Hello everyone!
I struggle with a SEM model based on a scale with 19 items and 3 factors (and 1 second order factor), which yields very bad fit indices. To adjust it, I believe I have to remove items with error correlations / cross loadings to make the model more parsimonious.
I have a question about Modification indices; should I rely more on the highest value in expected X2 change or the STDX E.P.C. value when I choose which items to remove?
Advice is appreciated!
Best
John-Kåre
Dear John -Kare,
I suspect you can anticipate my answer :D I would rather refrain from inspecting the modification indices at all and rather closely take a look onto
a) the standardized residuals (differences between Sigma and S)
b) the correlations among the variables
c) the question wording
The MI are "theoretically blind"; hence better use your advantage (the ability to re-specify your measurement theory in face of signs provided by the data).
Removing problematic items can be an option (and I admit I also do it as my primary goal is to measure my intended latent variable(s) ) - however, there's the danger, you could throw out the baby with the bathwater and eliminate a whole latent variable together with the respective item.
One example: I once had a four-item goal commitment scale. The CFA fit badly and I wanted to eliminate the worst loading item as I recognized that this item was a measure of "intention" ("I intend to do X") whereas the goal commitment scale rather focus on the strength of willingness.
Best,
Holger
Dear John -Kare,
I suspect you can anticipate my answer :D I would rather refrain from inspecting the modification indices at all and rather closely take a look onto
a) the standardized residuals (differences between Sigma and S)
b) the correlations among the variables
c) the question wording
The MI are "theoretically blind"; hence better use your advantage (the ability to re-specify your measurement theory in face of signs provided by the data).
Removing problematic items can be an option (and I admit I also do it as my primary goal is to measure my intended latent variable(s) ) - however, there's the danger, you could throw out the baby with the bathwater and eliminate a whole latent variable together with the respective item.
One example: I once had a four-item goal commitment scale. The CFA fit badly and I wanted to eliminate the worst loading item as I recognized that this item was a measure of "intention" ("I intend to do X") whereas the goal commitment scale rather focus on the strength of willingness.
Best,
Holger
Hi John-Kare
The 2009 paper by Saris made a convincing argument for the use of EPCs and the Whittaker paper is an interesting follow-up. However, in practice I have not found a 'rule' that works well for different models with different samples. Holger is the voice of common sense that says that we must use our substantive understanding of the model and our data to make sensible modifications. Unlike MIs our understanding of the model, the data, and theory is not positively related to sample size!
Mark
Saris, W. E., Satorra, A. and van der Veld, W. M. 2009. Testing structural equation models or d detection of misspecifications?. Structural Equation Modeling, 16: 561–582.
Tiffany A. Whittaker (2012) Using the Modification Index and Standardized Expected Parameter Change for Model Modification, The Journal of Experimental Education, 80:1, 26-44, DOI: 10.1080/00220973.2010.531299
Hi Mark,
thanks for bringing the Saris et al. paper to my mind. Yes, their point of taking power into consideration is very valuable. One problem I saw (please correct me) at that time was that their approach (as the MI generally) is focused on fixed parameters, while the true problem often lies within the postulated structure itself. In addition, even when following the MI based on power considerations, it implies the problem of capitalization on chance, as you follow the algorithm and apply "theory" in a post-hoc manner. I know that this is always the case in the case of a post-misfit-respecification but I prefer to think about a theoretical alternative without being pushed by the MI's in a certain direction.
I hope I do not appear too narrow minded. Of course, there are different ways to Rome and the respecified model is new hypothesis, anyway :)
Best,
Holger
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling