A Comparison between bias Approximations Applied to Bivariate VAR Models

Bivariate VAR models are Monte Carlo simulated and OLS estimated, The resulting biases are used to compare two alternative approximations to the bias. They are found to be equivalent for first-order models, whereas for second-order models Nicholls and Pope's approximation outperforms Tjostheim and Paulsen's approximation, although the difference is negligible. This being the case when approximations are based on true parameters as well as when they are based on estimates, any of the two approximation can be used to construct bias-reduced estimates as outlined in an earlier paper, leading to better estimates both in terms of bias and mean square error than the original OLS estimate. Finally, using a more correct version of Nicholls and Pope's approximation turns out to improve results only marginally for second-order models, and for first-order models they appear to deteriorate.