Dear Mosayeb Dashtpeyma ,

Factor extraction in EFA is a bit like a dance. In order to extract stable factors you are likely to lose a lot (maybe half) of your items along the way for reasons such as:

• High correlations (> 0.8; remove one from the pair using a qualitative argument)
• Low communality scores (< 0.2)
• Low scores on all the factors (< 0.3)
• Cross loading on several factors (greater than 3:4)

The eigenvalue > 1 rule tends to be too generous but it can help you as you remove items incrementally and work towards a stable solution. My advice would be to reduce the number of factors manually one by one from an initial solut derived using the eigenvalue rule.

Please see my study guide for more information.

I provided several criteria on that. Please refer to this article:

Article Korean Patient-Perceived Satisfaction Scale of Community-Bas...

Hi Mosayeb

You have a few options-

If you use maximum likelihood extraction you can get a chi-square test; keep extracting factors until the p value is > .05.

You can use Horn's parallel test using the guidance at http://global.oup.com/us/companion.websites/9780199734177/supplementary/factors/factors_a/

..or you could use other software, such as Mplus, that has a very comprehensive set of tools to do EFA in a much better way than SPSS.

Mark

Section 6.4 of http://www.personality-project.org/r/book/Chapter6.pdf discusses some methods. This is written assuming the person is doing analyses in R, but the conceptual and mathematical issues are relevant. Following on from Mark Shevlin 's suggestion, fa.parallel in the R package psych using the parallel test is described there.

Taking eigrnvalue> 1.0 is the de facto option for choosing which factors to retain in EFA. Principal axis factoring (PAF) with maximum likelihood works for confirmatory FA. You should not use it unless you are totally certain of what you are doing. Also, parallel analysis is fine as long as you have a large dataset (e.g., 200) AND that you have LOTS of time to waste. Otherwise, PLEASE stick to eigenvalue > 1.0.

A so-called "manual" method for extracting the number of factors you desire is as follows:

1. Execute the initial PCA/PAF. Identify the no. of factors with eigenvalues > 1.0 (presumably 17 in your case).

2. Assess factor loadings in rotated matrix. Eliminate items that have low factor loadings (< .30) and then execute FA again.

3. Reassess the number of factors you have. It ought to be fewer than 17. If not, check the intraclass correlations (Cronbach's alpha) of the items constituting the scale you desire and eliminate items with low intraclass r values. Also, check the communality table. Eliminate items with low community values (< .30).

4. After performing two rounds of FA and eliminating those aforementioned "bad" items, run a third round of FA by specifying the number of factors you desire (click the checkmark in the options window of SPSS). If you specify 5 factors, you will get five factors, and ideally, you should get moderate to high community values AND discriminant factor loadings in the rotated pattern matrix - that is, high loadings on the factor(s) representing the scale(s) you desire (> .50) and low factor loadings on the rest (< .30).

Thsee steps have been implemented before. Feel fit to improvise as you deem necessary.

Thank you so much Claire Su-Yeon Park

Thank you so much Mark Shevlin

Thank you so much Daniel Wright

Thank you so much Jimmy Y. Zhong

Dear Mosayeb Dashtpeyma ,

Factor extraction in EFA is a bit like a dance. In order to extract stable factors you are likely to lose a lot (maybe half) of your items along the way for reasons such as:

• High correlations (> 0.8; remove one from the pair using a qualitative argument)
• Low communality scores (< 0.2)
• Low scores on all the factors (< 0.3)
• Cross loading on several factors (greater than 3:4)

The eigenvalue > 1 rule tends to be too generous but it can help you as you remove items incrementally and work towards a stable solution. My advice would be to reduce the number of factors manually one by one from an initial solut derived using the eigenvalue rule.

Please see my study guide for more information.

Dear\ Mosayeb Dashtpeyma

pls read this article :-

• https://www.unimuenster.de/imperia/md/content/ew/personen/obermeier/methodisches_vorgehen
• http://hosted.jalt.org/test/PDF/Brown31.pdf
• https://www.youtube.com/watch?v=x3u1G5JAuiY

Good luck

Thank you so much Peter Samuels

Thank you so much Mohamed Abdallah Turky

Hello Mosayeb,

In my experience, and in agreement with Peter Samuels above, using the eigenvalue > 1 "rule", also known as the Kaiser criterion, almost always results in too many factors when conducting EFA. There are other ways of getting an idea about how many factors might lie within a data set. Some are described in the following article:

Fives, H., & Buehl, M. M. (2010). Examining the factor structure of the Teachers’ Sense of Efficacy Scale. Journal of Experimental Education, 78, 118–134.

I have found that the scree plot / test is often the most helpful method, although some "cleaning up" of items, as suggested by Peter S. above, might also be desirable in some situations.

Perhaps better than even the scree plot is parallel analysis. It seems that too few researchers are aware of it or think it would be too difficult to use. The following website is very useful and easy to work with:

https://analytics.gonzaga.edu/parallelengine/

One simply has to enter the number of participants and variables into the fields, and a number of eigenvalues are produced. It's necessary to draw up a parallel list with the websites eigenvalues on one side, and your own eigenvalues on the other side.

When the extent of those eigenvalues exceeds the eigenvalues produced in one's own EFA, the point has been reached at which no more factors are likely to be in the data.

If that explanation isn't clear, please feel free to get back and I'll provide additional explanations. There are some useful YouTube presentations about parallel analysis.