Hi all,
I'm validating a tool after translation and have a technical question.
I'll do confirmatory factor analysis (CFA) to test the goodness of the model fit. Previous studies conducted exploratory factor analysis (EFA) or principal component analysis (PCA), but didn't come up with consistent structure matrix of the tool. My questions are:
1. Is it appropriate to do PCA only (without factor analysis) to identify structure matrix? If doing PCA only is adequate, do I still need to split sample into half for PCA and CFA? Are there any papers to support the choice?
2. If conducting an EFA is necessary, can I still choose 'PCA' as Extraction ? Or I should choose 'Maximum Likelihood' as a factor analysis?
Many thanks!
Hi:
If previous studies have not yielded consistent factor structure results, it is better to perform exploratory analysis. As far as I know, PCA covers some weaknesses of the EFA, so you may use PCA. Then, you should perform CFA to see if the results fit the data. It is not necessary to split the data, though if you have enough data, you can do this.
An alternative to all these is ESEM analysis, which covers the weaknesses of both EFA/CFA. Let me know if you need further help.
Best,
Mojtaba
Hello Xiaoyi,
The short answer is yes, you may split a sample and use part of it for developing a proposed factor model and the held-out portion to run a confirmatory factor analysis.
You may wish to re-think the choice of using PCA as the extraction method for the exploratory phase. It presumes all observed variance is common variance, and that variable-specific variance and error variance do not exist (two presumptions which are seldom true in any research involving human performance or judgment. As well, PCA tends to bias estimated loadings upwards, so your CFA might yield contrary results simply because of this tendency (let alone sampling variation and other influences). Finally, since most covariance-based CFA programs are using maximum likelihood estimation for measurement model parameters, it might be wise to follow a similar extraction method during the EFA phase.
Good luck with your work.
Hello Xiaoyi,
First, be aware that a component "model" (better: component-oriented approach) such as PCA is something substantially different than a factor model (as in EFA or CFA). A componant approach is purely data driven and intends to sort items into higher order components that keep as much information as possible despite the reduction in dimensionality. It's a purely empiricist appraoch and just deals with data.
A factor model is a hypothesis about a data generating process. The model states that the factor is the cause of the item generation and, when using several items, conseuqently is the cause of the correlation pattern (as a common factor is a common cause and creates spurious correlations among the items).
Hence, before treating *data* you have to decide what you want to do. The literature (esp. in psychology) is full of uses of PCA while believing in factors. The whole personality research history is full of that.
Second, I personaly would not advise to split the sample. The aforementioned procedures rely on the correlation matrix and correlations / regression slopes that have a small variance across samples. Hence, if you devide your sample you will almost ever get two samples that although differing slightly in the exact pattern of values in the multidimensional space will have very similar linear relationships . Hence, "validating" "something" found in first sample with the second sample is a myth as it is the capitilization on chance issue. The only thing you get is a reduction in power and precision when you divide your sample.
With regard to the EFA-then-CFA part of the question this has been heavily debated for decades in the literature as well as here on RG.
I have some moderate opinion but would rather vote for a enriched replication of the model with new variables. This can be done in the same sample or a new. I admit that there may be some statistical traps I don't see and am open to be lectured.
Here are two threads in this regard
https://www.researchgate.net/post/Can_I_use_EFA_before_SEM_analysis
https://www.researchgate.net/post/Can_we_do_exploratory_and_confirmatory_factor_analysis_in_the_same_data_set2#view=5486170ad4c118047c8b477f
HTH
Holger
Mojtaba Elhami Athar David Morse Holger Steinmetz Thanks all for sharing your thoughts. Very helpful discussions!
Apologies for coming back to you so late! I wasn't well for a while and just started to picking up work.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling