|
|
 |
| Partial Measurement Invariance in GMM... |
 
|
| Message/Author |
|
|
|
|
| I ran a series of GMM models (1-5 classes) with BMI data; fit indices indicated that a 3 class model was the best fitting. I then ran models (classes 1-5) separately by gender to evaluate if the BMI trajectories differed by gender. For both genders, a three class model was identified as the best fitting, and the models yielded trajectories that were very similar for boys and girls. I then ran models to test for measurement invariance (using KNOWNCLASS), and found that the unrestricted model was a significant improvement over the restricted model. However, a close look at the data indicated that for the first and second class, BMI trajectories completely overlapped across gender. But for the third class, the trajectories visually differed by gender. So then I tested for partial invariance, in which I restricted the parameters for the first two classes across gender to be equal, and then allowed the parameters for the third class to remain free. Based on the loglikehood diff test, this partially restricted model was a significant improvement over the unrestricted model, which is great. However, even though this third class is statistically different by gender, it isn't meaningful interpretation-wise, and we plan to combine the 3rd class across genders. So, where do I go from here? I would like to test for diffs in covariates by class membership, but I am not sure if I should do so in the full sample 3-class model or in the partially restricted model. |
|
|
|
|
You may not need a 3rd class if you use the skew-t approach of the paper on our website
Muthén, B. & Asparouhov T. (2015). Growth mixture modeling with non-normal distributions. Statistics in Medicine, 34:6, 1041–1058. DOI: 10.1002/sim6388
download paper show abstract |
|
|
|
|
Actually, I have tried the skew-t approach and continuously run into errors saying:
WARNING: DUE TO A LOW DF ESTIMATE IN CLASS 2 THE ESTIMATED VARIANCE IN THAT CLASS IS INFINITY.
THE DISTRIBUTIONAL ASSUMPTIONS OF THE SKEW-T DISTRIBUTION MAY NOT BE APPROPRIATE.
I have successfully used skewnormal, and the three class model was still the best fitting.
Note that the skewness values for the raw BMI data do not exceed 2.0. |
|
|
|
|
Ok; good that you got at least the skewnormal to work out.
The question you posed is more of a substantive nature - you can go either way. |
|
|
|
|
| Okay, that's good to hear. I wasn't sure if there was a specific approach I should take. Thank you so much! |
|
|
| Back to top |
|
|
