The multivariate nonparametric Behrens–Fisher problem

• Published in: Journal of statistical planning and inference 2002-11, Vol.108 (1), p.37-53
• Main Authors: Brunner, Edgar, Munzel, Ullrich, Puri, Madan L.
• Format: Article
• Language: English
• Subjects: Behrens–Fisher problem; Exact sciences and technology; Mathematics; Multivariate analysis; Multivariate two-sample problem; Nonparametric hypothesis; Nonparametric inference; Ordered categorical data; Probability and statistics; Rank tests; Sciences and techniques of general use; Statistics; Ties
• ISSN: 0378-3758; 1873-1171
• DOI: 10.1016/S0378-3758(02)00269-0

In this paper, we consider the multivariate case of the so-called nonparametric Behrens–Fisher problem where two samples with independent multivariate observations are given and the equality of the marginal distribution functions under the hypothesis in the two groups is not assumed. Moreover, we do not require the continuity of the marginal distribution functions so that data with ties and, particularly, multivariate-ordered categorical data are covered by this model. A multivariate relative treatment effect is defined which can be estimated by using the mid-ranks of the observations within each component and we derive the asymptotic distribution of this estimator. Moreover, the unknown asymptotic covariance matrix of the centered vector of the estimated relative treatment effects is estimated and its L 2-consistency is proved. To test the hypothesis of no treatment effect, we consider the rank version of the Wald-type statistic (as used in Puri and Sen, Nonparametric Methods in Multivariate Analysis, Wiley, New York, 1971) and the rank version of the ANOVA-type statistic which was suggested by Brunner et al. [J. Amer. Statist. Assoc. 92 (1997) 1494–1502] for univariate nonparametric models. Simulations show that the ANOVA-type statistic appears to maintain the pre-assigned level of the test quite accurately (even for rather small sample sizes) while the Wald-type statistic leads to more or less liberal decisions. Regarding the power, none of the two statistics is uniformly superior to the other.