| Message/Author |
|
|
| Xiao-min Li posted on Sunday, November 16, 2014 - 7:47 am
|
|
|
Dear Doctors:
I'm a Chinese student majoring in psychology in Beijing Normal University and now I am writing message to ask for help.
My graduate teacher wants me to analyze 2-wave longitudinal data (all variables are continuous) to explore behavior patterns and their transition across time. I have download data and syntax examples provided by official website but just found that all examples are about categorical variables. I¡®m wondering that whether I can run LTA with continuous variable. And if it is ok, following are some questions puzzling:
1. Whether estimation method or other specification should be changed in contrast to categorical data.
2. According to examples, number of category should be set in advance, which parameter in output can I use to test its appropriateness except for BIC and AIC for my graduate teacher ask me to provide more evidence;
3. In Latent Class Analysis, savedata syntax can save cprob to help researchers make sure the class each individual belongs to, so is it also possible in latent transition analysis? If possible, how to write the syntax?
That's all. Thanks for your patience and kindness in advance.
Best regards! |
|
|
|
|
Yes, LTA can also use continuous variables as latent class indicators. In fact, all variable types are allowed, including latent variables, and combinations.
1. No.
2. Apart from BIC you can also use the RESIDUAL option of the OUTPUT command.
3. Try the same approach as for LCA. |
|
|
| Xiao-min Li posted on Sunday, November 16, 2014 - 5:17 pm
|
|
|
thanks for your answering, I adapt the syntax of Hidden Markov model to the following:
ANALYSIS: ALGORITHM=INTEGRATION;
TYPE=MIXTURE;
MODEL:
%OVERALL%
c2 on c1;
MODEL c1: %c1#1%
[h11$1] (1);
[h12$1] (2);
[h13$1] (3);
%c1#2%
[h11$1] (6);
[h12$1] (7);
[h13$1] (8);
%c1#3%
[h11$1] (11);
[h12$1] (12);
[h13$1] (13);
MODEL c2: %c2#1%
[h21$1] (1);
[h22$1] (2);
[h23$1] (3);
%c2#2%
[h21$1] (6);
[h22$1] (7);
[h23$1] (8);
%c2#3%
[h21$1] (11);
[h22$1] (12);
[h23$1] (13);
OUTPUT: TECH1 TECH8 TECH11 TECH14 TECH15
But warning is output:
*** WARNING in MODEL command
All variables are uncorrelated with all other variables within class.
Check that this is what is intended.
*** ERROR
The following MODEL statements are ignored:
* Statements in Class %C1#1% of MODEL C1:
[ H11$1 ]
Thanks! |
|
|
|
|
| You don't use the $1 notation with continuous variables. |
|
|
| Xiao-min Li posted on Monday, November 17, 2014 - 6:15 pm
|
|
|
| Got it, and model executed successfully, thanks very much! |
|
|
| Xiao-min Li posted on Monday, November 17, 2014 - 6:43 pm
|
|
|
| One more question, I-state objects analysis (ISOA) can also be endorsed to explore transition of pattern across time? Is there any difference between ISOA and LTA? In assumption or output? Thanks! |
|
|
|
|
| I don't know what ISOA is. Do you have a reference that describes it? |
|
|
| Xiao-min Li posted on Tuesday, November 18, 2014 - 6:26 pm
|
|
|
L. R. Bergman, B. M. El-Khouri (1999) Studying Individual Patterns of Development Using I-States as Objects Analysis (ISOA)Biometrical Journal 41 (1999) 6, 753-770.
Thanks! |
|
|
|
|
| I glanced at the article and although there are similarities in goals, there seems to be many differences in methodology between ISOA and LTA. For one, ISOA uses cluster analysis whereas LTA uses mixture modeling. But I can't say more about ISOA since I have not studied it. |
|
|
| Xiao-min Li posted on Thursday, November 20, 2014 - 4:05 am
|
|
|
| Thanks a lot! |
|
|
| weiyi cheng posted on Wednesday, January 25, 2017 - 8:07 pm
|
|
|
Hi, I'm running a LTA model with 3 classes at two time points. Somehow, I got same transition probabilities for each time 1 class category. That can't be right I suppose? What might cause this then? I appreciate your insight!
C1 Classes (Rows) by C2 Classes (Columns)
1 2 3
1 0.369 0.403 0.228
2 0.369 0.403 0.228
3 0.369 0.403 0.228 |
|
|
|
|
| We need to see your full output to say. Please send to Support along with your license number. |
|
|
|
|
Hi Dr. Muthen,
My objective is to carry out LTA with continuous variables as latent class indicators. I have performed all the three steps as per webnote 15 (Subheading - "Estimating latent transition analysis using the 3-step method"). I can't see any other model fit information except AIC and BIC in the output of the last step.
VARIABLE:NAMES ARE y1-y8 cl1 cl2;
USEVARIABLES ARE cl1 cl2;
CLASSES = c1(3) c2(3);
Nominal are cl1 cl2;
(Only showing a part of model commands just to let you know about the classes)
My questions are:
How can I see any other model fit information?
Also, if you could please suggest, is it important to consider model fit information in step 3 when I have already assessed the model fit of latent classes (c1 and c2 separately) in previous steps.
Thanks and regards,
Anshuman |
|
|
|
|
Q1: There are not really any very useful fit statistics for continuous mixture modeling like LTA. Instead you have to work with a series of model and compare them using BIC. That is, models that differ by e.g. the number of classes or differ by imposing and not imposing measurement invariance, etc.
Q2: It could be useful (by using neighboring models) but perhaps not necessary. |
|
|
|
|
Thank you so much for your reply
Prof. Muthen.
