We consider the power of unit root tests for different deviations of the initial observation from the deterministic component of the series. Following recent work highlighting the relative power performance of extant tests, we propose a new procedure based on a data-dependent weighted average of...