I am given regression model $y_i=x_i^T\beta+\varepsilon_i$ with heteroscedasticity.

Let $Avar(b)=E[x_ix_i^T]^{-1}E[\varepsilon_i^2x_ix_i^T]E[x_ix_i^T]^{-1}$

Hayashi's book states that regression coefficients are asymptotically normal:

$\sqrt{n}(b-\beta) \rightarrow^{D} N(0, Avar(b))$ therefore $b\sim N(\beta, Avar(b)/n)$ asymptotically for sufficiently large n.

Further, Hayashi states that the standard error of $b_k$ is precisely $\sqrt{Avar(b)_{kk}/n}$ where $Avar(b)_{kk}$ implies the kkth element of the diagonal.


Why do we use $\sqrt{Avar(b)_{kk}/n}$ as a measure of standard error rather than $\sqrt{V(b|X)_{kk}}$ where $V(b|X)$ is the conditional variance of b?

I am confused because earlier in the book he mentioned that the conditional variance should be used. For example, he specified that in the finite sample case under conditional homoscedasticity the standard error of $b_k$ was precisely:

$\sqrt{V(b|X)_{kk}}=\sqrt{\sigma^2(X^TX)^{-1}}$ with $\sigma^2=E[\varepsilon_i^2|X]$

  • 2
    $\begingroup$ Why do you write $V(b|X)$ and say 'unconditional' variance when clearly you are using notation for conditioning on $X$. Makes question a little confusing. $\endgroup$ – bomadsen May 17 at 16:27
  • $\begingroup$ If $\sigma^2$ is known, then use $V(b|X)$. $\endgroup$ – Bertrand May 18 at 13:18

So the asymptotic variance in the first expression that you give is the variance under heteroskedasticity as you note. The definition you gave of $V(b|X)$ is under homoskedasticity. As one of the comments states, we could use $V(b|X)$ (as you define it) under this assumption. It seems that Hayashi uses this notation because he wants to consider heteroskedastic errors.


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.