On the Panel Unit Root Tests Using Nonlinear Instrumental Variables

This paper re-examines the panel unit root tests proposed by Chang (2002). She establishes asymptotic independence of the t-statistics when integrable functions of lagged dependent variable are used as instruments even if the original series are cross sectionally dependent. From this rather remarkable result she claims that her non-linear instrumental variable (NIV) panel unit root test is valid under general error cross correlations for any N (the cross section dimension) as T (the time dimension of the panel) tends to infinity. We show that her claim is valid only if NlnT/square root of T to 0, as N and T to infinity, and this condition is unlikely to hold in practice, unless N is very small. The favourable simulation results reported by Chang are largely due to her particular choice of the error correlation matrix, which results in weak cross section dependence. Also, the asymptotic independence property of the t-statistics disappears when Chang's modified instruments are used. Using a common factor model with a sizeable degree of cross section correlations, we are able to show that Chang's NIV panel unit root test suffers from gross size distortions, even when N is small relative to T (for example N=5, T=100)