On Bootstrap Evaluation of Tests for Unit Root and Cointegration

Abstract: This thesis is comprised of five papers that all relate to bootstrap methodology in analysis of non-stationary time series.The first paper starts with the fact that the Dickey-Fuller unit root test using asymptotic critical value has bad small sample performance. The small sample correction proposed by Johansen (2004) and bootstrap are two effective methods to improve the performance of the test. In this paper we compare these two methods as well as analyse the effect of bias-adjusting through a simulation study. We consider AR(1) and AR(2) models and both size and power properties are investigated.The second paper studies the asymptotic refinement of the bootstrap cointegration rank test. We expand the test statistic of a simplified VECM model and a Monte Carlo simulation was carried out to verify that the bootstrap test gives asymptotic refinement.The third paper focuses on the number of bootstrap replicates in bootstrap Dickey-Fuller unit root test. Through a simulation study, we find that a small number of bootstrap replicates are sufficient for a precise size, but, with too small number of replicates, we will lose power when the null hypothesis is not true.The fourth and last paper of the thesis concerns unit root test in panel setting focusing on the test proposed by Palm, Smeekes and Urbain (2011). In the fourth paper, we study the robustness of the PSU test with comparison with two representative tests from the second generation panel unit root tests. In the last paper, we generalise the PSU test to the model with deterministic terms. Two different methods are proposed to deal with the deterministic terms, and the asymptotic validity of the bootstrap procedure is theoretically checked. The small sample properties are studied by simulations and the paper is concluded by an empirical example.