Tests for evaluating nonresponse bias in surveys

摘要

How do we tell whether weighting adjustments reduce nonresponse bias? If a variable is measured for everyone in the selected sample, then the design weights can be used to calculate an approximately unbiased estimate of the population mean or total for that variable. A second estimate of the population mean or total can be calculated using the survey respondents only, with weights that have been adjusted for nonresponse. If the two estimates disagree, then there is evidence that the weight adjustments may not have removed the nonresponse bias for that variable. In this paper we develop the theoretical properties of linearization and jackknife variance estimators for evaluating the bias of an estimated population mean or total by comparing estimates calculated from overlapping subsets of the same data with different sets of weights, when poststratification or inverse propensity weighting is used for the nonresponse adjustments to the weights. We provide sufficient conditions on the population, sample, and response mechanism for the variance estimators to be consistent, and demonstrate their small-sample properties through a simulation study.