Power of tests for unit roots in the presence of a linear trend

Dickey and Fuller (1981) suggested unit root tests for an autoregressive model with a linear trend and a fixed initial value. This model has nuisance parameters so later authors have often worked with a slightly different model with a random initial value in which nuisance parameters can be eliminated by an invariant reduction of the model. This facilitates computation of envelope power functions and comparison of the relative performance of different unit root tests. It is shown here that invariance arguments also can be used when comparing power within the model with fixed initial value. Despite the apparently small difference between the two models the relative performance of unit root tests turns out to be very different.