This paper proposes a new unit root test against a non-linear exponential smooth transition autoregressive (ESTAR) model. The new test is build upon the non-standard testing approach of Abadir and Distaso (2007) who introduce a class of modified statistics for testing joint hypotheses when one of the alternatives is one-sided. In a Monte Carlo study the popular Dickey-Fuller type test proposed by Kapetanios et al. (2003) is compared with the new test. The results suggest that the new test is generally superior in terms of power. An application to a real effective exchange rate underlines its usefulness.