Distribution-Free Tests for Sparse Heterogeneous Mixtures

Abstract: We consider the problem of detecting sparse heterogeneous mixtures from a nonparametric perspective, and develop distribution-free tests when all effects have the same sign. Specifically, we assume that the null distribution is symmetric about zero, while the true effects have positive median. We evaluate the precise performance of classical tests for the median (t-test, sign test) and classical tests for symmetry (signed-rank, Smirnov, total number of runs, longest run tests) showing that none of them is asymptotically optimal for the normal mixture model in all sparsity regimes. We then suggest two new tests. The main one is a form of Higher Criticism, or Anderson-Darling, test for symmetry. It is shown to be asymptotically optimal for the normal mixture model, and other generalized Gaussian mixture models, in all sparsity regimes. Our numerical experiments confirm our theoretical findings.