Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It only takes a minute to sign up.
Sign up to join this community
Suppose, I estimate a simple instrumental variable regression with 2SLS. The reduced form is
$Y = \alpha_0 + \alpha_1 Z+u,$
where $Z$ is a dummy (for simplicity). The first stage is
$X = \beta_0 + \beta_1 Z+v,$
and the second stage becomes
$Y = \gamma_0 + \gamma_1 \hat{X}+w$,
where $\hat{X}$ are fitted values. We have $\hat{\beta}_1=1$ so that the IV estimate becomes:
$\hat{\gamma_1}=\frac{\hat{\alpha_1}}{\hat{\beta_1}} = \hat{\alpha_1}$. Is this evidence that the exclusion restriction holds?
You are right in saying that the IV coefficient is proportional to the reduced form and, in the case of $\widehat{\beta}_1$ it is the same. However, this is not a test for whether the exclusion restriction holds. In fact, there is no formal way of testing this assumption which makes the use of instrumental variable techniques tricky. This is the reason as to why most papers in economics usually devote a great deal of space to making an argument for why their instrument is valid.
The first stage coefficient, or rather it's t-statistic (in the just-identified case, otherwise the F-statistic on the excluded instruments), only allows you to test for instrument relevance. This is also important because weak instruments will yield biased estimates.
The point about the exclusion restriction is to make an argument for the assumption that $\text{Cov}(Z,w)=0$ such that the instrument affects the outcome $Y$ only through the channel $X$ but not directly. Given that you never observe $w$ there is no direct way of testing this assumption.
