I'm busy reading through an econometrics textbook (page 147), and I don't understand the step

$$\mathrm {Var}\left(n^{\frac 12}\left(\hat\beta - \beta\right)\right) = \boldsymbol{A^{-1}}\sigma^2\boldsymbol{AA^{-1}} = \sigma^2\boldsymbol{A^{-1}}$$

The relevant part of the text is:

first bit second bit and so on.


Did we get the two $\boldsymbol{A^{-1}}$'s from the rule that $\mathrm{Var}(a\boldsymbol X) = a^2\mathrm{Var}\boldsymbol X$ ? And should the two $\boldsymbol{A^{-1}}$'s be on the right-hand side of the expression? Why should they be on the right, when $\left(\frac 1n\boldsymbol{X'X}\right)^{-1}$ is on the left?

Secondly, I don't understand what the "$d$ " means in $\overset d\to$.


$$\beta-\hat{\beta}=\beta-(X'X)^{-1}X'Y=\beta-(X'X)^{-1}X'(X\beta+\epsilon)$$ $$\beta-\hat{\beta}=\beta-(X'X)^{-1}(X'X)\beta+(X'X)^{-1}X'\epsilon$$ $$\beta-\hat{\beta}=(X'X)^{-1}X'\epsilon$$ $$\beta-\hat{\beta}=(n^{-1}X'X)^{-1}(n^{-1}X'\epsilon)$$

Now the $\overset d\to$ means "convergence in distribution", by the central limit theorem we have that if the error term is homocedastic $$\sqrt{n}(n^{-1}X'\epsilon)\overset d\to N(0,\sigma^2 A)$$ We know that $$(n^{-1}X'X)\overset p\to A^{-1}$$ and hence by Slutzky's theorem we have that $$\sqrt{n}(\beta-\hat{\beta})\overset d\to A^{-1}N(0,\sigma^2 A)=N(0,(A^{-1})\sigma^2 A (A^{-1})')$$

Since $(A^{-1})'=A^{-1}$ we end up with $$\sqrt{n}(\beta-\hat{\beta})\overset d\to N(0,\sigma^2 A^{-1})$$

Notice that in order to obtain this you need ergodic stationarity of the variables $X$ and $Y$ (to ensure the convergence to the population value) and homoskedasticity of the error term in order to express the variance/covariance matrix as the product of an scalar and a matrix (if the matrix $A$ has constant elements in its diagonal of course)

  • $\begingroup$ I'm not sure why left-multiplying a distribution by a matrix results in the variance being $M\Sigma M'$, where $M$ is the matrix and $\Sigma $ is the covariance matrix. $\endgroup$ – ahorn Mar 11 at 9:38
  • $\begingroup$ You mentioned the fact that $Var(aX)=a^2Var(X)$. Think that "multiplying" the ditribution is like saying that you are multiplying the random variable itself. So for instance if you have that $X\sim N(0,\sigma^2) $ then $aX\sim N(0,a^2\sigma^2)$ right? Well with matrices is the same thing, the difference is that you cannot just write $M^2$ (see quadratic forms). Instead you write $M\Sigma M'$. $\endgroup$ – RScrlli Mar 11 at 9:52
  • $\begingroup$ My question was more to do with this multiplication of the variance than with explaining the theorem. $\endgroup$ – ahorn Mar 11 at 10:05
  • $\begingroup$ The variance result you quoted is for univariate random variables. The one for vector-random variables is $\text{Var}(AX) = A\text{Var}(X) A'$ where here $\text{Var}$ is the variance-covariance matrix of its argument (e.g. see Wikipedia here). Even if you're not familiar with this, it's easy to confirm the correctness of the result element-by-element. Note that $A$ needn't be square, as long as its column-dimension matches the length of $X$ $\endgroup$ – Glen_b Mar 11 at 10:51

$\sigma^2$ is commutative: $$\boldsymbol {A^{-1}}\sigma^2\boldsymbol{AA^{-1}} = \sigma^2\boldsymbol{A^{-1}A A^{-1}} = \sigma^2\boldsymbol{A^{-1}}.$$

Even if the $\boldsymbol{A^{-1}}$'s were at the left-hand side of the expression, the commutativity of $\sigma^2$ will ensure that the result $\left(\sigma^2\boldsymbol{A^{-1}}\right)$ is true.

$\overset d\to$ means "the distribution converges to".


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.