Clustering SE in an Instrumental Variable (IV) regression yields to negative effects compared to "unclustered" IV regression - How's that?

[Andrea Maier]

Hello


today I performed an IV regression with two instruments. Both instruments are indicator variables (dummies). At some point, I felt the need to cluster the SE by individual - and all over sudden- the effects of the instrumented variables changed direction. I thought only Standard errors change.

How can that be?

I try to make a similar example using dataex, but so far I am loosing...

Thank you!

[FernandoRios]

can you provide the full output you get (that will include "first" option).
My guess is that the variable you are using for cluster does not identify the sample as your first model. in other words, it may have missing observations, which is causing the "change" in sign on your results.

[Hong Il Yoo]

I suspect that there are missing values in your cluster ID variable. So when you request cluster robust standard errors, Stata carries out estimation on a subset of the sample.