Title:A new graph-based two-sample test for multivariate and object data

Abstract:For comparing multivariate distributions when the dimension is moderate to high with a typical sample size, existing tests have low or no power either for location or scale alternatives. A new test based on adjacency between observations is proposed. It utilizes a common pattern in both types of alternatives and thus works under both scenarios. The new test is shown to exhibit substantial power gains in simulation studies. The asymptotic null distribution for the new test is also derived and the approximated $p$-value seems to be quite accurate for finite samples. The test is applied to assess covariate balance in a matched observational study on the effect of starting college at a two year versus four year college on ultimate educational attainment.