Powerful Nonparametric Asymptotic Tests for Change in the Mean with Reduced Type I Errors

We study numerically finite-sample power functions of nonparametric asymptotic tests for at most one change in the mean that are based on convergence in the distribution of sup- and integral functionals of an appropriately weighted and normalized tied-down partial sums process. For each test, a three-way trade-off is observed among its type I errors, power for detecting the change near the beginning or end of the sample, and power for detecting the change in the middle of the sample. By choosing suitable weight functions of a special form, we propose new sup- and integral tests that are shown to be nearly as powerful as the overall most powerful sup-test in the literature, regardless of where the change occurs in the sample. Moreover, the type I errors of the new tests are closer to the asymptotic significance level across various distributions and are lower and converge faster for distributions that are more asymmetric, heavy-tailed, or both.