Stata18 sem group-option

[Eva Matthaei]

Dear Statalist,

I have a question relating to Stata's sem framework and more precisely the integrated group-option. In the latest update from Stata 17 to Stata 18 the default settings when estimating an sem with the group-option seem to have changed. Stata now includes constants in paths with latent dependent variables. This means that the constraint "LatentEndogenous <- cons@0" seems to be only applied for the first group when performing group estimations. I cannot find any documentation on this change in the default settings, and it is not even mentioned in the reference manual at all (p. 597). Unfortunately the settings don't seem to be applied consistently as the output includes constants in some of the paths with latent dependent variables but not in others. Therefore, I would greatly appreciate any information on the change in default settings.

Thank you for your help.

Best wishes,
Eva

[Jeff Pitblado (StataCorp)]

We made no coding changes in sem for Stata 18.

Can you share an example that shows sem behaving differently between Stata 17 and Stata 18?

[Eva Matthaei]

Hi,
thank you so much Jeff for the very quick answer! I actually also sent this question to the support team and am truly amazed about the quick help. Here is the answer I got:

"We documented this changes in our whatsnew item. Please type

help whatsnew17

and locate the following item on update 13dec2021, it reads,

---

3. sem, when specified with latent variables in option means(), has improved starting values
for the latent means. The original starting values behavior uses a linear combination of
the observed means and the path starting values. The improved behavior uses the mean of
the latent variable's anchor, provided the anchor's intercept (_cons) is constrained to 0
and there are no other fixed-value constraints on paths from the latent variable to
observed variables.

---

When -group()- option is present, the means of all latent X variables are estimated by default, except the first group’s latent means are constrained to 0. With this change, the starting values code was distributing the latent mean starting values across the intercepts of latent Y variables having paths from the latent X variables."


As X and Y are not necessarily exclusive groups of variables in SEM , if I am correct the explanation on the group-option could be clarified to read:
When -group()- option is present, the means of all latent X, i.e., latent exogeneous, variables are estimated by default, except the first group’s latent means are constrained to 0. With this change, the starting values code was distributing the latent mean starting values across the intercepts of latent Y variables having paths from the latent X, i.e., latent exogeneous, variables.

Am I right?

With the explanation the different behavior now makes complete sense. I still wonder how I only got an update that was done in 2021 (?) when switching to Stata18. I update Stata regularly and installed Stata17 on my new computer in 2023. Anyway that's probably on me

Perhaps this could at some time be clarified in the documentation. The different -means- behavior in group models is clearly stated and has also been applied in earlier versions of Stata. This is why the new -_cons- behavior in group models came at a surprise to me.

[Jeff Pitblado (StataCorp)]

Hi Eva, your clarification looks correct to me.

One thing tech support left out of their reply was that the change in the starting values logic on the group means was not intended to affect the inclusion/exclusion of latent endogenous variable intercepts. While you can always force sem to include a latent endogenous variable's intercept by specifying _cons in the model specification, including it for you for a grouped model depending on the presence of latent exogenous variables in the linear form was not intended.

We are looking into restoring the original latent variable intercept behavior in a future update in Stata 18.