This paper proposes a normalizing transformation of the Dempster statistic for testing the equality of two mean vectors with unequal covariance matrices in high-dimensional settings. The distribution of the Dempster statistic is known to converge to a normal distribution as dimension p goes to infinity; however, its rate of convergence is not guaranteed. Therefore, normal approximation is often too loose for medium p settings or fails to capture the tail behavior of the resulting distribution. We developed a concept of normalizing transformation of a statistic based on the rate of convergence to normality and show that the rate of convergence to normality is improved by normalizing transformation of the Dempster statistic.
