st: Using SEM for a multi-level SUREG

[Robert Kraut <robert.kraut@cmu.edu>]

I'm trying to fit a regression model that is a combination of a lagged 
dependent variable regression, seemingly unrelated regression and a 
multi-level repeated measures design.

Each respondent completed a questionnaire three times and I have 
behavioral data about communication surrounding each questionnaire 
administration. I am predicting several psychological outcome 
variables-- social support (csupport), loneliness (clone) and 
satisfaction with life cswl).

The independent variables are some demographics (age, male), the lagged 
dependent variables  (lag_csupport, lag_clone & lag_cswl) and the amount 
of communication the respondent received in the month prior to a survey 
administration (comin & prepackagedin)

If the DVs were independent, I would model each dependent variable as a 
repeated-measures lagged dependent variable regression with multiple 
observations per respondent using the xtreg command:

   xtset (respondid)
   xtreg csupport lag_csupport male age comin prepackagedin
   xtreg clone lag_clone  male age comin prepackagedin
   xtreg cswl lag_cswl male age comin prepackagedin

However, the dependent variables and the lagged variables are moderately 
correlated with each other. Therefore, SUREG seems appropriate.

   sureg (csupport lag_csupport male age comout comin prepackagedout
   prepackagedin)  (clone lag_clone male age comout  comin
   prepackagedout prepackagedin)  (cswl lag_cswl male age comout  comin
   prepackagedout prepackagedin)

However, the SUREG command can't accommodate multiple observations per 
respondent.

Since you can use the SEM and GSEM commands to reproduce SUREG modeling 
and multilevel models, I thought I'd try to combine them:

The SEM equivalent of the SUREG command seems to work fine:

   sem ( csupport <- lag_csupport male age  comin prepackagedin ) (
   clone <- lag_clone male age  comin prepackagedin ) ( cswl <-
   lag_cswl male age  comin prepackagedin ), cov(e.csupport*e.clone)
   cov(e.csupport*e.cswl) cov(e.clone*e.cswl)

However, the equivalent GSEM command including respondid as the nesting 
variable doesn't converge after a 1000 iterations.

   gsem ( csupport <- lag_csupport male age  comin prepackagedin
   M1[respondid]) ( clone <- lag_clone male age comin prepackagedin
   M1[respondid]) ( cswl <- lag_cswl male age  comin prepackagedin
   M1[respondid]), cov(e.csupport*e.clone) cov(e.csupport*e.cswl)
   cov(e.clone*e.cswl)

Do you have any suggestions about how to analyze this data?

bob kraut

--
Robert Kraut
Herbert A Simon Professor of Human-Computer Interaction
Carnegie Mellon University, 3515 NSH
5000 Forbes Ave, Pittsburgh, PA 15213
(o) 412 268-7694  (f) 412 268-1266
web:www.cs.cmu.edu/~kraut   email:robertkraut@cmu.edu



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