I have used SEM with AMOS for my data for hypothesis testing and the model parameters came out good. I wanted to conduct a multigroup analysis with the same sample by categorizing them into two groups A and B to see if there is a difference between the two groups for the same hypothesis. although I am getting acceptable cmin/df, RMSEA, CFI values, my GFI and AGFI values remain below the threshold.
can anyone please help with the model parameters that I need to follow for multigroup analysis? my sample size for the two groups is 126 and 144 respectively. Even if I double my sample size (by duplicating the sample) to check if the sample size makes a difference, the model parameters remain the same.
The values of CFI, TLI, NFI should be taken into account as comparative indices, in addition to the RMSA model error and x2/gl for a good year.
Hi Swapna,
this may be not the thing you wanted to hear but I would like to disagree with the response by Cristian and vote for taking the chisquare test and its df seriously. From this point of view, it is possible that your basis model is causally misspecified which can be dealt with probably (and should).
Beyond that from all the fit indices, the relative indices (CFI, TLI) to me makes the least sense when your goal is to postulate, test, and intend to interpret the model with regard to causal effects. The logic behind the relative indices is of more information-theoretical/predictive point of view and asks the question "how much better (in terms of a reproduction of the data) is the target model compared with the worst possible model (i.e., the null model)?". This question is similar to the situation in which a car seller asks you "do you want to buy the car with one broken axis or two broken axes? Because: one broken axis is surely "better" than two broken axes"...Would you buy the car?
For a solid respecification of your model, you need "standardized residuals" (i.e., the differences between empirical correlations and model-implied correlations). As far as I know, AMOS does not provide these. If you invest a bit of time (=one day!) in R's lavaan package, I would offer to assist with a re-specification.
Recently, we did that here:
https://www.researchgate.net/post/Is_it_possible_to_perform_mediation_analysis_in_SEM_with_dependent_latent_variable#view=5ddd17dd0f95f1376a0388d1
Best,
Holger
Holger Steinmetz thank you for your answer. I definitely want to discuss further on this. your answer has helped clear a few of my doubts.
although it is possible that the base model is causally misspecified, thinking from the literature point of view, the causal relationships do make sense in my base model. however, there is always a chance that my assumptions can be wrong and there exist a better way to put the model.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS