| Message/Author |
|
|
| Jayden Nord posted on Friday, September 18, 2015 - 9:11 am
|
|
|
| How does Mplus estimate the interaction effect between two continuous latent variables when there are missing data patterns in the observed variables? |
|
|
|
|
| Based on the "use all available data" principle of "FIML", which is ML under the MAR assumption. |
|
|
| Jayden Nord posted on Friday, September 18, 2015 - 6:12 pm
|
|
|
| I know that LMS uses an EM algorithm. However, I am not well versed in EM algorithms. Am I correct to assume FIML occurs at the E step? |
|
|
|
|
| The E step concerns information about the latent variables, whereas the M step estimates the relationships between the latent and the observed vbles. |
|
|
| Jayden Nord posted on Saturday, September 19, 2015 - 11:58 pm
|
|
|
My apologies; I am still confused.
How is FIML applied to the LMS procedure? Or, how is the principle of "use all available data" applied to the LMS procedure?
From my readings of Moosbrugger et al (1997), the log likelihood functions used for LMS depend on compete data matrices of the observed variables. Perhaps I have misunderstood the functions? |
|
|
|
|
| The LMS procdedure (XWITH in Mplus) needs raw data. As long as an individual has data on at least one of the observed variables he/she gets to contribute to the model estimation. That is the ML under MAR principle. The LMS procedure is not different in this regard. |
|
|
| Back to top |
