|
Activity Number:
|
46
|
|
Type:
|
Invited
|
|
Date/Time:
|
Sunday, August 2, 2009 : 4:00 PM to 5:50 PM
|
|
Sponsor:
|
Section on Survey Research Methods
|
| Abstract - #303108 |
|
Title:
|
Power-Shrinkage: An Alternative Method for Dealing with Excessive Weights
|
|
Author(s):
|
Xiao-Li Meng*+ and Naihua Duan and Chih-nan Chen and Margarita Alegria
|
|
Companies:
|
Harvard University and Columbia University and Cambridge Health Alliance/Harvard Medical School and Harvard Medical School
|
|
Address:
|
Deaprtment of Statistics , Cambridge, MA, 02138,
|
|
Keywords:
|
Survey weights ; Winsorization ; Weight trimming ; Mental Health ; Service Utilization
|
|
Abstract:
|
Gelman (2007, Statistical Science) is absolutely correct that "Survey weighting is a mess." Dealing with excessive weights is a part of this mess, with the most popular method being the rather ad hoc Winsorization method (a.k.a. weight trimming). We propose a more principled method based on shrinking the log of weights, motivated by the expectation that imperfections (e.g., measurement errors) in forming the raw weights typically affect all sampling units multiplicatively. The new method amounts to raising the raw weight to a power a, essentially a compromise between using equal weights (a=0) and using the raw weights (a=1). We conduct theoretical and empirical investigations of this new method, in comparison with the Winsorization method. We illustrate the new method using data from the National Latino and Asian American Study (NLAAS) for mental health and service utilization.
|
