The small sample properties of tests on long-run coefficients in cointegrated systems are still a matter of concern to applied econometricians. We compare the performance of the Bartlett correction, the bootstrap and the fast double bootstrap for tests on ccointegration parameters in the maximum likelihood framework. We show by means of a theoretical result and simulations that all three procedures should be based on the unrestricted estimate of the cointegration vectors. The fast double bootstrap delivers superior size correction, whereas the Bartlett correction leads to the least loss of power. However all three perform much better than the asymptotic tests and differences between them are small.