A limitation of the current practice of pre-testing for unit roots before modelling cointegration is that unit root tests have low power in finite samples. As a result, series may be misclassified and estimation and inference dramatically affected. Phillips (1995) proposes a method based on the principle of fully modified least squares that can be used for estimation and inference whether the series are stationary, nonstationary or a mixture of both. Sheldon (1997) also proposes an IV method that deals with the case where the size of the root in the regressor is not known a priori. In an application to UK consumers' expenditure we show how the appropriate estimator may be chosen in the light of the available information.