Suppose we have the linear regression model given by $y=X\beta+\epsilon$, but we have a violation of assumptions where $X$, the regressor matrix, and $\epsilon$ are correlated. Also, suppose there is an observable collection of variables $Z$, such that $Z$ is uncorrelated with $\epsilon$, and $Z$ is correlated with $X$.

To derive the IV estimator $\beta_{IV}$ that is consistent for the true value of $\beta$, I am told that the first step is to premultiply the model $y=X\beta+\epsilon$ by $Z^T$, which ultimately yields $\beta_{IV}=(Z^TX)^{-1}Z^Ty$. My question is, what exactly is the intuition for premultiplying both sides by the instrument, $Z$?

  • $\begingroup$ Try to consider the special case of a single regressor $X$ and $Z$ is a dummy---you just get a (sub-)sample mean for $X$ and $Y$. In general, when instrumenting you are taking conditional averages of $y$ and $X$ conditional on $Z$. Since $Z$ is exogenous, conditioning on $Z$ washes out endogeneity. $X$ is randomly assigned---conditional on $Z$. The IV estimates are noisy because, informally, passing to conditional averages mean you now have a "smaller" sample---again consider the dummy IV case. $\endgroup$ – Michael Apr 13 '19 at 16:10

Intuitively, $Z^T\epsilon$ is expected to be $0$ (as the sample size increases) since they are uncorrelated. So, the equation becomes $$Z^Ty\approx Z^TX\beta+\underbrace{Z^T\epsilon}_{0}$$ which yields the solution you have by multipying each side by $(Z^TX)^{-1}$

For OLS (i.e. when $X$ and $\epsilon$ are uncorrelated), you multiply each side by $X^T$ instead of $Z^T$, and you'll have the typical solution $(X^TX)^{-1}X^Ty$.

| cite | 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.