1
$\begingroup$

Suppose we have $k_1$ exogenous variables $X_1$, $k_2$ endogenous variables $X_2$ and $k_2$ instruments $Z$ and the model $Y=X_1\beta_1+X_2\beta_2+e$. Let the 2SLS estimates be $(\hat\beta_1,\hat\beta_2). $If we directly run a regression with the instruments instead $Y=X_1\alpha_1+Z\alpha_2+e$, then the claim is that $\hat\alpha_1=\hat\beta_1$ i.e. the estimated coefficient on the exogenous variables are the same.

I'm having trouble formally proving the result as well as understanding the intuition. I tried applying the Frisch–Waugh–Lovell theorem so that $\hat\alpha_1=(X_1'M_ZX_1)'(X_1'M_ZY)$ where $M_Z=(I-Z(Z'Z)^{-1}Z')$ and then tried showing that $\hat\beta_1$ will be equal to that. However, applying the FWL theorem in the 2SLS context, $\hat\beta_1=(X_1'M_{\hat{X_2}}X_1)'(X_1'M_{\hat{X_2}}Y)$. This would require $M_{\hat{X_2}}=M_Z$. However, I can't see why the two should be equal. Furthermore, that would imply that the residuals from regressing $X_1$ on $Z$ and $X_1$ on $\hat{X}_2$ would be the same which doesn't seem to make sense either since the predicted value $\hat{X}_2$ uses both $X_1$ and $Z$. What am I missing here?

$\endgroup$
1
$\begingroup$

The OLS regression of $Y$ on $X$ decomposes dependent variable $Y$ into orthogonal components $$ Y = \hat{Y} + e $$ in $\mathbb{R}^n$, where $\hat{Y} = X \hat{\beta}$ is the fitted value and $e$ the residuals.

It's a basic fact of linear algebra that this orthogonal decomposition remains the same, for all regressor matrix $X$ with the same column space. The OLS $\hat{\beta}$ may change, but $\hat{Y}$ remains the same.

This can be extended to a subset of regressors.

Suppose $X = [X_1 \; Z]$ and you replace $Z$ by $\tilde{Z}$ so that the regressors in $\tilde{Z}$ has the same span as those in $Z$, i.e. $Z = \tilde{Z} W$ for some invertible $W$. Then \begin{align*} Y &= X_1 \hat{\beta}_1 + Z \hat{\beta}_2 + e \\ &= X_1 \hat{\beta}_1 + \tilde{Z} W \hat{\beta}_2 + e. \end{align*} Uniqueness of the orthogonal decomposition therefore tells you that regressing $Y$ on $[X_1 \; \tilde{Z}]$ gives OLS estimate $$ [\hat{\beta}_1 \; W \hat{\beta}_2]. $$ Notice that the coefficients for $X_1$ from the two regressions remains the same---$\hat{\beta}_1$.

The 2SLS example is a special case, where $Z$ consists of instruments and $\tilde{Z} = \hat{X}_2$ are the fitted values from the first stage regression of $X_2$ on $Z$. Note that $\tilde{Z}$ and $Z$ has the same column space if the model is exactly identified---in your notation, $k_2$ instruments for $k_2$ endogenous variables.

The statement is no longer true when model is over-identified---when the column space of $\tilde{Z}$ is a proper subspace of that of $Z$.

$\endgroup$

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.