Cross-lagged model PreviousNext
Mplus Discussion > Structural Equation Modeling >
Message/Author
 Lucia Salinas posted on Wednesday, April 22, 2015 - 5:04 am
Hello Drs Muthen,

I specified a cross-lagged model using MLR with two waves and 5 variables (1 = manifest, continuous + 4 = latent factors assuming strong measurement invariance). This model and moderations by age and gender (via multi-group) worked fine.

I was asked to rerun this model including only manifest, binary variables to assess the clinical meaningfulness. I'm wondering if it's adequate to specify such a model? If yes, is this specification done properly using WLSMV:

X1 ON SES;
X2-4 ON age gender SES;

Y1 ON SES;
Y2-4 ON age gender SES;

X1 WITH X2 X3 X4 X5;
X2 WITH X3 X4 X5;
X3 WITH X4 X5;
X4 WITH X5;

Y1 WITH Y2@0 X3@0 X4@0 X5@0;
Y2 WITH Y3@0 X4@0 X5@0;
Y3 WITH Y4@0 X5@0;
Y4 WITH Y5@0;

Y2 ON X2 X1;
Y1 ON X1 X2;
Y3 ON X3 X1;
Y1 ON X1 X3;
Y4 ON X4 X1;
Y1 ON X1 X4;
Y5 ON X5 X1;
Y1 ON X1 X5;

In addition, would you recommend allowing correlations between manifest variables at t2.

Thanks for your help in advance and kind regards!
 Bengt O. Muthen posted on Wednesday, April 22, 2015 - 2:05 pm
Looks ok although I don't understand the argument for dichotomizing to study clinical significance. Regarding correlating variables as t2, I would discuss on SEMNET.
 Lucia Salinas posted on Thursday, April 23, 2015 - 6:29 am
Thank you for your quick response. The variables assessed can be categorized into "normal" vs. "abnormal" scores on the basis of recommended cutoff scores for clinical diagnoses.
 Lisa van Zutphen posted on Wednesday, October 23, 2019 - 2:41 am
Dear Drs. Muthen,

I specified a cross-lagged panel model between three variables (X, Y, Z) and I have 5 time points. The participants come from two datasets and I would like to adjust the model for 2 confounding variables, age and the dataset. I am confused where I add these two confounders in my model.
Do I add them to all the paths (Y2 on X1 Z1 age dataset; Y3 on X2 Z2 age dataset; etc)?
Or only the association between these confounders and X Y Z at the first measurement (X1 on age dataset; Y1 age dataset; Z1 age dataset; Y2 on X1 Z1; Y3 on X2 Z2; etc.)? Or something completely else?

Kind regards,

Lisa
 Bengt O. Muthen posted on Wednesday, October 23, 2019 - 1:46 pm
Use the RI-CLPM and let the age and data set dummy influence the random intercepts. See

http://www.statmodel.com/RI-CLPM.shtml
 Lisa van Zutphen posted on Wednesday, October 23, 2019 - 11:40 pm
Thank you for your suggestion. Two of the three main variables in my model are dichotomous. Is the RI-CLPM also possible with dichotomous variables? And if so, do I have to change something to the syntax (other than stating these variables as categorical).

Kind regards, Lisa
 Bengt O. Muthen posted on Thursday, October 24, 2019 - 11:24 am
Which of your variables are dichotomous? RI-CLPM with categorical DVs needs special considerations and is perhaps more of a methods research topic at this point.
 Lisa van Zutphen posted on Thursday, October 24, 2019 - 11:58 am
Out of the three variables that are measured 5 times, XYZ, are X and Y dichotomous. Z is continuous.
For the two confounders (age and cohort): age is continuous and cohort is dichotomous.
 Bengt O. Muthen posted on Thursday, October 24, 2019 - 4:04 pm
Sounds like 2 of your 3 DVs that are repeatedly measured are dichotomous. You are then in a methods research area where the RI aspect of the RI-CLPM is the issue.
 Lisa van Zutphen posted on Friday, October 25, 2019 - 12:12 am
Oh, that's a pity. I dichotomized those two repeatedly measured DV's because there were no linear relations (thus the linearity assumption was violated).

If I would specify a CLPM without the RI, thus an old-fashioned CLPM, how would you than suggest that I correct my model for these two confounder?

Kind regards, Lisa
 Bengt O. Muthen posted on Sunday, October 27, 2019 - 4:11 pm
I would as you say add them to all the paths (Y2 on X1 Z1 age dataset; Y3 on X2 Z2 age dataset; etc)
 Lisa van Zutphen posted on Tuesday, October 29, 2019 - 8:46 am
Thank you, I have added the confounding variables to all the paths. This is my syntax:

VARIABLE:
NAMES = id cohort Y1 Y2 Y3 Y4 Y5 X1 X2 X3 X4 X5 Z1 Z2 Z3 Z4 Z5 age sex ;
MISSING IS ALL (-99); IDVARIABLE = id;
USEVARIABLES ARE cohort age X1 X2 X3 X4 X5 Z1 Z2 Z3 Z4 Z5 Y1 Y2 Y3 Y4 Y5;
CATEGORICAL = Y2 Y3 Y4 Y5 X2 X3 X4 X5;
ANALYSIS:
type=general; parameterization=THETA;
MODEL:
! Estimate the covariance between the observed variables at the first wave
Z1 with X1 Y1; Y1 with X1;
! Estimate the covariances between the residuals of the observed variables
Z2 with Y2 X2; Y2 with X2; Z3 with Y3 X3; Y3 with X3; Z4 with Y4 X4; Y4 with X4; Z5 with Y5 X5; Y5 with X5;
! Estimate the lagged effects between the observed variables
Z2 ON Z1 Y1 X1 cohort age; Z3 ON Z2 Y2 X2 cohort age; Z4 ON Z3 Y3 X3 cohort age; Z5 ON Z4 Y4 X4 cohort age;
X2 on Z1 Y1 X1 cohort age; X3 ON Z2 Y2 X2 cohort age; X4 ON X3 Z3 Y3 cohort age; X5 ON X4 Z4 Y4 cohort age;
Y2 ON Y1 Z1 X1 cohort age; Y3 ON Y2 Z2 X2 cohort age; Y4 ON Y3 Z3 X3 cohort age; Y5 ON Y4 Z4 X4 cohort age;

I checked with MOD INDICES if there would be interesting suggestions to improve the model. These included ON statements between the 3 DV's at the first measurement and the confounders, such as:

X1 ON AGE; Z1 ON COHORT; Z1 ON AGE; etc.

I find it difficult to determine if this makes sense in a CLPM. What would you suggest?
 Bengt O. Muthen posted on Tuesday, October 29, 2019 - 4:04 pm
I would regress Y1, X1, Z1 on all the confounding variables because the latter are background variables.
 Luisa Liekefett posted on Wednesday, August 26, 2020 - 7:21 am
Dear Profs Muthén,

I am estimating a cross-lagged model with 4 time points. The model fit is poor. The modindices command suggest that I add additional on statements between x4, x3, x2 and x1, and y4, y3, y2 and y1, respectively:

x4 on x2 x1;
x3 on x1;
y4 on y2 y1;
y3 on y1;

does it make sense to add these paths? Can I still interpret the cross-lagged effects similarly with these paths included in the model?

Thank you very much in advance and kind regards.
 Bengt O. Muthen posted on Wednesday, August 26, 2020 - 3:40 pm
I would recommend that you first try an RI-CLPM. See
http://www.statmodel.com/RI-CLPM.shtml
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: