Application of tube formula to distributional problems in.pdf

Application of tube formula to distributional problems in multiway layouts Satoshi Kuriki Institute of Statistical Mathematics and Akimichi Takemura Graduate School of Information Science and Technology University of Tokyo October, 2001 Abstract Recently, an integral geometric method called the tube method has been actively developed. The tube method gives us a powerful tool to tackle problems where con- ventional matrix theory such as the singular value decomposition cannot be applied. The aim of this paper is to survey several recent applications of the tube method to distributional problems in multiway layouts mainly by the authors. Null distribu- tions of test statistics of the following four testing problems are discussed: (i) A test for interaction in three-way layout based on the three-way analogue of the largest singular value. (ii) A test for multivariate normality by searching a nonnormal direction proposed by Malkovich and Afifi. (iii) Testing independence in ordered categorical data by maximizing row and column scores under order restriction. (iv) Detecting a change point in two-way layout with ordinal factors. Key words: change point, Gaussian random field, integral geometry, multilinear model, ordered categorical data, testing multivariate normality. 1 Introduction. Standard results of matrix theory play a major role in conventional multivariate statistical analysis or categorical data analysis. In particular when data are summarized as a two-way table, many methods of data analysis have been based on the singular value decomposition of the data matrix. However data may not be summarized as a single matrix. In a factorial design or cross-classification, data are usually obtained as a multiway layout. Moreover even when data are summarized as a two-way table, usual matrix theory is not necessarily 1 applicable. For example, in a two-way cross-classified table with ordinal row and column categories, ordinary matrix methods invariant with respect to permu