Modified lag augmented vector autoregressions
提出一种修正的滞后增广向量自回归估计方法，通过刀切法消除准渐近偏差，使Wald统计量渐近服从卡方分布，在有限样本下比传统方法更准确、对滞后长度误设更稳健。
This paper proposes an inference procedure for a possibly integrated vector autoregression (VAR) model. We modify the lag augmented VAR (LA-VAR) estimator to exclude the quasiasymptotic bias, which is associated with the term Op(T-1), using the jackknife method. The new estimator has an asymptotic normal distribution and then the Wald statistic to test for the parameter restrictions has an asymptotic chi-square distribut,ion. We investigate the finite sample properties of this approach by comparing with the LA-VAR approach. We find t,hat our modified LA-VAR (MLA-VAR) approach excels the LA-VAR approach in view of an accuracy of the empirical size and the robustness to the tnisspecification of the lag length. The MLA-VAR approach may be used when the researchers place importance on an accuracy of the size, and also be used to complement other testing procedures that may suffer from serious size distortion.