Limiting power of unit-root tests in time-series regression (Q756339)

The model considered in this paper is \(y_ t=x'_ t\beta +\eta_ t\), \(\eta_ t=\rho \eta_{t-1}+v_ t\), where \(y_ t\) is scalar, \(x_ t\) is a nonstochastic vector, and the \(v_ t\) are martingale differences satisfying certain moment conditions. To test \(H_ 0: \rho =1\) against \(H_ 1: \rho <1\) or \(H'_ 1: \rho >1\), six different test statistics are considered, all defined as ratios of quadratic forms in \({\hat \eta}{}_ t's\) (residuals from ordinary least squares estimation), or \({\tilde \eta}{}_ t's\) (residuals from generalized least squares estimation). Local alternatives with \(\rho =1-c/T\) (c a constant, T sample size) are studied. The authors conclude that ``the limiting powers of some unit- root tests are computed accurately under a sequence of local alternatives. The limiting powers are found to give a good approximation to the finite-sample powers. An extension of the model is also discussed allowing for autocorrelated errors in the AR(1) process''.