Abstract
This paper considers computer intensive methods to inference on cointegrating vectors in maximum likelihood cointegration analysis. The likelihood ratio test statistics used in the literature are known to have an asymptotic X2-distribution. However, previous simulation studies show that the size distortion of the test can be considerable for small samples. Typically the nominal significance level, say 5%, is much smaller than the attained actual level, and as a consequence, too many true null hypotheses will be rejected. It is demonstrated how a parametric bootstrap can be implemented, frequenly resulting in a nearly exact a level test. Furthermore, response surface regression is used to examine small sample properties of the asymptotic likelihood ratio test. The estimated equations can be used as approximate finite-sample corrections, allowing rough, but easily applied, corrections of the LR test.