I conducted ESEM and CFA on two different questionnaires, Big Five Inventory (BFI) and Symptom Check-List (SCL). The BFI gives 5 factors, E, A, C, N and O. The SCL gives 9 factors and a general G factor (Bifactor). I would like to test the path G on E A C N O. Whatever method I use (ESEM, CFA, ESEM within CFA) it does not work because of non-convergence.
Is it legitimate to replace the path analysis by a classical multiple regression performed on the factor scores yielded by the software? Thank you in advance for your answer.
Jean-Michel Petot I recommend that you check out the broader literature on classical bifactor models. Those can be troublesome in many ways--as isolated models and when it comes to predicting G and/or specific factors by other variables. See:
Article Anomalous Results in G-Factor Models: Explanations and Alternatives
Article Bifactor Models for Predicting Criteria by General and Speci...
Article Explaining General and Specific Factors in Longitudinal, Mul...
There are solutions to the problems/issues that you describe for your application that are discussed in these papers.
Although not totally inappropriate, there are some issues with your proposed strategy of using factor scores followed by multiple regression in replacement of latent variable simultaneous path analysis. Using factor scores reintroduces (estimation) error--the very thing that we're trying to avoid by using SEM/CFA. Moreover, multiple regression does not allow you to test overidentified path models, that is, path models that have > 0 degrees of freedom in the structural model.
Most latent variable software allow you to control some aspects of the estimation, and this may allow convergence (e.g., increase iterations).
Yes, this is entirely possible. One issue might be the "attenuation" of your effects due to unreliability in factor scores.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS