Suppose that we run the simple linear regression $Y = \alpha + \beta X + \epsilon$. I want to test whether the independent variable $X$ is exogenous. If the correlation between the independent variable $X$ and the residual of linear regression $\epsilon$ is almost zero, i.e. $cor(X, \epsilon) \approx 0$, can I then conclude that this simple test suggests that the independent variable $X$ is exogenous?
$\begingroup$
$\endgroup$
4
asked Dec 27, 2013 at 23:01
-
2$\begingroup$ the issue is that in practice you don't observe $\epsilon$ (only $\hat{\epsilon}$) so you can never compute $\mbox{cor}(X,\epsilon)$ $\endgroup$– user603Commented Dec 27, 2013 at 23:45
-
$\begingroup$ As said above, $\epsilon$ is unobservable so you need to rely on your own knowledge and use the common sense to decide about exogeneity. $\endgroup$– StatCommented Dec 27, 2013 at 23:48
-
2$\begingroup$ The correlation between regressors and residuals (not errors) in a linear regression model estimated by least squares is always zero. You cannot test exogeneity (conditional uncorrelatedness) without instrumental variables. $\endgroup$– tchakravartyCommented Dec 27, 2013 at 23:49
-
$\begingroup$ Thank you for your comments, as putting them all together answers my question. Is anyone willing to organize these together into an answer which I can accept? $\endgroup$– I Like to CodeCommented Dec 28, 2013 at 1:59
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
The linear regression model is $$ \boldsymbol{Y} = \mathbf{X}\boldsymbol{\beta} + \boldsymbol{\varepsilon} $$ together with the conditional uncorrelatedness assumption $\mathbb{E}( \mathbf{X}\boldsymbol{\varepsilon}) = \boldsymbol{0}$.
If estimation of the parameters $\boldsymbol{\beta}$ proceeds by least squares, then the first order conditions (normal equations) are $$ \begin{align} \mathbf{X}'\left(\boldsymbol{Y} - \mathbf{X}\hat{\boldsymbol{\beta}} \right) &= \boldsymbol{0} \\ \mathbf{X}'\hat{\boldsymbol{\varepsilon}} &= \boldsymbol{0} \end{align} $$
where the last line indicates that the correlation between the residuals and the regressors is always zero for a linear regression model estimated by least squares. Thus this cannot form the basis of a test for the unconditional uncorrelatedness assumption in the linear regression model.
In order to test the exogeneity assumption, you will need access to instrumental variables. See the Durbin-Wu-Hausman test.
answered Dec 28, 2013 at 15:40