科研成果详情

Multiple normal mean comparison without the equal-variance assumption is frequently encountered in medical and biological problems. Classical analysis of variance (ANOVA) requires the assumption of equal variances across groups. When variations across groups are found to be different, classical ANOVA method is essentially inapplicable for multiple mean comparison. Although various approximation methods have been proposed to solve the classical Behrens-Fisher problem, there exists computational complexity in approximating the null distributions of the proposed tests. In this paper we employ the theory of spherical distributions to construct a class of exact F-tests and a simple generalized F-test for multiple mean comparison. The methods in this paper actually provide a simple exact solution and a simple approximate solution to the classical Behrens-Fisher problem in the case of balanced sample designs. A simple Monte Carlo study shows that the recommended tests have fairly good power performance. An analysis on a real medical dataset illustrates the application of the new methods in medicine.