Thank you for this video Prof @QuantFish! I have collected data at two time points. At Time 1, there are 5 latent profiles, and at Time 2, there are 4 latent profiles. I am planning to conduct a latent transition analysis. However, most of the literature I’ve found only discusses methods for cases where the same number of profiles are identified at both time points. Do you know of any documentation or resources that explain how to perform a latent transition analysis when the number of profiles differs between time points? Thank you.
At least Mplus can handle multiple categorical latent variables (latent class variables) with different numbers of classes without any problems. Therefore, you should be able to specify an LTA model with different numbers of classes at each time point in the program. Best, Christian Geiser
How about the part labeled odds? we focused on diagonal part right? If I have other covariates. That diagonal odds showed how each variable infleunce class transition right? Thank you
Dear Dr. Geiser, thanks for the video. I got a question about the output missing. I ran a model with 8 continuous variables, both at T0 and T1 separately. I determined that the number of profiles was 3 for both timepoints. I got the report of transition probabilities but no transition probability odds report like your video showed. How can I get it? My Mplus version is 8.3. Thanks you very much.
Thank you very much for your informative videos, which have been very helpful for my own analyses. However, I have a question regarding the LTA. I conducted an LTA in Mplus and now I am getting the following error message: "ONE OR MORE MULTINOMIAL LOGIT PARAMETERS WERE FIXED TO AVOID SINGULARITY OF THE INFORMATION MATRIX. THE SINGULARITY IS MOST LIKELY BECAUSE THE MODEL IS NOT IDENTIFIED, OR BECAUSE OF EMPTY CELLS IN THE JOINT DISTRIBUTION OF THE CATEGORICAL LATENT VARIABLES AND ANY INDEPENDENT VARIABLES. THE FOLLOWING PARAMETERS WERE FIXED: Parameter 27, C2#3 ON C1#3, Parameter 25, C2#1 ON C1#3". Could you perhaps give me a tip on how to best proceed with my analyses? I assume I shouldn't interpret the output without further consideration, correct? Thanks in advance.
This often simply means that some of the transition probabilities were estimated to be exactly zero, which is a boundary value for probabilities. It is not necessarily a problem as long as these zeros are plausible (some transitions may be completely unlikely so zero could be plausible). I would check all transition probabilities carefully to make sure they all make sense. Best, Christian Geiser
Hello. Thank you for the clear explanation :) I have one question regarding what to report if there are differences between the results of (1) FINAL CLASS COUNTS AND PROPORTIONS FOR EACH LATENT CLASS VARIABLE BASED ON THE ESTIMATED MODEL (2)FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS PATTERNS BASED ON THEIR MOST LIKELY LATENT CLASS PATTERN in the LTA output. I initially planned to report (2), but I discovered that the LATENT TRANSITION PROBABILITIES in the output is only based on (1). Then, is it appropriate to report the FINAL CLASS COUNTS AND PROPORTIONS & LATENT TRANSITION PROBABILITIES based on the estimated model? (I'm confused because the membership in the cprob file I got and this result are different.)
Yes, in my opinion it is best to report the values that are based on the estimated model (i.e., the estimated model parameters) as these don't require individual estimation of class membership/assignment of individuals to classes based on their most likely patterns or class membership. Best, Christian Geiser
Hello Prof, thanks for your videos and great explanations. I have a question about LPA and LTA. I ran a model with 9 continuous variables, both at T0 and T1 separately. I determined that the number of profiles was 3 for both timepoints. I went on with the description of the groups and tested their differences with ANOVAs and other tests. However, when I went on with LTA testing I noticed that the distribution of the profiles were a bit different than the ones obtained in the LPA (e.g, in the LPA I obtained 0.39, 0.14 and 0.45 at T0, and in the LTA I obtained 0.39, 0.11 and 0.49 at T0; small differences also at t1). Since there is a slight difference that involves a few individuals, I ask, why this happens, and if I can keep on with relying on the results of the LPA that I ran initially. Thanks in advance!
It is difficult to say without seeing your data and analyses in detail. It could be due to a variety of factors including (but not limited to) the additional information gained from analyzing all indicators simultaneously in the LTA, local likelihood maxima, missing data, measurement invariance constraints included in the LTA model (if any). I would check to see whether the profiles look equivalent and whether there is any indication of local likelihood maxima. Best, Christian Geiser
@@QuantFish Thanks for your feedback. It looks like no issues are present: no missing values, and no local maxima. Probably the slight differences lie on the fact that LTA uses all the repeated measures together, thus increasing statistical power. If this is the case, does it make sense to rely on the profiles of the LPA models run separately and present the transition probabilities from the LTA in a paper, or you think it is better to describe the profiles as turned out by the LTA? Thanks
@@Paolo-tu2ft I would report the set of parameter estimates (class proportions, class profiles, and transition probabilities) from the final LTA model since it uses all available data in a single model. That way, you would not have to report parameter estimates from two different models, which could be confusing. Best, Christian Geiser
Dear Dr. Geiser - I am working on a LTA with longitudinal data, with 6 timepoints with several classes each (4,4,5,5,5,5). After aroung ~20 hours, Mplus has stopped running the model, and I am not given any log or output. I have of course tried the script with fewer classes/timepoint and it runs smoothly. Do you know what the issue could be due to? Mplus support is not super helpful. Thanks in advance.
