Improving Potential Reduction Scale a... PreviousNext
Mplus Discussion > Dynamic Structural Equation Modeling >
Message/Author
 Emilie Joly-Bura posted on Tuesday, June 30, 2020 - 8:22 am
Dear Mplus team,

I am working on AR(1) DSEM models. Some of my outcome variables at level2 are not normally distributed, but rather zero-inflated.
When I run my model, the PSR is absurdly high (728.312). In other models I obtain satisfactory PSR values, but still have problematic estimated parameters according to the autocorrelation plots.
My questions are :
1. What can I do to improve model convergence ?
2. Is there a way to treat zero-inflated data properly within the DSEM ?
3. How to deal with unsatisfactory diagnostic plots for autocorrelations (mainly for estimated parameters related to the zero-inflated variables)?


Here is a (truncated) input for a model with very high final PSR :

CLUSTER = Subject ;
USEVAR = (Level1 + Level2 VAR);
between = (Level2 VAR);
within = PM_window ;
LAGGED = logrt_PM(1) ;
TINTERVAL = Trial_PM(1) ;

ANALYSIS: TYPE = TWOLEVEL RANDOM ; ESTIMATOR = BAYES; PROCESSORS = 2; fbiter = (5000); BSEED = 41; THIN = 10; ALGORITHM = GIBBS(RW);

MODEL:
%WITHIN%
p_rt_PM | logrt_PM ON logrt_PM&1;
slowing | logrt_PM ON PM_window;

%BETWEEN%
TBPM_PM_ACC ON p_rt_PM slowing (Level2 VAR);
PM_ACC ON p_rt_PM slowing (Level2 VAR);
 Bengt O. Muthen posted on Wednesday, July 01, 2020 - 4:20 pm
There are many issues to consider, including:

Don't use Proc=2 (see Version History for 8.3)

You most likely don't need RW - if Mplus says you do, then you have most likely not specified the model correctly

Because of your random AR, make sure you have at least 10-20 time points for most subjects and that most subjects change over time

When you say zero-inflated data, perhaps you mean a strong floor effect (and not counts). Such as with negative affect. There are ways to handle this using two-part modeling and we are in the process of writing an applied paper on this.

If this doesn't help, send your output and data to Support along with your license number.
 Emilie Joly-Burra posted on Monday, July 06, 2020 - 3:37 am
Dear Dr. Muthén,

Thank you, it worked !
If I have convergence issues with subsequent models, I will contact the support service.

I do have counts data for one of my outcome variables (from 0 to 6). I'm looking forward to read your paper on the subject. In the meantime, did you already present how to treat these kind of data during one of the webinars or elsewhere ?

Kind regards,
 Bengt O. Muthen posted on Monday, July 06, 2020 - 4:45 pm
Two-part DSEM modeling has not been described yet in webinars but it is coming later this summer. Count modeling has several webinars but not with DSEM.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a public posting area. Enter your username and password if you have an account. Otherwise, enter your full name as your username and leave the password blank. Your e-mail address is optional.
Password:
E-mail:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: