A cautionary note on cointegration testing

Abstract

Inferences from the Johansen procedure regarding cointegration, and the magnitude, significance and even the sign of the estimated parameters of a familiar macroeconomic relation, are shown to be extremely sensitive to the treatment of its deterministic components and to the assumed lag structure. An unrestricted error-correction model yields unambiguous inferences and performs better in a range of tests.

Notes

¹ UK data are from Economic Trends Annual Supplement 1997. US data are from Survey of Current Business, May 1997.

² Details of these and other results reported but not documented are available on request.

³ The error-correction terms are defined here as (LC – f(LY, INF)). In MICROFIT4 the definition switches between this conventional form and its negative.