Hi Jonathan,

I have attached an article on EFA for you. It is pretty much concise, hands-on and easy to understand.

I hope this helps.

Hi Johnathan:

With translated scales, you need to be sure that your factor structure matches the structure from samples using the source language that it was translated from. This is usually done by testing for measurement equivalence/invariance (ME/I) using confirmatory factor analysis. Of course, to do this you would need a sample from the source language, either one you collected or one someone else collected if you could get someone to provide a data set for you. If you don't have both samples, but only your own sample, the next best thing would be to verify the structure using confirmatory factor analysis. You would see if the items fit the original structure by specifying the number of factors in the original scale and assigning each item to the factor is originally loaded on. Finally, you could use exploratory factor analysis by specifying the number of factors and seeing how well your structure seems to match the original. Unless you have factor analysis results using the original source language, you won't know how clean that structure might have been, so it could be that some items didn't behave well in the original language.

If I follow, when you rotate two factors, some of the items don't load as expected. This is not unusual when scales are translated, and it could be because some items were not translated well, or it could be due to cultural differences in how items are interpreted or responded to. For example, sometimes negatively worded items don't work well when translated, either for cultural or linguistic reasons.

You should not interpret the nonrotated results--only rotated.

It might be that you need to drop an item or two that doesn't behave well. In addition to the factor analysis, I would also compute coefficient alpha to be sure the scales have sufficient internal consistency, and what happens if the non-behaving item/s are deleted.

Here's a source that discusses issues in translation and ME/I approaches.

Spector, P. E., Liu, C., and Sanchez, J. I. (2015). Methodological and substantive issues in conducting multinational and cross-cultural research. Annual Review of Organizational Psychology and Organizational Behavior, 2, 101-131.

Dear Jonathan,

EFA is a complex process. In principle you can fix the number of extracted factors. A factor needs at least 3 items loading on it greater than 0.4. If items cross load or have a low maximum loading they can be removed and the analysis re-run. The sample size is also an issue. I would recommend trying an orthogonal rotation before trying an oblique one.

I wanted to thank everyone for your response!

Dear Paul, that is an interesting article indeed. Thank you for your insights. Here are two more questions/concerns (if you do not mind):

1. Would it be possible to do the following:

a) compare the 1 factor solution (the factor loadings from the original study in English) with the solution on the translated scale I get using an EFA but with the option "fixed number of factors" where I pick "1", in order to compare by using Tucker's congruence coefficient?

b) exclude the variable that shows zero correlations with the factor or less than 3?

2. The main problem I have when excluding too many variables is that it decreases my variance-so is it possible to remain some of the variables although they might potentially belong to this "second factor"-due to the fact that, the purpose of my study is not to validate the measure? When removing that item my alpha coefficient is 0.843. Removing other items would still increase my alpha a bit, but I think that is relatively okay.