Is the classical Wald test always suitable under response-adaptive randomization?