Cointegration: how short is the long run?

Abstract When testing for cointegration and faced with short sample periods, it is common for researchers to turn to more frequently sampled observations, that is, to move from annual to quarterly or monthly data, in order to increase the number of observations. We argue that such a gain in the degrees of freedom is more apparent than real. Essentially, cointegration is a long-run concept and hence requires long spans of data to give tests for cointegration much power rather than merely large numbers of observations. Using Monte Carlo simulations, we demonstrate this for four popular tests for cointegration.