We investigate the probability of rejecting the I(1) hypothesis when u
nit root tests are applied to fractionally integrated time series. Esp
ecially the augmented Dickey-Fuller test performs poorly. Our analytic
al arguments are supported by Monte Carlo experiments.