EFA Demonstration

This is a demonstration of how to do an exploratory factor analysis in PASW SPSS statistics. This demo was made for the DM students at CWRU, but may be useful to anyone seeking to learn to do an EFA.

[gobadboygo]

watching these ten years later! 100x more helpful than my lecturers, cheers mate!

[evitajames]

Clinical psych phd student here. THANK YOU for this video, and all your others for that matter. So helpful.

[Gaskination]

I include all latent continuous variables (except control variables) in the EFA. You do not have to have a perfectly clean Pattern Matrix. My recommendation is to have loadings on each factor that average out to around 0.700 and make sure no cross loadings are within 0.200 (absolute value). So, if item1 loaded on factor1 at 0.730 and on factor2 at 0.500, that's fine. Hope this helps.

[Gaskination]

That is a fabulous question, and the answer is: yes, but. Yes, you can certainly use the same data for EFA and CFA. But, if you have a sufficiently large dataset, you would achieve higher rigor by using two randomly selected subsamples of your data for each. This would demonstrate greater validity and reliability.

[Gaskination]

Principal Axis Factoring (PAF) is an extraction method which considers only common variance (places communality estimates on diagonal of correlation matrix). I've never heard of Principle Factor Analysis, but I have used Principle Components Analysis which is another extraction method that considers all of the available variance (common + unique) (places 1’s on diagonal of correlation matrix). PAF is preferred in SEM because it accounts for covariation, whereas PCA accounts for total variance.

[creativeschool]

I said to my husband thatI James Gaskin and he didn’t like that! Lol. Thanks GURU. Now I can’t wait to go back and run my stats.

[kanetakc]

will look at it now-- and thank you so much for all the help, it has been so difficult to sift through all the literature and figure out how to go about this

[Gaskination]

I recommend always doing an EFA, even with established items, because you are using them with a new context and new set of data. Doing them one at a time is problematic because when you move on to the CFA, you will still run into problems. However, if you simply cannot get the EFA to work, you may need to examine normality and outliers first. Often kurtosis will throw off the EFA.

[Gaskination]

Oh, haha. Initially, these videos were intended just for my students at Case Western Reserve University. So "next time" meant "the next time I see you in class", not, "in my next video". Sorry for the confusion.

[reave]

great video. helped me understand at least the core points of EFA. cheers!

[ZMQ]

Your vidoes are fantastic. Thank you very much. Do you happen to have a version which would allow for testing multilevel effects? Thank you.

[hdawod]

Dear Dr.James,
Thanks a lot for your very helpful demonstrations.
some researchers skip EFA for adapted questionnaire and they argue that EFA is just for developed measurements.
So could you please provide me with a reference for what you stated above because I want to cite that in my thesis. I am doing EFA for my adapted instrument because of the different context, subject and new set of data.

[Gaskination]

I do not know of a way to do a 2nd order factor analysis in SPSS. If you want to jimmy-rig it, you could just do a first order analysis first, then create composites out of the first order factors (so that there is only one composite to represent each factor) and then do a new factor analysis with just the composites. This would be a quasi-2nd order factor analysis. I don't know of any literature that does this as a precedence. Usually 2nd order factor analysis occurs in the CFA.

[janelleveenstra]

Hello! This may be beyond the scope of this video, but I initially did a Monte Carlo Parallel Analysis on my data in order to determine how many factors I should extract. After running an EFA (along with your example) and setting the fixed number of factors to 2, my 'total variance explained' chart says that both factors only make up approx. 34% of the variance within my data. You mentioned in the video that one should at least aim for 50%. What do you suggest I do with such a low percentage since my data should contain 2 factors? 

[Gaskination]

No. You use the EFA to inform the CFA, but you do not then need to then go back to the EFA. You are good.

[hadiyasrebdoost]

Dear James
I have a question about EFA and second order CFA . At first EFA:
1) assume that we want to discover the relationship between two concepts with two separate questionnaire, for example customer satisfaction and employee engagement . Should we do EFA for each questionnaire (concept) separately or we should do an EFA for whole questions?
2)In second order CFA, we assume the example in your demo for second order structure, is it necessary to do a CFA for the model without second order factors with all first order factors? or we should do a CFA for the elements of second order factors separately and after that do a CFA for the model with second order factors?
Best regards

[liliasosa]

Hi James.
Thanks for the videos, they are really interesting and clarify our doubts. I do not understand why it is better to use principal axis factoring in the  factorial exploratory analysis and not the classical method of principal components. 
I am currently doing this analysis as conducted under PLS my structural equations model and really I want to do a good analysis before proceeding. You could explain a little more about it, would greatly appreciate it. 

Regards

[Gaskination]

I have a video to show how to do this. It's called "Imputing Composite Variables in AMOS"

[jrmboya]

Dear Dr Gaskin
This is extremely a good video, very illustrative. I am asking if I can send to you my questionnaire designed for factor analysis for your comments. I am doing a PhD study on competitiveness here in Tanzania
Joh

[kanetakc]

hi there i'm running an efa as follows:

Items with primary factor loadings ≥.40 (including values that rounded to .4) and secondary factor loadings ≤.30 and those that did not load on more than one factor were retained. Items not meeting these criteria were removed one at a time.

My question is - should i eliminate items in a particular order e.g. items with loadings lower than .3 should be eliminated before those with cross loadings higher? I hope this makes