SEM using panel data PreviousNext
Mplus Discussion > Structural Equation Modeling >
Message/Author
 Prabakar posted on Tuesday, November 08, 2011 - 7:59 am
How to do an SEM model with panel data, i dont want to do ordinary regression for each panel/category
 Bengt O. Muthen posted on Wednesday, November 09, 2011 - 5:30 pm
When you say panel data, I interpret that as each subject being measured at more than one time point. The analysis depends on whether or not you want to model the development over time, or simply take into account the across-time correlations for each subject. In the former case you do longitudinal modeling such as growth modeling.
 Jean-Samuel Cloutier posted on Monday, February 20, 2012 - 6:30 am
Dear Professor,

I have a model of 5 measured variables (x1, x2, x3 , x4 and x5) and 3 latent variables (f1 f2 and f3).

Obaserved variable are being measured at 7 time point for every of the 750 subjects.

x1-x5 and f1 and f2 are time specific (ie indicator of performance for each measuring time)

F3 is a variable that capture the persistence of the agregate performance of each subject.

I am not interested in the effect of time. The purpose is to validate a multidimentional performance construct that evolve over time.

!here i use 3 time points

F11 BY x11@1 x21* x31*
F21 BY F11@1 x11* x21*

F12 BY x12@1 x22* x32*
F22 BY F12@1 x12* x22*
F32 BY F21@1 F22*

F13 BY x13@1 x23* x33*
F23 BY F13@1 x13* x23*
F32 BY F21@1 F22* F23*
 Jean-Samuel Cloutier posted on Monday, February 20, 2012 - 6:33 am
The question is:

How should I model this?

Best,
Jean-Samuel
 Linda K. Muthen posted on Monday, February 20, 2012 - 1:57 pm
Your approach seems reasonable.
 Kathleen Crum posted on Thursday, October 29, 2015 - 4:01 pm
Hello,
I am trying to run a panel analysis using two waves of longitudinal data. MPlus is installed on my Macbook computer. However, MPlus does not run the analysis (or at least it does not show the usual "working" window I expect when running a complex analysis). Instead, it opens a completely blank output window. Is this an issue with the syntax I am using, or a software problem?
Thank you so much for your help!
Kathleen
 Linda K. Muthen posted on Friday, October 30, 2015 - 10:37 am
Please send the input and data to support@statmodel.com.
 Grace Song posted on Monday, December 14, 2015 - 2:15 am
Hi, Bengt O. Muthen, just add on your comments. You mentioned:

The analysis depends on whether or not you want to model the development over time, or simply take into account the across-time correlations for each subject. In the former case you do longitudinal modeling such as growth modeling.

What if in the later case? I just want to control the unobserved variables for each subject.

I would like to conduct SEM with panel data, not looking at the time effect, but to control the unobserved variables for each subject.

Pls kindly advice.

Thank you!
 Bengt O. Muthen posted on Monday, December 14, 2015 - 6:13 pm
You can do that using Type=Complex wiht cluster = id to just adjust the SEs or you can do it using Type=Twolevel where you have random intercepts for the DVs that have variances and covariances estimated on Between and that captures the correlation due to repeated measures within the same individual.
 Anshuman Sharma posted on Friday, September 15, 2017 - 11:37 pm
Hi,

Could you please suggest how I can perform panel data regression( I need to take into account the across-time correlations for each subject). I saw chapter eight which discusses this issue, however, I am still not confident.
I have something like following data
ID timeInstant Com Speed Spacing
1 /1 /1 20 / 10
1 /2 /1 22 12
2 /1 /0 26 8
2 /2 /1 28 15
3 /1 /0 23 13
3 /2 /0 22 17
. /. /. . .
. /.
. /. /. . .
Model: Com ON Speed Spacing;
I gave it a try but I am not sure I have done it correctly or not:

VARIABLE: NAMES ARE ID Time Com Speed Spacing;
USEVARIABLES ARE Com Speed Spacing; NOMINAL IS Com;
Cluster=ID;
within=Speed Spacing;
analysis: type = twolevel random;
estimator=ml;
MODEL: %within%
Com ON Speed TimeH;


Thanks,
Anshuman
 Anshuman Sharma posted on Friday, September 15, 2017 - 11:39 pm
% please ignore the above post
Hi,

Could you please suggest how I can perform panel data regression( I need to take into account the across-time correlations for each subject). I saw chapter eight which discusses this issue, however, I am still not confident.
I have something like following data
ID timeInstant Com Speed Spacing
1 1 1 20 10
1 2 1 22 12
2 1 0 26 8
2 2 1 28 15
3 1 0 23 13
3 2 0 22 17
. . . . .
. .
. . . . .
Model: Com ON Speed Spacing;
I gave it a try but I am not sure I have done it correctly or not:

