On (bootstrapped) cointegration tests in partial systems

As applied cointegration analysis faces the challenge that (a) potentially relevant variables are unobservable and (b) it is uncertain which covariates are relevant, partial systems are often used and potential (stationary) covariates are ignored. Recently it has been argued that a nominally significant cointegration outcome using the bootstrapped rank test (Cavaliere, Rahbek, and Taylor, 2012) in a bivariate setting might be due to test size distortions when a larger data-generating process (DGP) with covariates is assumed. This study reviews the issue systematically and generally finds noticeable but only mild size distortions, even when the specified DGP includes a large borderline stationary root. The previously found drastic test size problems in an application of a long-run Phillips curve (inflation and unemployment in the euro area) appear to hinge on the particular construction of a time series for the output gap as a covariate. We conclude that the problems of the bootstrapped rank test are not severe and that it is still to be recommended for applied research.