A1341
Title: Information-geometric viewpoint on partitioning of test statistics for symmetry in contingency tables
Authors: Tomoyuki Nakagawa - Meisei University & RIKEN Center for Brain Science (Japan) [presenting]
Takeru Matsuda - University of Tokyo & RIKEN Center for Brain Science (Japan)
Kouji Tahata - Tokyo University of Science (Japan)
Abstract: Numerical and asymptotic partitioning of goodness-of-fit statistics has been investigated for numerous models in contingency tables. The purpose is to explore how goodness-of-fit statistics for symmetry in contingency tables can be partitioned. The symmetry model is dually flat and can be characterized as the intersection of an e-flat submodel and an orthogonal m-flat submodel. Based on this result, the Wald statistic for the symmetry model can be exactly partitioned into components corresponding to the submodels. On the other hand, there are very few models for which the likelihood ratio test statistic for the symmetry model can be exactly partitioned. The focus is on the relationship between information geometry and the exact partitioning of goodness-of-fit statistics for symmetry in contingency tables.
