Cross-lagged Panels
Message/Author
 Natasha Mendoza posted on Wednesday, November 14, 2012 - 8:44 am
Hi all,

I am attempting a cross-lagged panel path analysis.

I have 4 panels (i.e., data collection waves). Each panel has 2 measured variables: 1 continuous and 1 count. I'm trying to predict one with the other over time.

I understand that it is convention to correlate the errors within each panel. When I treat both variables as continuous, I can correlate them. However, when I specify the count variable, I am unable to correlate the errors (and it really should be treated as a count because zero inflated).

Is there another way to correlate the errors or is it impossible because of the variable type?

Much Thanks!
 Linda K. Muthen posted on Thursday, November 15, 2012 - 9:10 am
You need to use the BY option to specify covariances when numerical integration is required because each covariance requires one additional dimension of integration. Following is how you would specify this:

f BY u@1 y;
f@0; [f@0];

 Nassim Tabri posted on Wednesday, March 30, 2016 - 9:05 am
Hi Dr. Muthen,

What would the Mplus syntax look like if one wanted to include within lag correlations between three or more categorical DVs in a cross-lagged model?

Would the following syntax capture the within lag correlations between three DVs (u, y, and z)?

f BY u@1 y z;
f@0; [f@0];

f1 By y@1 z;
f1@0; [f1@0];

OR, would each within lag correlation require 1 latent variable like this:

f BY u@1 y;
f@0; [f@0];

f1 By u@1 z;
f1@0; [f1@0];

f2 By y@1 z;
f2@0; [f2@0];

 Bengt O. Muthen posted on Wednesday, March 30, 2016 - 6:38 pm
The second alternative is right. But don't fix the factor variance at zero - fix it at one. You also want to make these 3 factors uncorrelated (and uncorrelated with everything else if that needs adjustment - check TECH1).
 Nassim Tabri posted on Friday, April 01, 2016 - 6:48 am
Thank you Dr. Muthen!
 Simon Coulombe posted on Saturday, July 28, 2018 - 10:21 am
Hi,
I'm testing a simple cross-lagged model involving the inter-influence between one indicator variable and one latent construct (with three indicator variables) across two measurement times.

The model already includes the autoregressive link between the latent construct at the two measurement waves.
However, the modification indices suggest to add correlations (i.e., WITH links) between each of the three indicator variables of the latent construct at time 1 and at time 2 (i.e., indicator 1 at T1 WITH indicator 1 at T2; indicator 2 at T1 WITH indicator 2 at T2; indicator 3 at T1 WITH indicator 3 at T3)?

Should I avoid doing this since the latent construct at T1 and T2 is already linked (ON pathway: autoregressive link)?

I tried doing it and the fit of the model improved a lot.

Thank you

Simon
 Bengt O. Muthen posted on Saturday, July 28, 2018 - 5:18 pm
I think it is perfectly ok to add those 3 correlations. These measurement error correlations are different in nature from the auto-correlation of the constructs.