Kim Betts posted on Wednesday, June 22, 2016 - 4:10 am
I have run a FMA with 3 classes and 1 factor with full non-invarance. I have tried exporting the resulting factor scores and class probabilities for further analysis (so covariates don't effect factor scores/classes and to save time).
I was expecting to receive three factors (one from each class) which I would then weight by the posterior probabilities of the corresponding class, and use to undertake regression in another program.
However, the savedata results in only two factors:
Could you please explain how my three non-invariant classes, each with their own factor, result in only two factors and how I would use these for further regression (perhaps in some combination with posterior probs?)
F is the factor score when mixed over the classes. C_F is the factor score for the most likely class.
Kim Betts posted on Wednesday, June 22, 2016 - 5:31 pm
Thank you for the reply.
One follow-up question. I want to see if the 'severity' (i.e., factor scores) differently relate to covariates across different classes. If I wanted to save/export these factors to another program for regression, do you suggest I use F weighted by the posterior probability of the class of interest, e.g. (where x1...xk are covariates):
F = x1..xk [weight=CPROB1]
(to derive estimates of F in class 1 on x's)
Or should I use C_F and restrict analyses to the most likely class, e.g.,:
C_F = x1...xk if C=1
(once again to derive estimates of F in class 1 on x's)
I assume that the binary outcome variable you mention is not one of the factor indicators - this should work fine; if not, send output to Support along with your license number.
If the binary outcome variable is one of the factor indicators, saying y on f would represent their relationship twice and would give a non-identified model.
Kim Betts posted on Monday, June 27, 2016 - 11:49 pm
Yes the binary outcome is not in anyway related to the measurement part of the FMA (and is not a factor indicator).
I seem to have gotten it working (seems to work better with continuous outcomes and if the factor within the classes is invariant).
However when using the first method (outcome on F) the class membership proportions change quite a lot (from 40% in my pathological class to 30%)and the entropy changes a lot also. While, when I use the second method (F on outcome) the class membership and entropy stays the same as when I compute the measurement model by itself (a situation which I'm sure reviewers will prefer).
Is there a reason that we would expect the first method to change the measurement structure while the second to not? I don't necessarily need to use the first method, but it would be nice to justify the use of the second as leading to more stable measurement model because...
Asparouhov, T. & Muthén, B. (2014). Auxiliary variables in mixture modeling: Three-step approaches using Mplus. Structural Equation Modeling: A Multidisciplinary Journal, 21:3, 329-341. The posted version corrects several typos in the published version. An earlier version of this paper was posted as web note 15. Download appendices with Mplus scripts.
Asparouhov, T. & Muthén, B. (2014). Auxiliary variables in mixture modeling: Using the BCH method in Mplus to estimate a distal outcome model and an arbitrary second model. Web note 21.