Complex mixture (multilevel growth mi... PreviousNext
Mplus Discussion > Growth Modeling of Longitudinal Data >
Message/Author
 Dave Flora posted on Friday, March 28, 2003 - 11:19 am
I would like to estimate a growth mixture model with clustered data. The Addendum to the Mplus User's Guide suggests this is possible with TYPE = COMPLEX MIXTURE. However, I'm unclear on the correct syntax for the MODEL commands, specifically with respect to integrating the %WITHIN% and %BETWEEN% statements with the %OVERALL% and specific class portions of the model. For now I am just trying to get a one-class model to get the syntax right.
My (simplified) code is like this:
USEVARIABLES ARE family y1-y5;
cluster = family;
classes = c(1);
analysis:
type = complex mixture;
model:
%overall%
%c#1%
%within%
iw by y1-y5@1;
sw by y1@0 y2@1 y3@2 y4@3 y5@4;
%between%
ib by y1-y5@1;
sb by y1@0 y2@1 y3@2 y4@3 y5@4;
[y1-y5@0 ib sb];

This leads to an error message about the %within% statement being illegal.
I realize this is a new method for Mplus and appreciate the many capabilites of the software -- thanks for any and all help!
 Linda K. Muthen posted on Friday, March 28, 2003 - 11:31 am
TYPE=COMPLEX MIXTURE is not the same as TYPE=TWOLEVEL MIXTURE (which is not yet available). TYPE=COMPLEX MIXTURE adjusts the standard errors and fit statistics for clustering. It does not allow modeling of between and within parts of the model. So if you include CLUSTER in your VARIABLE command and get rid of the %BETWEEN% and %WITHIN% parts of your MODEL command, you should be fine.
 Anonymous posted on Wednesday, November 19, 2003 - 8:16 pm
I am trying to look at several things and wondering if I could incorporate all in one model or I need several models to address each question. In our study students from several schools (14) were randomly assigned to two groups (control and intervention). Measures were obtained at baseline, post 1 week, post 3 months , and post 6 months. Three different outcomes (X,Y,Z) were assessed at each time point (1-4).

Q1: X1-->Y1-->Z1 (baseline)
The fit of this model and how well this model fits at subsequent time points (CFA with repeated measures)

Q2: How intervention and baseline factors affect the pre and post measures (X,Y,Z)(Growth modeling)

Q3: How we can examine the effect of different schools on the outcome measures and the above models (Multilevel analysis)

If it is possible to do these in one model, I would greatly appreciate your help in obtaining the correct syntax. If not, could you please let me have the syntax for Q1.
Thanks a million.
 Linda K. Muthen posted on Thursday, November 20, 2003 - 9:18 am
You could run one model that contains all components of your model. I would, however, not start there. I would always build a model up. With only 14 schools, multilevel modeling may not work well. We recommend 30-50 cluster units and never less than 20. If Q1 refers to a multiple indicator factor model with repeated measures over time, that is, if x1, y1, and z1 are latent variables, see Example 22.4 in the Mplus User's Guide and ignore the growth part of the model.
 Anonymous posted on Thursday, November 20, 2003 - 11:43 am
Thanks so much for your response. I would like a follow-up on the first part of the previous question. I did look at the example 22.4 and ran a model as follows (all variables below are latent):

Model 1 (at baseline)
z1 on x1 y1
y1 on x1

Model 2 (at time 2)
z2 on x2 y2
y2 on x2

Model 3 (at time 3)
z3 on x3 y3
y3on x3

This would give us parameter estimates and t-statics for each path. However, I would like to test the equality of the model parameters / fit at each time point. How do I test each model to determine wether the relationships among x, y, and z are statistically the same or different at each time point.
I have run all the basic models (ie. the above models and growth models on each outcome separately). Now, how do I incorporate these two into one model?
Thank you so much again.
 Linda K. Muthen posted on Thursday, November 20, 2003 - 5:12 pm
You just put them together in one MODEL command, for example,

Model:
z1 on x1 y1
y1 on x1

z2 on x2 y2
y2 on x2

z3 on x3 y3
y3on x3

and then set equalities and do a series of chi-square difference tests. Following is how you would set equalities for y on x.

Model:
z1 on x1 y1
y1 on x1 (1)

z2 on x2 y2
y2 on x2 (1)

z3 on x3 y3
y3on x3 (1)
 Nancy Beauregard posted on Tuesday, November 13, 2012 - 12:02 pm
Dear Dr Muthen,
I have a cross-sectional dataset where workers (N=2000) are nested in organizations (N=60). I'd like to perform a LCA on individual-level health outcomes (y1-y5), adjusted for individual-level covariates(x1-x10). I do not wish to explain organizational variability in latent classes, but simply to adjust the estimates for organizational membership (non independance of observations, see below org variable). To account for the hierarchical nature of the data, I'd run the following syntax:

VARIABLE: NAMES ARE org y1-y5 x1-x10;
USEVARIABLES ARE y1-y5 x1-x10;
CLUSTER = org;
CLASSES = c(2);
CATEGORICAL = y1-y5;
ANALYSIS: TYPE=MIXTURE COMPLEX;
STARTS = 500 10; STITERATIONS = 20;
MODEL: %OVERALL%
C#1 ON x1-x10;
OUTPUT: CINTERVAL TECH8 TECH11;

My question is thus the following. Is there a diagnostic test to assess whether the clustering effect is statistically significant ? The rationale here being that if the clustering effect is not significant, a simple one-level LCA would yield the same results.


