A Beginner's LTA question PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
 Lucy Barnard posted on Thursday, August 30, 2007 - 4:01 pm
If you don't exactly know how your classes move across time but you are pretty sure that they do in fact transition, would it possible to examine a saturated LTA model using a random half of your sample? Then, take the significant paths and confirm this model with the random second half of the sample?

Many thanks,

Lucy Barnard
 Linda K. Muthen posted on Friday, August 31, 2007 - 8:54 am
If you have a large enough sample, you could randomly split it and do the analysis in both halves to see if you get the same results. I would not fix any parameters.
 Lucy Barnard posted on Saturday, September 01, 2007 - 9:46 am
Is there any way to do multiple latent class analyses within the same model and then compare the results? I basically want to see if the latent class structure I observe at the first time point is stable across time.
 Linda K. Muthen posted on Saturday, September 01, 2007 - 1:28 pm
As a first step in a latent transition analysis, you should do a latent class analysis for each timepoint to see how many classes fit the data at each timepoint. The dissertation on the website by Karen Nylund describes the steps to take to carry out a latent transition analysis. I suggest taking a look at it.
 Raheem Paxton posted on Friday, June 19, 2009 - 2:50 pm

I not sure if this is a crazy question or not, but here it is. First let me say that I have cross-sectional data, but the questionnaire that Iím looking at asked participants to respond to questions related to their past body weight and physical activity behavior at multiple time points (e.g., age 20 to 25, age 15- 20). As expected this data also asks questions about current behavior. Can I use LTA to determine how participants transition from old health behavior to their current behavior? Again, sorry but Iím really not sure. Thanks
 Bengt O. Muthen posted on Friday, June 19, 2009 - 6:01 pm
I think so, with the usual caveat about "telescoping effects".
 Julia Lee posted on Saturday, June 02, 2012 - 12:25 pm
Hi Dr. Muthen,

I am using LTA mover-stayer modeling for continuous variables. Is it possible for movers to "move" within the same class for a LTA mover-stayer modeling for continous variable?

The mover and stayer means for my output are different. For 1 1 1 (movers class 1 class 1), the class count is 98.54, proportion is .18. The essential question is whether the 98 students moved (since they are classified as movers) within the same class and mixture modeling in Mplus decided that they had only "moved" within a particular threshold. Did the students make just enough improvements to remain in Class 1 but did not move to Class 2 because of a particular threshold? Conceptually, wouldn't this type of movers (e.g. 1 1 1) be considered "movers who moved within the same class" and aren't they (logically) portraying the actual stability of group membership even though 69% of the students were classified as movers?
1 1 1 98.54240 0.18914
1 1 2 0.00000 0.00000
1 2 1 56.89545 0.10920
1 2 2 162.25686 0.31143
2 1 1 56.41806 0.10829
2 1 2 0.04694 0.00009
2 2 1 0.00487 0.00001
2 2 2 146.83542 0.28183

Thank you in advance.
 Bengt O. Muthen posted on Saturday, June 02, 2012 - 1:39 pm
It doesn't look like your results for movers (first class=1) indicate that they behave like movers since most stay in the same class (classes 1 1 1 and 1 2 2 are the most frequent). You don't move if you stay in the same class over time.
 Julia Lee posted on Saturday, June 02, 2012 - 7:00 pm
So the classes 1 1 1 and 1 2 2, although being classified as "movers," were conceptually stayers? Then they are not any different from the stayers (i.e., Classes 2 1 1 and 2 2 2). Is it possible to have a large number of classified as movers (61%) yet, they are 1 1 1 or 2 2 2? Did I miss something? I appreciate your insight. Thank you very much.

Latent Class
Variable Class

C......... 1 317.69470 0.60978
.......... 2 203.30531 0.39022
C1.........1 155.00745 0.29752
...........2 365.99255 0.70248
C2.........1 211.90700 0.40673
...........2 309.09302 0.59327
 Bengt O. Muthen posted on Sunday, June 03, 2012 - 9:40 am
Well, you have 10.9% in 1 2 1, so those people are moving. It is the 0% in 1 1 2 that don't want to move - perhaps going from 1 to 2 is a hard transition to make (an example would be knowing less at time 2 than at time 1). I assume that you haven't specified your model to have the constraint that nobody is in the 1 1 2 cell.
 Julia Lee posted on Sunday, June 24, 2012 - 5:29 pm
Hi Dr. Bengt Muthen,

My question is regarding the LTA mover-stayer output on LATENT TRANSITION PROBABILITIES BASED ON THE ESTIMATED MODEL.

I am conducting a LTA mover-stayer with continuous variables. Is the output for the lower panel here for the movers only and hence the diagonal matrix refers to the stabiliity of the movers only (and not inclusive of the stayers' stability)? I am assuming measurement invariance. Stayers have zero probability of moving classes and movers have some probability of moving classes. I have different means for movers and stayers.

