Proceedings of the Twentieth Conference on the Design of Experiments in Army Research Development and Testing. Part II. Ranking and Selection Procedures.

摘要

In recent years statisticians have become increasingly concerned with the meaningful formulation and solution of certain multiple-decision problems which arise in experimentation. Thus, for example, when an experimenter conducts tests to compare the performances of several competing categories of items, his ultimate objective often is to select the category (or categories) which is (are) best, goodness being measured in terms of a particular parameter (e.g., the population mean or the population variance) associated with the random variable being observed. To accomplish this the experimenter requires a statistical decision procedure which will tell him how many observations to take, how to take these observations, and based on these observations which population(s) to choose; the decision procedure should have the property that the probability of an incorrect selection (or more generally, the risk or expected loss) is controlled at some specified level. In response to the need for such decision procedures, research statisticians have been studying various possible appropriate formulations of these problems, and have developed a body of statistical methodology to cope with them. The procedures have come to be referred to as ranking and selection procedures. The purpose of this paper is to introduce the reader to these procedures, to describe some of them and the philosophy underlying their use, and to discuss their properties.