The importance of long run structure for impulse response analysis in VAR models

This paper examines the finite sample accuracy of impulse response functions in VAR models when cointegration is present but not imposed in estimation. It finds that in typical applications there can be substantial biases at the short horizon, as well as the long. These results suggest that, in contrast to common practice, before conducting impulse response analysis researchers should analyse the cointegrating properties of the data and restrict the VAR (i.e. estimate a reduced rank VAR) when the data are cointegrated. But there are practical dangers associated with following this advice. When the data are “nearly” cointegrated then not just does the VAR in levels provide misleading inference about long-run impulse responses but so does the reduced rank VAR.