Harnessing paradata and multilevel multiple imputation when analysing survey data: a case study

Abstract

Missing data (attrition and non-response) are a feature of most surveys especially longitudinal/panel studies. And many such studies now have multilevel designs and hence multilevel data structures. Recent advances in imputation methodology now offer social researchers opportunities to address issues of missing data in a statistically principled way. Paradata can offer great insights in understanding the nature and causes of missingness and can be used to construct auxiliary variables to be included in imputation models. In this paper we present multilevel multiple imputation which has recently extended MI to incorporate multilevel data, making it a flexible and robust strategy for many research settings. We illustrate the procedures by analysing data drawn from a longitudinal study of prisoners. We show how paradata of that study was instrumental in guiding our approach and subsequent analysis.