Tracy Kline posted on Sunday, September 09, 2007 - 4:49 pm
I am currently using version 4.1 and have been asked to build a certain prediction model. I was curious if this was possible in Mplus.
Basically, I am running a structural equation model with 5 latent variables (each with somewhere between 3 and 6 categorical indicators) and one of the latent variables is the outcome of several others. I have analyzed different groups with this model and the fit is always more than adequate. The factor scores from these runs are being utilized in another analysis that I am not a major part of.
1) Currently, I am being asked to run this model again, but without the indicators of the outcome variable in order to 'predict' what the factor score will be (and then predict the values of the indicators).
2) I have also been asked to provide standard errors on the factor scores in the original model.
3) Finally, I was asked to produce observed values, predicted values and residuals from the 'prediction' model.
I have never tried to use Mplus in this manner, is this analysis and output possible?
1. You can estimate the exogenous factor scores from the observed variables except the indicators of the distal latent variable (say f2). Then you can use the estimated regression of f2 on earlier factors from the full model to estimate f2 values.
2. Mplus does not provide standard errors for factor scores when there are multiple factors and categorical outcomes.
3. Once you have estimated the f2 values, the full model can be used to estimate to observed indicator values.
Tracy Kline posted on Monday, October 01, 2007 - 5:28 am
Thank you so much Dr. Muthen.
I do have more questions.
From your response to point #1...
Is there a way in Mplus to calculate the estimated regression of F2, should it be done by hand, or is there another program you would recommend? Also, is there a particular reference for using this method? I will need to include that in my document.
From point #3...
What do you recommend as the best method to estimate the observed indicator values?
This is a big topic and we cannot get into statistical consulting. What I can say is that it is well-known that estimated factor scores don't behave as true scores in their relationships to other variables. I don't know of any specific references relevant tp your question - anyone? To choose the best method for your questions, I would recommend sitting down with an SEM expert as a consultant.