Tracy Witte posted on Wednesday, June 16, 2010 - 7:30 am
Based upon multiple published EFA studies on an assessment instrument, I specified a CFA model with 2 factors. The resulting factors are highly intercorrelated (standardized correlation = 0.92). Of course, I was concerned that there was limited discriminant validity with these factors. However...
1) Fixing the covariance of the factors to 1 (and thereby testing a one-factor solution) results in a significantly worse fitting model
2) When I run an SEM with various other measures as DV's, the factors differentially predict these DV's in a predictable, theoretically compelling way.
Am I justified in retaining the 2-factor solution, since it's consistent with prior studies (although these prior studies do not typically have such high inter-factor correlations) and the factors differentially predict other DV's? Or, as the differential prediction possibly an artifact of multicollinearity?
It sounds like your items may not be valid measures of the items used in the studies you are looking at. I would do an EFA to see how your items behave. If you have cross-loadings, you could consider ESEM. See Examples 5.24 to 5.27 in the Version 6 user's guide.
Tracy Witte posted on Wednesday, June 16, 2010 - 8:31 am
So even though the measure is the exact same measure as the previous studies, and the samples are similar in nature, it's necessary to do an EFA? I only have one sample with 227 people, so I wanted to avoid doing an EFA (since it has already been done by 5 other studies with this same measure), as I won't be able to independently replicate the results with a subsequent CFA. (thank you for your help!)
If the measures are exactly the same and your results are different than all of the other studies, it would seem that somehow your sample is different given that you are not able to replicate the results of the other studies.