Panel Data Unit Root Tests with an Application

Abstract

Using several different unit root/stationarity tests on single time series Konya (2000) found the logarithm of real GDP of most OECD countries behaving as a random walk during the last four decades. This outcome, however, might be due to the generally low power of these tests. The aim of this paper is to reconsider this issue by exploiting the extra information provided by the combination of the time-series and cross-sectional data and the subsequent power advantages of panel data unit root tests. We apply the tests advocated by Levin and Lin (1993), Im, Pesaran and Shin (1997) and Maddala and Wu (1999). The joint unit root null hypothesis cannot be rejected for the whole panel, however, after having dropped the least likely stationary series from the panel, the Im, Pesaran and Shin (1997) and Maddala and Wu (1999) tests can reject the null for the remaining sub-panels. Since these tests are not valid under cross-correlated error terms, we repeat them with data specific bootstrap critical values taking contemporaneous cross-correlation into account. This way, both tests can reject the unit root null hypothesis even for the whole panel. Yet, they cannot suggest how many and which particular panel members are stationary, so these results are not really informative. For this reason finally we apply the unit root test of Breuer, McNown and Wallace (1999) with bootstrap critical values. This procedure makes use of the panel data setting and seemingly unrelated regressions, but performs separate unit-root tests on each panel member. The results are markedly different from the conventional univariate Dickey-Fuller test results, and they support trend stationarity in the case of four countries: Australia, Japan, the Netherlands and Switzerland.