2001 Fiscal Year Final Research Report Summary
Study of applying efficient resampling methods in discriminatn analysis
Grant-in-Aid for Scientific Research (C)
|Allocation Type||Single-year Grants |
|Research Institution||Nagasaki University (2001)|
Chiba University (1999-2000)
HONDA Masayuki Nagasaki University, School of Medicine Hospital and Clinics, Professor -> 長崎大学, 医学部・附属病院, 教授 (10143306)
KONISHI Sadanori Kyushu University, Graduate School of Mathematics, Professor, 大学院・数理学研究科, 教授 (40090550)
NAKANO Masataka Mie University, School of Medicine, Professor, 医学部, 教授 (00114306)
TAGURI Masaaki Chiba University, Faculty of Science, Professor, 理学部, 教授 (10009607)
YAMANOBE Yuji Nagasaki University, School of Medicine Hospital and Clinics, Lecturer, 医学部・附属病院, 講師 (40284690)
SHIBATA Yoshisada Nagasaki University, School of Medicine, Professor, 医学部, 教授 (40010954)
|Project Period (FY)
1999 – 2001
|Keywords||DISCRIMINANT ANALYSIS / ERROR RATE / RESAMPLING METHOD / BOOTSTRAP METHOD / MONTE-CARLO SIMULATION / VARIANCE REDUCTION METHOD|
Theoretical study of efficient resampling methods and applying them to real problems was investigated. Research papers about several approaches that reduced fluctuations introducing resampling methods were collected. The efficiency and limits of those approaches were summarized through several numerical experiments.
We focused on the bias of apparent error rate, which means a difference between an apparent error rate and an actual error rate. An actual error rate is a goal of estimation problem in discriminant analysis. We investigated the efficiency of an approach eliminating the influence function when the error rate was decomposed.
We have continued to study several approaches of numerical experiments (Monte-Carlo experiments). We checked statistical program packages such as S, S'PLUS and SPSS. Those packages were considered be lack of several kinds of efficient resampling methods for real problems to be analyzed.
We joined the international conference of IFCS-2000 which was held at Belgium in 2000 and the domestic conference of Grant-in-Aid for Scientific Research (A) about Statistical Prediction in 2000. Head investigator Honda and investigator Konishi was presented our research papers at those conferences.
We also studied nonlinear discriminant analysis which outperformed discrimination power based on kernel function and neural network approach. Construction of nonlinear discriminant function introduced criterion of model selection, which is based on information criterion and estimation method based on normalized technique.
Research Products (14 results)