129 – SPEED: Topics in Statistical Methods and Applications
Methods and Issues in Trimming Extreme Weights in Sample Surveys
Frank Potter
Mathematica Policy Research
Yuhong Zheng
Mathematica Policy Research
In survey sampling practice, unequal sampling weights (the inverse of the selection probabilities) can be both beneficial and deleterious. Extreme variation in the sampling weights can result in excessively large sampling variances when the data and the selection probabilities are not positively correlated. In addition, extreme variation in the weights can result from unplanned subsampling, nonresponse adjustments, or post-stratification. In some survey situations, the survey statistician may impose a trimming strategy for excessively large weights. Because of the weight trimming, the survey statistician will usually expect an increased potential for a bias in the estimate and a decrease in the sampling variance. The ultimate goal of weight trimming is to reduce the sampling variance more than enough to compensate for the possible increase in bias and, thereby, reduce the mean square error. In this presentation, I will discuss current methods used to identify the appropriate trimming values and provide guidance on selecting the final trimming level, which may be different from the values suggested by the algorithms.