Dear all,
After a number of unsuccessful trials, I am hoping someone has encountered
this problem before, or can refer me to a good source.
I am trying to estimate the effect of a high school program on student
outcomes. The equations of interest are:
(1) y1 = y2 + program + demog + family + state
(2) y2 = y1 + program + demog + family + z1 + z2
(3) program = parents_occ + demog + family + state
Where y1 and y2 are student outcomes; demog, family and state are various
control variables, z1 and z2 are instrumental variables. In addition, the
variable "program" is binary, and y2 is categorical and sequential
(ordered).
Outcomes y1 and y2 are simultaneously determined. In addition, students
self-select into the program, and this selection can be instrumented with
parent's occupation - the program has a vocational nature.
My question is: would 3SLS confound the two types of endogeneity here
(simultaneity and self-selection)? Using reg3 produced results that were all
over the place.
Also, how should I deal with the binary nature of "program" in a 3SLS
estimation? Is there a better way to estimate the system?
I also have two other options:
1. I can plausibly argue that the program affects y1 via y2. I am not sure
how to set up and estimate the system under this assumption, but I am
thinking of splitting up the system and estimate (1) and (3) separately from
(2) and (1) and somehow incorporate fitted values from one system into
another (and adjust standard errors).
2. I can get another year of data, and have a two-period panel. Perhaps I
can sort out the selection problem by first differencing and then applying
2SLS to the first differenced outcomes.
Thanks for any ideas on how to tackle this problem.
Elda
Elda Pema
Assistant Professor of Economics
GSBPP, NPS
Monterey, CA 93943
[email protected]
[email protected]
<<attachment: winmail.dat>>