Have you tried a model with measurement invariance constraints (equal conditional response probabilities/equal means ) across time? Oftentimes, an time-invariant class model is easier to fit because it does not have as many free parameters that need to all be estimated. In general, it is actually better to have more time points since this adds information to the estimation procedure. But an unconstrained model may be difficult to fit because of the large number of free parameters (especially in a smaller sample). Best, Christian Geiser
thank you for your sharing. I have a question to ask you. After I use latent transition analysis to obtain the transition probability matrix, can I use probability to identify one or several subgroups that are meaningful based on the research context? Then conduct a difference analysis based on the baseline data of the subgroups that have undergone transformation and those that have not transformed. Is this statistical idea acceptable?
I'm not exactly sure what you mean. You can add covariates to an LTA model to examine whether those are related to (and can predict) class membership and/or transition probabilities. See: ruclips.net/video/7crrDHnszEE/видео.html Best, Christian Geiser
@@QuantFish Thank you for your answer! I have another question to ask you. Because the potential transformation analysis is based on vertical research, what are the requirements for the sample of this statistical method? How should the sample quantity be calculated? thank you for your reply!
Thank you for your reply. I'm sorry that my expression was not standardized enough. The question I would like to address is: LTA is a statistical technique that typically relies on longitudinal data collected at a minimum of two time points. In a more rigorous approach, it is common practice to first conduct LCA/LPA on the data, followed by latent transition analysis. My inquiry pertains to the calculation of sample size for such longitudinal studies.🤔@@QuantFish Best, Yuqi
Thank you for this video Prof @QuantFish! I have collected data at two time points. At Time 1, there are 5 latent profiles, and at Time 2, there are 4 latent profiles. I am planning to conduct a latent transition analysis. However, most of the literature I’ve found only discusses methods for cases where the same number of profiles are identified at both time points. Do you know of any documentation or resources that explain how to perform a latent transition analysis when the number of profiles differs between time points?
Thank you.
At least Mplus can handle multiple categorical latent variables (latent class variables) with different numbers of classes without any problems. Therefore, you should be able to specify an LTA model with different numbers of classes at each time point in the program.
Best,
Christian Geiser
hi Dr. Geiser. i wonder that how do we expalin transitiion odds in mplus output
How about the part labeled odds? we focused on diagonal part right? If I have other covariates. That diagonal odds showed how each variable infleunce class transition right? Thank you
Dear Dr. Geiser, thanks for the video. I got a question about the output missing. I ran a model with 8 continuous variables, both at T0 and T1 separately. I determined that the number of profiles was 3 for both timepoints. I got the report of transition probabilities but no transition probability odds report like your video showed. How can I get it? My Mplus version is 8.3. Thanks you very much.
I'm not sure why you didn't get the transition probability odds. I would ask the Mplus support team (support@statmodel.com).
Best, Christian Geiser
Thank you very much for your informative videos, which have been very helpful for my own analyses. However, I have a question regarding the LTA. I conducted an LTA in Mplus and now I am getting the following error message: "ONE OR MORE MULTINOMIAL LOGIT PARAMETERS WERE FIXED TO AVOID SINGULARITY OF THE INFORMATION MATRIX. THE SINGULARITY IS MOST LIKELY BECAUSE THE MODEL IS NOT IDENTIFIED, OR BECAUSE OF EMPTY CELLS IN THE JOINT DISTRIBUTION OF THE CATEGORICAL LATENT VARIABLES AND ANY INDEPENDENT VARIABLES. THE FOLLOWING PARAMETERS WERE FIXED: Parameter 27, C2#3 ON C1#3, Parameter 25, C2#1 ON C1#3". Could you perhaps give me a tip on how to best proceed with my analyses? I assume I shouldn't interpret the output without further consideration, correct? Thanks in advance.
This often simply means that some of the transition probabilities were estimated to be exactly zero, which is a boundary value for probabilities. It is not necessarily a problem as long as these zeros are plausible (some transitions may be completely unlikely so zero could be plausible). I would check all transition probabilities carefully to make sure they all make sense.
Best, Christian Geiser
@@QuantFish Thank you very much for your quick and helpful reply!
