A class of procedures that consistently classify the stochastic compon
ent of a time series as being integrated either of order zero [I(0)] o
r one [I(1)] are proposed for general I(0) or I(1) processes and polyn
omial or piecewise linear detrending. Large-sample Bayesian inference
is free of nuisance parameters describing short-run dynamics and requi
res specifying priors only on the point hypotheses 'I(0)' and 'I(1)' t
hereby avoiding problematic choices of parametric priors over roots an
d nuisance parameters. Applied to the Nelson-Plosser (1982) data with
linear detrending, these procedures largely support Nelson and Plosser
's original inferences. With piecewise-linear detrending these data ar
e typically uninformative, producing Bayes ratios close to one.