Suppose we want to estimate \begin{equation} y = \alpha_0 + x_1 \beta + c'\gamma + \varepsilon \end{equation} where $\beta$ is the variable of interest and $c$ is a vector of controls. We suspect $E[x\varepsilon]\neq0$ but we have an instrument $z$ such that $E[z\varepsilon]=0.$ My question is - as long as IV is valid - should we care that there could be an endogenous control $c_1$ such that $E[c_1\varepsilon]\neq0$? Why or why not?

My hunch is that it doesn't because as long as IV is valid (it satisfies inclusion-exclusion principle) the correlation between controls and unobservables should not lead to the inconsistency of the estimate of $\beta$.


In general, if $c_1$ is endogenous, you need instruments for $c_1$ as well.

Example: Even when $x_1$ is exogenous, if $c_1$ is endogenous and $x_1$ and $c_1$ are correlated, then OLS (which is the IV estimator using $x_1$ as IV for $x_1$) is generally inconsistent.

Exceptions exist. For example, when exactly identified, if $(Ezc') (Ecc')^{-1} (Ecu)=0$, then your IV estimator is consistent, I guess. (Can derive this using standard technique involving the law of large numbers.)

| improve this answer | |

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.