We know that sometimes the error variable in a regression framework may be correlated with explanatory variables; this happens e.g. when we omit an important predictor from our study, getting the well known "omitted variable bias" issue.

Now, suppose that in our case study $(Y,X_1,X_2)$ are jointly distributed but we're aware only of $X_1$.

Let's assume for simplicity that $E[Y|X_1,X_2]$ and $E[Y|X_1]$ are both linear.

We collect $n$ data , $i=1,...,n$ ; In our Hp. the data generating process is in the $long$ form: $$y_i=\beta_0+\beta_1x_{i1} +\beta_2x_{i2}+\epsilon_i$$

our regression model is in the $short$ form: $$y_i=\Gamma_0 +\Gamma_1x_{i1} +u_i$$

Let us recall $\mathbf{X}=(1,X_1)$ and $\mathbf{\Gamma }=(\Gamma_0,\Gamma_1$), such that $y=\mathbf{X}^t\mathbf{\Gamma}+u$.

If we take $\mathbf{\Gamma}$ as $\mathbf{\Gamma}=\left({E[\mathbf{X}\mathbf{X^t}]}\right )^{-1}E[\mathbf{X}y]$ we end up with $E[\mathbf{X}u]=0$, in fact:

$u=y-\mathbf{X^t}\mathbf{\Gamma} \\ E[\mathbf{X}u]=E[\mathbf{X}y-\mathbf{X}\mathbf{X^t}\mathbf{\Gamma}]=E[\mathbf{X}y-\mathbf{X}\mathbf{X^t}\left({E[\mathbf{X}\mathbf{X^t}]}\right )^{-1}E[\mathbf{X}y]]=E[\mathbf{X}y]-E[\mathbf{X}\mathbf{X^t}]\left({E[\mathbf{X}\mathbf{X^t}]}\right )^{-1}-E[\mathbf{X}y]=0.$

But $E[\mathbf{X}u]=0$ means $cov(\mathbf{X},u)=0.$

This means that with this particular choice of $\mathbf{\Gamma} $ the endogeneity issue fades out,providing we're able to compute the aforementionend expectation. Clearly there's something wrong with my argument but I can't figure out what it is.


Your Answer

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

Browse other questions tagged or ask your own question.