I have a question. I noticed that model B of strict invariance is equal to the scalar invariance model. In fact, the outputs of the scalar model and strict model B are the same, with the same degrees of freedom and index values. Can you explain why? Thank you very much.
Another super helpful video. May I ask how does the syntax change if there are more than two groups (e.g., year of undergrad study) and across time points (e.g. repeated measures at two-time points) ?
In this situation it is best to switch from a CFA framework to an IRT framework where missing data for binary indicators as you describe can be assumed MAR. Essentially, when you change the estimator to MLR and specify the items/indicators as categorical you are invoking an IRT 2PL model. Now the estimates are logits.
Hi, thanks for the very informative walk through. After running the difftest option in Mplus through the various levels of measurement invariance, what are the chi-square values for the nested models beyond the configural model? The output states that the chi-square test of model fit value is uninterpretable. Do we simply add the value of the chi-square test for difference testing to the prior model's chi-square? A numerical example, given the output, would be most appreciated. Thanks!
In our example data, the chi-square for Weak invarinace = 32.189 with df = 7. The DIFFTEST, comparing Weak to Configural = 7.874 with df = 3 and associated p value for the nested model test = .049. As for Strong invarinace, its chi-square = 44.925 with df = 18. The DIFFTEST, comparing Weak to Strong = 11.517 wit df = 11 and associated p value for this nested model comparison test = .401. Importantly, you will notice the chi-square DIFFTEST when using WLSMV estimation is not simply the difference in the nested models chi-square values and the same is true for computing the difference in df when WLSMV is used. As you will notice in the Mplus out, it will refer you to the Mplus technical appendices to see how they are computed within Mplus. So, the DIFFTEST is done behind the scenes within Mplus.
This is true, but when testing for measurement invariance you walk through a series of weak constraints (factor loadings held equal across groups) to a more constrained model to allow you to make more sensible comparisons under the assumption that the measurement model is stable/invariant across the two groups. Thus, giving you more confidence that the comparisons you are making between groups are more true effects and not confounded by measurement artifacts.
in this measurement invriance analysis we are trying to test the assumption that certain parameters are equal across groups. If they're not equal then significance will be found and then the parameter has to be relaxed or freely estimated.
I have a question. I noticed that model B of strict invariance is equal to the scalar invariance model. In fact, the outputs of the scalar model and strict model B are the same, with the same degrees of freedom and index values. Can you explain why? Thank you very much.
This video was so helpful. Thanks so much for sharing the video and the link to resources.
Thanks so much for the sharing. I wonder any updates with measurement invariance for longitudinal continuous data?
Another super helpful video.
May I ask how does the syntax change if there are more than two groups (e.g., year of undergrad study) and across time points (e.g. repeated measures at two-time points) ?
I also want to know this
Can we use MLR estimator for a large sample size (>4000) with binary indicators. basically I wish to accommodate the missing data through FIML.
In this situation it is best to switch from a CFA framework to an IRT framework where missing data for binary indicators as you describe can be assumed MAR. Essentially, when you change the estimator to MLR and specify the items/indicators as categorical you are invoking an IRT 2PL model. Now the estimates are logits.
Michael Toland thanks a lot for clarification of my concept. Can WLSMV accommodate for missing data in any way for the outcome variables.
Unfortunately, pairwise deletion or pairwise present data is the default technique for handling missing data in Mplus when WLSMV is selected.
Hi, thanks for the very informative walk through. After running the difftest option in Mplus through the various levels of measurement invariance, what are the chi-square values for the nested models beyond the configural model?
The output states that the chi-square test of model fit value is uninterpretable. Do we simply add the value of the chi-square test for difference testing to the prior model's chi-square? A numerical example, given the output, would be most appreciated. Thanks!
In our example data, the chi-square for Weak invarinace = 32.189 with df = 7. The DIFFTEST, comparing Weak to Configural = 7.874 with df = 3 and associated p value for the nested model test = .049. As for Strong invarinace, its chi-square = 44.925 with df = 18. The DIFFTEST, comparing Weak to Strong = 11.517 wit df = 11 and associated p value for this nested model comparison test = .401. Importantly, you will notice the chi-square DIFFTEST when using WLSMV estimation is not simply the difference in the nested models chi-square values and the same is true for computing the difference in df when WLSMV is used. As you will notice in the Mplus out, it will refer you to the Mplus technical appendices to see how they are computed within Mplus. So, the DIFFTEST is done behind the scenes within Mplus.
How can the factor loadings be equal in each group, some factors will load more and some less. can you please explain this point?
This is true, but when testing for measurement invariance you walk through a series of weak constraints (factor loadings held equal across groups) to a more constrained model to allow you to make more sensible comparisons under the assumption that the measurement model is stable/invariant across the two groups. Thus, giving you more confidence that the comparisons you are making between groups are more true effects and not confounded by measurement artifacts.
in this measurement invriance analysis we are trying to test the assumption that certain parameters are equal across groups. If they're not equal then significance will be found and then the parameter has to be relaxed or freely estimated.
Thank you so much for sharing. This is super helpful!