Escolha do valor crítico em modelos VAR restritos para análise da resposta ao impulso

Abstract

The correct parametrization of vector autoregression (VAR) models plays a central role on the accuracy of impulse response estimates. The incorrect inclusion of lagged temporal relationships among variables causes biased estimators for the responses to impulses on the variables. This work shows that the bias increases with the number of non-existent relationships included in the model and it also increases with the number of true relationships not entered in the model. An intensive simulation study is performed to explore the balance between the overall alpha level, in the light of Bonferroni's correction, and the statistical power for identifying each of the actual correlations. With this, a rule of thumb is offered for selecting the alpha level and sample size required to bound the bias of impulse response estimates under desired tolerances.