Power-Shrink Excessive Weights Preferred to Trimming

• Activity Number: 534
• Type: Contributed
• Date/Time: Thursday, August 10, 2006 : 10:30 AM to 12:20 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Naihua Duan and Xiao-Li Meng and Chihnan Chen*+ and Margarita Alegria
• Companies: University of California, Los Angeles and Harvard University and Boston University and Cambridge Health Alliance

Abstract

Large-scale surveys typically produce raw weights with very large variations. A standard approach is to trim, a way to reduce potentially huge variances of various survey estimators. The trimming method is popular mainly because of its simplicity. In this paper we investigate a more principled method by shrinking the variance of the log of the weights, which is equivalent to introducing a power parameter p in [0,1] for the weights. Our investigation suggests that p=1/2, that is, taking square root is a good general-purpose choice according to the mean-squared error criterion. This power shrinking method, besides preserving the ranking of the raw weights, provides a natural way to deal with measurement errors and outliers in the raw weights. We demonstrate the use of this method with the National Latino and Asian American Study (NLAAS) on mental health.