Abstract:
|
As nonresponse continues to increase, weighting adjustments increasingly rely on implicit (or explicit) mathematical models for explaining nonresponse. Although it is crucial to understand this relationship and its impact on survey estimates, the literature that describes the different weighting adjustments is dispersed and sometimes contradictory. We propose a more unified way for understanding these relationships. We present an expression for nonresponse bias for estimates computed using nonresponse adjusted weights. We explain why this expression may be preferred for examining survey statistics under the Total Survey Error framework. We argue that weighting for nonresponse should be seen as an estimation task and once the statistical models have been identified, classical statistical tools such as goodness of fit and model diagnostics can be used to evaluate the quality of nonresponse adjusted weights. This approach enables us to evaluate the effect of model misspecification from incorrect functional forms or from omitted variables in the nonresponse models.
|