Hi Navya,

according to the underlying theory, are two factors plausible or should it be one? If you have any prior assumptions about item-factor-distribution, choose CFA instead of EFA. How did you run the EFA (method of factor extraction and retention, rotation, etc.)?

Best

Marcel

Thank you for your prompt response, Marcel Grieger

So I am making a new scale.

The 4 positively worded items were to denote satisfaction.

The 3 negatively worded items were to denote frustration.

Theoretically, satisfaction and frustration are separate.

High frustration is not equal to low satisfaction.

And low frustration is not equal to high satisfaction.

And in EFA they behave as expected: 2 separate factors.

But, in the CFA dataset, satisfaction and frustration items are loading onto the same single factor.

So high frustrations is coming equal to low satisfaction and vice-versa.

I am working in Stata.

Using "factor, pcf" and varimax rotation.

I cannot use PCA or PAF only.

"factor, pcf" combines the two (Mooi et al., 2017)

Navya Kumar : What happens if you specify the CFA with the two theoretically distinct factors (i.e., a 2-factor model)? Does the 2-factor model fit better than a 1-factor model and how highly are the two factors correlated in the 2-factor solution? A one factor solution only makes sense if the correlation of the factors is very close to |1|.

I'd say EFA with Varimax (i.e., orthogonal) rotation makes little sense in your case because the two dimensions/factors are likely strongly (negatively) correlated. Rotating the factors to be uncorrelated (orthogonal) which is implied by Varimax will likely result in misleading estimates of loadings etc.

I agree with Marcel Grieger that CFA is preferable in your case as you have a clear theory regarding these factors and their measures.

Thank you for your response, Christian Geiser

CFA with two distinct factors (SATISFACTION and FRUSTRATION) gives a better model fit with other constructs (OUTCOMES) included and paths drawn.

But the inter-factor correlation is too high: 0.71. And the way they behave with other constructs is all wrong. The paths take all the wrong signs.

When I load it all onto one factor (SATISFACTION + REVERSE-CODED FRUSTRATION), the model fit reduces but the paths take the right signs.

I understand your point about EFA and orthogonal rotation.

Thank you!

The sign issue is strange. Could it be due to the coding/recoding of the items? Since it looks like you recoded one set of items (frustration), it does make sense for all loadings and the latent correlation to be positive in the 2-factor solution.

A latent correlation of .71 is substantial, but it is still quite far from 1.0 (indicating that there is a lack of unidimensionality--satisfaction and reverse coded frustration are not on the same dimension). A 2-factor solution is therefore likely more appropriate.

Thank you, Christian Geiser