Jack Noone posted on Monday, October 15, 2007 - 8:04 pm
I have a latent variable which is indicated by three categorical variables. I am assuming that the underlying latent variable is also categorical but how do i make this decision?
If i want to do a CFA with categorical and continuous latent variables along with covariates must i undertake latent class mixture modelling? I am not really concerned with how many classes are uncovered but the relationships between the factors and between the factors and covariates.
Factors as in factor analysis are continuous whether their factor indicators are continuous or categorical. If you have both continuous and categorical factor indicators that represent a set of factors as in factor analysis, you do not need mixture modeling.
Categorical latent variables are used to represent unobserved heterogeneity.
Jack Noone posted on Monday, October 15, 2007 - 11:17 pm
Thank you for clearing this up for me (and so quickly too!)
Jack Noone posted on Monday, October 15, 2007 - 11:29 pm
Actually i have one more question. the standardised coefficients between a factor and its categorical indictors are fairly straight forward to interpret - One standard deviation change in the latent variable results in a standard deviation change in the indicator (well i assume this is correct!). WHat about when the indicators are categorical? I understand that these are logistic or probit regressions but how do you interpret these standardised coefficients? I have done an ordinal regression analyses in the past and have converted the parameter estimates to an odds ratio - is this a good option in the case of categorical indicators?
Also, i understand that the R2 values (for categorical indicators) can not be interpreted as perentage variance explained (which makes sense)- however i got the impression that one could use the residual variance values to calculate expected probabilities of group membership. Do you know if this is correct?
Also could you point me towards a good source for learning about interpreting (and reporting) threshold values?
I really am a novice here so i appreciate your help!
Regarding the questions of the first 2 paragraphs, see the Snijders & Bosker Multilevel Analysis book in reference to R-square for continuous latent response variables underlying categorical outcomes. This draws on McKelvey & Zavoina (1975) in J of Math Soc. This is how Mplus computes R-square with categorical outcomes.
Learning about thresholds can be done in conjunction with ordered polytomous regression described in one of Agresti's categorical data analysis books from Wiley.
And, to learn more you can come to the March 2008 Mplus course at Johns Hopkins described on our home page.
Jack Noone posted on Thursday, October 18, 2007 - 11:40 pm
thanks for your help, a little bit far from New Zealand though - I'll keep an eye out for the video.
linda beck posted on Friday, July 18, 2008 - 1:58 pm
I have several basic questions...
I want to do a LGMM. My observed indicators are skewed (skewness ranges around 1 at 4 measurement points). Is it possible to do LGMM with MLR-estimator? I've heard about problems using LMR-Test with skewed observed indicators, is that true?
And finally: Assume I've found a quadratic curve in a single class analysis fits the data best. Should one use this quadratic "setting" throughout the entire mixture analysis, or should one postulate linear curves in some trajectories instead. I've found some authors using the quadratic setting derived from single class analysis also for obviously not quadratic trajectories and consequently getting not sig. quadratic means for some trajectories. How does that influence the process of finding the number of trajectories in LGMM?
I don't think it hurts the process to use a quadratic throughout - if some classes are linear, that comes out as a special case as you say with insignificant quadratic mean.
linda beck posted on Monday, July 21, 2008 - 6:38 am
thank's a lot for your reply! There is a discrepancy between my LMRT and BIC. BIC prefers far more groups (4-5) then LMRT (2 groups). I thought that could have something to do with the skewness of my observed indicators. Do you have any other ideas?
I have a LTA that has been running for many hours, and is not close to finishing. I would like to stop it, but see the output so that I can inspect what's going on in the model. I tried control-C once before and got no output. Is there any way to stop the run and get preliminary output?