For standard scale development, you must perform CFA after EFA. The CFA should be conducted on a new data set after conducting EFA.

For low CFI value, try to detect any problem with "modification index" and factor loadings.

Regarding the number of items for each factor, there is no "standard" rule. However, the rule-of-thump suggests that it should be minimum three. Yet you can go with two items for one factor.

One more important issue is that, you also need to test the predictive validity of your proposed scale in structural model after performing CFA.

Hope that helps!

Dear Khan Taufique,

Thanks for your reply. What do you mean with 'problem with modification index and how can I solve that problem? And how can I perform predictive validity?

Meryem

Dear Meryem,

The issue of conducting CFA or EFA has no simple answer. For a simple start, I suggest to read the following article 

http://www2.sas.com/proceedings/sugi31/200-31.pdf

Hope this can help you.

Dear Hesham Fathy Gadelrab,

Thank you for the article, it is inderdaad useful!

Meryem

Dear Meryem,

I am not sure what software package you've used for CFA. If you use AMOS, in the output file you will see one Tab on the left titled "Modification Indices" under which the error covariances are displayed. If you find any unusual high error covariance value between two error terms, you can connect those two error terms in your AMOS graphics to make them free which will result in an increase in model fit (e.g., CFI) values.

For Predictive Validity: For example, you have two factors (e.g., 'knowledge' having 3 times and 'attitude' having 3 items) that affect 'intention'. You have to run a structural model linking two independent variables (i.e., knowledge and attitude) and one dependent variable (i.e., intention) to confirm how strong these scales in predicting 'intention'.

Hope this clarifies.

Dear Meryem

there are many issues here, using CFA you must exam many models and test them to find the best one, so you might test a general factor with 5 independent factors, then a g- factor with 5 dependent factors, then as Khan mentioned you must take a look for the modification indices which tell you that there are covariances between them. Note that all of this must be theoretically based.

Meryem,

As you can see, you have kicked over a hornet's nest. As noted by others in this thread, EFA and CFA look like the same procedure, but they have different purposes. EFA is akin to the traditional factor analysis procedures you are taught in multivariate stats classes. More specifically, it is principal axis factoring. This fact distinguishes EFA from mere '..data reduction! Sources of error be damned!'

such as is the purpose of principal components factoring.

CFA is a more touchy subject. EFA is often viewed as an activity of instrument development. CFA is an activity of theory testing. CFA is run after EFA, yes. But CFA assumes that the theory you are using is correct. This is what Rana Momani is alluding to in the above response.

If you are using an accurate translation of a well-studied assessment, exploratory analyses shouldn't point to many needed modifications. Poor fit at the EFA step may indicate translation issues or a host of other assessment-related problems (or theory problems). By the time you get to theory-testing (CFA),

you should be fairly confident that poor fit is isolated to issues of theory, not assessment.

You aren't going to learn the intricacies of model-testing from a reserachgate.net thread. By now, there are literally hundreds of books on path analysis/structural equation modeling: see authors David Kaplin, Randall Schumacher, Tenko Raykov, Rick Hoyle, Barbara Byrne, Ralph Mueller, etc. These authors and others have produced very accessible first books on the subject that should give you a great start on understanding the chore you have taken upon yourself.

Best of luck!

Mark

Performing EFA is always odd . It means you do not have an idea of the structure of the concept and sub-concepts at all. Why did you make the questions? You had some idea about the relation between items and the dimensions you were modelling or not? If you have designed items to belong to a speciic (sub)-concept test it as a single factor using ML instead of another option. I hardly ever use factor-analysis, but use IRT models instead as they are appropriate for ordinal items. Factor-analysis demands at least interval. In R under the Rcmdr you can not perform a factor-analysis with ordinal data.

In factor-analysis a minimum of three items is required for a factor. Otherwise you can not test the fit of the model. The correlations between items set limits on the rest of the correlations between other items. With only two items there are no additional correlations to set a limit for the correlation. You have just one correlation. With three items you have three correlations. The first two correlations imply a limit for the third one. It is an application of the axiom of Pythagoras (c2 = a2 + b2).

Kindly find the attached file, In which I mentioned a Journey form EFA to CFA

Article An Easy Approach to Exploratory Factor Analysis: Marketing P...

Thank you Noor Ul Hadi

Thank you for your contribution Chalamalla Srinivas

No, you can not have a factor with only two items. You need at least three items to have a factor. The third item sets a limit on the correlations. Only then it is possible to test if the inequality it the correlations between a,b and a,c are in accord with the correlation b,c. The axiom of Pythagoras sets a limits to the correlation between b,c when a,b, and a,c are given.

Dear Peter Moorer, 

Could you please give reference?

Thank you.

Meryem, It is implied as factoranalysis can be represented geographically (see http://tqmp.org/Content/vol09-2/p079/p079.pdf, pag 82). You can also read it in Harman, H.H. (1976). Modern factor analysis  or Jöreskog, K.G., & Sörbom, D. (1979).

Advances in factor analysis and structural equation models. Cambridge, MA:

Abt Books.  Factors are needed to reproduce the correlation or covariance matrix. The structures of the factors and their between structure sets limits on the covariance matrix. When using a test of fit, it will indicate how well the implied theoretical structures can reproduce the correlation/covariance matrix. For this to be tested you will need at least three correlations and thus three items. The third will be limited by the other two. 

Thank you very much !!

Thank you Chalamalla

I think that only EFA or CFA  is sufficient for construct validity. Although some researchers do not accept this.

But there are some studies in the literature that only use CFA for measure construct validity 

ex=

Henry, J. D., & Crawford, J. R. (2005). The short‐form version of the Depression Anxiety Stress Scales (DASS‐21): Construct validity and normative data in a large non‐clinical sample. British Journal of Clinical Psychology,44(2), 227-239.

For scale that is developed based on a theoretical framework which detailed the domains . I don't recommend EFA rather start with CFA, if your items do not support CFA, revise your items.

A nice presentation of the relationships of PCA, EFA and CFA was described by Matsunaga in this paper.

Matsunaga, M. (2010). How to Factor-Analyze Your Data Right: Do's, Don'ts, and How-To's. International journal of psychological research, 3(1), 97-110.