We have an unbalanced multilevel longitudinal dataset with students
nested in schools over time. We have variables of interest for both
students (such as their position in the test score distribution) and
school (such as their accountability scores). We have large N and
small T (T=6 at most).
We want to estimate the differential effects of a lagged school
characteristic on kid’s test score through interactions of lagged
distance to grade level and the school characteristic of interest.
DV – score_t
Exogenous – S_t-1
Endogenous
test score_t-1
belowgl_t-1
abovegl_t-1
S_t-1*belowgl_t-1
S_t-1*abovegl_t-1
Where
score_t is the kid’s test score at time t
s_t-1 is whether the school that kid was enrolled in a time t had a
particular characteristic in the prior year
belowgl_t-1 is whether the kid was well below grade level in the prior year
above gl_t-1 is whether the kid was well above grade level in the prior year
** Note that the interactions S_belowgl and S_abovegl involve both kid
level and school level variables.
We would like to run use the Arello-Bond estimator to run something like this:
xtabond score l(1/1).sanc, lags(1) endogenous(belowgl abovegl
S_belowgl S_abovegl, lag(1,.))
The problem is that we need to be able to utilize lags for the
school-level variables that are based on the school's previous year
status, but our dataset is student level yearly data. Therefore, if a
student changes schools or appears in a school for the first time, the
lag for a school variable will be based on the student's prior school,
not the lag of the school’s prior measurement. This can be worked
around by hard-coding the lagged variables in a school level dataset
and then merging these variables back into the student level dataset.
However, we would like to use the xtabond regression command but, as
far as we can tell, the lagged variables cannot be hard-coded, so the
lagged variables will not be correct to the school level. Thus, we
need a way to specify the xtabond regression command to fit our
multilevel data for the lags, or another work-around to code a new
variable and "trick" Stata. Or will xtabond2, xtdpdsys, or xtdp work
better for us?
Thanks in advance for your thoughts!
Doug Lauen
UNC-Chapel Hill
USA
[email protected]
*
* 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/
