Auxiliary variables and adjustments for missingness in longitudinal studies

Different sorts of auxiliary variables � variables measured at previous waves, frame variables and paradata - can be used to improve the accuracy of response propensity models, and to enhance adjustments for missing longitudinal data. All these variables are used in this paper when constructing iterative probability weights, carrying out multiple imputations, and specifying models that jointly model a substantive process and the missingness mechanism. Data from the first two waves of the UK Millennium Cohort Study are used to illustrate the potential value of auxiliary variables. We find that the accuracy of response probability models � as measured by the area under the Receiver Operating Characteristic curve � is improved by the inclusion of frame variables and paradata but these variables have rather little effect when adjusting the chosen longitudinal estimates. There is, however, evidence to suggest that unobserved variables are correlated with the outcome of interest and with the probability of being a respondent at wave two.