|
Abstract:
|
High nonresponse constitutes a serious threat to the quality of sample surveys. The magnitude of the nonresponse is of importance, but not always the most problematic issue. In terms of bias for the resulting estimators of e.g. population totals, other phenomena may matter more. If we have access to reliable auxiliary information, calibration methods may adjust for a "nonrepresentative" response set. The auxiliary information is usually linked to both the study variable and the underlying nonresponse mechanism.
Previously suggested calibration estimators under nonresponse include "GREG-type" estimators. However, without using a model for response propensities, there is not much room for improvement on this "standard" weighting. We propose instead to use estimators inspired by the design-based optimal regression estimator. It turns out that there is some freedom for alternative weighting within this family of estimators. The Poisson sampling design will be used to illustrate theoretical and empirical results, where the emphasis is put on the reduction of bias.
|
