Intercomparing residuals to find outliers in randomized blocks

Abstract

The results of experiments laid out in a randomized complete block often appear to contain one or more unusual values. Techniques for detecting such exotic values, understanding of alternative descriptions of exoticity, one or the other of which applies rather often, leading to what might be done when exotic values are detected, can all be of practical use. This study was designed to investigate such a two-way data set: (a) by using for detection the combination of a median polish to highlight unusual values and an analysis on its residuals that is guided by the order statistic medians from a unit half-Gaussian distribution, (b) by exploring an alternate analysis on a logarithmic scale, and (c) by either Winsorizing or applying missing-value techniques to those values found to be exotic. It details the processes and the ways in which the results may be interpreted by application to data from an experiment conducted by the Rubber Research Institute in Malaysia. This article provides a concise practical tool for the investigation of such data, leading into more appropriate statistical analysis.