Model specification: Connecting indicators with other variables in the model

34 views
Skip to first unread message

CSM

unread,
Mar 28, 2018, 6:09:34 AM3/28/18
to lavaan
Dear all,

I have a small question about model specification. I specified a SEM model based on theoretical connections among the variables (two latent variables, with three indicators each, plus three observed variables). After this specification, fitting measures were acceptable. After analysing modification indices, I observed that by adding two correlations in the model, the fitting was much improved, indicating really good fitting to the data.

My question is about these two correlations. They can be written in the form: D ~~ X1; D ~~ Y1, where D is a dummy variable (ordered, not indicator), and X1 and Y1 are indicators of two different latent variables, say X and Y. Regressions between D and X, and D and Y are nonsignificant. Theoretically, the following relations are meaningful: D ~ X1 and Y1 ~ D, and so, I would prefer to have these two regressions instead of the corresponding correlations. However, after replacing the two correlations by these two regressions, I got an error due to D ~ X1 (but not from Y1 ~ D).

First question: Since I have never seen SEM models where indicators are connected with "exterior" variables of the model, I would like to know if the model including "D ~~ X1; Y1 ~ D" (correlation + regression) is well specified/ statistically meaningful. If not, what about including "D ~~ X1; D ~~ Y1" (correlation + correlation)? Notice that in lavaan both cases are allowed and converge.

Second question: Although it is a general SEM question, I would be grateful if you can tell me why is it senseless to specify the regression D ~  X1.

Best regards,
 csm

Terrence Jorgensen

unread,
Mar 28, 2018, 7:53:07 AM3/28/18
to lavaan
First question: Since I have never seen SEM models where indicators are connected with "exterior" variables of the model, I would like to know if the model including "D ~~ X1; Y1 ~ D" (correlation + regression) is well specified/ statistically meaningful. If not, what about including "D ~~ X1; D ~~ Y1" (correlation + correlation)? Notice that in lavaan both cases are allowed and converge.

Second question: Although it is a general SEM question, I would be grateful if you can tell me why is it senseless to specify the regression D ~  X1.

You are almost specifying a well-known class of models called MIMIC models, which can be used to test for measurement equivalence/invariance (i.e., measurement parameters do not depend on D) and differential item functioning (DIF; i.e., whether a particular indicator's measurement parameter(s) do(es) differ across levels of D).  Here is an example application:


Essentially, yes, you can regress X1 and Y1 on D to test the relationships you theoretically expect.  But you also regress indicators (e.g., one at a time, or maximum all but 1 indicator so the model will be identified) on D to test for DIF.  Note, you regress indicators on D, you do not simply correlate them with D.  That way, factor loadings are estimated as the effect of the factor on the indicator, controlling for the effect of D.  If the effect of D on an item is substantial/significant (i.e., that item has "uniform DIF"), you need to make sure that estimate is in your model in order for the effect of D on the latent factor to be unbiased.  If none of them are, then you can leave the direct effects of D on indicators fixed to zero, and conclude scalar/strong invariance.

I have an upcoming book chapter I can share that shows how to test this in lavaan, as well as whether D and the latent factor interact to effect the indicator (i.e., to test metric/weak invariance).  I can email a copy if you write me a private email.

Terrence D. Jorgensen
Postdoctoral Researcher, Methods and Statistics
Research Institute for Child Development and Education, the University of Amsterdam

CSM

unread,
Mar 28, 2018, 9:49:13 AM3/28/18
to lavaan
Dear Prof. Terrence Jorgensen,

Great! You went directly to the point; that's exactly what I was looking for.
I will be grateful if you could send me a copy of your book chapter. I'm sending a private message from my email box.
Thank you very much for your valuable help!

Best,
celia
Reply all
Reply to author
Forward
0 new messages