UNIT-ROOT TESTS WITH CONDITIONAL HETEROSKEDASTICITY

This paper considers the finite-sample accuracy (size) of the Dickey-Fuller unit root tests when the errors are conditionally heteroskedastic. We consider the specific case that the error variance follows a GARCH (1, 1) model. The Dickey-Fuller tests tend to overreject in the presence of GARCH errors, but the problem is not very serious except when the variance process is nearly degenerate (in the sense that the ratio of the GARCH intercept to the initial variance is near zero) and the volatility parameter is large.