VARIABLE: NAMES ARE ID Time Com Speed Spacing;
USEVARIABLES ARE Com Speed Spacing; NOMINAL IS Com;
Cluster=ID;
within=Speed Spacing;
analysis: type = twolevel random;
estimator=ml;
MODEL: %within%
Com ON Speed TimeH;
 Kyle Levesque posted on Tuesday, October 24, 2017 - 9:35 pm
Hello,
A bit of advice for a longitudinal panel model.
I completed the measurement invariance testing... and partial strict invariance was reached. Good fitting models thoughout the invariance testing.
When I finally get down to business for adding the auto-regressive paths and cross-lag paths, I encounter standardized (STDYX) coefficients for some of the autoregressive paths that are greater than 1.00. Have I missed something in specifying constraints?

Based on Little (2013) book, I used the fixed factor method of scaling, so the Time 1 constructs are assigned a variance of 1 and intercept of 0. With all the constraints in place, should I be looking at the unstandardized coefficients for the paths between latent variables. That is, are the unstandardized coefficients "standardized" in panel models because the T1 constructs are scaled to 1?

Thanks,
KL
 Bengt O. Muthen posted on Wednesday, October 25, 2017 - 10:11 am
Standardized paths greater than 1 can happen with highly correlated variables.

For the time 2 factor I assume you are fixing the residual variance at 1, not the full variance. The unstandardized paths are only standardized if all variables involved have (full) variances of 1.
 Kyle Levesque posted on Wednesday, October 25, 2017 - 10:34 am
Thank you! This is all very helpful for my self-teaching of Mplus.

Lastly, what is the distinction between fixing Time 2 residual value at 1, versus the full variance? (all I know is the @1 function)

My syntax follows that of an example in Little (2013), which uses the fixed factor scaling method to test for measurement invariance. Namely:
Pos1@1 Pos2* Pos3*;
[Pos1@0 Pos2* Pos3*];
Neg1@1 Neg2* Neg3*;
[Neg1@0 Neg2* Neg3*];

Are you saying that once the autoregressive paths are specified, the factors at time 2 and time 3 receive the "@1" as well, to fix their residual variance?
 Bengt O. Muthen posted on Wednesday, October 25, 2017 - 10:45 am
The statements you show are fine as they are (I assume that Pos and Neg are factors): The factor variances at time 2 and time 3 are not fixed at 1 but are freely estimated as they should be.
 Christoph Schaefer posted on Saturday, October 26, 2019 - 6:14 pm
Dear Professors,

Q1: I have panel data, each subject being measured at two time points. Could I still conduct a regression and take into account the across-time correlations for each subject, even though there are only two points?

Q2: Would the following input be correct for a fixed effects model (with a time-invariant individual effect that is allowed to be correlated with the predictors)?

VARIABLE:
CLUSTER = time;

ANALYSIS:
TYPE = complex;

Model:
y on x1 x2;
 Christoph Schaefer posted on Saturday, October 26, 2019 - 6:43 pm
P.S.: I am only interested here in the regression of y on x1 x2, not in the growth on the within-level. (Each of the variables y, x1, and x2 was measured at two time points for each individual. Number of indivuals is about 200.) The core of Q2 is whether the input specification is sufficient to allow the time-invariant individual effect to be correlated with the predictors.

And: Above, I should have typed:

Cluster = id;
 Bengt O. Muthen posted on Sunday, October 27, 2019 - 4:20 pm
You can use Type=Twolevel with Cluster = id and use Define to group-mean center x1 and x2. For the relationship with allowing a time-invariant individual effect correlated with the predictors, see the paper that we have posted on our website under Recent Papers:

Hamaker, E. & Muthén, B. (2019). The fixed versus random effects debate and how it relates to centering in multilevel modeling. Accepted for publication in Psychological Methods.
 Christoph Schaefer posted on Monday, October 28, 2019 - 9:43 am
Many thanks.
Can I do that even though I have only two time points per individual/cluster?
Would it then look like this?

VARIABLE:
[...]
WITHIN = x1 x2;
BETWEEN = x1m x2m;
CLUSTER = id;
DEFINE: CENTER x1 x2 (GROUPMEAN);
ANALYSIS: TYPE = TWOLEVEL;

MODEL: %WITHIN%
y on x1 x2;
%BETWEEN%
y on x1m x2m;
 Bengt O. Muthen posted on Monday, October 28, 2019 - 9:55 am
I think this is identified for 2 time points because you don't have Y_t ON Y_t-1. The within variance of Y and covariance between Y and X1 X2 identify the Y residual variance and the 2 slopes. Same on between. Or, looked at another way, using wide format, it is a random intercept only growth model.
 Christoph Schaefer posted on Monday, October 28, 2019 - 10:04 am
Thank you so much.
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: