On the accuracy of a normal approximation to the power of the mantel-haenszel procedure