System tests for cointegration proposed by Stock and Watson (1988), Johansen (1988), and Bewley and Yang (1995) are compared using Monte Carlo experiments that include overspecification of the lag length and data-generating processes with moving average disturbances. Both AIC and SIC are used to select the lag length of the approximating vector autoregressions. The three tests considered are found to have very different characteristics, and each dominates in some portion of the parameter space.