I've read the discussions about how to treat endogenous interaction terms when using an instrumental variable approach (2SLS) and followed the idea of Andy.

However, in my case there is just one instrument which is used for the endogenous independent variable (X) and the interaction term which consists of the endogenous independent variable (standardized) and one of my exogenous controls (also standardized)

Using Stata v15.1 this results in:

ivregress 2sls Y exog1 exog2 exog3 i.industry i.year (X Xegog1 = inst1 inst1exog1), vce (cluster isin)

However, following this approach my r-squared is going down dramatically (from >20% to less then 2%) just due to the "handmade" interaction-instrument.

Maybe there is an obvious answer to the declining R-squared or even some misspecification in terms of e.g. the robust standard-errors which might cause this? If someone could help this is highly appreciated.

  • $\begingroup$ Hey all, I still didn't manage to find a solution, but maybe my last approaches could trigger an answer as it might lead to a more detailed problem description. When I run the first stage regression manually - just with the interaction term, which means without the single endogenous variable - everything looks fine. T-stats are comparable to the OLS results and r-squared similar. However, running then ivregress with exact the same variables, this leads even to different coefficients, which I can absolutely not explain. As to my understanding, ivregress should not impact the coefficients. $\endgroup$ – Sebastien Apr 2 at 10:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.