Abstract
We examine the properties of several residual-based cointegration tests when long-run parameters are subject to multiple shifts driven by an unobservable Markov process. Unlike earlier study, which considered one-off deterministic breaks, our approach has the advantage of allowing for an unspecified number of stochastic breaks. We illustrate this issue by exploring the possibility of Markov switching cointegration in the stock price-dividend relationship and showing that this case is empirically relevant. Our subsequent Monte Carlo analysis reveals that standard cointegration tests are generally reliable, their performance often being robust for a number of plausible regime shift parameterizations.