Hypothesis testing with error correction models

Grant and Lebo (2016) and Keele et al. (2016) clarify the conditions under which the popular general error correction model (GECM) can be used and interpreted easily: In a bivariate GECM the data must be integrated in order to rely on the error correction coefficient, alpha(1)*, to test cointegration and measure the rate of error correction between a single exogenous x and a dependent variable, y. Here we demonstrate that even if the data are all integrated, the test on alpha(1)* is misunderstood when there is more than a single independent variable. The null hypothesis is that there is no cointegration between y and any x but the correct alternative hypothesis is that y is cointegrated with at least one but not necessarily more than one of the x's. A significant alpha(1)* can occur when some i(1) regressors are not cointegrated and the equation is not balanced. Thus, the correct limiting distributions of the righthand-side long-run coefficients may be unknown. We use simulations to demonstrate the problem and then discuss implications for applied examples.