Bernardo Rota: Comparisons of some weighting methods for nonresponse adjustment

Auxiliary information is vital when weighting for nonresponse adjustments in estimation. Both sample and population level information has shown to be useful and yield approximately equal results in large samples. One alternative weighting procedure is calibration, where weights are defined to replicate auxiliary variable population totals, known or estimated using sample level information. There are several functional forms of weights suggested in the literature. Chang and Kott (2008) suggest weights defined by using the reciprocal of a response probability function, assumed known up to model parameters. This paper studies the properties of calibration estimators when the response probability is assumed known. The particular focus is on the difference between population and sample level information. For the Chang and Kott (2008) estimator, sample level information is shown to be superior for estimation of response model coefficients. Results from a simulation study suggest a two-step procedure, utilizing sample level information for model coefficient estimation and calculation of initial weights in a first step, and then calibration estimation of the study variable total is performed in a second step.

In collaboration with Thomas Laitila, Department of Statistics, Örebro university; Department of Research and Development, Statistics Sweden.