Trend stationarity versus long-range dependence in time series analysis

Empirically, it is difficult to offer unequivocal judgment as to whether many real economic variables are fractionally integrated or trend stationary. The objective of this paper is to study the effects of spurious detrending of a nonstationary fractionally integrated NFI(d), d∈(12,32). With respect to the performance of the traditional least squares estimators and tests we prove that the estimated time trend coefficient is consistent but that the corresponding t-Student test diverges. We also analyze a local version in the frequency domain of least squares. We are able to show the consistency of this estimator and that, after conveniently adjusting variance estimates, its t-ratio has a well-defined but nonstandard limiting distribution. Nonetheless, in this latter case it is possible to obtain a set of critical values giving rise to the correct size for any given d∈(12,32).