ModelBased Alternatives to Trimming Survey Weights Michael R. Elliott and Roderick J.A. Little Abstract: In sample surveys with unequal probabilities of inclusion, units are often weighted by the inverse of the probability of inclusion to avoid biased estimates of population quantities such as means. Highly disproportional sample designs yield large weights, which can result in weighted estimates that have a high variance. Weight trimming reduces large weights to a fixed cutpoint value and adjusts weights below this value to maintain the untrimmed weight sum. This approach reduces variance at the cost of introducing some bias. An alternative approach uses randomeffects models to induce shrinkage across weight strata. We compare these two approaches, and introduce extensions of each: a compound weight pooling model that allows Bayesian averaging over estimators based on different trimming points, and a weight smoothing model based on a nonparametric spline function for the underlying weight stratum means. The latter method performs well in simulations as compared with alternative estimators. Methods are also applied to estimates of depression using weighted data from the National Comorbidity Survey. Keywords: Sample surveys inference, sampling weights, unit nonresponse adjustments, randomeffects models, nonparametric regression
