Can a Chow test be run on a dataset which has autocorrelation and/or heteroscedasticity? Will the F-stat give accurate results?
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
Yes, provided an appropriate robust version of the statistic is used.
The Chow test is a special case of a general Wald test testing multiple restrictions, whose test statistic (in the $\chi^2$ version, which unlike the $F$ statistic version is justified, asymptotically, without error normality) is given by $$ W=n(Rb-r)'\left[R\widehat{\mathrm{Avar}}(b)R'\right]^{-1}(Rb-r). $$ Here, $R\beta=r$ is the null we test (coefficient equality of two subvectors in the case of the Chow test), and $b$ is the OLSE. Now, the test statistic will be asymptotically robust to heteroskedasticity and autocorrelation if you use a robust (aka Eicker-White) estimate $\widehat{\mathrm{Avar}}(b)$ of the asymptotic variance of the OLS estimator.
answered Apr 27, 2016 at 6:45
-
$\begingroup$ Actually I'm a bit of a novice, I'm just running 3 regressions,one each pre and post the structural break point, and one for the whole period. I am then calculating the F-stat. So will this approach work in the presence of heteroscedasticity/autocorreation? $\endgroup$ Commented Apr 27, 2016 at 7:43
-
$\begingroup$ It is not fully transparent how exactly you compute the test statistic, but my guess is no, as you do not seem to use robust asymptotic variance estimators $\endgroup$ Commented Apr 27, 2016 at 8:29