Dear SEM experts, I will appreciate your help on this:
My model contains 4 predictors. the outcome is a latent construct (built of 3 manifests). Fit indices (CFI, TLI, RMSEA) for the CFA show a poor fit.
I am specifically interested in one of the predictors - let's call it X4 - and its relation to the outcome. Checking the structural model, X4's effect is not statistically significant (but anyway the model has a poor fit). Comparing this model ( Chi-square difference test) to a nested model where the effect of X4 is omitted from the model (thus differentiated by 1 df) shows the alternative model with just 3 predictors is superior.
This is not surprising. My question is what can be understood about this particular predictor X4? can I use these analyses to demonstrate X4 's effect on the outcome is not statistically significant?
I hesitate because the model showed poor fit - so can I even consider its results, and the results of its structural model?
Thank you in advance for any help,
& take care
Hello Rany,
There are many reasons why a model may not perform well with a given data set. These include: (a) the model is wrong; (b) one or more of the observed measures are inappropriate or technically poor; and (c) sampling error.
I think that your task would be to come up with a defensible explanation that squares with the relevant theory as to why exclusion of "X4" from the model makes sense (and not just because of the fit indices that result from its omission).
You probably know that model-data fit indices rely on how close the values of estimated parameters can reproduce the observed relationships among the variables in the data set. If it were the case that X4 had peculiar relationships with the other variables that the model did not capture, then of course, excluding X4 makes it "easier" to show overall model-data fit (all other things equal). Inspection of individual standardized residuals for each of the relationships among the manifest/observed variables may be helpful in understanding whether the difference between the two model outcomes resolves to something as simple as this.
Good luck with your work.
Hello Rany,
beyond what David said, do I understand you correctly (please always provide a path diagram) that you have 4 observed predictors while your DV is the only latent variable. If yes than the misfit points most likely to the misfit of the DV.
Excluding one X also reduces the restrictions for the model to fit (or in other words the barriers, the algorithm has to pass to fit the data).
Perhaps you can provide more info. Please see here some modest attempt to formulate a netiquette for research gate:
https://www.researchgate.net/post/Netiquette_proposal_for_questions_on_structural_equation_modeling
All the best,
--Holger
please check your items of each IV and DV. after update ask one of the people who does not form your field to respond and it will come back to normal.
100% sure.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling