# What can go wrong if I use normal 2sls IV regression for count data?

My dependent variable is count data and for explanatory analysis we used a negative binomial regression. One variable is endogenous and we would like to instrument it. However, in Stata there is only one (user-written) program for IV regressions when the dependent variable is count data by nature, namely ivpois. It employs GMM, though. For R, there are even less, namely zero, packages available that deal directly with such IV regressions.

Therefore I thought about using standard 2SLS methods (ivreg() for R or ivreg/ivreg2 for Stata). Can anything go wrong if we do so?

• Yes: in particular, such a regression could imply negative responses, even within the range of the data. May 29, 2015 at 13:22
• So, given the limited supply, we should definitely stick with ivpois for Stata? May 29, 2015 at 13:27
• The best advice is to write your own instrumented negative binomial regression routine if you think that is needed. But the focus here on CV is not what software to use but what statistical problem you are facing. May 29, 2015 at 13:29