I have it on pretty good authority that it's pretty difficult to obtain solutions for LCA and growth mixture models (GMMs) even when you have as many as 500 cases. My source (a former post-doc of Bengt Muth閚, an author of MPlus) was/is working with a student on GMM simulations, and he has found that unless the data are *very* well-behaved, you need several hundreds of cases to get a stable model. With your sample size and 30(!) variables, I'm not surprised to hear about your problem. Is there any way you can drop and/or combine some of your covariates or otherwise simplify your model?

Good luck!

Mike

Dear Users,

I am desperately trying to fit a zero-inflated Poisson model with latent classes by using SAS. Can anyone help?

I have a data set that has 399 observations in it. In data set, I have 4 dependent variables, 25 covariates and one control variable. I have got problem when I tried to run latent class analysis with Mplus. With Mplus, classification based on means seems OK. Classification based on means, variance and residual, however, had converge problem. Outcomes in dependent variables are categorical and range 0 - 6 and clearly zero-inflated. I tried run PROC TRAJ in SAS (Jones et al.,) but this procedure doest work on 8.0. I need run SAS for following models,

Models will be 1) Yi =3DMean + F1 + ei F1=3D x1+x1+.......x25

2) Yi =3DMean + F1 + Controlvariable + ei F1=3D x1+x1+.......x25

Thanks in advance,

Hayretti

[此贴子已经被作者于2005-10-28 8:58:49编辑过]

If you are doing a true latent class analysis, then you don't know which groups the consumers are in and you wish to make inferences about it. Latent class analysis is an inferential cluster analysis in a sense. Mplus will do it. Freeware, WINLTA, stands for Windows latent transition analysis, is available from Penn State through Linda Collins. http://methodology.psu.edu/downloads/winlta.html. Even though its called latent transition analysis, it will also do latent class analysis.

Paul R. Swank, Ph.D. Professor, Developmental Pediatrics Medical School UT Health Science Center at Houston

I am trying to estimate a latent class choice model. I think there are two groups of consumers 1 and 2 and they derive different forms of utilities from choosing a product. Assume there are J products. If a consumer belongs to type I, he derives utility of the form U(1,j)= function1 (product j); If she belongs to type II, he/she derives utility of the form U(2,j)= function2 (product j);Now, I don't whether a consumer is of type I or tye II. I would like to use a binary logit to model the probability of a consumer belonging to either of the segment.Can PROC NLMIXED handle this? If so, can you suggest a code?

I'm sure that PROC NLMIXED can do it. But, if you're really working with latent variablers models, then why don't you take a look at PROC CALIS, which is designed with such problems in mind.

HTH, David -- David Cassell, CSC Cassell.David@epa.gov Senior computing specialist mathematical statistician

THIS IS FROM THE ASA METHODOLOGY SECTION IN RELATION TO LATENT CLASS ANALYSIS
We are extremely pleased to announce that Jeroen Vermunt is the first
recipient of the Leo Goodman Award of the ASA Methodology Section.

The Leo Goodman Award, established in 2004 by the Methodology Section
of the American Sociological Association, recognizes contributions to
sociological methodology made by a scholar who is no more than 10
years past Ph.D.  The Goodman award is made every other year.

This year's award committee consists of Scott Eliason, Adrian Raftery,
and Yu Xie (chair), has awarded the 2005, and the first, Leo Goodman
Award, to Jeroen Vermunt of Tilburg University, The Netherlands.
Since he received his PhD in 1996, Vermunt has made outstanding
contributions to sociological methodology as well as innovative uses
of sociological methods in substantive research.

Jeroen Vermunt’s primary methodological contributions are in the area
of categorical data analysis, with particular attention to latent
heterogeneity.  Using a latent class analysis approach, he has
incorporated into loglinear event history analysis methods for
handling missing data, unobserved heterogeneity, censoring, and
measurement error.  He has also successfully applied the same approach
to classification and clustering analysis, and multi-level and random
coefficient models for categorical data.  In his recent work, he has
made original and important contributions to the analysis of ordered
data with different flexible constraints.

An important feature of Vermunt’s work is that he has made his
methodological work accessible for social scientists by developing
user-friendly computer programs: LEM and Latent Gold (the later of
which he jointly developed with Jay Magidson).  He has taught and
presented workshops worldwide, and has collaborated and published
together with many substantively interested social science
researchers.

In sum, in the nine years since his PhD, Jeroen Vermunt has
established himself as a major contributor to sociological methodology
and a distinguished scholar of social science research.  To recognize
his important contributions in his early career, the Methodology
Section of the American Sociological Association is pleased to present
Jeroen Vermunt with the 2005 Leo Goodman Award.

One of the things that happens in science is separate discovery of mathematically / logically identical approaches with different terminology or the same terms with different meanings. Nominal level variables with two legitimate values have been called dichotomies, binary, quantal, dummy, or flag variables. Usually a "sample", is a set of entities / cases / elements that are selected via a probability based process, but sometimes a sample is a single entity/case/element selected via a probability based process. Binary variable can mean a dichotomous variable or an internal computer representation of any number. If I am reading Hector's post correctly, LCA can be a grouping / ordination / clustering of variables. Now that I try harder to recall, I think it was in the 70's, most likely at a joint meeting of the Classification and Psychometric Societies that I saw a presentation that showed each of the Lazarfeld approaches was formally identical to an approach that was called by some other name is psych. I posted to class-l asking about this and will post a summary of responses here if I get some. Art Art@DrKendall.org Social Research Consultants Inside the Washington, DC beltway.

Not really, LCA clusters (in the everyday, not the mathematical meaning of the word) cases, not variables, even though some discernable clustering of variables exists, if you look at the parameter estimates.

While this may be true -- LCA is in the end a special case of a log-linear model -- the interpretation of different mathematically identical models may be very different. The LCA model makes particularly sense, because it is straightfoward to interpret its parameters. If you would write the same thing in log-linear model terms, it would be much harder to interpret, as you would have to perform cumbersome mathematical transformations. Thomas -- thomas koenig http://www.lboro.ac.uk/research/mmethods/staff/thomas/index.html

I would like to second this recommendation: It seems to me currently the most comprehensive textbook and it is fairly readable. First, I would read two primers, though. I would start with Lazarsfeld's original formulation, which is even more readable and then continue with McCutecheon's book in Sage's green methodology series A supplementary source is: http://ourworld.compuserve.com/homepages/jsuebersax/index.htm Thomas -- thomas koenig http://www.lboro.ac.uk/research/mmethods/staff/thomas/index.html

Paul Lazarsfeld proposed latent structure analysis in the 1950s, and Lazarsfeld and Henry's book appeared in 1968. At the time, the estimation methods for latent structure analysis and latent class analysis were not very good. Leo Goodman wrote his classic articles in the 1970s, which were implemented in the freeware program MLLSA. There have been advances since then, and now you can estimate latent class models in Latent Gold, Mplus, or the LEM freeware.

The point of your quote is that you must consider the measurement levels of the indicators and the latent variables. Traditional factor analysis works with continuous indicators and continuous latents; latent trait models (such as item response models) work with categorical indicators and continuous latents; latent profile models work with continuous indicators and categorical latents; and latent class models work with categorical indicators and categorical latent variables. Because the traditional factor model has continuous indicators and assumes a normal framework, it works with covariances. The traditional latent class model has categorical indicators, which can be understood as a multi-way table. Ton Heinen has a book on latent class and latent trait models that elaborates on these various models. Anthony Babinec

useful,thanks

It is really funny when i read the announcement located in the 11th floor.~~.