The finite sample properties of LM‐type linearity tests based on the discussion in Granger (1995) are examined. The tests are constructed based on regression models which contain stationary linear and nonlinear functions of nonstationary variables, thus generalizing standard linear cointegrating equations. Power and size simulations are promising, suggesting that the tests are worthy of further examination, and an illustrative empirical example shows that some form of nonlinear error‐correction may be useful for explaining the evolution of U.S. money stock in a simple vector autoregression framework.