TESTING FOR COINTEGRATION IN LINEAR-QUADRATIC MODELS

This article evaluates the finite-sample performance of various tests for cointegration by Monte Carlo methods. The evaluation takes place within the linear quadratic model. The results indicate sharp differences in the ability of the tests to detect cointegrating relations, especially when the cost-of-adjustment term and the number of regressors are large. Although no single test dominates for all the parameter settings considered, overall the augmented Dickey-Fuller, Z(alpha), and Z(t) tests of Phillips seem the most reliable in terms of test size and power.