Many thanks, very much appreciated !

Nancy Beauregard
Assistant professor
School of Industrial Relations, University of Montreal
 Linda K. Muthen posted on Tuesday, November 13, 2012 - 1:21 pm
You can run this with and without taking clustering into account and see if there is a difference in the standard errors.
 Tracy Waasdorp posted on Thursday, March 21, 2013 - 12:05 pm
I am running a multilevel GMM. I found that in a 1-level GMM a 3 class model was best. I would now like to examine how school level factors are associated with the classes (but not changing or informing the makeup of those classes). I am trying to use the starting values from the 1-level GMM in the multilevel GMM, however I keep running into errors. I have also tried using starts=0 and optseed from the 1-level GMM in the multilevel GMM but this also resulted in errors. I am pasting a portion of my input. Thank you in advance for your assistance.

Model:
%WITHIN%
%OVERALL%
i s | t111db@0 t211db@.5 t311db@1.5 t411db@2.5 t511db@3.5;

...HERE IS WHERE I POSTED THE ENTIRE SVALUE OUTPUT FROM THE 1-level GMM DUE TO POSTING LIMITS I HAD TO CUT THIS....

%BETWEEN%
%OVERALL%

ib sb | t111db@0 t211db@.5 t311db@1.5 t411db@2.5 t511db@3.5;
[ib@0];

T111DB-T511DB@0;

ib sb on T0M_smob T0M_Tenr T0FacTo t005strati;

c#1 c#2 on T0M_smob T0M_Tenr T0FacTo t005strati;
 Linda K. Muthen posted on Thursday, March 21, 2013 - 1:35 pm
Please send your output and license number to support@statmodel.com.
 Megan Perez posted on Monday, April 06, 2020 - 6:07 pm
I am running a growth mixture model that includes parent dyads (mother and father from the same family). I am trying to determine the best method to account for the non-independence in the outcome. I originally planned to conduct a two-level model; however, I have 70+ families and my research question is not interested in evaluating differences between families, but rather I just want to account for the non-independence. Would you recommend TYPE = COMPLEX or TYPE = TWOLEVEL for this type of analysis
 Bengt O. Muthen posted on Tuesday, April 07, 2020 - 3:16 pm
I would recommend the single-level, wide approach that many use - see papers listed on our website under Papers, Dyadic Analysis.
 Stefan Diestel posted on Saturday, April 11, 2020 - 3:04 am
Dear Mrs. and Mr. Muthén,

I hope my question finds you well.

Currently, I test a multilevel LGM with a within-level predictor (x1-x3). The parameters of the within-relations were specified to vary across the between-units: the within-relations can be predicted by between-level variables (similar to cross-level interactions).

I would like to clarify two questions:
1). Which kind of centering is appropriate for the predictor x1-x3 at the within-level? I suggest group-mean centering.
2). Which estimator would be appropriate, when analyzing cross-level moderating effects? I suggest ALGORITHM = INTEGRATION.

Thank you for considering my request!

Kind regards,
Stefan

DEFINE:

CENTER x1-x3 (Groupmean);
CENTER A B (Grandmean);

Inter = A*B;

ANALYSIS:
TYPE = TWOLEVEL RANDOM;
ALGORITHM = INTEGRATION;

MODEL:

%WITHIN%

iw_y sw_y | y1@0 y@1 y3@2 y4@3;

y1-y4 (1);

sw_y ON x1;

s | y1 ON x1;
t | y2 ON x2;
u | y3 ON x3;

%BETWEEN%

ib_y sb_y | y1@0 y@1 y3@2 y4@3;

y1-y4@0;

ib_y sb_y s t u ON A B Inter;
 Bengt O. Muthen posted on Saturday, April 11, 2020 - 5:45 pm
1) Group-mean centering is good here.

2) ML (or MLR) is fine. Bayes is also possible.
 Megan Perez posted on Monday, April 13, 2020 - 12:22 pm
Thank you for your suggestion single-level and on the dyadic analysis papers. I have reviewed the papers and found the Planalp, et al. (2016) very helpful.

I wanted to verify, that since my research question is not concerned with between or within cluster differences, but rather I want to adjust for the non-independence of data due to clustering of family participants, then TYPE = COMPLEX more appropriately adjusts the standard errors and fit statistics for clustering for this research question. Thank you!
 Bengt O. Muthen posted on Monday, April 13, 2020 - 5:27 pm
That's fine.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: