Autocorrelation robust tests are notorious for suffering from size distortions and power problems. After revisiting recent analytical finite sample results concerning size and power properties of such tests, we investigate under which conditions the size of autocorrelation robust tests can be controlled by an appropriate choice of critical value.