Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

RE: st: Latent class estimation


From   "Verkuilen, Jay" <[email protected]>
To   <[email protected]>
Subject   RE: st: Latent class estimation
Date   Wed, 5 Dec 2007 12:45:26 -0500

Phil Schumm wrote:

>>>I would like to offer an alternative opinion.  RE references that are

accessible, Bartholomew's "Latent Variable Models and Factor  
Analysis" (1987, Charles Griffin & Co.) offers a good theoretical  
introduction, and Skrondal and Rabe-Hesketh's "Generalized Latent  
Variable Modeling" (2005, Chapman & Hall) is superb in terms of both  
its theoretical and applied presentation.  The latter should be read  
in any case without question.<<<

Bartholomew's book had a second edition (co-authored with Martin Knott).


There's also a Stata-specific book Multilevel and Longitudinal Modeling
Using Stata by Skrondal and Rabe-Hesketh. I would recommend seeing
http://www.gllamm.org for example code, the GLLAMM manual, etc. 


>>>The program -gllamm- (type -ssc install gllamm-) is capable of  
estimating latent class models in a very flexible manner, and the  
manual available with the program (as well as the book cited above)  
offer several worked examples.  It can be slow for datasets with a  
large number of unique covariate patterns, but this has to be  
evaluated relative to the time required to learn how to use another  
program and to move your data and results back and forth.  It's  
certainly an excellent place to start and to run some preliminary  
analyses; you can always then move to another piece of software if  
speed becomes a problem.<<<

Yeah, LatentGOLD, LEM, and Mplus would all be the ones I'd figure as
places to go if GLLAMM doesn't work out, or else coding things up in
software like winBUGS. There might be an R package of some sort that
does this; there are so many I can't possibly keep up with them all.

IMO it is wise to do the analysis in multiple software packages---these
models are tricky. Mixture regression is particularly known for having
multiple optima problems and run the serious risk of capitalizing on
chance, especially in "exploratory" mode when there aren't covariates
predicting class membership or other things to constrain the model. 

Jay

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index