On the size distortion of stationarity tests