Auxiliary variables and missingness weighting adjustment

Abstract

In panel studies, the survey organization creates and releases non-response weights for use by data analysts. As a general practice after wave 1, survey statisticians use substantive survey variables measured in previous waves to estimate response propensity models from which the weights are derived. However, auxiliary variables such as frame variables, paradata and interviewer observations may also be used in this regard. In this paper, we investigate whether the inclusion of auxiliary variables in the response propensity models will improve the accuracy of the model and hence results in more efficient weighting. We used data from wave 1 and 2 of the British Household Panel Survey. Our findings indicate that if auxiliary variables are used together with substantive survey variables, the accuracy of the response propensity model will be improved and the set of longitudinal weights will be more precise. Thus, weighted estimates may become more precise and more significant.