Re: st: Latent Growth Curve SEM GUI

[vwiggins@stata.com (Vince Wiggins, StataCorp)]

Carter Rees <carterrees@gmail.com> writes

> I have the basic latent growth curve model with 4 time points set up
> using the SEM GUI. I would like to add one time invariant covariate
> (gender) to predict the growth factors (intercept and slope) using
> the GUI. However, each time I try to add this variable either as
> latent or observed I get the error message "Cannot add path. Target
> variable will become endogenous and is correlated with exogenous
> variables". How do I add this via the GUI?
>
> My goal is to replicate the results of an example from MPlus so I
> can get a better feel for what the Stata code will look like and how
> the GUI works. No troubles fitting the LGC without the covariate but
> the covariate is stumping me.

I fear my response will not make much sense to anyone who does not look
at the SEM diagram Carter refers to in a later posting in the thread,

> Slide 86 is what I am trying to do (not the same data but the model
> spec is the same). But, I only want to use the observed variable
> female. Syntax is on slide 87.
>
> http://www.statmodel.com/download/Topic3-v.pdf

I suspect what Carter has done is create the top of the diagram first.
That is he has created a measurement model with 4 measurements
(math7-math10) and with two latent factors (i and s) <step 1>.  I
believe Carter then correlated the two latent variables (i and s) by
adding a covariance path between the variables <step 2>.  Then Carter
added an observed variable (female) that he considers a determinant of
i and s <step 3>.  He then drew a path from female to i (or s) <step 4>.  
At this last step a warning box notified him that,

      "Cannot add path. Target variable will become endogenous
       and is correlated with exogenous variables."

You cannot model the correlation between an endogenous variable and any
other variable.  Such correlations are a side-effect of the model, not
something that can be directly modeled.  By adding a path pointing to
i, Carter made that variable endogenous, but in step 2 he had
previously correlated i and s.  The SEM Builder cannot know what Carter
intends.  Does he want the correlation deleted, does he want it moved
to the error for i?

The important thing to know is that the model Carter is specifying from
the diagram does not correlate i and s.  It correlates the errors from
i and s.  Carter never desired to correlate the variables themselves.

If I am guessing right, what Carter needs to do is delay step 2 until
after he has drawn the paths from female to i and s.  Then, he can add
a covariance path between the errors for i and s.



-- Vince 
   vwiggins@stata.com

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/