Abstract
Recently researchers have begun to develop methods to detect stationarity or cointegration when the underlying error correction process follows a threshold model. In the context of one type of threshold model, Pippenger and Goering found that the power of unit root or cointegration tests decay substantially under many economically relevant parameterization. However, in a recent article Balke and Fomby conclude that linear tests work 'reasonably well when threshold cointegration is present' and based on these findings propose a modelling methodology under threshold cointegration. This study examines the threshold processes they analysed under a wide range of parameterizations and finds that linear unit root and cointegration tests display relatively low power even under large sample sizes. Hence, contrary to Balke and Fomby linear cointegration tests appear to be an inappropriate test under threshold cointegration.