Stratification and the Adjusted Chi-Square Statistic: Application to Analysis of Employment Discrimination Data

摘要

In this paper we describe and illustrate the concept of #x201c;stratification of data#x201d; to enhance inferential validity when evaluating rates in employment discrimination lawsuits. Stratification is a means of avoiding confounding effects when classes of employees are distributed differently with respect to variables that correlate with the main variable of interest (e.g., promotions). The summary chi-square statistic developed by Mantel and Haenszel may be helpful when data are stratified, but a threshold consideration of second-order interactions is required. Two alternatives to chi-square are considered, Fisher's exact test and Wilcoxon's signed-rank test. The problem of choosing the appropriate significance level is also considered, with a recommendation. Data from actual lawsuits are used in the illustrations.