C Classes (Rows) by C1 Classes (Columns)
1 .......0.690...0.310
2 .......0.278...0.722
C1 Classes (Rows) by C2 Classes (Columns)
1....... 0.845...0.155
2....... 0.000...1.000

I appreciate your response. Thank you.
 Linda K. Muthen posted on Monday, June 25, 2012 - 10:35 am
No, it is for both the movers and the stayers.
 Oxnard Montalvo posted on Thursday, September 06, 2012 - 11:00 pm
I was wondering whether the LTA parameterization 2 outlined in webnote 13 is equivalent to the model in Proc LTA (SAS), where x is includes as a covariate for both time 1 status and the transition probabilties?

i.e. in SAS Proc LTA

ITEMS u11 u12 u13 u14 u21 u22 u23 u24;
SEED 123456;
Covariates1 x;
Covariates2 x;

 Bengt O. Muthen posted on Friday, September 07, 2012 - 8:45 am
Please send/point to documentation that shows the PROC LTA parameterization.
 Oxnard Montalvo posted on Friday, September 07, 2012 - 5:29 pm
Hi Bengt,
The Proc LTA parameterization details are at http://methodology.psu.edu/downloads/proclcalta
in the User's Guide. Thankyou.
 Bengt O. Muthen posted on Friday, September 07, 2012 - 5:34 pm
Please point to a specific document and a specific page.
 Oxnard Montalvo posted on Friday, September 07, 2012 - 8:43 pm
Sorry, document is the User's Guide
v 1.2.7, pages 7-8
 Bengt O. Muthen posted on Monday, September 10, 2012 - 1:33 pm
Yes, it looks like the Mplus parameterization 2 of Web Note 13 is the same as used in Proc LTA.
 Jane Smith posted on Wednesday, March 05, 2014 - 7:45 am
Drs. Muthen,

I've run an LTA with a second order effect, an interaction and a covariate, and am trying to interpret the transitions. Specifically, how are the transitions under 'LATENT TRANSITION PROBABILITIES BASED ON THE ESTIMATED MODEL' different than those given by the LTA calculator--which I know are based on the levels of the covariates. In other words, how are the covariates/interaction allowed to influence the estimated model transitions? I've read webnote 13, but am unable to resolve my question.
 Linda K. Muthen posted on Thursday, March 06, 2014 - 2:07 pm
Please send the output and your license number to support@statmodel.com.
 Annie Desrosiers posted on Monday, January 19, 2015 - 7:36 am
Dear. Dr. Muthen

I recently started using LTA and have some questions regarding this technique. I have measured subjects at two time points but made a classification with LCA at Time 1. Is it possible to use the classification at time 1 to re-score the subjects with the same variable at Time 2 and export that new classification to see if the subjects have changed classification between the two time points ? If yes, do you perhaps have an example how to do this?

Thank you in advance.
 Bengt O. Muthen posted on Monday, January 19, 2015 - 9:07 am
I don't understand what you are suggesting. How is it different from doing LCA at each time point and comparing the classifications? Perhaps you are thinking of holding the measurement parameters of time 2 equal to time 1.
 Annie Desrosiers posted on Monday, January 19, 2015 - 9:36 am
Thanks for your quick reply, I prefer not to perform a LCA for each time point because I want to maintain the same classes.
Indeed I would like to hold on to the measurement parameters of time 1 and apply them to time 2. More specifically, I would like to maintain the orginal classes and see which of my participants moved between the classes at time 2.
 Bengt O. Muthen posted on Monday, January 19, 2015 - 10:41 am
You can fix all the time 2 parameters at the estimated values for time 1 so that there are no parameters to be estimated for time 2. That will give you a time 2 classification which can be different than at time 1 due to different time 2 outcomes.
 Annie Desrosiers posted on Tuesday, January 20, 2015 - 11:36 am
Thank you so much for your help.
I have been searching the Web for some examples, but havent come across similar approaches. Do you perhaps have a syntax I could build on? This would be very helpful because first of all, I am not sure where to find the parameters in my LCA output of time 1 and second, how to fix the parameters for LCA at time 2 for each class?
 Bengt O. Muthen posted on Tuesday, January 20, 2015 - 12:20 pm
Use the SVALUES output option which gives you the final estimates. You will see which estimates go with time 1. Then change * (free) to @ (fixed) globally when copying this to the Mplus input for running time 2.
 Daniel Lee posted on Friday, July 27, 2018 - 11:26 am
Hi Dr. Muthen,

I have conducted an LTA with a covariate (RACISM) to determine if RACISM influences the transitional probabilities of the latent profiles across two time points. Specifically, there were 3 profiles in both time points. I was wondering if you could help me interpret just one of the coefficients below (I excluded the s.e., p-value, and t-statistic). I know that the value is in log odds, but I'm not quite sure how to interpret this value conceptually. I appreciate your help as always!

Latent Class Pattern 1 1

C2#1 ON
RACISM 0.382
C2#2 ON
RACISM 0.194
 Bengt O. Muthen posted on Friday, July 27, 2018 - 12:16 pm
See Web Note 13. And also our Short Course Topic 10 which has a part focusing on LTA.
 Daniel Lee posted on Friday, July 27, 2018 - 4:33 pm
Thank you! That cleared things up. I have one more question. I was wondering why the p-value for the odds ratio is different from the p-values for the log odds. Thank you!
 Bengt O. Muthen posted on Friday, July 27, 2018 - 4:53 pm
They have different sampling distributions - the log odds is closer to normal. You use the p-value corresponding to the one you are interested in - the odds ratio or the logodds ratio.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message