Methods for Extreme Weights in Sample Surveys Methods for Extreme Weights in Sample Surveys

In survey sampling practice, planned and unplanned variation in the sampling weights can result in inflated sampling variances. As a result, extreme sampling weights are sometimes trimmed to reduce the sampling variance. However, when sampling weights are trimmed, a bias can be introduced into the survey estimates. The goal of sampling weight trimming is to reduce the sampling variance while avoiding the introduction of substantial bias: thereby reducing the mean square error of the point estimates. Because of the bias potential, the coverage probability for interval estimates can also be affected. The two currently used procedures are basically ad hoc, with little theoretical basis and no empirica~ evaluations of the effects of their weight trimming. This research summarizes the two currently used procedures, provides an analytical framework for weight trimming, and proposes three alternative procedures. In this research an empirical investigation of the weight trimming procedures uses 200 replicated samples of 100 units from a data base for 2,989 counties in the U.S. The effect of each procedure is measured in terms of the sampling variance, the bias introduced, the estimated mean square error, and the coverage probability for interval estimates.