st: Allowing for interaction year effects, using the XTABOND2 command

["Lovisa Persson" <lovisa.persson@nek.uu.se>]

Hello everyone,

I am estimating a panel data model on consumption and income, I want to know
the effect of a change in income on changes in consumption.
I use difference GMM, with lagged levels of income as instruments, and in
practice I use xtabond2 command to estimate the model.

This is how I write my main command...

xtabond2 consumption income dum1-dum19, gmm(income, laglimits(3 5))
iv(dum1-dum19) robust noleveleq artests(3)

dum1-dum19 is a full set of year dummies. I have data on years 1993 through
2011.

What I want to do now is to allow for the coefficient on "income" to be
different for different time periods.  More specifically, in my case I want
to see if the coefficient on income is different before and after a reform I
am studying. The reform was implemented in 2000. What I then did was to
create a new dummy variable.

generate afterdum=0
replace afterdum=1 if year==2000 
replace afterdum=1 if year ==2001

and so on...up to 2011

Then I interacted this dummy variable with "income"

generate afterinter=afterdum*income

And then I estimated the following model.

xtabond2 consumption income afterinter afterdum dum1-dum19, gmm(income
afterinter, laglimits(3 5)) iv(afterdum dum1-dum19) robust noleveleq
artests(3)

It doesn't matter if I include the "afterdum" variable or not since the year
dummies are already included. It does not matter for the results if
afterinter is included in the gmm() specification either.

The problem is that the coefficient on "afterinter" is not the same as the
difference between two estimations where I have split the sample at the time
for the reform. I split the sample and estimate the coefficient after the
reform in 2000 by dropping all years prior to 1997, since the first year I
estimate is 2000 and with the first instrument lag being t-3 I need to keep
years from 1997. Then I split the sample by dropping all variables at 2000
and after, and then estimate the coefficient before the reform. The
difference between these two estimations should be the same as when I create
and interaction variable and test the difference between the two time
periods. Why are the differences not the same? How am I doing this wrong?

Thank you for your kind help,
Lovisa


*
*   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/