Regards,
Anshuman |
|
|
|
|
| When conducting an LTA with categorical variables, response probabilities are generally set to be equal across time by placing $1 after each parameter. I want to compare models with full invariance and full noninvariance, but my LTA has continuous indicators. How does one fix the item-response probabilities to be equal across two timepoints when working with continuous indicators? I'm assuming this is done using "@" rather than "$", but I'm uncertain which value(s) should follow "@". Is it a single/standard value? Are there values that are output when running the freely estimated model? If so, where are these values in the output? |
|
|
|
|
Continuous indicators work with intercepts, so for one class:
[y1-y5] (1-5); |
|
|
|
|
| That worked well. Thank you for the quick response! |
|
|
| EH posted on Monday, March 11, 2019 - 3:18 am
|
|
|
Dear Dr. Muthén,
I am running a 4-class LTA with two latent variables for two timepoints (CLASSES = T1(4) T2(4);) I tried to run the two LCA's in one script in order to check measurement invariance later on.
I defined the latent variables for Model T1 and for Model T2 (%OVERALL% MODEL T1: PET1 by 1T1, 2T1, 3T1; MWT1 by 4T1, 5T1, 6T1; %T1#1 (2,3 and 4)% [MWT1]; [PET1];
MODEL T2: PET2 by 1T2, 2T2, 3T2; MWT2 by 4T2, 5T2, 6T2; %T2#1 (2,3 and 4)% [MWT2]; [PET2];
Yet I always get the same warning: *** ERROR in MODEL command
No OVERALL or class label for the following MODEL statement(s):
PET1 BY
Yet I should define my latent variables under the MODEL T1 and T2 commands, right?
Looking forward for your answer! |
|
|
|
|
After saying e.g.
Model T1:
you want to add
%t1#1%
.....
%t1#2%
....
See UG examples. |
|
|
| EH posted on Wednesday, March 13, 2019 - 11:55 am
|
|
|
Dear dr Muthén,
Thank you for your response. Unfortunately I cannot find a UG example using latent variables as indicators. To check for measurement invariance I specified my latent factors under the %overall% command.
Can I check measurement invariance over time and of classes over time?
Because running a model in which I specified (metric and scalar) measurement invariance for the variables and constrained the means of the classes per time indicated to be equal did not seem to work:
CLASSES = T1(4) T2(4);
%overall%
PET1 by 11U-14U (1-4);
MWT1 by 15U-17U (5-7);
PET2 by 21U-24U (1-4);
MWT1 by 25U-27U (5-7);
[11U - 21U] (111);
...
MODEL T1:
%T1#1%
[MWT1*] (A);
[PET1*] (B);
PET1 WITH MWT1*;
%T1#2%
[MWT1*] (C);
[PET1*] (D);
PET1 WITH MWT1*;
...
MODEL T2:
%T2#1%
[MWT3*](A);
[PET3*](B);
PET3 WITH MWT3*;
... |
|
|
|
|
This looks like a good start. Note, however, that for identification factor means need to be fixed at zero in at least one place - so in one class at one time point.
If this doesn't help, send your output to Support along with your license number. |
|
|
|
|
Dr.Muthen,
Hello!I am now conducting a study about students' academic motivation. The data were collected at 2 waves.
First, I ran LPA for 2 two time points respectively, the results indicated that a 3 class model should be chosen for each time point. But when I ran LTA, the results indicated that c1 should have 2 classes because the first class has only 3 cases(0.5% of the population) and should be ignored, while c2 should have 3 classes and each class has enough cases for becoming significant.
What should I do now? Or how should I interpret these results? Thank you for shedding light on my questions!
User Wen |
|
|
|
|
| I would keep the 3 classes at both time points because you are interested in seeing how many transition from the small class at time 1 to other classes at time 2. I assume that you impose measurement invariance. |
|
|
|
|
| Thanks for your advice! I will try to interpret the results. |
|
|
|
|
Dr.Muthen,
Hello! Two more questions, need your help.Thank you in advance!
I ran a LTA including an interation of X and C1 in the model, in accordance with your instructions in the LTA webnote in 2011. My data's covariate has 3 categories, is there any modification I have to specify in the syntax? I specified categorical=x but the program incicated that only dependent variables should be specified in categorical command. Without specification of the 3 categories of covariate, I think the programme may treat it as continuous? The regression estimates of different classes of c1,c2 on x don't make any sense for me as I want to see how each category of X influence them.
I see that your webnotes get the transition probabilities with different values of the binary covariate, I also need them for 3 categories, but can't see the results in the output. Do I need to calculate them by hand? |
|
|
|
|
| When your covariate has 3 categories, you need to create 2 dummy variables and those 2 variables should both have different effects for the different C1 categories. Because the dummy variables are covariates, they should not be put on the Categorical list. They should be treated as continuous like we usually do in regression analysis. |
|
|
|
|
Dear support,
I am running a two-wave LTA model with continuous indicators of latent classes. One of the steps in such an analysis involves constraining parameters between waves. I wonder which parameters should be held equal. Only means? Or means and variances?
Also, how can I test if such constraints are reasonable and do not impact model fit? LTA with continuous indicators does not print chi-square value, only log-likelihood and information criteria indexes. Is there some established way of testing model-fit decrease in such models?
Thank you! |
|
|
|
|
Constrain only the means.
First apply across-time constraints on the means for all indicators. Then free the constraint for one indicator at a time and compute a chi-square difference test as 2 times the loglikelihood difference with df equal to the diff in number of parameters. |
|
|
|
|
| Thank you!! |
|
|
| Back to top |
