Viktor--
Yes, you would have problems if you followed the outlined
strategy--for one, estimated SEs would be wrong, and I don't think you
have any proof of consistency for the coefficient. But you can
estimate
y2 = y1 + x1 + Z1 + Z2 + e2
using your notation (which shows linear models, not nonlinear ones),
and the coefficient on y1 will measure the impact on y2 of the
component of y1 orthogonal to x1 and Z1 as you desire. In Stata, you
would write the equation as
nbreg y2 y1 x1 Z1 Z2
of course. However, given that y1 takes on 4 discrete values, you can
tab y1, gen(dy)
nbreg y2 dy2 dy3 dy4 x1 Z1 Z2
and you could include interactions between the {dy1 dy2 dy3 dy4} and
{x1, Z1}variables for more flexibility in your specification.
On 6/15/07, Viktor Slavtchev <[email protected]> wrote:
Dear,
I have followed question.
I know that y1 depends on x1 in the way:
y1 = x1 + Z1 + e1.
However, I also am interested in the impact of both y1 as well as x1 on
y2. y2 is given by:
y2 = y1 + x1 + Z2 + e2.
So, I have in mind a reduced form model for y2:
y2 = e1 + x1 + Z2 + e2.
But:
y1={0;1;2;3} Likert-scale variable, hence I am going for ologit in order
to estimate y1 = x1 + Z1 + e1.
y2 is an overdispersed count variable; therefore I have to use negbin.
Are there any problems if I estimate the reduced form y2 = e1 + x1 + Z2
+ e2, knowing that in the first step ologit has been used and in the
second step I have to use negbin?
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
