On the Sensitivity of Unit Root Inference to Nonlinear Data Transformations

In this paper we analyze the sensitivity of unit root inference to nonlinear transformations through Bayesian techniques. We make joint inference about the Box-Cox transformation, which includes the cases yt and log(yt), and the unit root. When we apply our method to the fourteen Nelson-Plosser series, we find that unit root inference can be very sensitive to the transformation chosen and that the usual practice of taking logs is not always warranted.