Spectral density bandwidth choice: source of nonmonotonic power for tests of a mean shift in a time series

Data dependent bandwidth choices for zero frequency spectral density estimators of a time series are shown to be an important source of nonmonotonic power when testing for a shift in mean. It is shown that if the spectral density is estimated under the null hypothesis of a stable mean using a data dependent bandwidth (with or without prewhitening), non-monotonic power appears naturally for some popular tests including the CUSUM test. On the other hand, under some fixed bandwidth choices, power is monotonic. Empirical examples and simulations illustrate these power properties. Theoretical explanations for the power results are provided