysis yield different results sometimes and can produce phantom-factors. But this depends on your theoretic approach.
Best, René
Mattiullah, Jan Rasmus Böhnke's explanation is very good.
Your scree plot shows that 10 components are sufficient to explain the correlations between the items. This is the exploratory analysis. If you think, that a second-order structure would be appriate, go ahead and test it. This would be either a confirmatory analysis based on theory or it would be another exploratory analysis if you used modification indices or patterns of error covariances or whatever to modify your model.
Model fit criteria a quite o.k., so that there is no problem. I do not see a conflict, as these are two different methods and both seem to be justified by theory.
Kind regards, Karin
Dear Mr. Butt,
I appreciate the detailed explanation from Prof Böhnke and much can be learned from his detailed write up.
Before i say further, please understand that this is exploratory factor analysis only and confirmatory factor analysis is different and is done using structural equation modelling.
Coming to your results, though the initial results based on Eigen values indicate the presence of many factors but when we look at the screen plot, we see the basically the plot is twisted four times. So even though technically we are able to get ten different factors, yet they may be overlapping which is visible through the screen plot. As per the screen plot, ideally, we could be having four factors, as we see four twists in the plot.
You can make this argument, that screen plot indicates four factors, as the base and then re run the factor analysis by constraining the number of factors to four. And then see the number of factors as well as the percentage of variance explained by it.
Once you have the four factors identified, then you can go to confirmatory analysis if you have theoretical basis of relationship between the four factor. For that software such as AMOS will be used.
Please feel free to write again if i can be of any use. I am attaching a write-up on factor analysis. Go through it if time permits. It will help in improving your overall understanding related to factor analysis.
Best wishes
There's a lot of good advice offered in the previous comments, so I'll try to avoid duplication. Though Cattell was convinced that the scree plot was superior to other methods for numbers of factors to retain, more recent evidence from simulation studies suggests that parallel analysis and minimum average partial (map) correlation methods work more reliably. That's on the exploratory side. Did you run the CFA on the same data set as the EFA? Ideally, the CFA would be run on a new or different data set.
I will say that neither the scree plot nor the eigenvalues would make a case for a four-factor solution, in my opinion. One factor or 10 factors, perhaps. The other possibility is that a higher-order factor solution is required. Based on the variables you're working with, what makes the most sense?
To me, the scree plot suggested one massive factor and lot of smaller factors. It's probably for that reason that your confirmatory analysis shows a one-factor solution to yield fit indices indistinguishable from those of a four-factor solution. Would a 10-factor solution be better? In terms of reproducing observed relationships among your data set, the answer will always be 'yes.' From a parsimony standpoint, if a simpler factor model works comparably to a more complex model, it should be preferred. In your case, that would suggest a one-factor model is adequate to account for the observed relationships. *However*, note that shifting models, either because of parsimony or fit concerns, means that you've abandoned the hypothesized model as no longer viable, and your chore will now be to come up with suitable justification for the amended model.
The "what is called conflict" could be due to the type and method selected for factor analysis, for example in EFA which method and under rotation or not. What was the criteria to retain/ drop items in each factor, and if the decision was based on eigenvalue or number of factors
in CFA, what items assinged to latent variables,
to give more clarifications, please read the attached 2 papers
good luck
I found this paper interesting although I could not understand it well...
Confirmatory Analysis of Exploratively Obtained Factor Structures
van Prooijen, Jan-Willem; van der Kloot, Willem A. Educational and Psychological Measurement, v61 n5 p777-92 Oct 2001
I will advise you to use parallel analysis to determine the number of factors especially when your data is ordinal. Using SPSS will overestimate the eigenvalues and actual factors might not be determined. Also, CFA is to confirm the structure coming from EFA. I hope this helps.
- Badges
- Science topic
- Similar topics
- Comment