All items in the model should be included. Pay attention to the dimension reduction technique and follow proper procedures.

Depending on your objective is to validate an instrument or evidence a theoretical model, if it is instrumental you have to know if it is a multidimensional variable that presents related factors to know how to use the rotation method.

Hi ! it depends on the constructs. whether the constructs are belong to the same source or constructs were adopted from various sources to develop the model. Thanks

Phuong Lam, although you might have received some advice above that is helpful, I think you might attract additional helpful advice if you were to provide a little more information about what you sre measuring, e.g., what your main variables are.

At the moment, you have used words such as factor, item, test, and variable - and different people can use those words with their own, unique, meanings.

Apart from that, EFA (exploratory factor analysis) might not be the best way to determine the number of factors in a data set, particularly if the Kaiser criterion (eigenvalues > 1 is used). A procedure known as parallel analysis is sometimes more effective for doing that - after which EFA is likely to provide additional useful information.

A scree test is also often a good way of identifying the number of factors in a data set - and, in my experience, almost always better than the Kaiser criterion.

Hi, I think you need to apply EFA to all items especially if you are using them for the first time (I mean if you are the developer of these items), also you need to apply EFA to all items in case you used spss, however, if the items had been adopted for previous studies then better you apply CFA using Amos instead of spss.

Robert Trevethan Thanks for your advice. Because I found it unreasonable to input all the items in the model for EFA. For example, I analyse the relationship between Psychological Capital (PsyCap) and Employee Innovative Behavior (EIB). PsyCap is a multidimentional variable including Hope, Resilience, Self-efficacy, Optimism with 24 items in total. EIB is unidimentional variable with 8 items. If I put 32 items together in EFA and 1 item from EIB moves to one of the four factors of PsyCap? Should we remove it? Because the meaning of this item does not reflect the measuring value of main variable.

With data analysis, we do not only look at the number but also consider the meaning of the items. By this I mean, before we apply EFA, should we make sure all the items belong to 1 variable? Thus, even if 1 item moves from this factor to the other factor, it still remain in 1 multidimensional variable - still measure for 1 variable.

Salem Alfagira I agree. Because EFA is factor-exploring analysis. So is it necessary to do it again with the variable that we already know the sub-dimention? I found many research which applied only CFA. However, in my research, the variables I adopted from previous study. I tried 2 ways of analysis - one with EFA first, then CFA and the other with CFA only. Unfortunately, the results were different. For first method, the relationship between variables was not supported, for the second method, it is contrast.

Phuong Lam, hello: From what you are now revealing, I think you have two variables: PsyCap (with four dimensions / factors) and EIB (with only one dimension).

Sorry, but I don't really understand your second paragraph up above, but sure, it's important to look at the meaning of the items. However, EFA can help to reveal whether your participants ascribe certain meanings to the items. In my experience, items that belong on certain factors in one sample don't always belong on the same factors in a different sample.

I would check the factor structure of each variable just to be sure that what you think is the structure of the variables is actually the structure in your data. If you think there is reasonable cause to doubt that, I'd be inclined to use EFA with the two variables. If you think you can be sure about the factor structure of your variables but you'd simply like to "declare" they have retained that structure, you could aim for confirmatory factor analysis, which can't be done with SPSS unfortunately.

Maybe it's safest for you to do EFA.

Incidentally, given the factors within PsyCap, I'd use an oblique rotation (e.g., oblimin) and perhaps obtain a scree plot at the start to get an idea of the lay of the land.

Every good wish with your research. And if you'd like to ask more questions, please feel free.

If cross loading appears, than yes you need to remove that variable from the construct. Check any book on EFA and you will find thresholds,etc.

I think it's necessary to be aware that words such as item, variable, and construct can easily be used in overlapping and, sometimes, vague ways. If so, that can lead to confusion.

Robert Trevethan my concern is content validity, nomological validity. If we input all the items in the model for EFA (in my case: the items of unidimentional variable - PsyCap - are mixed with the items of multidimentional variable - IEB), we violate the content validity and nomological validity at the beginning. For example, if an item from EIB named X1 moves to PsyCap 's sub-dimension with the loading factor higher than 0,5 after EFA, we have to accept it in new factor although its content does not related to PsyCap? This can lead to bias in bias in measurement.

In investigating the relationship between 2 variables, I tried 2 different ways of data analysis: one with EFA and CFA and the other with CFA only, then used SEM to analysed the relationship. The 2 results were completely different. The measurement of each variables I adopted from previous research so should I do CFA only?

George Onofrei you mean item from variable? Because variable and construct are completely different aspect. As Robert Trevethan said, we should be careful when using these words.

Phuong Lam, I'd be happy to keep trying to help you, but there seems to be some fuzziness going on. Three days ago you said that PsyCap is multidimensional and EIB is unidimensional. In your post a couple above here (a few minutes ago), they are the other way around.

Ha ha, with respect, are you being careful enough with how you are thinking? :-)

Furthermore, I'm not sure in places whether you are asking a question or making a statement / declaration.

From a "distance" (i.e., without being really familiar with what you are trying to do), if your two scales have been designed to measure different things, I would begin by doing a separate CFA for each of them and seeing if you get good model fit (I assume you are aware of the set of criteria you might use to determine that). Then, if you don't get model fit, I'd be inclined to turn to EFAs to see where the "problems" lie.

I hope that makes sense to you and is helpful.

Robert Trevethan I am sorry for confusing you. It is my mistake. My correction: Psycap is Multidimentional Variable and IEB is Unidimentional variable. That is why in my example I mentioned that one item moved to PsyCap's sub-dimention.

This is neither my statement nor declaration, just questions to open my mind. I read many book and discussed with many people but the answers were not the same and that made me confused.

I highly appreciate your patience with my questions. It is not easy to find someone who is willing to discuss and share his knowledge. Thank you very much.

Phuong Lam, good to hear from you - and I appreciate that we all mix things up at times.

I am not surprised that you have been getting different messages from different books and different and people. Unfortunately, not everything is black and white, even in quantitative methods / stats. It's just like real life.

I'd heed the advice of those whom you feel you can trust most, and, where there are discrepancies simply indicate (say, if you're writing a thesis or journal article) what you did and why.

Please feel free to get back if I might be able to help any more.