|
Activity Number:
|
479
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 5, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Survey Research Methods
|
| Abstract - #304284 |
|
Title:
|
A Simulation Study of Treatments of Influential Values in a Monthly Retail Trade Survey
|
|
Author(s):
|
Mary H. Mulry and Broderick Oliver*+
|
|
Companies:
|
U.S. Census Bureau and U.S. Census Bureau
|
|
Address:
|
4600 Silver Hill Road, Suitland, MD, 20746,
|
|
Keywords:
|
M-Estimation ; Clark Winsorization ; Outliers
|
|
Abstract:
|
While influential values are rare in economic surveys, they are problematic when they occur. An influential value is a value that is correct but whose weighted contribution has an excessive effect on an estimated total. Currently the U.S. Census Bureau uses weight trimming to address influential values. Mulry and Feldpausch (2007) showed that two methodologies, M-Estimation and Clarke Winsorization, showed potential for improvements in detection and treatment of influential values in the Monthly Retail Trade Survey, a program that uses the Hidiroglou-Berthelot algorithm to identify potential outliers on a flow basis prior to final identification of influential values. The simulation assesses the effectiveness of the two methods under controlled (but realistic) scenarios by examining the statistical properties of the treated estimates obtained using each method over repeated sampling.
|
