Regression-based analysis of cointegration systems

Two estimation procedures dominate the cointegration literature: Johansen's maximum likelihood inference on vector autoregressive error correction models and estimation of Phillips' triangular forms. This latter methodology is essentially semiparametric, focusing on estimating long run parameters by means of cointegrating regressions. However, it is less used in practice than Johansen's approach, since its implementation requires prior knowledge of features such as the cointegrating rank and an appropriate set of non-cointegrated regressors. In this paper we develop a simple and automatic procedure (based on unit root and regression-based cointegration testing) which, without imposing a parametric specification for the short run components of the model, provides an estimator of the cointegrating rank and data-based just-identifying conditions for the cointegrating parameters which lead to a Phillips' triangular form. A Monte Carlo analysis of the properties of the estimator and an empirical application are also provided.