Questions about LCGA with continuous ... PreviousNext
Mplus Discussion > Growth Modeling of Longitudinal Data >
 Kiki van Broekhoven posted on Wednesday, February 22, 2017 - 4:38 am

I am running a LCGA on a continuous variable assessed at 7 consecutive time points. I have a few questions left after consulting the MPlus manual.

- I have included only those participants who have filled out the assessments at time point 1 and at least one other time point. Is that sufficient, or should participants have filled out more than 2 measurement points?

- The continuous variable that I use is left skewed with an abundance of scores equal to 0. Therefore I have tried to run a LCGA with a censored distribution. However, entropy values are quite less optimal with a censored model than with a "normal" model. For example: entropy in 3-class "normal" model = 0,834; entropy in 3-class censored model = 0,783. Should I consider using a non-censored model despite the fact that the continuous variable I use is left skewed?

- In the Mplus syntax, do I have to state somewhere that my variable is a continuous variable? Right now, in the non-censored model I state nothing about the measurement level of my variable and in the censored model I only state "censored=eds1-7(b)". I can find commands in the manual for categorical variables, count variables etc but not for continuous variables.
 Bengt O. Muthen posted on Wednesday, February 22, 2017 - 12:30 pm
q1: You can use all people who have at least one time point.

q2: I would use a censored model - or a two-part model - because it is more suitable for the data. The entropy consideration is secondary (you have to take what you get).

q3: No, continuous is the default.
 Kiki van Broekhoven posted on Friday, February 24, 2017 - 6:35 am
Thank you very much for your quick reply!

I have one additional question: through the syntax TECH7 I can request the statistics for each class. I am interested in the Mean scores for each time point in each of the classes, which indeed are given. But is there also a possibility to request accompanying standard deviations and confidence interval's for each of these time points, for each class?

Thank you in advance!
 Bengt O. Muthen posted on Friday, February 24, 2017 - 4:59 pm
 Kiki van Broekhoven posted on Monday, February 27, 2017 - 7:56 am
Okay, thank you.

I am now deciding the optimal number of classes. I consulted BIC, LMR-LRT, and BLRT.

The BLRT is significant at the p<.001 level for all models I run (1-7 class models). The LMR-LRT p-value is .47 for the 4 versus 3 class model, but after that again becomes significant for the 5 versus 4 class model: p=.004. Does this mean that I could also explore the possibility of using the 5-class model, or should I already have stopped after I found that the 4-class model (compared to the 3-class model) was not significant based on the LMR-LRT value? (and should I as such stick with the 3-class model?)

Thank you in advance.
 Bengt O. Muthen posted on Monday, February 27, 2017 - 6:48 pm
You should stop when you first find the non-significance. So use 3 classes.
 Kiki van Broekhoven posted on Tuesday, June 13, 2017 - 8:19 am

I have another question about using a left skewed continuous dependent variable in GMM. The variable I use is obviously skewed to the left, but opposed to the variable that I was talking about above (post: posted on Wednesday, February 22, 2017 - 4:38 am), there is not an abundance of scores equal to 0. For the currenct variable, the values range from 0-26 with a median, mode, and interquartile range of 5 (which also indicates the skewness).

For this variable, would it also be appropriate to use CENSORED = x1-x3(b)
to indicate that the scores are censored with a floor effect, even though for this variable there maybe isn't a "true" floor effect as there is not an abundance of scores equal to zero?

Thank you in advance.
 Bengt O. Muthen posted on Tuesday, June 13, 2017 - 9:10 am
No, don't use censored for this. Instead, use a skew-t distribution for the outcome as shown in the paper on our website:

Muthén, B. & Asparouhov T. (2015). Growth mixture modeling with non-normal distributions. Statistics in Medicine, 34:6, 1041–1058. DOI: 10.1002/sim6388
 Kiki van Broekhoven posted on Wednesday, June 14, 2017 - 4:40 am
Thank you for your reply, this was very helpful.

Would this syntax setup be correct? (mainly with regard to the DISTRIBUTION = SKEWT) or are there additional things that I should specify?

MISSING = all (999);
CLASSES = c(1);
STARTS = 20 4;
i s | TPDSNA1@0 TPDSNA2@1 TPDSNA3@2;
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