SEM for longitudinal data PreviousNext
Mplus Discussion > Structural Equation Modeling >
 Carl-Etienne Juneau posted on Friday, March 11, 2011 - 2:05 pm
Dear Dr. Muthén,
Dear Dr. Muthén,

I've just found out about this forum. I've read 2-3 threads, and I must say I'm impressed by the level of service you provide to the community. My hat off to you.

I have a question. But before, a few observations:

1. Longitudinal data sets are hierarchical (observations nested within subjects).

2. Therefore, in longitudinal data sets, observations are not independent.

3. SEM treats all observations as independent.

4. Therefore, SEM is not appropriate for longitudinal data.

My questions:

a. Is this reasoning correct?

b. If yes, are multilevel SEM the solution?

c. If no, do I need to take special measures when specifying my models to account for longitudinal data?

Thanks so much!

Carl-Etienne Juneau
PhD candidate in public health
Université de Montréal
 Linda K. Muthen posted on Monday, March 14, 2011 - 4:15 pm
a. No.
c. Taking a multivariate approach to growth modeling, with data in the wide format where each time point is represented by one variable, takes into account the non-independence of observations due to repeated measures.
 Virginia Warner posted on Friday, March 18, 2011 - 2:42 pm
What about correlations within families?
 Linda K. Muthen posted on Friday, March 18, 2011 - 2:58 pm
If you have sampled family members from a random set of families, generally one would use multilevel modeling with family as the cluster variable. If there are not too many family members, one could take a multivariate approach as described in:

Khoo, S.T. & Muthn, B. (2000). Longitudinal data on families: Growth modeling alternatives. Multivariate Applications in Substance use Research, J. Rose, L. Chassin, C. Presson & J. Sherman (eds.), Hillsdale, N.J.: Erlbaum, pp. 43-78.
 Andrea Norcini Pala posted on Friday, April 19, 2013 - 12:43 am

I have three time point measurment of biomarkers about Illness progression.
does it make sense to you using these three measurements as indicators of a latent dimension? Namely, let's say I have three assessements of viral load (baseline, time 1 and time 2), may I consider them indicators of a latent dimension "viral load change over the time"?

if so, can you suggest papers where this approach has been adopted?

Thank you very much,
 Bengt O. Muthen posted on Friday, April 19, 2013 - 5:05 pm
If you have a notion of illness progression, perhaps a growth model would be suitable. That explores the individual variation in both the level and the change over time. See our Topic 3 handout and video on our website.

A growth model is a specific kind of a latent variable model.
 Andrea Norcini Pala posted on Saturday, April 20, 2013 - 12:24 am
Thank you very much!
 Margarita  posted on Thursday, October 13, 2016 - 9:09 am
Dear Dr. Muthen,

I was reading the manual and I have 2 questions, if you have the time:

1. Assuming one has established measurement invariance across time, when proceeding to longitudinal SEM (panel/cross-lagged models) with latent variables, should BOTH the factor loadings and thresholds constrained to be equal across time points? In an example in the manual about latent growth modelling, both factor loadings and thresholds are held equal across time, but should that also be applied to cross-lagged models?

2. In a cross-lagged model with 3 time points and 2 latent variables should theta parameterisation be used?

I appreciate your help.
 Bengt O. Muthen posted on Thursday, October 13, 2016 - 1:28 pm
1. Cross-lagged models typically don't imply a mean structure and in that case the threshold invariance is not absolutely necessary - but I would still apply it.

2. With WLSMV that would be good.
 Margarita  posted on Thursday, October 13, 2016 - 1:41 pm
Thank you for your time and guidance!
 Alaine Garmendia posted on Monday, July 24, 2017 - 3:31 am

We have a large longitudinal and multilevel database with more than 500 companies. In this database we have people related indicators collected by surveys(invididual level) and also company related indicators (group level). It is longitudinal because we have repeated measures of most of the companies. Some of the companies have done the survey 2 times and others 3 times. We want to analyze the relationship among the people related indicators and company related indicators from a longitudinal perspective. The "problem" is that not all the companies are measured at the same times, supposing this that maybe at time 1 we have 50 company measures and at time 2 100 companies (but maybe this 100 do not include the previous 50 or yes) and at time 3 we can have 70 companies which some of them are the same as the measured at time 1 and others at time 2. Do you understand what I mean? How can we model this in order to conduct a longitudinal analysis?

Thank you very much in advance for your time and guidance!

 Bengt O. Muthen posted on Monday, July 24, 2017 - 4:22 pm
You can do this via missing data for those time points that a company wasn't measured.

Or, you can take the multiple-cohort approach of UG ex6.18.

There is a UG example on twolevel longitudinal (3-level analysis done as twolevel wide format).
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