This paper considers tests for a unit root when the innovations follow a near-integrated GARCH process. We compare the asymptotic properties of the likelihoodratio statistic with that of the least-squares based Dickey-Fuller statistic. We first useasymptotics where the GARCH variance process is stationary with fixed parameters,and then consider parameter sequences such that the GARCH process converges to adiffusion process. In both cases, we find a substantial asymptotic local power gain ofthe likelihood ratio test for parameter values that imply heavy tails in theunconditional innovation distribution. An empirical application to the term structureof interest rates in the Netherlands illustrates the proposed procedures.