Hello. Thank you for the clear explanation :) I have one question regarding what to report if there are differences between the results of (1) FINAL CLASS COUNTS AND PROPORTIONS FOR EACH LATENT CLASS VARIABLE BASED ON THE ESTIMATED MODEL (2)FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS PATTERNS BASED ON THEIR MOST LIKELY LATENT CLASS PATTERN in the LTA output. I initially planned to report (2), but I discovered that the LATENT TRANSITION PROBABILITIES in the output is only based on (1). Then, is it appropriate to report the FINAL CLASS COUNTS AND PROPORTIONS & LATENT TRANSITION PROBABILITIES based on the estimated model? (I'm confused because the membership in the cprob file I got and this result are different.)
Yes, in my opinion it is best to report the values that are based on the estimated model (i.e., the estimated model parameters) as these don't require individual estimation of class membership/assignment of individuals to classes based on their most likely patterns or class membership.
Best, Christian Geiser
Hello Prof, thanks for your videos and great explanations. I have a question about LPA and LTA. I ran a model with 9 continuous variables, both at T0 and T1 separately. I determined that the number of profiles was 3 for both timepoints. I went on with the description of the groups and tested their differences with ANOVAs and other tests. However, when I went on with LTA testing I noticed that the distribution of the profiles were a bit different than the ones obtained in the LPA (e.g, in the LPA I obtained 0.39, 0.14 and 0.45 at T0, and in the LTA I obtained 0.39, 0.11 and 0.49 at T0; small differences also at t1). Since there is a slight difference that involves a few individuals, I ask, why this happens, and if I can keep on with relying on the results of the LPA that I ran initially. Thanks in advance!
It is difficult to say without seeing your data and analyses in detail. It could be due to a variety of factors including (but not limited to) the additional information gained from analyzing all indicators simultaneously in the LTA, local likelihood maxima, missing data, measurement invariance constraints included in the LTA model (if any).
I would check to see whether the profiles look equivalent and whether there is any indication of local likelihood maxima.
Best, Christian Geiser
@@QuantFish Thanks for your feedback. It looks like no issues are present: no missing values, and no local maxima. Probably the slight differences lie on the fact that LTA uses all the repeated measures together, thus increasing statistical power. If this is the case, does it make sense to rely on the profiles of the LPA models run separately and present the transition probabilities from the LTA in a paper, or you think it is better to describe the profiles as turned out by the LTA? Thanks
@@Paolo-tu2ft I would report the set of parameter estimates (class proportions, class profiles, and transition probabilities) from the final LTA model since it uses all available data in a single model. That way, you would not have to report parameter estimates from two different models, which could be confusing.
Best, Christian Geiser
@@QuantFish many thanks for your support!
Dear Dr. Geiser - I am working on a LTA with longitudinal data, with 6 timepoints with several classes each (4,4,5,5,5,5). After aroung ~20 hours, Mplus has stopped running the model, and I am not given any log or output. I have of course tried the script with fewer classes/timepoint and it runs smoothly. Do you know what the issue could be due to? Mplus support is not super helpful.
Thanks in advance.
Have you tried a model with measurement invariance constraints (equal conditional response probabilities/equal means ) across time? Oftentimes, an time-invariant class model is easier to fit because it does not have as many free parameters that need to all be estimated.
In general, it is actually better to have more time points since this adds information to the estimation procedure. But an unconstrained model may be difficult to fit because of the large number of free parameters (especially in a smaller sample).
Best,
Christian Geiser
thank you for your sharing. I have a question to ask you. After I use latent transition analysis to obtain the transition probability matrix, can I use probability to identify one or several subgroups that are meaningful based on the research context? Then conduct a difference analysis based on the baseline data of the subgroups that have undergone transformation and those that have not transformed. Is this statistical idea acceptable?
I'm not exactly sure what you mean. You can add covariates to an LTA model to examine whether those are related to (and can predict) class membership and/or transition probabilities.
See: ruclips.net/video/7crrDHnszEE/видео.html
Best, Christian Geiser
Thank you for your reply! ❤I will seriously study what you mentioned. Have a great day!@@QuantFish
@@QuantFish Thank you for your answer! I have another question to ask you. Because the potential transformation analysis is based on vertical research, what are the requirements for the sample of this statistical method? How should the sample quantity be calculated? thank you for your reply!
@@kennethmorris4657 I'm not sure what you mean by "vertical research" or "sample quantity."
Best, Christian Geiser
Thank you for your reply.
I'm sorry that my expression was not standardized enough.
The question I would like to address is: LTA is a statistical technique that typically relies on longitudinal data collected at a minimum of two time points. In a more rigorous approach, it is common practice to first conduct LCA/LPA on the data, followed by latent transition analysis. My inquiry pertains to the calculation of sample size for such longitudinal studies.🤔@@QuantFish
Best, Yuqi