Some Extensions of Regression Based Cointegration Analysis

Martin Wagner (TU Dortmund)
Monday, November 24, 2014 - 2:00pm
Spandauer Straße 1, Room 23

The analysis of cointegrating relationships in a regression framework is typically carried out using modified least squares estimators that employ corrections for the effects of endogeneity and error serial correlation to obtain limiting distributions that allow for asymptotic standard inference. Several such estimation procedures are available in the literature. We discuss extensions of such approaches along two dimensions. On the one hand we discuss the applicability of modified least squares estimators in cointegrating regressions that are linear in parameters, but nonlinear in I(1) variables. Typical examples of such relationships are (environmental) Kuznets curves or translog cost or production functions. Especially the latter poses new challenges for estimation, inference and specification analysis when using e.g. the so-called fully modified OLS or the integrated modified OLS estimation principles. On the other hand we discuss estimation when it is not the functional form that is allowed to be more general, but the dynamic properties of the variables. In particular we discuss cointegration analysis based not on summed up i.e. integrated stationary but integrated locally stationary processes. We consider localized fully modified OLS estimation for this case. Some empirical illustrations complement the theory.