5
$\begingroup$

Why is the F-test for overall significance (OLS regression analysis) invalid when residuals are heteroscedastic? Is there a way to calculate it in a consistent way under heteroscedasticity? Is there any function in R to accomplish that?

$\endgroup$
  • 1
    $\begingroup$ I think it can be calculated in R through waldtest function. Example: waldtest(model_restricted, model_unrestricted, vcov=vcovHC(model_unrestricted)) Is it correct? $\endgroup$ – Baumann Oct 22 '13 at 18:59
  • 1
    $\begingroup$ About the invalidity of the test, the Wooldridge (Introductory Econometrics, 2nd edition, p. 253) says: "If heteroskedasticity is present, this version of the [F] test is invalid. The heteroskedasticity-robust version has no simple form, but it can be computed using certain statistical packages." And the mystery continues! $\endgroup$ – Baumann Oct 22 '13 at 19:05
4
$\begingroup$

Using the F-statistic formula from Hayashi:

$F=\frac{\left(Rb-r\right)^T\left[R\left(X^T X\right)^{-1}R^T\right]^{-1}\left(Rb-r\right)/ \rho}{s^2}$

(notation: we're testing the hypothesis $R\beta = r$, $R\in\mathbb{R}^{\rho\times K}$, rank($R$)=$\rho$, $s^2$ is the variance of the error term)

We can immediately see why this won't apply to the heteroskedastic case--in particular, we definitely don't expect $s^2$ to be a constant.

Intuitively, we know there's a (nontrivial) relationship between $t$ and $F$ statistics--so given that $t$ statistics are heteroskedasticity-dependent, we should expect $F$-statistics to be so as well.

See here (pp 3-4) for why the Wald test you suggested is correct & hence the waldtest function in R is appropriate.

$\endgroup$

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.