3
$\begingroup$

I'm looking at the influence of an additional regressor in an OLS-model and on the adjusted $\bar{R}^2$. What I have to proove is that $\bar{R}^2$ rises if and only if the square of the respective t-statistic is bigger than 1. I found a solution to the proof in "Greene - Econometric Analysis, (Chapter 3, exercise 9)" and managed to replicate the steps more or less to get to the final result, which is as follows:

$\frac{b_k^2(x_k'M_1x_k)}{(s^2)}>1$,

where $b_k$ is the coefficient of the additional regressor in the long model and $s^2$ is its estimated variance. In my understanding the squared t-statistic of this regressor should just be

$\frac{b_k^2}{(s^2)}$.

How do I interpret the rest of the nominator $(x_k'M_1x_k)$? Is this even the right proof I'm looking for?

The complete solution I'm looking at is given in this PDF (p5, ex. 9): pages.stern.nyu.edu/~wgreene/Text/Greene_6e_Solutions_Manual.pdf

$\endgroup$
2
$\begingroup$

Recall that the t-statistic on the $k$-th coefficient (for testing $\beta_k=0$ assuming homoskedasticity) in a linear regression is given by $$ t=\frac{b_k}{\sqrt{s^2(X'X)^{-1}_{kk}}}, $$ where $(X'X)^{-1}_{kk}$ denotes the $k$-th diagonal element of $(X'X)^{-1}$, so that the squared t-ratio is $$ t=\frac{b_k^2}{s^2(X'X)^{-1}_{kk}}. $$ This is the same as your $$ \frac{b_k^2(x_k'M_1x_k)}{s^2}=\frac{b_k^2}{s^2/(x_k'M_1x_k)} $$ due to the Frisch-Waugh-Lovell theorem discussed in for example Chapter 3 of Greene's textbook, or more specifically/directly, as an application of the partitioned inverse lemma:

For $A\;(m \times m)$, $B\;(m \times n)$, $C\;(n \times m)$ and $D\;(n \times n)$ we have that, provided the respective inverses exist, $$ \left(% \begin{array}{cc} A & B \\ C & D \\ \end{array}% \right)^{-1}= \left(% \begin{array}{cc} A^{-1} + A^{-1}B(D-CA^{-1}B)^{-1}CA^{-1} & -A^{-1}B(D-CA^{-1}B)^{-1} \\ -(D-CA^{-1}B)^{-1}CA^{-1} & (D-CA^{-1}B)^{-1} \\ \end{array}% \right). $$ For $A=X_1'X_1$ (with $X_1$ the other regressors except $x_k$), $B=X_1'x_k$, $C=x_k'X_1$ and $D=x_k'x_k$, we can plug in for the lower-right element of the inverse to directly obtain that $$ (D-CA^{-1}B)^{-1}=(X'X)^{-1}_{kk}=1/(x_k'M_1x_k). $$

$\endgroup$

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.