I'm still confused which is more viable as saw different stories appy different one. I'm developing a new model to test relationship between antecedents and consequences of a construct using multiple regression. Is EFA for each original dimension and assessing their leading, validity and reliability correct?
Think about what you're doing with Exploratory Factor Analysis, Eias.
You are using the correlations (or covariances) of your variables to see if you can group your variables into factors that summarise most of the information contained in those variables. With social research we usually rotate the factors so that factor loadings are high on some variables and low on the others. In this way, we find variables that are highly correlated with each other (factors), and the factors have very low correlations between each other.
This is where theory is important. EFA, on its own, takes no account of the social theory that created your measures - it will find factors whether it makes sense to or not.
If current practice and theory tells you that several items make up, say, three distinct dimensions then there is no need to do EFA at all. You already know what the dimensions are!
On the other hand, you may want to test whether your instrument (questionnaire) works as expected with your sample. Then you would not run an Exploratory Factor Analysis; you would run a Confirmatory Factor Analysis (CFA). With CFA, you decide upfront what the factor structure should look like (theory), and then you test to see if reality (your data) corresponds well with your theory. If you are using SPSS through your university then you will also have access to a very good program called AMOS that works very well with EFA.
While you're doing that, for a very quick check, you could indeed do as you suggest: run EFA only those items that are supposed to be manifestations of just one dimension, and then see if you get only one factor, as expected.
Another quick check would be to run only those items that are supposed to be manifestations of two or more related dimensions. For example, items measuring social attitudes in one set; items measuring personal values in another set. Run EFA to see if you get rotated factor patterns that match your theoretical expectations.
Do not make the mistake of running EFA on all of your variables at once. It's especially a bad idea to run EFA on variables that you expect to be antecedents with variables that you expect to be consequences. If you do, then you will get factors that are a confusing mix of both antecedents and consequences, and they will have very poor relationships when you get to your regression model - because you made them that way. Again, think about what you're doing when you use EFA.
Hi Hume,
As a quick check, I ran PCA with Varimax rotation on each dimension individually hoping to extract only one factor. The results were almost satisfactory except for two cases:
1. One of the consequences dimensions constitutes of 9 questions generated two components with all communalities are above 0.5. However, I ran again with the five dimensions constituting the consequences as you instructed and extracted 5 factors with accepted communalities and variance.
2. One of the antecedents dimension constitutes of 5 questions generated one components but with total variance of only 48%, looking at communalities there are 3 items below the threshold of 0.5. There's no problem however if I run all related dimensions.
All that looks good to me, Eias.
EFA on the items that represent related dimensions or a multifaceted construct is the more common approach, and it looks like you have demonstrated consistent results.
Dear Prof. Hume,
I have been reading your response on the above question. I clearly understand that when a strong theory and constructs are already established we do not need to undertake combined EFA but a quick checking for each dimension is appropriate. However how do I justify that choice from theoretical background and reference perspective? Any books, articles , suggestions? Looking forward to your response
I'm not sure what would be the best source to explain everything, Rajat. For me, it's just logic (and quite a lot of experience). The usual citation for all things multivariate is:
Hair, Black, Babin & Anderson (2013) Multivariate Data Analysis 7ed, Pearson
I understand that an 8th edition is about to come out. Your library should have this, or an earlier edition, on the shelf.
Dear Prof Hume.
what will be if u use EFA to analysis but the variable is dimension only one variable. so i dont get a rotated component matrix.