DIPARTIMENTO DI SCIENZE STATISTICHE Seminar MULTIVARIATE SCREENED NORMAL CLASSIFICATION ANALYSIS H YOUNG -M OON KIM Konkuk University, Seoul, Korea December 9, 2014 14.30 Aula Cucconi Abstract www.stat.unipd.it/fare-ricerca/seminari via C. Battisti, 241 - 35121 Padova, Italia - tel +39 049 8274168 - fax +39 049 8274170 - dipstat@stat.unipd.it DIPARTIMENTO DI SCIENZE STATISTICHE UNIVERSITÀ DEGLI STUDI DI PADOVA MULTIVARIATE SCREENED NORMAL CLASSIFICATION ANALYSIS HYOUNG-MOON KIM PROFESSOR AT THE DEPARTMENT OF APPLIED STATISTICS COLLEGE OF COMMERCE AND ECONOMICS, KONKUK UNIVERSITY, SEOUL, KOREA Abstract In many real problems, we encounter the situation of screening. We consider the multivariate screening scheme where underlying distribution is jointly normal distribution. Following this assumptions, we derive a new classification rule which usually outperforms the classical LDA or QDA when underlying joint distribution of screening random vector and observation vector is truly normal distribution. Resulted classification rule is nonlinear in observation vector so an approximate linear classification rule is obtained. Based on this linear classification analysis, we obtain several aspects of this rule including error rates based on total probability of misclassification, effect of misspecified screening interval, and comparison with the classical normal classification. ML estimates via ECM algorithm is clearly derived. The suggested multivariate screened classification rules are illustrated in detail using some numerical examples. 2