197 – Contributed Poster Presentations: Survey Research Methods Section
Misspeci fied Sampling Weights in Weight-Smoothing Methods
Xia Li
University of Maryland
Eric Slud
University of Maryland
Modifications of weights due to calibration, trimming, sometimes in multiple stages, are very common in survey analysis. It is typical to work with modified as opposed to design/inverse-inclusion-probability weights, especially in publicly released survey data. Various weight-smoothing methods (Pfeffermann-Sverchkov 1999, Zheng-Little 2003, Beaumont 2008) have been proposed to improve the efficiency of the Horvitz-Thompson (HT) and Generalized Regression (GREG) estimators of survey totals. These methods depend on correctness of model relationships between the survey attribute y, covariate x, and y and the given weights w. Little is known about the impact of treating modified weights as design weights, for example when the model assumptions connecting x, y, w might also be misspecified. It is, therefore, important to evaluate the performances of these three methods under different modified weights and misspecified models. In this simulation study, we generate finite frame populations from superpopulation models, simulate misspecified models, and quantified mis-calibrations of the weights to compare GREG and HT results with estimators based on the three weight-smoothing methods.