A few questions: Are these theoretically distinct constructs? What does your theory predicts regarding their relationship/strength of association? How high is the correlation between the factors in a 2-factor CFA solution? Does a 2-factor CFA solution fit better than a 1-factor solution? The factors would have to be correlated 1.0 in the population to justify a 1-factor model.

A high correlation in a 2-factor model may or may not indicate a lack of discriminant validity--depending on your theory. For example, depression and anxiety are theoretically distinct constructs/show qualitatively different symptoms; yet they are often comorbid disorders. So it is not uncommon or unexpected for depression and anxiety factors in CFA to be correlated .7 or .8 (or even higher). You would typically still not combine symptoms of depression and anxiety into a single factor because these are theoretically distinct constructs. In other situations, correlations like this may be seen as too high and as indicating insufficient discriminant validity. I hope this helps.

Christian Thank you so much for the answer.

yes they are theoretically distinct. for example: one item is meauring the loyalty by asking if the patient will recommend the hospital, the other is measuring the image regarding the hospitals’ services.

the image is the independent and the loyalty is the dependent here

however in EFA they load together .

in CFA i tried to separate them into 2 different components (loyalty and image components) at the same model, RMSEA was 1.2

the correlation between items of the 2 components are higher than .65

running CFA for image items alone lowers RMSEA below .08

also should I report all these details why i chose to separate the scales in my manuscript?

Your RMSEA values seem exorbitantly high. They should be below .05 for a decent model. Are you sure they are correct? RMSEA rewards "parsimony", and so it may take on a better (lower) value for the 1-factor model, but that may be misleading. What do the chi-square fit test values look like for these models?

Also, without knowing more details about your data/model, it's difficult to provide useful feedback. For example, how many items are on each factor? What is the sample size? A bad fit could be due to having a large number of items per factor that lack unidimensionality within one or both factors. This may be a separate issue to address (misspecification of the measurement model).

Besides that, it would not surprise me if the two dimensions/constructs were highly correlated. It makes sense that people are very likely to recommend hospitals for which they rate the image as high/positive, and that they tend not to recommend hospitals with a bad image.

For the experience model

In EFA: KMO = 0.808 (Chi square = 2961.3, degrees of freedom = 630)

-Bartlett's test P-value

You have a lot of items and the mediocre fit results may be an indication that your items could be rather heterogeneous (measuring multiple dimensions, several major and minor factors and/or method/wording factors). A good strategy may be to try to identify the "best" items that are a core reflection of your dimensions of interest and use only those to have a cleaner model.

Christian

thank you

yes i believe this is the problem

However, i am validating the Balanced scorecard which has 5 perspectives, under each perspective i could identify different constructs

actually i find it hard to delete or ignore any perspective

so i thought why not to split items depending on patient real experience items and the patient attitude items then to identify the constructs of each scale alone

the cfa and efa results became much better, also the constructs became more logic

my theory is that the 5 perspectives of balanced scorecard ( financial, customer, internal, external, and innovation and knowledge) experiences have effect on the balanced scorecard 5 perspectives images, which lead to patient satisfaction and loyalty.

so do you think it is not a good idea to plit the item based on dependent and independent items then to perform the construct validation for each separately?