Comparing Methods for Developing Stratum Boundaries in Audit Populations

• Activity Number: 466
• Type: Contributed
• Date/Time: Thursday, August 7, 2003 : 10:30 AM to 12:20 PM
• Sponsor: Section on Survey Research Methods
• Author(s): David E. McGrath*+ and Eric T. Falk
• Companies: Ernst & Young, LLP and Ernst & Young, LLP
• Keywords: stratification ; stratum boundary ; audit population ; Neyman allocation ; skewed data

Abstract

Empirical tests were conducted to compare four methods for determining stratum boundaries for Ernst & Young's (E&Y) audit populations. As with other economic data, audit populations generally contain highly skewed numeric variables that present special problems. Prior to this experiment, E&Y used the cumulative square root of the frequency method as the initial determination of stratum boundaries. When paired with Neyman allocation, this method proves unsatisfactory because it often creates boundaries where one highly variable, large-dollar stratum receives a disproportionate amount of the overall sample. This often creates undesirably small sample sizes in other strata, thus limiting the use of separate ratio and regression estimators. To conduct the experiment, we stratified all populations by a highly skewed dollar variable using the cumulative square root of the frequency as a control group, and the following three methods as experimental groups: equal dollars per stratum, equal NhSh in each stratum, Lavalle/Hidiriglou method. We selected the method producing the smallest relative precision of the design variable for a fixed sample size.