Elliott (1998) has shown that inference on cointegrating coefficients is not robust to local deviations from the unit root. When unit roots are incorrectly imposed, the resulting Vector Error Correction Model is misspecified. Quasi-maximum likelihood in this misspecified model leads to asymptotically mixed normal estimators, but centered around the pseudo-true value, which is related to the concept of long-run multiplier (as discussed by Pesaran, 1997). However, in near-integrated systems, there is no unique stationary linear combination, so that the identification of appropriate long-run effects has to be addressed. We follow the approach of Johansen (2005) and focus on the long-run relative impulse response. We also address the invalidity of bootstrap inference (as noted earlier by Stock and Watson, 1996), and explore the effectiveness of the double bootstrap.