Cointegration requires all the variables in the system to have exact unit roots; accordingly it is conventional for researchers to test for a unit root in each variable prior to a cointegration analysis. Unfortunately, these unit root tests are not powerful. Meanwhile, conventional cointegration methods are not at all robust to slight violations of the requirement that each variable have a unit root. In this paper I show how this difficulty may be circumvented by instrumenting the regressors in the cointegrating regression by deterministic polynomial time trends or by artificially generated random walks.