Dear Statalisters:
I am estimating negative binomial and Poisson models on a panel data set
using the -re- and -fe- options to model random and fixed effects. My
dataset consists of country-years, and the possible outcomes on the
dependent variable range from 0 to 7, with most values at 0 or 1. The
dependent variable is the number of new treaties signed in a particular
year.
My issue is that using the fixed effects estimator causes about 1/4 of
my sample countries to drop from the sample because over the sample
period they never sign any treaties. This causes a problem of perfect
collinearity between the fixed effects estimator and the dependent
variable, the latter of which never varies from zero, and the former of
which is also by nature "fixed" across all time periods.
My results are "better" using the FE estimator in the sense that the
hypothesized relationship is stronger and more consistent. But I am
wary of the potential for bias due to the dropping of the all-zero
countries.
Is there a recommended way for arbitrating between the two approaches
that takes into account the issue of dropped "all-zero" units? I have
tried estimating the -fe- model, than following that estimation with an
-re- model that includes only the observations used in the -fe- model,
and here too the results between the two estimators differ
significantly.
Is my best bet to perform some exploratory analysis of the dropped
"all-zero" countries to examine whether the dropped countries differ
significantly across the main explanatory variable than the non-dropped
countries? And if so, what is the most scientific way of going about
this?
Jason Webb Yackee, PhD Candidate; J.D.
Fellow, Gould School of Law
University of Southern California
[email protected]
Cell: 919-358-3040
*
* 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/
