University of New South Wales

Lag Length and Mean Break in Stationary VAR

Email address: m.yang@unsw.edu.au

Keywords: Autoregression lag length; Mean break; Information criteria; Consistency

JEL Classifications: C12,C13,C32

Abstract:
Break tests are often performed in practice as a specification check on a model selected by an information criterion. In the context of a vector autoregression (VAR) model, when the lag length is selected ignoring a possible break in the mean, it is not clear whether the subsequent break tests have desired size and power properties since the estimated lag length may be inconsistent. We show that the estimated lag length is at worst biased upwards asymptotically if a break in the mean is ignored. Thus the large-sample size and power properties of the break tests are not affected by conditioning on the estimated lag length. The above modelling approach is compared with two other approaches: testing the break prior to selecting the lag length and determining the break and lag length simultaneously. Monte Carlo experiments show that the third approach tends to select a “spurious” break too often and the first two approaches exhibit stable performances.

PDF file of paper: yang.pdf

Session: Topics in Time Series

Time: Sunday, 8 July, 8am - 9:30am

Room: B