Optimal classifier for multivariate rectangle-screened normal data classification

This paper discusses the classification procedures which make provision for the case where the interest of an investigator is to classify a multidimensional screened normal observation into one of two or more populations. A number of problems are thoroughly investigated in order to render the procedures optimal. These include multidimensional screened normal models that describe the evolutions of the screened observations from different populations, derivation of an optimal classification rule and its linear approximation, and estimation of the rules via the expectation–maximization algorithm with a sequence of conditional maximization steps. The efficiency of the classification rules is examined by using simulation studies. Using real data of bank employee’s salary from the IBM SPSS Statistics 19, we illustrate the empirical relevance of the screened classification.